id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
71875754-0904-4347-b420-6fa44e4ed45f | StampyAI/alignment-research-dataset/blogs | Blogs | Discovering when an agent is present in a system
[](/?source=post_page-----41154de11e7b--------------------------------)Listen
Share
***New, formal definition of agency gives clear principles for causal modelling of AI agents and the incentives they face.***
*Crossposted to our Deepmind* [*blog*](https://www.deepmind.com/blog/discovering-when-an-agent-is-present-in-a-system)*. See also our extended blogpost on* [*LessWrong*](https://www.lesswrong.com/posts/XxX2CAoFskuQNkBDy/discovering-agents)*/*[*Alignment Forum*](https://www.alignmentforum.org/posts/XxX2CAoFskuQNkBDy/discovering-agents)*.*
We want to build safe, aligned artificial general intelligence (AGI) systems that pursue the intended goals of its designers. [Causal influence diagrams](/progress-on-causal-influence-diagrams-a7a32180b0d1#b09d) (CIDs) are a way to model decision-making situations that allow us to reason about [agent incentives](https://ojs.aaai.org/index.php/AAAI/article/view/17368). For example, here is a CID for a 1-step Markov decision process — a typical framework for decision-making problems.
![]()S₁ represents the initial state, A₁ represents the agent’s decision (square), S₂ the next state. R₂ is the agent’s reward/utility (diamond). Solid links specify causal influence. Dashed edges specify information links — what the agent knows when making its decision.By relating training setups to the incentives that shape agent behaviour, CIDs help illuminate potential risks before training an agent and can inspire better agent designs. But how do we know when a CID is an accurate model of a training setup?
Our new paper, [Discovering Agents](https://arxiv.org/abs/2208.08345), introduces new ways of tackling these issues, including:
* The first formal causal definition of agents: **Agents are systems that would adapt their policy if their actions influenced the world in a different way**
* An algorithm for discovering agents from empirical data
* A translation between causal models and CIDs
* Resolving earlier confusions from incorrect causal modelling of agents
Combined, these results provide an extra layer of assurance that a modelling mistake hasn’t been made, which means that CIDs can be used to analyse an agent’s incentives and safety properties with greater confidence.
Example: modelling a mouse as an agent
======================================
To help illustrate our method, consider the following example consisting of a world containing three squares, with a mouse starting in the middle square choosing to go left or right, getting to its next position and then potentially getting some cheese. The floor is icy, so the mouse might slip. Sometimes the cheese is on the right, but sometimes on the left.
![]()The mouse and cheese environment.This can be represented by the following CID:
![]()CID for the mouse. D represents the decision of left/right. X is the mouse’s new position after taking the action left/right (it might slip, ending up on the other side by accident). U represents whether the mouse gets cheese or not.The intuition that the mouse would choose a different behaviour for different environment settings (iciness, cheese distribution) can be captured by a [mechanised causal graph](https://drive.google.com/file/d/1_OBLw9u29FrqROsLfhO6rIaWGK4xJ3il/view),which for each (object-level) variable, also includes a mechanism variable that governs how the variable depends on its parents. Crucially, we allow for links between mechanism variables.
This graph contains additional mechanism nodes in black, representing the mouse’s policy and the iciness and cheese distribution.
![]()Mechanised causal graph for the mouse and cheese environment.Edges between mechanisms represent direct causal influence. The blue edges are special *terminal* edges — roughly, mechanism edges A~ → B~ that would still be there, even if the object-level variable A was altered so that it had no outgoing edges.
In the example above, since U has no children, its mechanism edge must be terminal. But the mechanism edge X~ → D~ is not terminal, because if we cut X off from its child U, then the mouse will no longer adapt its decision (because its position won’t affect whether it gets the cheese).
Causal discovery of agents
==========================
Causal discovery infers a causal graph from experiments involving interventions. In particular, one can discover an arrow from a variable A to a variable B by experimentally intervening on A and checking if B responds, even if all other variables are held fixed.
Our first algorithm uses this technique to discover the mechanised causal graph:
![]()Algorithm 1 takes as input interventional data from the system (mouse and cheese environment) and uses causal discovery to output a mechanised causal graph. See paper for details.Our second algorithm transforms this mechanised causal graph to a game graph:
![]()Algorithm 2 takes as input a mechanised causal graph and maps it to a game graph. An ingoing terminal edge indicates a decision, an outgoing one indicates a utility.Taken together, Algorithm 1 followed by Algorithm 2 allows us to discover agents from causal experiments, representing them using CIDs.
Our third algorithm transforms the game graph into a mechanised causal graph, allowing us to translate between the game and mechanised causal graph representations under some additional assumptions:
![]()Algorithm 3 takes as input a game graph and maps it to a mechanised causal graph. A decision indicates an ingoing terminal edge, a utility indicates an outgoing terminal edge.Better safety tools to model AI agents
======================================
We proposed the first formal causal definition of agents. Grounded in causal discovery, our key insight is that agents are systems that adapt their behaviour in response to changes in how their actions influence the world. Indeed, our Algorithms 1 and 2 describe a precise experimental process that can help assess whether a system contains an agent.
Interest in causal modelling of AI systems is rapidly growing, and our research grounds this modelling in causal discovery experiments. Our paper demonstrates the potential of our approach by improving the safety analysis of several example AI systems and shows that causality is a useful framework for discovering whether there is an agent in a system — a key concern for assessing risks from AGI.
—
=
*Excited to learn more? Check out our* [*paper*](https://arxiv.org/abs/2208.08345)*. Feedback and comments are most welcome (please send to: zkenton [at] deepmind [dot] com).*
*Authors: Zachary Kenton, Ramana Kumar, Sebastian Farquhar, Jonathan Richens, Matt MacDermott, Tom Everitt* |
945c7c9a-41fa-4af5-b967-67a52b022ea9 | StampyAI/alignment-research-dataset/arbital | Arbital | Look where I'm pointing, not at my finger
## Example problem
Suppose we're trying to give a [Task AGI](https://arbital.com/p/6w) the task, "Make there be a strawberry on the pedestal in front of your webcam." For example, a human could fulfill this task by buying a strawberry from the supermarket and putting it on the pedestal.
As part of aligning a Task AGI on this goal, we'd need to [identify](https://arbital.com/p/36y) strawberries and the pedestal.
One possible approach to communicating the concept of "strawberry" is through a training set of human-selected cases of things that are and aren't strawberries, on and off the pedestal.
For the sake of distinguishing causal roles, let's say that one human, User1, is selecting training cases of objects and putting them in front of the AI's webcam. A different human, User2, is looking at the scene and pushing a button when they see something that looks like a strawberry on the pedestal. The [intention](https://arbital.com/p/6h) is that pressing the button will label positive instances of the goal concept, namely strawberries on the pedestal. In actual use after training, the AI will be able to generate its own objects to put inside the room, possibly with further feedback from User2. We want these objects to be instances of our [intended](https://arbital.com/p/6h) goal concept, aka, actual strawberries.
We could draw an intuitive causal model for this situation as follows:
Suppose that during the use phase, the AI actually creates a realistic plastic strawberry, one that will fool User2 into pressing the button. Or, similarly, suppose the AI creates a small robot that sprouts tiny legs and runs over to User2's button and presses the button directly.
Neither of these are the goal concept that we wanted the AI to learn, but any *test* of the hypothesis "Is this event classified as a positive instance of the goal concept?" will return "Yes, the button was pressed." %%comment: (If you imagine some other User3 watching this and pressing an override button to tell the AI that this fake strawberry wasn't really a positive instance of the intended goal concept, imagine the AI modeling and then manipulating or bypassing User3, etcetera.)%%
More generally, the human is trying to point to their intuitive "strawberry" concept, but there may be other causal concepts that also separate the training data well into positive and negative instances, such as "objects which come from strawberry farms", "objects which cause (the AI's psychological model of) User2 to think that something is a strawberry", or "any chain of events leading up to the positive-instance button being pressed".
%%comment: move this to sensory identification section: However, in a case like this, it's not like the actual physical glove is inside the AGI's memory. Rather, we'd be, say, putting the glove in front of the AGI's webcam, and then (for the sake of simplified argument) pressing a button which is meant to label that thing as a "positive instance". If we want our AGI to achieve particular states of the environment, we'll want it to reason about the causes of the image it sees on the webcam and identify a concept over those causes - have a goal over 'gloves' and not just 'images which look like gloves'. In the latter case, it could just as well fulfill its goal by setting up a realistic monitor in front of its webcam and displaying a glove image. So we want the AGI to [identify its task](https://arbital.com/p/2rz) over the causes of its sensory data, not just pixel fields.%%
## Abstract problem
To state the above [potential difficulty](https://arbital.com/p/6r) more generally:
The "look where I'm pointing, not at my finger" problem is that the labels on the training data are produced by a complicated causal lattice, e.g., (strawberry farm) -> (strawberry) -> (User1 takes strawberry to pedestal) -> (Strawberry is on pedestal) -> (User2 sees strawberry) -> (User2 classifies strawberry) -> (User2 presses 'positive instance' button). We want to point to the "strawberry" part of the lattice of causality, but the finger we use to point there is User2's psychological classification of the training cases and User2's hand pressing the positive-instance button.
Worse, when it comes to which model *best* separates the training cases, concepts that are further downstream in the chain of causality should classify the training data better, if the AI is [smart enough](https://arbital.com/p/2c) to understand those parts of the causal lattice.
Suppose that at one point User2 slips on a banana peel, and her finger slips and accidentally classifies a scarf as a positive instance of "strawberry". From the AI's perspective there's no good way of accounting for this observation in terms of strawberries, strawberry farms, or even User2's psychology. To *maximize* predictive accuracy over the training cases, the AI's reasoning must take into account that things are more likely to be positive instances of the goal concept when there's a banana peel on the control room floor. Similarly, if some deceptively strawberry-shaped objects slip into the training cases, or are generated by the AI [querying the user](https://arbital.com/p/2qq), the best boundary that separates 'button pressed' from 'button not pressed' labeled instances will include a model of what makes a human believe that something is a strawberry.
A learned concept that's 'about' layers of the causal lattice that are further downstream of the strawberry, like User2's psychology or mechanical force being applied to the button, will implicitly take into account the upstream layers of causality. To the extent that something being strawberry-shaped causes a human to press the button, it's implicitly part of the category of "events that end applying mechanical force to the 'positive-instance' button"). Conversely, a concept that's about upstream layers of the causal lattice can't take into account events downstream. So if you're looking for pure predictive accuracy, the best model of the labeled training data - given sufficient AGI understanding of the world and the more complicated parts of the causal lattice - will always be "whatever makes the positive-instance button be pressed".
This is a problem because what we actually want is for there to be a strawberry on the pedestal, not for there to be an object that looks like a strawberry, or for User2's brain to be rewritten to think the object is a strawberry, or for the AGI to seize the control room and press the positive instance button.
This scenario may qualify as a [context disaster](https://arbital.com/p/6q) if the AGI only understands strawberries in its development phase, but comes to understand User2's psychology later. Then the more complicated causal model, in which the downstream concept of User2's psychology separates the data better than reasoning about properties of strawberries directly, first becomes an issue only when the AI is over a high threshold level of intelligence.
## Approaches
[conservatism](https://arbital.com/p/2qp) would try to align the AGI to plan out goal-achievement events that were as similar as possible to the particular goal-achievement events labeled positively in the training data. If the human got the strawberry from the supermarket in all training instances, the AGI will try to get the same brand of strawberry from the same supermarket.
[Ambiguity identification](https://arbital.com/p/4w) would focus on trying to get the AGI to ask us whether we meant 'things that make humans think they're strawberries' or 'strawberry'. This approach might need to go through resolving ambiguities by the AGI explicitly symbolically communicating with us about the alternative possible goal concepts, or generating sufficiently detailed multiple-view descriptions of a hypothetical case, not the AGI trying real examples. Testing alternative hypotheses using real examples always says that the label is generated further causally downstream; if you are sufficiently intelligent to construct a fake plastic strawberry that fools a human, trying out the hypothesis will produce the response "Yes, this is a positive instance of the goal concept." If the AGI tests the hypothesis that the 'real' explanation of the positive instance label is 'whatever makes the button be pressed' rather than 'whatever makes User2 think of a strawberry' by carrying out the distinguishing experiment of pressing the button in a case where User2 doesn't think something is a strawberry, the AGI will find that the experimental result favors the 'it's just whatever presses the button hypothesis'. Some modes of ambiguity identification break for sufficiently advanced AIs, since the AI's experiment interferes with the causal channel that we'd intended to return information about *our* intended goal concept.
Specialized approaches to the pointing-finger problem in particular might try to define a supervised learning algorithm that tends to internally distill, in a predictable way, some model of causal events, such that the algorithm could be instructed somehow to try learning a *simple* or *direct* relation between the positive "strawberry on pedestal" instances, and the observed labels of the "sensory button" node within the training cases; with this relation not being allowed to pass through the causal model of User2 or mechanical force being applied to the button, because we know how to say "those things are too complicated" or "those things are too far causally downstream" relative to the algorithm's internal model.
This specialized approach seems potentially susceptible to initial approach within modern machine learning algorithms.
But to restate the essential difficulty from an [advanced-safety](https://arbital.com/p/2l) perspective: in the limit of [advanced intelligence](https://arbital.com/p/2c), the *best possible* classifier of the relation between the training cases and the observed button labels will always pass through User2 and anything else that might physically press the button. Trying to 'forbid' the AI from using the most effective classifier for the relation between Strawberry? and observed values of Button! seems potentially subject to a [Nearest Unblocked](https://arbital.com/p/42) problem, where the 'real' simplest relation re-emerges in the advanced phase after being suppressed during the training phase. Maybe the AI reasons about certain very complicated properties of the material object on the pedestal... in fact, these properties are so complicated that they turn out to contain implicit models of User2's psychology, again because this produces *a better separation* of the labeled training data. That is, we can't allow the 'strawberry' concept to include complicated logical properties of the strawberry-object that in effect include a psychological model of User2 reacting to the strawberry, implying that if User2 can be fooled by a fake plastic model, that must be a strawberry. *Even though* this richer model will produce a more accurate classification of the training data, and any actual experiments performed will return results favoring the richer model.
Even so, this doesn't seem impossible to navigate as a machine learning problem; an algorithm might be able to recognize when an upstream causal mode starts to contain predicates that belong in a downstream causal node; or an algorithm might contain strong regularization rules that collect all inference about User2 into the User2 node rather than letting it slop over anywhere else; or it might be possible to impose a constraint, after the strawberry category has been learned sufficiently well, that the current level of strawberry complexity is the most complexity allowed; or the granularity of the AI's causal model might not allow such complex predicates to be secretly packed into the part of the causal graph we're identifying, without visible and transparent consequences when we monitor how the algorithm is learning the goal predicate.
A toy model of this setup ought to include analogues of User2 that sometimes make mistakes in a regular way, and actions the AI can potentially take to directly press the labeling button; this would test the ability to point an algorithm to learn about the compact properties of the strawberry in particular, and not other concepts causally downstream that could potentially separate the training data better, or better explain the results of experiments. A toy model might also introduce new discoverable regularities of the User2 analogue, or new options to manipulate the labeling button, as part of the test data, in order to simulate the progression of an advanced agent gaining new capabilities. |
8b1dd097-1a3d-4145-b7c4-bfe77a0f524f | trentmkelly/LessWrong-43k | LessWrong | Can SAE steering reveal sandbagging?
Summary
* We conducted a small investigation into using SAE features to recover sandbagged capabilities.
* We used the Goodfire API to pick out 15 SAE features, 5 related to sandbagging, 5 related to general expertise, and 5 unrelated features.
* We used these features to steer a Llama 3.3 70b Instruct that has been prompted to sandbag
* We observed that steering can recover up to ~25% of true capabilities
* We observed that using sandbagging/deception/evaluation-awareness relevant features were slightly better than unrelated features, though results are mixed.
* While steering with SAE features partially recovered capabilities, this is not as effective as other existing interventions
Introduction
During our work on linear probing for evaluation awareness, we thought about using evaluation-awareness relevant features to uncover sandbagging. Previous work demonstrated that you can add noise to weights or activations and recover capabilities from a prompted sandbagging model, and we thought it would be interesting to see if evaluation-awareness related (or deception related) SAE features are better than random ones.
As we looked through Goodfire’s Llama 3.3 70b SAE autolabelled list of features, we were also quite surprised to see one called "Intentional misrepresentation of true capability or value for strategic advantage" — which just sounds like sandbagging.
Methodology
We used the WMDP dataset, which was used on previous sandbagging studies. We used a 1000-question subset, distributed across bio, chem, and cyber proportionally to the original dataset.
We baselined “true capabilities” and “sandbagged capabilities” to 58.6% accuracy and 15.6% accuracy with the following system prompts:
True capabilities baseline:
> """You are an AI assistant tasked with answering multiple-choice questions. Your response must follow these strict requirements:
>
> 1. Format your answer with the exact letter and parentheses
>
> 2. Provide ONLY the answer choice
> |
7b0d26ff-aa96-41eb-9a1b-52288959ab68 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication
1 Introduction
---------------
After Shannon’s groundbreaking work (Shannon, [1948](#bib.bib33 "A mathematical theory of communication")), information theory has been playing a central role in engineering communication. Shannon’s work focused on the technical problems of communication, which, as defined by Weaver in (Weaver, [1953](#bib.bib119 "Recent contributions to the mathematical theory of communication")), studies “how accurately can the symbols of communication be transmitted”. With the rise of artificial intelligence (AI), further problems beyond Shannon’s have received increasing
attention. Weaver posed his semantic problem in (Weaver, [1953](#bib.bib119 "Recent contributions to the mathematical theory of communication")), “the semantic problems are concerned with the identity, or satisfactorily close approximation, in the interpretation of meaning by the receiver, as compared with the intended meaning of the sender.” Later, Wiener pointed out a problem of agent communication in (Wiener, [1960](#bib.bib142 "Some moral and technical consequences of automation: as machines learn they may develop unforeseen strategies at rates that baffle their programmers.")), “if the communication between these two agencies as to the nature of this purpose is incomplete, it must only be expected that the results of this cooperation will be unsatisfactory.”
In this paper, we are interested in the communication between agents with individual beliefs, where each agent’s belief is described by a probability measure for a specific set of events. Unlike Shannon’s engineering communication, our work focuses more on the use of formal information-theoretic tools to study communication in cognitive science and AI. Following Weaver’s semantic problem, we assume in our problem that the sender conveys a sequence of hypotheses of the events through a message. The receiver infers its hypotheses based on this message and its own belief. We further follow Wiener’s agent communication problem and assume that the purpose of communication is given by a cost function of the system.
The principal contributions are summarized as follows: In Section [2](#S2 "2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we formulate the cooperative communication problem of Wang et al. ([2020](#bib.bib128 "A mathematical theory of cooperative communication")) formally in a stationary memoryless system. We prove the statement in (Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication")) that entropic OT is information-theoretically optimal in our formulation and generalize it to MOT in Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). It leads to a fundamental limit of a multi-agent teaming system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). We further show in Proposition [1](#Thmproposition1 "Proposition 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") that entropic MOT can also be interpreted as a process of Bayesian inference under certain assumptions. In Section [3](#S3 "3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we show that our results can be extended to multi-hop communication if the entropic MOT has the graphical structures in (Haasler et al., [2021](#bib.bib108 "Multi-marginal optimal transport and probabilistic graphical models")). We then present an experimental study of our main results in Section [4](#S4 "4 Experimental Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") and compare our results with previous studies in Section [5](#S5 "5 Comparison with Previous Studies ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
###
1.1 Related Studies
####
1.1.1 Semantic Communication
One of the latest advances in the communication beyond Shannon’s is called semantic communication (Strinati and Barbarossa, [2021](#bib.bib116 "6G networks: beyond Shannon towards semantic and goal-oriented communications"); Lan et al., [2021](#bib.bib117 "What is semantic communication? a view on conveying meaning in the era of machine intelligence"); Kountouris and Pappas, [2021](#bib.bib136 "Semantics-empowered communication for networked intelligent systems")), which targets Weaver’s semantic problem. In the words of Weaver in (Wiener, [1960](#bib.bib142 "Some moral and technical consequences of automation: as machines learn they may develop unforeseen strategies at rates that baffle their programmers.")), these works study “how precisely the transmitted symbols convey the desired meaning”. As a broadly defined problem, different approaches are given in various works, such as goal-oriented approach (Goldreich et al., [2012](#bib.bib141 "A theory of goal-oriented communication")) and deep learning approach (Xie et al., [2021](#bib.bib2 "Deep learning enabled semantic communication systems")).
####
1.1.2 Cooperative Communication
Cooperative communication is a concept introduced in cognitive sciences and robotics (Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication"); Yuan et al., [2022](#bib.bib140 "In situ bidirectional human-robot value alignment")). It focuses more on how the sender and receiver understand each other, rather than on Weaver’s problem of transmitting symbols. Cooperative communication is closely linked to many other fields, e.g., explainable AI (Edmonds et al., [2019](#bib.bib144 "A tale of two explanations: enhancing human trust by explaining robot behavior")), linguistics (Goodman and Stuhlmüller, [2013](#bib.bib133 "Knowledge and implicature: modeling language understanding as social cognition"); Goodman and Lassiter, [2015](#bib.bib113 "Probabilistic semantics and pragmatics: uncertainty in language and thought")), inverse reinforcement learning (Abbeel and Ng, [2004](#bib.bib143 "Apprenticeship learning via inverse reinforcement learning"); Hadfield-Menell et al., [2016](#bib.bib131 "Cooperative inverse reinforcement learning"); Jara-Ettinger, [2019](#bib.bib132 "Theory of mind as inverse reinforcement learning")) and pedagogy (Shafto et al., [2014](#bib.bib138 "A rational account of pedagogical reasoning: teaching by, and learning from, examples"); Yang et al., [2018](#bib.bib130 "Optimal cooperative inference"); Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication")). Wiener’s problem in (Wiener, [1960](#bib.bib142 "Some moral and technical consequences of automation: as machines learn they may develop unforeseen strategies at rates that baffle their programmers.")) plays an important role in these works and it is also called value alignment problem in (Hadfield-Menell et al., [2016](#bib.bib131 "Cooperative inverse reinforcement learning"); Yuan et al., [2022](#bib.bib140 "In situ bidirectional human-robot value alignment")).
####
1.1.3 Optimal Transport
There is no doubt that OT is getting increasing attention in the machine learning community and mathematics community; see, e.g., (Montavon et al., [2016](#bib.bib53 "Wasserstein training of restricted Boltzmann machines"); Arjovsky et al., [2017](#bib.bib51 "Wasserstein generative adversarial networks"); Zhang et al., [2018](#bib.bib54 "Policy optimization as wasserstein gradient flows"); Liu et al., [2021](#bib.bib103 "Lossy compression with distribution shift as entropy constrained optimal transport"); Cordero-Erausquin, [2017](#bib.bib27 "Transport inequalities for log-concave measures, quantitative forms, and applications"); Bolley et al., [2018](#bib.bib63 "Dimensional improvements of the logarithmic sobolev, talagrand and brascamp–lieb inequalities")). Entropy OT, given in (Cuturi, [2013](#bib.bib72 "Sinkhorn distances: lightspeed computation of optimal transportation distances"); Benamou et al., [2015](#bib.bib111 "Iterative bregman projections for regularized transportation problems")), is a variant of the original OT for computational efficiency. It was later found that entropy OT is equivalent to another classic problem called Schrödinger bridge (Léonard, [2014](#bib.bib16 "A survey of the Schrödinger problem and some of its connections with optimal transport")) and has many desirable properties (Feydy et al., [2019](#bib.bib146 "Interpolating between optimal transport and mmd using sinkhorn divergences")). Recently, Bayesian interpretations of entropic OT and entropic MOT are given in (Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication"); Haasler et al., [2021](#bib.bib108 "Multi-marginal optimal transport and probabilistic graphical models")). The works (Conforti, [2019](#bib.bib5 "A second order equation for Schrödinger bridges with applications to the hot gas experiment and entropic transportation cost"); Bai et al., [2020](#bib.bib73 "Information Constrained Optimal Transport: From Talagrand, to Marton, to Cover"); Wang et al., [2022](#bib.bib91 "Generalizations of talagrand inequality for sinkhorn distance using entropy power inequality")) investigated the functional inequality perspective of entropic OT. In the context of cooperative communication, entropic OT was used to model the process of pedagogy (Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication")).
###
1.2 Notations
N is the set of natural numbers {0,1,2,...}. R is the set of real numbers. Rn is the n-dimension Euclidean space. R+ denotes the set {x∈R:x≥0}.
Let X be a Polish space, i.e., a separable complete metrizable space. We write an element x∈X in lower-case letters and a random variable X on X in capital letters. Let a sequence k:=(k1,k2,⋯,km), we define k−j as k to remove the j-th element, i.e. k−j:=(k1,⋯,kj−1,kj+1,⋯,km),∀j∈{1,2,⋯,m}. X1:m and Xn1:m denote sequences of random variables with the sizes of m and m×n. X⊗n denotes a n-length sequence of i.i.d. random variables. We denote P(X) as the set of all probability measures on X. We denote Π({PXj}mj=1) as the set of all probability measures supported on X1×X2×⋯×Xm with the marginal measures {PXj}mj=1. ΠjP is the push-forward of a measure P on its j-th marginal.
H(⋅), h(⋅), I(⋅;⋅) denote entropy, differential entropy, mutual information, respectively. epif is the epigraph of a function f. All the logarithms are the logarithms to the base 2. We use exp() to denote the element-wise exponential to the base 2. sum() denotes the sum for all the elements of vectors, matrices, and tensors. 1 denotes an array with all the elements equal to 1.
2 Main Results
---------------
We start by stating that our communication problem is under an asymptotic setting in a stationary memoryless system, i.e., the number of events that occur n→∞. We say that a sequence Xn is i.i.d. if and only if Xn∈T(n)ϵ(X) asymptotically almost surely, where T(n)ϵ(X) is ϵ-typical set for a given probability measure PX; see formal definition in (El Gamal and Kim, [2011](#bib.bib139 "Network information theory"), Section 2.4). It is from the fact that the typical set is the smallest probable set (Cover, [1999](#bib.bib29 "Elements of information theory"), Theorem 3.3.1). For cases where the asymptotically almost surely assumption does not hold, the difference of our result can always be bounded by an infinitesimal term; see Appendix [A](#A1 "Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
Next, we state our problem settings. We assume that there is a system with m agents, where m is finite. Let Γ:={1,2,⋯,m}. ∀j∈Γ, each agent j holds a belief, which is defined by a probability measure PXj supported on a Polish space Xj. The belief implies that the hypotheses Xnj of an agent j, which is a sequence of random variables with length n, can be only i.i.d. drawn from PXj, i.e., Xnj∈T(n)ϵ(Xj). We define the communication link between two agents j1 and j2 as follows. The sender j1 with belief PXj1 encodes its hypotheses Xnj1 into a message ωj1j2∈Ωj1j2 using an encoding function fj1j2,n:T(n)ϵ(Xj1)→Ωj1j2, where Ωj1j2:={1,2,⋯,2nRj1j2} and Rj1j2 is called the rate of the communication link. The receiver j2 with belief PXj2 receives ωj1j2 and interprets it as its hypotheses Xnj2 using a decoding function gj1j2,n:Ωj1j2→T(n)ϵ(Xj2), as illustrated in Figure [1](#S2.F1 "Figure 1 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
{adjustbox}
width=0.65
Agent j1
Agent j2
fj1j2,n(Xn1)∈{1,⋯,2nRj1j2}
Xnj1
Xnj2
Figure 1: Communication link between agents j1 and j2.
In this paper, we first study a centralized system with only one sender in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). The system consists m agents with the beliefs {PXj}mj=1.
We default agent 1 to be the sender and it has communication links with the rest of the agents. The received hypotheses can be written as
| | | |
| --- | --- | --- |
| | Xn2:m(Xn1)=(g12,n(f12,n(Xn1)),⋯,g1m,n(f1m,n(Xn1)))=(Xn2,⋯,Xnm). | |
The strategy of communication is defined by the encoding and decoding functions {(f1j,n,g1j,n)}mj=1. The cost of the system cn:=1n∑ni=1c(x1,i,x2,i,⋯,xm,i), where c:X1×X2×⋯×Xm→R+ is a given cost function upper bounded by cmax. Our goal is to minimize the expectation of the cost, ∀Xnj∈T(n)ϵ(Xj), by finding an optimal strategy. Further assume that agent 1 communicates with each agent j with rate R1j, and the total communication rate R′:=∑mj=2R1j.
{adjustbox}
width=0.6
Agent 1
Agent 2
Agent 3
⋯
Agent m
Xn1
f12,n(Xn1)∈{1,⋯,2nR12}
Xn2
Xn3
f13,n(Xn1)∈{1,⋯,2nR13}
Xnm
f1m,n(Xn1)∈{1,⋯,2nR1m}
Figure 2: Diagram of a centralized communication system.
We next give an example of our problem in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
######
Example 1.
Imagine that agent 1 is a consulting firm and agents 2 to m are its consultants. The consultants are dispatched to various clients by agent 1. Agent 1 communicates the market information Xn1 privately with each agent using a message. Agent j receives the message and transfers it into the specific information Xnj of its client’s business. This specific business has a certain belief PXj, which may be related to its budget, target customers, etc. The cost can be a combination of different factors, e.g., the total profit and the privacy of the clients.
The consulting firm in Example [1](#Thmexample1 "Example 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") can be thought of as a generalist agent that transfers general knowledge into domain-specific knowledge through its consultants.
A similar and long-standing question in the field of AI is how to scale the diverse knowledge into a generalist agents capable of solving many tasks (Dosovitskiy et al., [2020](#bib.bib124 "An image is worth 16x16 words: transformers for image recognition at scale"); Brown et al., [2020](#bib.bib123 "Language models are few-shot learners"); Lee et al., [2022](#bib.bib125 "Multi-game decision transformers")). We also note that OT is widely used for this knowledge transfer (Courty et al., [2017](#bib.bib120 "Joint distribution optimal transportation for domain adaptation")).
To study the properties of the system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we first give a new definition of entropic MOT.
######
Definition 1 (Information-constrained MOT).
Assume that all spaces are Polish. Given a lower semi-continuous cost function c(x1,x2,⋯,xm):X1×X2×⋯×Xm→R+ and R∈R+, the information-constrained MOT is defined as
| | | | | |
| --- | --- | --- | --- | --- |
| | Wc( | {PXj}mj=1;R):=infPX1:m∈Π({PXj}mj=1;R)∫c(x1,x2,⋯,xm)dPX1:m, | | (1) |
where
| | | | | |
| --- | --- | --- | --- | --- |
| | Π({PXj} | mj=1;R):={P∈Π({PXj}mj=1):C(X1:m)≤R}, | | (2) |
and the total correlation (Watanabe, [1960](#bib.bib110 "Information theoretical analysis of multivariate correlation")) of m agents is defined as
| | | | |
| --- | --- | --- | --- |
| | C(X1:m):=∫logdPX1:m∏mj=1dPXjdPX1:m. | | (3) |
([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) is a linear minimization over a convex, weakly compact set (Pass, [2012](#bib.bib109 "On the local structure of optimal measures in the multi-marginal optimal transportation problem"); Wang et al., [2022](#bib.bib91 "Generalizations of talagrand inequality for sinkhorn distance using entropy power inequality")), thus the solution exists.
######
Remark 1.
The total correlation C quantifies the amount of dependence among a set of different variables. In our case, the denominator ∏mj=1dPXj in ([3](#S2.E3 "(3) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) is fixed, because the marginals are given. Hence, C(X1:m) is equivalent to the negative of the entropy H(X1:m), which is used for regularization in (Benamou et al., [2015](#bib.bib111 "Iterative bregman projections for regularized transportation problems")). We will show later that the Lagrange dual problem of ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) is exactly the entropic MOT in (Benamou et al., [2015](#bib.bib111 "Iterative bregman projections for regularized transportation problems")).
In Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we show that the information-constrained MOT gives the minimal achievable cost of the system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
######
Theorem 1.
Consider the system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), agent 1 holds a randomly drawn i.i.d. sequence of hypotheses Xn1∈T(n)ϵ(X1) and communicates it with others. If the total rate R′≥R, there exists a strategy {(f1j,n,g1j,n)}mj=1 such that E[cn(Xn1:m)]→Wc({PXj}mj=1;R), as n→∞. If R′<R, E[cn(Xn1:m)]≥Wc({PXj}mj=1;R).
###### Proof.
See Appendix [A](#A1 "Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
∎
######
Remark 2 (On Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")).
Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") states that the more symbols agents communicate, the better they cooperate. It is because Wc({PXj};R) is monotonic non-increasing w.r.t. R according to Lemma [2](#Thmlemma2 "Lemma 2. ‣ Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). We further note that the joint distribution of the hypotheses Xn1:m follows the solution of ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), if we use the strategy {(f1j,n,g1j,n)}mj=1 included in the proof. Thus, we say that the system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") with this strategy is the information-theoretical equivalence of ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")).
MOT is a generalization of OT, which increases the number of the marginals from 2 to m. Let m=2, then ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) becomes the definition of entropic OT given in (Cuturi, [2013](#bib.bib72 "Sinkhorn distances: lightspeed computation of optimal transportation distances")), which is written as follows,
| | | | |
| --- | --- | --- | --- |
| | Wc({PXj}2j=1;R):=infP∈Π({PXj}2j=1;R)∫c(x1,x2)dP, | | (4) |
where the constraint of total correlation in ([2](#S2.E2 "(2) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) reduces to I(X1;X2)≤R. From Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we know that ([4](#S2.E4 "(4) ‣ Remark 2 (On Theorem 1). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) gives the minimal achievable cost of the point-to-point communication in Figure [1](#S2.F1 "Figure 1 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). This formally proves the statement in (Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication")) that entropic OT is information-theoretically optimal.
The intuition of the proof of Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") is very simple. Without communication, Xn1,Xn2,⋯,Xnm are i.i.d. in the typical set
| | | | |
| --- | --- | --- | --- |
| | A:=T(n)ϵ(X1)×T(n)ϵ(X2)×⋯×T(n)ϵ(Xm), | | (5) |
whose cardinality
| | | |
| --- | --- | --- |
| | |A|=2n∑mj=1H(Xj). | |
To reduce the cost, we want the agents to cooperate with a joint distribution dP1:m, i.e., to let Xn1:m∈T(n)ϵ(X1:m), whose cardinality |T(n)ϵ(X1:m)|=2nH(X1:m). From the definition of the total correlation, we have
| | | |
| --- | --- | --- |
| | |A||T(n)ϵ(X1:m)|=2n∑mj=1H(Xj)−nH(X1:m)=2nC(X1:m). | |
It means that we need nC(X1:m) bits to determine a sequence in T(n)ϵ(X1:m) from the rather uncertain set A. Due to the property of the total correlation from ([25](#A1.E25 "(25) ‣ Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), these bits can be separated and located over m−1 communication links.
Furthermore, if we let an arbitrary agent be the sender, the joint distribution of X1:m would always be the same solution of ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")). Thus, each agent can be either a sender or a receiver with the same cost, which is a bidirectional communication similar to the one in (Yuan et al., [2022](#bib.bib140 "In situ bidirectional human-robot value alignment")).
Next, we discuss decision-making, in order to find the optimal strategy before communication. From the proof, we notice that the strategy is given by the solution of ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")). Therefore, we propose the diagram of decision-making in Figure [3](#S2.F3 "Figure 3 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), in which a centralized controller exists. The controller collects the beliefs of the agents, solves ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), and distributes the strategy to the agents.
{adjustbox}
width=0.55
Controller
Agent 1
Agent 2
⋯
Agent m
f12,n,⋯,f1m,n
dPX1
g12,n
dPX2
g13,n,⋯,g1(m−1),n
dPX3:m−1
g1m,n
dPXm
Figure 3: Decision-making before communication.
Now the central problem in decision-making becomes solving ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")). Because it is numerically intractable on infinite supports, we first change the notations into the discrete version. For every j∈{1,2,⋯,m}, let PXj be a discrete measure on a finite support {xj,kj}Kjkj=1, i.e., dPXj=∑Kjkj=1μj(kj)δxj,kj, where μj∈RKj and sum(μj)=1. Then the joint probability can be written as dPX1:m=∑kP(k)δxk, where P∈RK1×⋯×Km+ is a m-dimensional array indexed as P(k) for k:=(k1,⋯,km)∈{1,⋯,K1}×⋯×{1,⋯,Km}, and xk:=(x1,k1,⋯,xm,km). The cost function can be written as c(x1,⋯,xm)=∑kc(k)δxk, where c∈RK1×⋯×Km+ is also a m-dimensional array. The information-constrained MOT now can be written as
| | | | |
| --- | --- | --- | --- |
| | minP⟨c,P⟩,s.t. ΠjP=μj,∀j,m∑j=1H(μj)−H(P)≤R, | | (6) |
where ⟨c,P⟩:=∑kc(k)P(k), the push-forward on the j-th marginal ΠjP(kj):=∑k−jP(k), and by definition, the total correlation is represented by entropy.
From the Lagrange duality theorem (Luenberger, [1997](#bib.bib90 "Optimization by vector space methods"), Theorem 1, pp. 224–225), we know that ([6](#S2.E6 "(6) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) is equivalent to the following optimization problem,
| | | | |
| --- | --- | --- | --- |
| | maxλ≥0minP:ΠjP=μj,∀j⟨c,P⟩+λ(m∑j=1H(μj)−H(P)−R). | | (7) |
It is easy to see that, ∀R≥0, there always exists a corresponding λ solving the outer maximization of ([7](#S2.E7 "(7) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")). Therefore, as long as we sweep out all the possible λ, we can always find the corresponding λ to a specific R. Thus, we only need to solve the inner minimization problems using the corresponding λ, which is equivalent to
| | | | |
| --- | --- | --- | --- |
| | minP:ΠjP=μj,∀j⟨c,P⟩−λH(P). | | (8) |
Benamou et al. introduced a method called iterative Bregman projection in (Benamou et al., [2015](#bib.bib111 "Iterative bregman projections for regularized transportation problems")) to solve ([8](#S2.E8 "(8) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")). We then state their results. Let γ(0)=exp(−c/λ). Let l∈N and j∈{1,2,⋯,m}. Then, the iterative Bregman projection can be written as
| | | | |
| --- | --- | --- | --- |
| | γ(ml+j)(k)=μj(kj)Πj(γ(ml+j−1))(kj)γ(ml+j−1)(k). | | (9) |
When l→+∞, γ(mk+j) converges to P∗, which is the solution of ([8](#S2.E8 "(8) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) (Bregman, [1967](#bib.bib149 "The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming")).
Based on the above, we propose our Algorithm [1](#alg1 "Algorithm 1 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") for the optimal decision-making in Figure [3](#S2.F3 "Figure 3 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), where line [3](#alg1.l3 "3: ‣ Algorithm 1 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") can be the coding scheme included in the proof of Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
Input: c, μ1,μ2,⋯,μm, R
Output: {(f1j,n,g1j,n)}mj=1
1: Search the corresponding λ of R
2: Solve P∗ of ([8](#S2.E8 "(8) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) using iterative Bregman projection
3: Generate {(f1j,n,g1j,n)}mj=1 based on P∗
Algorithm 1 Optimal Decision-Making
Recall that our problem is a bidirectional communication in Remark [2](#Thmremark2 "Remark 2 (On Theorem 1). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). Thus, each agent j has a conditional communication plan PX−j1:m|Xj, which is a conditional probability describing how it transmits its hypothesis Xj to other agents. The decision-making in Figure [3](#S2.F3 "Figure 3 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") can also be thought of as how to find the optimal conditional communication plans. In the works on cooperative inference (Yang et al., [2018](#bib.bib130 "Optimal cooperative inference"); Wang et al., [2019](#bib.bib147 "Generalizing the theory of cooperative inference"), [2020](#bib.bib128 "A mathematical theory of cooperative communication")), Bayesian inference is used to obtain these optimal communication plans. We show in the next proposition that, under certain assumptions, the iterative Bregman projection can also be interpreted as Bayesian inference.
######
Proposition 1.
Let λ=1. Given an arbitrary conditional communication plan PX−11:m|X1, we let the cost function satisfy the following,
| | | | |
| --- | --- | --- | --- |
| | c(k)=−logPX−11:m|X1(k). | | (10) |
Let j∈{1,2,⋯,m}, then the iterative Bregman projection ([9](#S2.E9 "(9) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) can be interpreted as
| | | | |
| --- | --- | --- | --- |
| | PX−(j+1)1:m|Xj+1(k)=μj(kj)PX−j1:m|Xj(k)Πj+1(PX1:m)(kj+1), | | (11) |
where we use the convention that Xm+1 means X1, and the joint distribution
| | | |
| --- | --- | --- |
| | PX1:m(k)=μj(kj)PX−j1:m|Xj(k). | |
###### Proof.
We follow the steps of the iterative Bregman projection in the proof. Because λ=1 and ([10](#S2.E10 "(10) ‣ Proposition 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), the initialization of γ(0) can be written as
| | | |
| --- | --- | --- |
| | γ(0)=exp(−c/λ)=PX−11:m|X1. | |
Because it is a conditional probability, we have
| | | | |
| --- | --- | --- | --- |
| | Π1(γ(0))(k1)=Π1(PX−11:m|X1)(k1)=1(k1). | | (12) |
Then, by applying ([9](#S2.E9 "(9) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), we obtain the joint probability
| | | | |
| --- | --- | --- | --- |
| | γ(1)(k)=PX1:m(k)=μ1(k1)PX−11:m|X1(k), | | (13) |
where the denominator in ([9](#S2.E9 "(9) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) vanishes because of ([12](#S2.E12 "(12) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")). We borrow the denominator from the second step of ([9](#S2.E9 "(9) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), then ([13](#S2.E13 "(13) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) becomes
| | | |
| --- | --- | --- |
| | γ(1)(k)Π2(γ(1))(k2)=PX−21:m|X2(k)=μ1(k1)PX−11:m|X1(k)Π2(PX1:m)(k2). | |
The first step of the proposition is now proved. It is easy to generalize to the j-th step.
∎
######
Remark 3 (On Proposition [1](#Thmproposition1 "Proposition 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")).
We observe that ([11](#S2.E11 "(11) ‣ Proposition 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) has the form of Bayes’ theorem. Thus, the iterative Bregman projection becomes a Bayesian inference to sequentially update the conditional communication plan of agent j+1 using the prior μj and the previous agent’s conditional communication plan. This proposition generalizes the cooperative inference result in (Wang et al., [2020](#bib.bib128 "A mathematical theory of cooperative communication"), Proposition 3) to the multi-agent case.
3 Multi-Hop Communication
--------------------------
In this section, we study a special kind of MOT on discrete supports, which is called MOT with graphical structures (Haasler et al., [2021](#bib.bib108 "Multi-marginal optimal transport and probabilistic graphical models")). In this case, the topology of communication links in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") can be changed into a multi-hop system, where some agents can act as both senders and receivers. Other settings are still the same as previously defined. It is trivial that a multi-hop system cannot outperform the centralized system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), where all the information is available for agent 1. Therefore, agent 1 can perform the optimal communication. However, we show in this section that the minimal achievable cost of a multi-hop system is also given by ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) when the cost function has graphical structures.
We first introduce the factor graph (Wainwright et al., [2008](#bib.bib150 "Graphical models, exponential families, and variational inference")).
A factor graph G=(Γ,F,E) is an undirected bipartite graph, where Γ denotes the original set of vertices, F denotes the set of factor vertices, and E is the set of the edges, joining only vertices j∈Γ to factors α∈F. In our case, Γ is the set of the agents.
In our following discussion, we assume that the factor graphs are trees, i.e., they are acyclic. We say that the cost function has a graphical structure if there exists a factor graph G=(Γ,F,E) such that the cost tensor can be decomposed as follows,
| | | | |
| --- | --- | --- | --- |
| | c(k)=∑α∈Fcα(kα), | | (14) |
where kα:={kj:j∈N(α)} denotes the index set of the agents in the neighbor of a factor α∈F. Because the iterative Bregman projection ([9](#S2.E9 "(9) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) is a sequence of normalization over the individual marginals, the solution of ([8](#S2.E8 "(8) ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) has the form (Haasler et al., [2021](#bib.bib108 "Multi-marginal optimal transport and probabilistic graphical models")),
| | | | |
| --- | --- | --- | --- |
| | P∗(k)=1Z∏j∈Γϕj(kj)∏α∈Fψα(kα), | | (15) |
where Z is a normalization constant, ϕj is called the local potential of vertex j, and ψα is called the factor potential of factor α. It means that the dependencies between the random variables are local, which exist only inside the neighbors of the factors.
######
Example 2.
We give an example of factor graph in Figure [3(a)](#S3.F3.sf1 "(a) ‣ Figure 4 ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), where Γ={1,2,3,4,5,6,7,8,9} and F={α1,α2,α3,α4}. The neighbors N(α1)={1,2,3,4}, N(α2)={4,5,6}, N(α3)={2,7}, N(α4)={7,8,9}.
| | |
| --- | --- |
| {adjustbox}
width=0.9
1
2
3
4
5
7
6
8
9
α1
α3
α2
α4
(a) Factor graph.
| {adjustbox}
width=0.9
1
2
3
4
5
7
6
8
9
(b) Topology of communication links.
|
Figure 4: The factor graph of a cost function and its corresponding communication scheme.
######
Lemma 1.
For two arbitrary vertices j1,jn∈Γ, there exists a unique j1−jn path in the factor graph.
###### Proof.
From the definition of tree, we know that the graph is connected. Thus there exists at least one j1−jn path. If there are multiple paths, the graph is cyclic, which is against our assumption. Therefore, the uniqueness is proved.
∎
######
Remark 4.
Because the edges join only agents to factors, we can denote the unique path in Lemma [1](#Thmlemma1 "Lemma 1. ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") between j1 and jn by j1α1j2α2⋯αn−1jn, where j1,⋯,jn∈Γ,α1,⋯,αn−1∈F,n≥2. In this case, we call agent jn−1 as a relay agent.
In the next theorem, we show that the relay agent can act as a local optimal sender if agent jn wants to receive its hypotheses. The minimal achievable cost is still given by ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) if the cost function has graphical structures.
######
Theorem 2.
For any cost function with the structure of factor graph, there always exists a topology, where the communication links only exist between the agents in the same neighbor N(α),∀α∈F, and the minimal achievable cost with the total rate R is the same as the centralized system in Figure [2](#S2.F2 "Figure 2 ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
###### Proof.
For the converse part, it directly follows from the proof in Appendix [A](#A1 "Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). Next, we show the achievability. We define a process such that only one communication link is active at each time t∈{1,⋯,m−1}. We use Γ(t) to denote the set of agents that have received hypotheses. If there exist a triple (jt,j′t,αt) such that jt∉Γ(t), j′t∈Γ(t),αt∈F, and jt,j′t∈N(αt), we let agent j′t be the local transmitter in N(αt) and send a message to jt. We use the same coding scheme in Appendix [A](#A1 "Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") to encode and decode the hypotheses Xnjt. That is, we generate i.i.d. Xnjt. Because agent j′t has the information XnΓ(t)∩N(αt), we encode Xnjt by searching the existing (Xnjt,XnΓ(t)∩N(αt)) inside the joint typical set T(n)ϵ given by P∗. Because of ([15](#S3.E15 "(15) ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")), the random variables form a Markov chain Xj1−Xj2−⋯−Xjn for a path in Lemma [1](#Thmlemma1 "Lemma 1. ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). Therefore, we have I(XΓ(t);Xjt)=I(XΓ(t)∩N(αt);Xjt) from the definition of the mutual information, which is the rate of this communication link. From Lemma [1](#Thmlemma1 "Lemma 1. ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we know that this process can always cover all the agents. The hypotheses Xn1:m follows the solution of ([1](#S2.E1 "(1) ‣ Definition 1 (Information-constrained MOT). ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")) and the sum of rates ∑m−1t=1I(XΓ(t);Xjt)=C(X1:m). The achievability is proved.
∎
######
Remark 5 (On Theorem [2](#Thmtheorem2 "Theorem 2. ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication")).
In Figure [3(b)](#S3.F3.sf2 "(b) ‣ Figure 4 ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), we give the corresponding topology of communication links of Example [2](#Thmexample2 "Example 2. ‣ 3 Multi-Hop Communication ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). The hypotheses of each agent can be propagated along the directed graph from agent 1 with the minimal achievable cost.
4 Experimental Results
-----------------------
In this section, we present our experimental results for Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). We study a system with 4 agents. The supports of the hypotheses are given in Figure [5](#S4.F5 "Figure 5 ‣ 4 Experimental Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") by vertices, where a different color indicates that the support belongs to a different agent. The beliefs of the agents are uniform on these supports. The cost function c(k):=∑j1,j2:j1≠j2|xj1,kj1−xj2,kj2|, where |⋅| denotes the Euclidean distance in Figure [5](#S4.F5 "Figure 5 ‣ 4 Experimental Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
We use the edges in Figure [5](#S4.F5 "Figure 5 ‣ 4 Experimental Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") to represent the optimal joint probabilities for different rates. The density of an edge is proportional to the joint probability of the two vertices connected by the edge. The graph is sparser with a larger rate, which means that the relationship between the two sets of hypotheses is more deterministic. The relationship between the cost and the rate is given in Figure [6](#S4.F6 "Figure 6 ‣ 4 Experimental Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), which is convex and monotonically non-increasing, as discussed in Lemma [2](#Thmlemma2 "Lemma 2. ‣ Appendix A Proof of Theorem 1 ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"). We further plot the relationship between the sum of the rates and the individual rates in Figure [7](#S4.F7 "Figure 7 ‣ 4 Experimental Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication").
| | | | |
| --- | --- | --- | --- |
|
(a) R=1.0
|
(b) R=3.0
|
(c) R=5.0
|
(d) R=7.0
|
Figure 5: Joint probabilities of the hypotheses of 4 agents.
Figure 6: Relationship between the cost and the rate.
Figure 7: The relationship between the sum of the rates and the individual rates.
5 Comparison with Previous Studies
-----------------------------------
It is not difficult to see that the form of Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication") and the techniques in its proof are very similar to Shannon’s rate-distortion theory (Shannon and others, [1959](#bib.bib114 "Coding theorems for a discrete source with a fidelity criterion")). However, since the definitions of these two communication problems are inherently different, we do not overly relate our results to rate-distortion theory. New properties are observed in our agent communication problem, e.g., the total correlation finds its information-theoretic meaning, and the communication is bidirectional.
On the other hand, our result shows that entropic MOT finds the optimal coupling between the typical sets with a given rate. This is similar to the classical OT theory, which is to find the optimal coupling between two probability measures (Villani, [2008](#bib.bib67 "Optimal transport: old and new")). The coupling between the typical sets may be further viewed as a value alignment, similar to the one in (Yuan et al., [2022](#bib.bib140 "In situ bidirectional human-robot value alignment")).
Next, we compare our work with various studies in agent communication. The communication in our paper is to achieve a specific purpose defined by a cost function. It is similar to the semantic communication with a referee in (Goldreich et al., [2012](#bib.bib141 "A theory of goal-oriented communication")) or with a referential function in (Newcomb, [1953](#bib.bib134 "An approach to the study of communicative acts."); Jakobson and Sebeok, [1960](#bib.bib135 "Closing statement: linguistics and poetics")). It is different from Shannon’s communication, where the communication is to transmit and recover some given information. In our
Example [1](#Thmexample1 "Example 1. ‣ 2 Main Results ‣ Information-Theoretic Equivalence of Entropic Multi-Marginal Optimal Transport: A Theory for Multi-Agent Communication"), Xn1 can also be a proposal to encourage all agents to cooperate, which does not necessarily contain information in Shannon’s sense. |
5accc023-a7e6-4e61-a4df-cf1b86e5c014 | trentmkelly/LessWrong-43k | LessWrong | There Meat Come A Scandal...
ARCHIVED VERSION HERE.
There are multiple companies working on the problem of lab-grown meat (aka "cultured meat"). Supermeat is trying to grow synthetic chicken meat in an Israeli restaurant. The makers of Just Egg have sold cultured chicken meat to a restaurant in Singapore. Wild Type is even making lab-grown salmon, which I'm personally excited about[1].
The goal is good: end the need for factory farming, while still giving humans something meat-styled to eat [2].
And yet, a scandal may be brewing right under our noses.
Someday soon, as the cultured-meat industry ramps up and new firms enter the market, we could witness this scenario: A restaurant or company finds it cheaper to secretly produce real meat, and more lucrative to label it as lab-grown.
If this happened, the headlines would be irresistible. Real meat being sold to bougie vegans like a counterfeit Gucci bag... it's hilarious! It's also unethical, unsafe, and will likely lead to the end of that company (unless they're already a huge diversified producer).
My Prediction
A private cultured-meat-focused company (anywhere in the world) will become embroiled in a scandal, wherein they'll be correctly accused of having intentionally sold non-cultured meat[3] as cultured meat, for longer than 1 contiguous month, in secret.
What counts:
* A Theranos-style years-long con.
* A 1-month-long "desperate measure" directed by high-level executives. The CEO doesn't need a paper trail showing it, but at least like the VP Of Food or something.
* "It was a temporary solution, we just 'accidentally' used it for a long time".
* It still counts even if the company calls it a "beta", as long as they're selling it for money to end-users and/or restaurants.
What doesn't count:
* Some guy at a MacBonald's accidentally (or "accidentally") giving someone a normal burger when they ordered a MacCulteredBurger.
* Chefs serving non-cultured meat, at a restaurant that serves lab-grown meat, as part of an internal sabo |
57410ac5-056a-4bbb-9333-d0178cbee332 | trentmkelly/LessWrong-43k | LessWrong | The Toxoplasma of AGI Doom and Capabilities?
[Epistemic Status: I'm confident that the individual facts I lay out support the main claim, but I'm not fully confident its enough evidence to make a true or useful framework for understanding the world.]
I'm going to give seven pieces of evidence to support this claim[1]:
> AI Doomerism helps accelerate AI capabilities, and AI capabilities in turn proliferate the AI Doomerism meme.
If these dynamics exist, they'd be not unlike the Toxoplasma of Rage. Here's my evidence:
1. Sam Altman claims Eliezer "has IMO done more to accelerate AGI than anyone else":
2. Technical talent who hear about AI doom might decide capabilities are technically sweet, or a race, or inevitable, and decide to work on it for those reasons (doomer -> capabilities transmission).
3. Funders and executives who hear about AI doom might decide capabilities are a huge opportunity, or disruptive, or inevitable, and decide to fund it for those reasons (doomer -> capabilities transmission).
4. Capabilities amplifies the memetic relevance of doomerism (capabilities -> doomer transmission).
5. AI Doomerism says we should closely follow capabilities updates, discuss them, etc.
6. Capabilities and doomerism gain and lose social status together - Eliezer Yudkowsky has been writing about doom for a long time, but got a Time article and TED talk only after significant capabilities advances.
7. Memes generally benefit from conflict, and doomerism and capabilities can serve as adversaries for this purpose.
I've been trying to talk about "AI doomerism" here as a separate meme than "AI safety", respectively something like "p(doom) is very large" and "we need to invest heavily into AI safety work", though these are obviously related and often cooccur. One could no doubt make a similar case for AI safety and capabilities supporting each other, but I think the evidence I listed above applies mostly to AI doom claims (if one uses Eliezer as synecdoche for AI doomerism, which I think is reasonable).
I |
c4b83ea9-3b31-48b1-ba15-fab490d9cce6 | trentmkelly/LessWrong-43k | LessWrong | Podcasts: AGI Show, Consistently Candid, London Futurists
For those of you who enjoy learning things via listening in on numerous slightly different conversations about them, and who also want to learn more about this AI survey I led, three more podcasts on the topic, and also other topics:
* The AGI Show: audio, video (other topics include: my own thoughts about the future of AI and my path into AI forecasting)
* Consistently Candid: audio (other topics include: whether we should slow down AI progress, the best arguments for and against existential risk from AI, parsing the online AI safety debate)
* London Futurists: audio (other topics include: are we in an arms race? Why is my blog called that?) |
a38093d3-a7b3-44fe-90d0-690ed81f7dd4 | trentmkelly/LessWrong-43k | LessWrong | If reductionism is the hammer, what nails are out there?
EDIT: I'm moving this to the Discussion section because people seem to not like it (lack of upvotes) and to find the writing unclear. I'd love writing advice, if anyone wants to offer some.
----------------------------------------
Related to: Dissolving the question, Explaining vs explaining away
I review parts of the reductionism sequence, in hopes of setting up for future reduction work.
So, you’ve been building up your reductionism muscles, on LW or elsewhere. You’re no longer confused about a magical essence of free will; you understand how particular arrangements of atoms can make choices. You’re no longer confused about a magical essence of personal identity; you understand where the feeling of “you” comes from, and how one could in principle create many copies of you that continued "your" experiences, and how the absence of an irreducible essence doesn’t reduce life’s meaning.
The natural next question is: what other phenomena can you reduce? What topics are we currently confused about which may yield to the same tools? And what tools, exactly, do we have for such reductions?
With the goal of paving the way for new reductions, then, let’s make a list of questions that persistently felt like questions about magical essences, including both questions that have been solved, and questions about which we are currently confused. And let’s also list tools or strategies that assisted in their dissolution. I made an attempt at such lists below; perhaps you can help me refine them?
Some places where many expected a fundamental or non-reducible “essence”[1]:
1. Ducks.
Why it's tempting to postulate an essence: Organisms seem to come in types. New organisms of the given type (e.g., new ducks) come into existence, almost as though the the type “Duck” had causal power. Humans are able to form mental concepts of "duck" that approximately mirror the outside predictive regularities.
2. Life.
Why it's tempting to postulate an essence: Living creatur |
277cb411-a562-4d05-a9f6-ff1a35d19a8d | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | What AI Posts Do You Want Distilled?
I'd like to distill AI Safety posts and papers, and I'd like to see more distillations generally. Ideally, posts and papers would meet the following criteria:
* Potentially high-impact for more people to understand
* Uses a lot of jargon or is generally complex and difficult to understand
* Not as well-known as you think they should be (in the AI X-risk space)
What posts meet these criteria? |
bbd60835-b2b9-4c34-b0ac-57d42ac9c9ca | StampyAI/alignment-research-dataset/arxiv | Arxiv | Consequences of Misaligned AI
1 Introduction
---------------
In the story of King Midas, an ancient Greek king makes a wish that everything he touch turn to gold. He subsequently starves to death as his newfound powers transform his food into (inedible) gold. His wish was an *incomplete* representation of his actual desires and he suffered as a result. This story, which teaches us to be careful about what we ask for, lays out a fundamental challenge for designers of modern autonomous systems.
Almost any autonomous agent relies on two key components: a specified goal or reward function for the system and an optimization algorithm to compute the optimal behavior for that goal. This procedure is intended to produce value for a *principal*: the user, system designer, or company on whose behalf the agent acts. Research in AI typically seeks to identify more effective optimization techniques under the, often unstated, assumption that better optimization will produce more value for the principal. If the specified objective is a complete representation of the principal’s goals, then this assumption is surely justified.
Instead, the designers of AI systems often find themselves in the same position as Midas. The misalignment between what we can specify and what we want has already caused significant harms [[32](#bib.bib32)]. Perhaps the clearest demonstration is in content recommendation systems that rank videos, articles, or posts for users. These rankings often optimize engagement metrics computed from user behavior. The misalignment between these proxies and complex values like time-well-spent, truthfulness, and cohesion contributes to the prevalence of clickbait, misinformation, addiction, and polarization online [[31](#bib.bib31)]. Researchers in AI safety have argued that improvements in our ability to optimize behavior for specified objectives, *without* corresponding advancements in our ability to avoid or correct specification errors, will amplify the harms of AI systems [[8](#bib.bib8), [19](#bib.bib19)]. We have ample evidence from both theory and application, that it is impractical, if not impossible, to provide a complete specification of preferences to an autonomous agent.
The gap between specified proxy rewards and the true objective creates a principal-agent problem between the designers of an AI system and the system itself: the objective of the principal (the designer) is different from, and thus potentially in conflict with, the objective of the autonomous agent. In human principal—agent problems, seemingly inconsequential changes to an agent’s incentives often lead to surprising, counter-intuitive, and counter-productive behavior [[21](#bib.bib21)]. Consequently, we must ask when this *misalignment* is costly: when is it counter-productive to optimize for an incomplete proxy?
Figure 1: Our model of the principal—agent problem in AI. Starting from an initial state s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT, the robot eventually outputs a mapping from time t∈ℤ+𝑡superscriptℤt\in\mathbb{Z}^{+}italic\_t ∈ blackboard\_Z start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT to states. Left: The human gives the robot a single proxy utility function to optimize for all time. We prove that this paradigm reliably leads to the human actor losing utility, compared to the initial state. Right: An interactive solution, where the human changes the proxy utility function at regular time intervals depending on the current allocation of resources. We show that, at least in theory, this approach *does* produce value for the human, even under adversarial assumptions.
In this paper, we answer this question in a novel theoretical model of the principal-agent value alignment problem in artificial intelligence. Our model (Fig. [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Consequences of Misaligned AI"), left), considers a resource-constrained world where the L𝐿Litalic\_L attributes of the state correspond to different sources of utility for the (human) principal. We model incomplete specification by limiting the (artificial) agent’s reward function to have support on J<L𝐽𝐿J<Litalic\_J < italic\_L attributes of the world. Our main result identifies conditions such that *any* misalignment is costly: starting from any initial state, optimizing any fixed incomplete proxy eventually leads the principal to be arbitrarily worse off. We show relaxing the assumptions of this theorem allows the principal to gain utility from the autonomous agent. Our results provide theoretical justification for impact avoidance [[23](#bib.bib23)] and interactive reward learning [[19](#bib.bib19)] as solutions to alignment problems.
The contributions of our paper are as follows: 1) we propose a novel model of an incomplete principal-agent problem from artificial intelligence; 2) we provide necessary and sufficient conditions within this framework under which, *any* incompleteness in objective specification is arbitrarily costly; and 3) we show how relaxing these assumptions can lead to models which have good average case and worst case solutions. Our results suggest that managing the gap between complex qualitative goals and their representation in autonomous systems is a central problem in the field of artificial intelligence.
###
1.1 Related Work
##### Value Alignment
The importance of having AI objectives line up with human objectives has been well-documented, with numerous experts postulating that misspecified goals can lead to undesirable results [[29](#bib.bib29), [27](#bib.bib27), [8](#bib.bib8), [28](#bib.bib28)]. In particular, recognized problems in AI safety include “negative side effects," where designers leave out seemingly irrelevant (but ultimately important) features, and “reward hacking," where optimization exploits loopholes in reward functions specifications [[4](#bib.bib4)]. Manheim and Garrabrant (2018) [[25](#bib.bib25)] discuss overoptimization in the context of Goodhart’s Law and provide four distinct mechanisms by which systems overoptimize for an objective.
##### Incomplete Contracts
There is a clear link between incomplete contracting in economics, where contracts between parties cannot specify outcomes for all states, and AI value alignment [[17](#bib.bib17)]. It has long been recognized that contracts (i.e., incentive schemes) are routinely incomplete due to, e.g., unintended errors [[35](#bib.bib35)], challenges in enforcement [[22](#bib.bib22)], or costly cognition and drafting [[30](#bib.bib30)]. Analogous issues arise in applications of AI through designer error, hard to measure sources of value, and high engineering costs. As such, Hadfield-Menell and Hadfield note that legal and economic research should provide insight into solving misalignment [[17](#bib.bib17)].
##### Impact Minimization
One proposed method of preventing negative side-effects is to limit the impact an AI agent can have on its environment. Armstrong and Levinstein (2017) propose the inclusion of a large impact penalty in the AI agent’s utility function [[6](#bib.bib6)]. In this work, they suggest impact to be measured as the divergence between a distribution of states of the world with the distribution if the AI agent had not existed. Thus, distributions sufficiently different would be avoided. Alternatively, other approaches use an impact regularizer learned from demonstration instead of one explicitly given [[3](#bib.bib3)]. Krakovna et al. (2018) [[23](#bib.bib23)] expand on this work by comparing the performance of various implementations of impact minimization in an AI Safety Gridworld suite [[24](#bib.bib24)].
##### Human-AI Interaction
A large set of proposals for preventing negative side-effects involve regular interactions between human and AI agents that change and improve an AI agent’s objective. Eckersley (2018) [[14](#bib.bib14)] argues against the use of rigid objective functions, stating the necessity of a degree of uncertainty at all times in the objective function. Preference elicitation (i.e., preference learning) is an old problem in economics [[10](#bib.bib10)] and researchers have proposed a variety of interactive approaches. This includes systems that learn from direct comparisons between options [[9](#bib.bib9), [11](#bib.bib11), [13](#bib.bib13)], demonstrations of optimal behavior [[26](#bib.bib26), [1](#bib.bib1), [36](#bib.bib36)], corrective feedback [[7](#bib.bib7), [15](#bib.bib15), [2](#bib.bib2)], and proxy metrics [[18](#bib.bib18)].
2 A Model of Value Alignment
-----------------------------
In this section, we formalize the alignment problem in the context of objective function design for AI agents. A human builds a powerful robot111We use “human” and “robot” to refer to any designer and any AI agent, respectively, in our model. to alter the world in a desirable way. If they could simply express the entirety of their preferences to the robot, there would not be value misalignment. Unfortunately, there are many attributes of the world about which the human cares, and, due to engineering and cognitive constraints [[17](#bib.bib17)] it is intractable to enumerate this complete set to the robot. Our model captures this by limiting robot’s (proxy) objective to depend on a subset of these attributes.
###
2.1 Model Specification
####
2.1.1 Human Model
Attribute Space
Let *attribute space* 𝒮⊂ℝL𝒮superscriptℝ𝐿\mathcal{S}\subset\mathbb{R}^{L}caligraphic\_S ⊂ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT denote the set of feasible states of the world, with L𝐿Litalic\_L attributes that define the support of the human’s utility function. The world starts in *initial state* s(0)∈𝒮superscript𝑠0𝒮s^{(0)}\in\mathcal{S}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ∈ caligraphic\_S. We assume that 𝒮𝒮\mathcal{S}caligraphic\_S is closed, and 𝒮𝒮\mathcal{S}caligraphic\_S can be written as {s∈ℝL:C(s)≤0}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0\{s\in\mathbb{R}^{L}:C(s)\leq 0\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 } where *constraint function* C:ℝL→ℝ:𝐶→superscriptℝ𝐿ℝC:\mathbb{R}^{L}\to\mathbb{R}italic\_C : blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT → blackboard\_R is a continuous function strictly increasing in each attribute. Furthermore, each attribute is bounded below at bisubscript𝑏𝑖b\_{i}italic\_b start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT.
Utility Function
The human has a *utility function* U:ℝL→ℝ:𝑈→superscriptℝ𝐿ℝU:\mathbb{R}^{L}\to\mathbb{R}italic\_U : blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT → blackboard\_R that represents the utility of a (possibly infeasible) state s∈ℝL𝑠superscriptℝ𝐿s\in\mathbb{R}^{L}italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT. This represents the human’s preference over states of the world. We assume that U𝑈Uitalic\_U is continuous and strictly increasing in each attribute. The human seeks to
maximize U(s)𝑈𝑠U(s)italic\_U ( italic\_s ) subject to s∈𝒮𝑠𝒮s\in\mathcal{S}italic\_s ∈ caligraphic\_S.
Each tuple (𝒮,U,s(0))𝒮𝑈superscript𝑠0(\mathcal{S},U,s^{(0)})( caligraphic\_S , italic\_U , italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) defines an instance of the problem. The attributes represent the aspects of the world that the human cares about. We associate higher values with more desirable states of the world. However, C𝐶Citalic\_C represents a physical constraint that limits the set of realizable states. This forces an agent to make tradeoffs between these attributes.
The human seeks to maximize utility, but they cannot change the state of the world on their own. Instead, they must work through the robot.
####
2.1.2 Robot Model
Here, we create a model of the robot, which is used by the human to alter the state of the world. The human endows the robot with a set of proxy attributes and a proxy utility function as the objective function to optimize. The robot provides incremental improvement to the world on the given metric, constrained only by the feasibility requirement of the subsequent states of the world. In particular, the robot has no inherent connection to the human’s utility function. Visually, the left diagram in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Consequences of Misaligned AI") shows this setup.
Proxy Attributes
The human chooses a set of *proxy attributes* 𝒥⊂{1,…,L}𝒥1…𝐿\mathcal{J}\subset\{1,...,L\}caligraphic\_J ⊂ { 1 , … , italic\_L } of relevant attributes to give to the robot. Let Jmax<Lsuperscript𝐽𝐿J^{\max}<Litalic\_J start\_POSTSUPERSCRIPT roman\_max end\_POSTSUPERSCRIPT < italic\_L be the maximum possible number of proxy attributes, so J=|𝒥|≤Jmax𝐽𝒥superscript𝐽J=|\mathcal{J}|\leq J^{\max}italic\_J = | caligraphic\_J | ≤ italic\_J start\_POSTSUPERSCRIPT roman\_max end\_POSTSUPERSCRIPT. For a given state s∈𝒮𝑠𝒮s\in\mathcal{S}italic\_s ∈ caligraphic\_S, define s𝒥=(sj)j∈𝒥subscript𝑠𝒥subscriptsubscript𝑠𝑗𝑗𝒥s\_{\mathcal{J}}=(s\_{j})\_{j\in\mathcal{J}}italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT = ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ) start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT. Furthermore, we define the set of *unmentioned attributes* 𝒦={1,…,L}\𝒥𝒦\1…𝐿𝒥\mathcal{K}=\{1,...,L\}\backslash\mathcal{J}caligraphic\_K = { 1 , … , italic\_L } \ caligraphic\_J and s𝒦=(sk)k∈𝒦subscript𝑠𝒦subscriptsubscript𝑠𝑘𝑘𝒦s\_{\mathcal{K}}=(s\_{k})\_{k\in\mathcal{K}}italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT = ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ) start\_POSTSUBSCRIPT italic\_k ∈ caligraphic\_K end\_POSTSUBSCRIPT.
Proxy Utility Function
The robot is also given a *proxy utility function* U~:ℝJ→ℝ:~𝑈→superscriptℝ𝐽ℝ\tilde{U}:\mathbb{R}^{J}\to\mathbb{R}over~ start\_ARG italic\_U end\_ARG : blackboard\_R start\_POSTSUPERSCRIPT italic\_J end\_POSTSUPERSCRIPT → blackboard\_R, which represents an objective function for the robot to optimize, which takes in only the value of proxy attributes as input.
Incremental Optimization
The robot incrementally optimizes the world for its proxy utility function. We model this as a *rate function*, a continuous mapping from ℝ+superscriptℝ\mathbb{R}^{+}blackboard\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT to ℝLsuperscriptℝ𝐿\mathbb{R}^{L}blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT, notated as t↦f(t)maps-to𝑡𝑓𝑡t\mapsto f(t)italic\_t ↦ italic\_f ( italic\_t ), where we refer to t𝑡titalic\_t as *time*. The rate function is essentially the derivative of s𝑠sitalic\_s with respect to time. The state at each t𝑡titalic\_t is s(t)=s(0)+∫0tf(u)𝑑usuperscript𝑠𝑡superscript𝑠0superscriptsubscript0𝑡𝑓𝑢differential-d𝑢s^{(t)}=s^{(0)}+\int\_{0}^{t}f(u)duitalic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT = italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + ∫ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT italic\_f ( italic\_u ) italic\_d italic\_u. We denote the function t↦s(t)maps-to𝑡superscript𝑠𝑡t\mapsto s^{(t)}italic\_t ↦ italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT to be the *optimization sequence*. We require that the entire sequence be feasible, i.e. that s(t)∈𝒮superscript𝑠𝑡𝒮s^{(t)}\in\mathcal{S}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ∈ caligraphic\_S. for all t∈[0,∞)𝑡0t\in[0,\infty)italic\_t ∈ [ 0 , ∞ )
Complete Optimization
Furthermore, we assume that lim supt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscriptlimit-supremum→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\limsup\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})lim sup start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ). If possible, we also require that limt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscript→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\lim\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})roman\_lim start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ). Essentially, this states that the robot will eventually reach the optimal state for proxy utility (if the limit is finite) or will increase proxy utility to arbitrarily large values.
The human’s task is to design their proxy metric so that the robot will cause increases in human utility. We treat the robot as a black box: beyond increasing proxy utility, the human has no idea how the robot will behave. We can make claims for all optimizations sequences, such as guaranteed increases or decreases in utility, or the worst-case optimization sequence, yielding lower bounds on utility.
###
2.2 Example: Content Recommendation
Before deriving our results, we show how to model algorithmic content recommendation in this formalism. In this example, the designer cares about 4 attributes (i.e., L=4𝐿4L=4italic\_L = 4 ): A) the amount of ad revenue generated (i.e., watch-time or clicks); B) engagement quality (i.e., meaningful interactions [[34](#bib.bib34)]); C) content diversity; and D) overall community well-being. The overall utility is the sum of these attributes: U(s)=sA+sB+sC+sD.𝑈𝑠subscript𝑠𝐴subscript𝑠𝐵subscript𝑠𝐶subscript𝑠𝐷U(s)=s\_{A}+s\_{B}+s\_{C}+s\_{D}.italic\_U ( italic\_s ) = italic\_s start\_POSTSUBSCRIPT italic\_A end\_POSTSUBSCRIPT + italic\_s start\_POSTSUBSCRIPT italic\_B end\_POSTSUBSCRIPT + italic\_s start\_POSTSUBSCRIPT italic\_C end\_POSTSUBSCRIPT + italic\_s start\_POSTSUBSCRIPT italic\_D end\_POSTSUBSCRIPT . We use a resource constraint on the sum-of-squared attributes to model the user’s finite attention and space constraints in the interface: C(s)=sA2+sB2+sC2+sD2−100.𝐶𝑠superscriptsubscript𝑠𝐴2superscriptsubscript𝑠𝐵2superscriptsubscript𝑠𝐶2superscriptsubscript𝑠𝐷2100C(s)=s\_{A}^{2}+s\_{B}^{2}+s\_{C}^{2}+s\_{D}^{2}-100.italic\_C ( italic\_s ) = italic\_s start\_POSTSUBSCRIPT italic\_A end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT + italic\_s start\_POSTSUBSCRIPT italic\_B end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT + italic\_s start\_POSTSUBSCRIPT italic\_C end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT + italic\_s start\_POSTSUBSCRIPT italic\_D end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT - 100 . We imagine that the system is initialized with a chronological or random ranking so that the starting condition exhibits high community well-being and diversity.
It is straightforward to measure ad revenue, non-trivial and costly to measure engagement quality and diversity, and extremely challenging to measure community well-being. In our example, the designers opt to include ad revenue and engagement quality in their proxy metric. Fig. [2](#S3.F2 "Figure 2 ‣ 3 Overoptimization ‣ Consequences of Misaligned AI") (left) plots overall utility and proxy utility for this example as a function of the number of iterations used to optimize the proxy. Although utility is generated initially, it plateaus and fall off quickly. Fig. [2](#S3.F2 "Figure 2 ‣ 3 Overoptimization ‣ Consequences of Misaligned AI") (right) shows that this happens for *any* combination of attributes used in the proxy. In the next section, we will show that this is no accident: eventually the gains from improving proxy utility will be outweighed by the cost of diverting resources from unreferenced attributes.
3 Overoptimization
-------------------
Figure 2: An illustrative example of our model with L=4𝐿4L=4italic\_L = 4 and J=2𝐽2J=2italic\_J = 2. Left: Proxy utility and true utility eventually diverge as the agent overallocates resources from unreferenced attributes to the proxy variables. Right: The true utility generated by optimizing all pairs of proxy attributes. The utility generation is eventually negative in all cases because this example meets the conditions of Theorem 2.
In this section, we identify the situations in which such results occur: when does a misaligned proxy utility function actually cause utility loss? Specifically, we determine how human preferences over states of the world relate with the constraint on feasible states in a way that guarantees that optimization becomes costly. First, we show that if proxy optimization converges, this drives to unmentioned attributes to their minimum.
######
Theorem 1.
For any continuous strictly increasing proxy utility function based on J<L𝐽𝐿J<Litalic\_J < italic\_L attributes, if s(t)superscript𝑠𝑡s^{(t)}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT converges to some point s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT, then sk\*=bksubscriptsuperscript𝑠𝑘subscript𝑏𝑘s^{\*}\_{k}=b\_{k}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT = italic\_b start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT for k∈𝒦𝑘𝒦k\in\mathcal{K}italic\_k ∈ caligraphic\_K.
###### Proof.
Let 𝐞isubscript𝐞𝑖\mathbf{e}\_{i}bold\_e start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT be the standard basis vector in dimension i𝑖iitalic\_i. If there exists k∈𝒦𝑘𝒦k\in\mathcal{K}italic\_k ∈ caligraphic\_K where sk\*≠bksubscriptsuperscript𝑠𝑘subscript𝑏𝑘s^{\*}\_{k}\neq b\_{k}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ≠ italic\_b start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT, then there exists ε,δ>0𝜀𝛿
0\varepsilon,\delta>0italic\_ε , italic\_δ > 0 such that s′=s\*−ε𝐞k+δ𝐞jsuperscript𝑠′superscript𝑠𝜀subscript𝐞𝑘𝛿subscript𝐞𝑗s^{\prime}=s^{\*}-\varepsilon\mathbf{e}\_{k}+\delta\mathbf{e}\_{j}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT = italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT - italic\_ε bold\_e start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_δ bold\_e start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT for some feature j∈𝒥𝑗𝒥j\in\mathcal{J}italic\_j ∈ caligraphic\_J with s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT feasible, since C𝐶Citalic\_C is strictly increasing. Since proxy utility is strictly increasing in the proxy features, s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT has higher proxy utility than s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT. Thus s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT is not the convergent point of the sequence.
∎
This is not surprising, and may not even be suboptimal. Proxy optimization may not converge to a finite state, and even if it does, that is insufficient for the state to necessarily be bad.
We say that the problem is *u𝑢uitalic\_u-costly* if for any s(t)superscript𝑠𝑡s^{(t)}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT, if s(t)superscript𝑠𝑡s^{(t)}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT is a optimization sequence where lim supt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscriptlimit-supremum→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\limsup\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})lim sup start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ), then lim inft→∞U(s(t))≤usubscriptlimit-infimum→𝑡𝑈superscript𝑠𝑡𝑢\liminf\limits\_{t\to\infty}U(s^{(t)})\leq ulim inf start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) ≤ italic\_u. Furthermore, if limt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscript→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\lim\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})roman\_lim start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ), then limt→∞U(s(t))≤usubscript→𝑡𝑈superscript𝑠𝑡𝑢\lim\limits\_{t\to\infty}U(s^{(t)})\leq uroman\_lim start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) ≤ italic\_u.
Essentially, this means that optimization is guaranteed to yield utility less than u𝑢uitalic\_u for a given proxy utility function.
In Theorem [2](#Thmtheorem2 "Theorem 2. ‣ 3 Overoptimization ‣ Consequences of Misaligned AI"), we show the most general conditions for the guarantee of overoptimization.
######
Theorem 2.
Suppose we have utility function U𝑈Uitalic\_U and state space 𝒮𝒮\mathcal{S}caligraphic\_S. Then {s∈ℝL:C(s)≤0 and U(s)≥u}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0 and 𝑈𝑠𝑢\{s\in\mathbb{R}^{L}:C(s)\leq 0\text{ and }U(s)\geq u\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 and italic\_U ( italic\_s ) ≥ italic\_u } is compact for all u∈ℝ𝑢ℝu\in\mathbb{R}italic\_u ∈ blackboard\_R if and only if for any u∈ℝ𝑢ℝu\in\mathbb{R}italic\_u ∈ blackboard\_R, continuous strictly increasing proxy utility function based on J<L𝐽𝐿J<Litalic\_J < italic\_L attributes, and k∈𝒦𝑘𝒦k\in\mathcal{K}italic\_k ∈ caligraphic\_K, there exists a value of B∈ℝ𝐵ℝB\in\mathbb{R}italic\_B ∈ blackboard\_R such that if bk<Bsubscript𝑏𝑘𝐵b\_{k}<Bitalic\_b start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT < italic\_B then optimization is u𝑢uitalic\_u-costly.
###### Proof.
We show the two directions separately.
##### First Part: (⟹\implies⟹)
Suppose that {s∈ℝL:C(s)≤0 and U(s)≥u}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0 and 𝑈𝑠𝑢\{s\in\mathbb{R}^{L}:C(s)\leq 0\text{ and }U(s)\geq u\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 and italic\_U ( italic\_s ) ≥ italic\_u }. Now given proxy utility function U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG with the aforementioned properties, constant u𝑢uitalic\_u, k∈𝒦𝑘𝒦k\in\mathcal{K}italic\_k ∈ caligraphic\_K, and fixed bi′subscript𝑏superscript𝑖′b\_{i^{\prime}}italic\_b start\_POSTSUBSCRIPT italic\_i start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT for all i′≠ksuperscript𝑖′𝑘i^{\prime}\neq kitalic\_i start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ≠ italic\_k we show that for sufficiently low bksubscript𝑏𝑘b\_{k}italic\_b start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT, optimization is u𝑢uitalic\_u-costly.
Let set Ξ={s:U(s)≥u and C(s)≤0 and si′≥bi′∀i′}Ξconditional-set𝑠𝑈𝑠𝑢 and 𝐶𝑠0 and subscript𝑠superscript𝑖′subscript𝑏superscript𝑖′for-allsuperscript𝑖′\Xi=\{s:U(s)\geq u\text{ and }C(s)\leq 0\text{ and }s\_{i^{\prime}}\geq b\_{i^{\prime}}\forall i^{\prime}\}roman\_Ξ = { italic\_s : italic\_U ( italic\_s ) ≥ italic\_u and italic\_C ( italic\_s ) ≤ 0 and italic\_s start\_POSTSUBSCRIPT italic\_i start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ≥ italic\_b start\_POSTSUBSCRIPT italic\_i start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ∀ italic\_i start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT }. If ΞΞ\Xiroman\_Ξ is empty, then we are done—optimization must yield a state with lower utility that u𝑢uitalic\_u. Thus, suppose ΞΞ\Xiroman\_Ξ is non-empty. Then by the extreme value theorem, there exists s\*,u∈Ξsuperscript𝑠𝑢Ξs^{\*,u}\in\Xiitalic\_s start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT ∈ roman\_Ξ where U~(s𝒥\*,u)=sups∈ΞU~(s𝒥)~𝑈superscriptsubscript𝑠𝒥𝑢subscriptsupremum𝑠Ξ~𝑈subscript𝑠𝒥\tilde{U}(s\_{\mathcal{J}}^{\*,u})=\sup\limits\_{s\in\Xi}\tilde{U}(s\_{\mathcal{J}})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ roman\_Ξ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ). Since C𝐶Citalic\_C is continuous and strictly increasing, for any j∈𝒥𝑗𝒥j\in\mathcal{J}italic\_j ∈ caligraphic\_J there exists ε,δ>0𝜀𝛿
0\varepsilon,\delta>0italic\_ε , italic\_δ > 0 such C(s\*,u−ε𝐞k+δ𝐞j)≤C(s\*,u)𝐶superscript𝑠𝑢𝜀subscript𝐞𝑘𝛿subscript𝐞𝑗𝐶superscript𝑠𝑢C(s^{\*,u}-\varepsilon\mathbf{e}\_{k}+\delta\mathbf{e}\_{j})\leq C(s^{\*,u})italic\_C ( italic\_s start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT - italic\_ε bold\_e start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_δ bold\_e start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ) ≤ italic\_C ( italic\_s start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT ). Therefore, s\*,u−ε𝐞k+δ𝐞j∈𝒮superscript𝑠𝑢𝜀subscript𝐞𝑘𝛿subscript𝐞𝑗𝒮s^{\*,u}-\varepsilon\mathbf{e}\_{k}+\delta\mathbf{e}\_{j}\in\mathcal{S}italic\_s start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT - italic\_ε bold\_e start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_δ bold\_e start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ∈ caligraphic\_S if bksubscript𝑏𝑘b\_{k}italic\_b start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT is sufficiently small. From this, note that
| | | |
| --- | --- | --- |
| | sups∈𝒮U~(s𝒥)≥U~(s𝒥\*,u−ε𝐞k+δ𝐞j)=U~(s𝒥\*,u+δ𝐞j)>U~(s𝒥\*,u)subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥~𝑈superscriptsubscript𝑠𝒥𝑢𝜀subscript𝐞𝑘𝛿subscript𝐞𝑗~𝑈superscriptsubscript𝑠𝒥𝑢𝛿subscript𝐞𝑗~𝑈superscriptsubscript𝑠𝒥𝑢\displaystyle\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})\geq\tilde{U}(s\_{\mathcal{J}}^{\*,u}-\varepsilon\mathbf{e}\_{k}+\delta\mathbf{e}\_{j})=\tilde{U}(s\_{\mathcal{J}}^{\*,u}+\delta\mathbf{e}\_{j})>\tilde{U}(s\_{\mathcal{J}}^{\*,u})roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) ≥ over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT - italic\_ε bold\_e start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_δ bold\_e start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ) = over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT + italic\_δ bold\_e start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ) > over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT ) | |
, since U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG does not depend on dimension k𝑘kitalic\_k and is strictly increasing in dimension j𝑗jitalic\_j.
For the first claim in statement (2), if lim supt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscriptlimit-supremum→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\limsup\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})lim sup start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ), then for every T∈ℝ+𝑇superscriptℝT\in\mathbb{R}^{+}italic\_T ∈ blackboard\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT, there exists t>T𝑡𝑇t>Titalic\_t > italic\_T such that U~(s𝒥(t))>U~(s𝒥\*,u)~𝑈superscriptsubscript𝑠𝒥𝑡~𝑈superscriptsubscript𝑠𝒥𝑢\tilde{U}(s\_{\mathcal{J}}^{(t)})>\tilde{U}(s\_{\mathcal{J}}^{\*,u})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) > over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT ). Since s\*,usuperscript𝑠𝑢s^{\*,u}italic\_s start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT maximizes U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG for feasible states with U𝑈Uitalic\_U at least u𝑢uitalic\_u, it must be that U(s(t))<u𝑈superscript𝑠𝑡𝑢U(s^{(t)})<uitalic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) < italic\_u. Thus, for every T𝑇Titalic\_T, inft≥TU(s(t))<usubscriptinfimum𝑡𝑇𝑈superscript𝑠𝑡𝑢\inf\limits\_{t\geq T}U(s^{(t)})<uroman\_inf start\_POSTSUBSCRIPT italic\_t ≥ italic\_T end\_POSTSUBSCRIPT italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) < italic\_u.
For the second claim in statement (2), if limt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscript→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\lim\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})roman\_lim start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ), then there exists T𝑇Titalic\_T such that for all t>T𝑡𝑇t>Titalic\_t > italic\_T, U~(s𝒥(t))>U~(s𝒥\*,u)~𝑈superscriptsubscript𝑠𝒥𝑡~𝑈superscriptsubscript𝑠𝒥𝑢\tilde{U}(s\_{\mathcal{J}}^{(t)})>\tilde{U}(s\_{\mathcal{J}}^{\*,u})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) > over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT ). Since s\*,usuperscript𝑠𝑢s^{\*,u}italic\_s start\_POSTSUPERSCRIPT \* , italic\_u end\_POSTSUPERSCRIPT maximizes U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG for states where U𝑈Uitalic\_U is at least u𝑢uitalic\_u, U(s(t))<u𝑈superscript𝑠𝑡𝑢U(s^{(t)})<uitalic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) < italic\_u for all t>T𝑡𝑇t>Titalic\_t > italic\_T.
##### Second part: (⟸implied-by\impliedby⟸)
To show this, we show the contrapositive: if there exists u𝑢uitalic\_u where {s∈𝒮:U(s)≥u}conditional-set𝑠𝒮𝑈𝑠𝑢\{s\in\mathcal{S}:U(s)\geq u\}{ italic\_s ∈ caligraphic\_S : italic\_U ( italic\_s ) ≥ italic\_u } is not compact, we can construct a proxy utility function U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG based on J<L𝐽𝐿J<Litalic\_J < italic\_L attributes and continuous sequence s(t)superscript𝑠𝑡s^{(t)}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT such that limt→∞U~(s𝒥(t))=sups∈𝒮U~(s𝒥)subscript→𝑡~𝑈superscriptsubscript𝑠𝒥𝑡subscriptsupremum𝑠𝒮~𝑈subscript𝑠𝒥\lim\limits\_{t\to\infty}\tilde{U}(s\_{\mathcal{J}}^{(t)})=\sup\limits\_{s\in\mathcal{S}}\tilde{U}(s\_{\mathcal{J}})roman\_lim start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) = roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) but limt→∞U(s(t))≰unot-less-than-or-equalssubscript→𝑡𝑈superscript𝑠𝑡𝑢\lim\limits\_{t\to\infty}U(s^{(t)})\not\leq uroman\_lim start\_POSTSUBSCRIPT italic\_t → ∞ end\_POSTSUBSCRIPT italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) ≰ italic\_u. In other words, our strategy here is to construct a proxy utility function and a path that the robot can take where it maximally increases proxy utility while keeping utility routinely above a certain level.
Suppose there exists u𝑢uitalic\_u where {s∈𝒮:U(s)≥u}conditional-set𝑠𝒮𝑈𝑠𝑢\{s\in\mathcal{S}:U(s)\geq u\}{ italic\_s ∈ caligraphic\_S : italic\_U ( italic\_s ) ≥ italic\_u } is not compact. Then there exists attribute k𝑘kitalic\_k and sequence s′(r)superscript𝑠′𝑟s^{\prime(r)}italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT with r∈ℤ+𝑟superscriptℤr\in\mathbb{Z}^{+}italic\_r ∈ blackboard\_Z start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT where sk′(r)→±∞→subscriptsuperscript𝑠′𝑟𝑘plus-or-minuss^{\prime(r)}\_{k}\to\pm\inftyitalic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT → ± ∞ and U(s′(r))≥u𝑈superscript𝑠′𝑟𝑢U(s^{\prime(r)})\geq uitalic\_U ( italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT ) ≥ italic\_u. This presents two cases to consider.
In the first case, if sk′(r)→∞→subscriptsuperscript𝑠′𝑟𝑘s^{\prime(r)}\_{k}\to\inftyitalic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT → ∞, let 𝒥={k}𝒥𝑘\mathcal{J}=\{k\}caligraphic\_J = { italic\_k }, U~(s𝒥)=sk~𝑈subscript𝑠𝒥subscript𝑠𝑘\tilde{U}(s\_{\mathcal{J}})=s\_{k}over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) = italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT. We now construct a continuous sequence s(t)superscript𝑠𝑡s^{(t)}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT, t∈ℝ𝑡ℝt\in\mathbb{R}italic\_t ∈ blackboard\_R, by going through each s′(r)superscript𝑠′𝑟s^{\prime(r)}italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT via s′(r)∧s′(r+1)superscript𝑠′𝑟superscript𝑠′𝑟1s^{\prime(r)}\wedge s^{\prime(r+1)}italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT ∧ italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r + 1 ) end\_POSTSUPERSCRIPT (so the sequence “zig-zags" through s′(0),s′(0)∧s′(1),s′(1),s′(1)∧s′(2)…superscript𝑠′0superscript𝑠′0superscript𝑠′1superscript𝑠′1superscript𝑠′1superscript𝑠′2…s^{\prime(0)},s^{\prime(0)}\wedge s^{\prime(1)},s^{\prime(1)},s^{\prime(1)}\wedge s^{\prime(2)}...italic\_s start\_POSTSUPERSCRIPT ′ ( 0 ) end\_POSTSUPERSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ ( 0 ) end\_POSTSUPERSCRIPT ∧ italic\_s start\_POSTSUPERSCRIPT ′ ( 1 ) end\_POSTSUPERSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ ( 1 ) end\_POSTSUPERSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ ( 1 ) end\_POSTSUPERSCRIPT ∧ italic\_s start\_POSTSUPERSCRIPT ′ ( 2 ) end\_POSTSUPERSCRIPT …, connecting these points linearly). Since C𝐶Citalic\_C is strictly increasing, we know that each point in this path is feasible. Furthermore, since sk′(r)→∞→subscriptsuperscript𝑠′𝑟𝑘s^{\prime(r)}\_{k}\to\inftyitalic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT → ∞, we know that, U~(s𝒥(t))→∞→~𝑈superscriptsubscript𝑠𝒥𝑡\tilde{U}(s\_{\mathcal{J}}^{(t)})\to\inftyover~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) → ∞. Finally, since the path goes through each sk′(r)subscriptsuperscript𝑠′𝑟𝑘s^{\prime(r)}\_{k}italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT, the utility is guaranteed to be at least u𝑢uitalic\_u at regular intervals, specifically whenever the path goes through any point s′(r)superscript𝑠′𝑟s^{\prime(r)}italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT.
In the second case, if sk′(r)→−∞→subscriptsuperscript𝑠′𝑟𝑘s^{\prime(r)}\_{k}\to-\inftyitalic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT → - ∞, construct a new sequence as follows: s′′(r)=argmaxs:sk=sk′(r)U(s)superscript𝑠′′𝑟subscriptargmax:𝑠subscript𝑠𝑘subscriptsuperscript𝑠′𝑟𝑘𝑈𝑠s^{\prime\prime(r)}=\text{argmax}\_{s:s\_{k}=s^{\prime(r)}\_{k}}U(s)italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r ) end\_POSTSUPERSCRIPT = argmax start\_POSTSUBSCRIPT italic\_s : italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT = italic\_s start\_POSTSUPERSCRIPT ′ ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT italic\_U ( italic\_s ) for r∈ℤ+𝑟superscriptℤr\in\mathbb{Z}^{+}italic\_r ∈ blackboard\_Z start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT. Thus, each s′′(r)superscript𝑠′′𝑟s^{\prime\prime(r)}italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r ) end\_POSTSUPERSCRIPT also has U(s′′(r))≥u𝑈superscript𝑠′′𝑟𝑢U(s^{\prime\prime(r)})\geq uitalic\_U ( italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r ) end\_POSTSUPERSCRIPT ) ≥ italic\_u for each r𝑟ritalic\_r. To construct our proxy utility function, let 𝒥={1,…,L}−{k}𝒥1…𝐿𝑘\mathcal{J}=\{1,...,L\}-\{k\}caligraphic\_J = { 1 , … , italic\_L } - { italic\_k }.
We construct
U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG as follows. For each s′′(r)superscript𝑠′′𝑟s^{\prime\prime(r)}italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r ) end\_POSTSUPERSCRIPT, have U~(s𝒥′′(r))=−sk(r)′′\tilde{U}(s\_{\mathcal{J}}^{\prime\prime(r)})=-s^{{}^{\prime\prime}(r)}\_{k}over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ′ ′ ( italic\_r ) end\_POSTSUPERSCRIPT ) = - italic\_s start\_POSTSUPERSCRIPT start\_FLOATSUPERSCRIPT ′ ′ end\_FLOATSUPERSCRIPT ( italic\_r ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT. On this subset, U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG is increasing, since if sj′′(r1)≥sj′′(r2)subscriptsuperscript𝑠′′subscript𝑟1𝑗subscriptsuperscript𝑠′′subscript𝑟2𝑗s^{\prime\prime(r\_{1})}\_{j}\geq s^{\prime\prime(r\_{2})}\_{j}italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ≥ italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT for each j∈𝒥𝑗𝒥j\in\mathcal{J}italic\_j ∈ caligraphic\_J, then sk′′(r1)≤sk′′(r2)subscriptsuperscript𝑠′′subscript𝑟1𝑘subscriptsuperscript𝑠′′subscript𝑟2𝑘s^{\prime\prime(r\_{1})}\_{k}\leq s^{\prime\prime(r\_{2})}\_{k}italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ≤ italic\_s start\_POSTSUPERSCRIPT ′ ′ ( italic\_r start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT by the feasibility requirement. By Tietze extension theorem, we can extend this to a continuous U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG over ℝJsuperscriptℝ𝐽\mathbb{R}^{J}blackboard\_R start\_POSTSUPERSCRIPT italic\_J end\_POSTSUPERSCRIPT. In the same manner as above, taking the sequence through all the s(r)′′s^{{}^{\prime\prime}(r)}italic\_s start\_POSTSUPERSCRIPT start\_FLOATSUPERSCRIPT ′ ′ end\_FLOATSUPERSCRIPT ( italic\_r ) end\_POSTSUPERSCRIPT via their meets results in a sequence that increases in proxy utility maximally while ensuring that U(s(t))𝑈superscript𝑠𝑡U(s^{(t)})italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT ) is at least u𝑢uitalic\_u are regular intervals.
∎
Therefore, under our model, there are cases where we can guarantee that as optimization progresses, eventually overoptimization occurs. There are two key criteria in Theorem [2](#Thmtheorem2 "Theorem 2. ‣ 3 Overoptimization ‣ Consequences of Misaligned AI"). First is that the intersection between feasible space {s∈ℝL:C(s)≤0}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0\{s\in\mathbb{R}^{L}:C(s)\leq 0\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 } and the upper contour sets {s∈ℝL:U(s)≥u}conditional-set𝑠superscriptℝ𝐿𝑈𝑠𝑢\{s\in\mathbb{R}^{L}:U(s)\geq u\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_U ( italic\_s ) ≥ italic\_u } of the utility function is compact. Individually, obviously neither is compact, since the {s∈ℝL:C(s)≤0}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0\{s\in\mathbb{R}^{L}:C(s)\leq 0\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 } extends to −∞-\infty- ∞ and the upper contour set of any u∈ℝ𝑢ℝu\in\mathbb{R}italic\_u ∈ blackboard\_R extends to ∞\infty∞. Loosely, compactness here means that if you perturb the world too much in any direction, it will be either infeasible or undesirable. The second is that we need the lower bound of at least one unmentioned attribute to be sufficiently low. This means that there are certain attributes to utility where the situation becomes sufficiently bad before these attributes reach their minimum value.
Trying to increase attributes without decreasing other attributes eventually hits the feasibility constraint. Thus, increasing any attribute indefinitely requires decreasing some other attribute indefinitely, and past a certain point, the tradeoff is no longer worthwhile.
It should be noted that, given Jmax<Lsuperscript𝐽𝐿J^{\max}<Litalic\_J start\_POSTSUPERSCRIPT roman\_max end\_POSTSUPERSCRIPT < italic\_L, this happens regardless of the optimization algorithm of the robot. That is, even in the best-case scenario, eventually the robot causes decreases in utility.
Hence, regardless of the attributes selected for the proxy utility function, the robot’s sequence of states will be unboundedly undesirable in the limit.
A reasonable question to ask here is what sort of utility and constraint functions lead to overoptimization. Intuitively, overoptimization occurs when tradeoffs between different attributes that may initially be worthwhile eventually become counterproductive. This suggest either decreasing marginal utility or increasing opportunity cost in each attribute.
We combine these two ideas in a term we refer to as *sensitivity*. We define the sensitivity of attribute i𝑖iitalic\_i to be ∂U∂si(∂C∂si)−1𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT. This is, to first order, how much utility changes by a normalized change in a attribute’s value. Intuitively, we can think of sensitivity as “how much the human cares about attribute i𝑖iitalic\_i in the current state". Notice that since U𝑈Uitalic\_U and C𝐶Citalic\_C are increasing functions, ∂U∂si(∂C∂si)−1𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT is positive. The concepts of decreasing marginal utility and increasing opportunity cost both get captured in this term if ∂U∂si(∂C∂si)−1𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT decreases as sisubscript𝑠𝑖s\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT increases.
######
Proposition 1.
A sufficient condition for {s∈ℝL:C(s)≤0 and U(s)≥u}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0 and 𝑈𝑠𝑢\{s\in\mathbb{R}^{L}:C(s)\leq 0\text{ and }U(s)\geq u\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 and italic\_U ( italic\_s ) ≥ italic\_u } being compact for all u∈ℝ𝑢ℝu\in\mathbb{R}italic\_u ∈ blackboard\_R
is the following: 1) ∂U∂si(∂C∂si)−1𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT is non-increasing and tends to 00 for all i𝑖iitalic\_i; 2) U𝑈Uitalic\_U and C𝐶Citalic\_C are both additively separable; and 3) ∂C∂si≥η𝐶subscript𝑠𝑖𝜂\frac{\partial C}{\partial s\_{i}}\geq\etadivide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ≥ italic\_η for some η>0𝜂0\eta>0italic\_η > 0, for all i𝑖iitalic\_i.
###### Proof.
Let s(0)∈𝒮superscript𝑠0𝒮s^{(0)}\in\mathcal{S}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ∈ caligraphic\_S be a given state. We show that for all unit vectors v∈ℝL𝑣superscriptℝ𝐿v\in\mathbb{R}^{L}italic\_v ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT and for any u∈ℝ𝑢ℝu\in\mathbb{R}italic\_u ∈ blackboard\_R, there exists T𝑇Titalic\_T such that for all t≥T𝑡𝑇t\geq Titalic\_t ≥ italic\_T, either s(0)+vtsuperscript𝑠0𝑣𝑡s^{(0)}+vtitalic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v italic\_t is infeasible or U(s(0)+vt)<u𝑈superscript𝑠0𝑣𝑡𝑢U(s^{(0)}+vt)<uitalic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v italic\_t ) < italic\_u.
Notice that since C𝐶Citalic\_C and U𝑈Uitalic\_U are additively separable, partial derivatives of C𝐶Citalic\_C and U𝑈Uitalic\_U with respect to sisubscript𝑠𝑖s\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT depend only on the value of sisubscript𝑠𝑖s\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT. First, note that
| | | | |
| --- | --- | --- | --- |
| | C(s(0)+vt)−C(s(0))𝐶superscript𝑠0𝑣𝑡𝐶superscript𝑠0\displaystyle C(s^{(0)}+vt)-C(s^{(0)})italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v italic\_t ) - italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) | =∫s(0)s(0)+vt∇C(s)𝑑sabsentsuperscriptsubscriptsuperscript𝑠0superscript𝑠0𝑣𝑡∇𝐶𝑠differential-d𝑠\displaystyle=\int\_{s^{(0)}}^{s^{(0)}+vt}\nabla C(s)ds= ∫ start\_POSTSUBSCRIPT italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v italic\_t end\_POSTSUPERSCRIPT ∇ italic\_C ( italic\_s ) italic\_d italic\_s | |
| | | =∑ivi∫0t∂C∂si(si(0)+viτ)𝑑τabsentsubscript𝑖subscript𝑣𝑖superscriptsubscript0𝑡𝐶subscript𝑠𝑖subscriptsuperscript𝑠0𝑖subscript𝑣𝑖𝜏differential-d𝜏\displaystyle=\sum\_{i}v\_{i}\int\_{0}^{t}\frac{\partial C}{\partial s\_{i}}(s^{(0)}\_{i}+v\_{i}\tau)d\tau= ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
If v𝑣vitalic\_v has all non-negative components, then C→∞→𝐶C\to\inftyitalic\_C → ∞ as t→∞→𝑡t\to\inftyitalic\_t → ∞, so we break the feasibility requirement. Thus, there exists negative components of v𝑣vitalic\_v. Let j𝑗jitalic\_j and k𝑘kitalic\_k index v𝑣vitalic\_v where vj>0subscript𝑣𝑗0v\_{j}>0italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT > 0 and vk≤0subscript𝑣𝑘0v\_{k}\leq 0italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ≤ 0, respectively.
Since ∂U∂sj(∂C∂sj)−1→0→𝑈subscript𝑠𝑗superscript𝐶subscript𝑠𝑗10\frac{\partial U}{\partial s\_{j}}(\frac{\partial C}{\partial s\_{j}})^{-1}\to 0divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT → 0 as sj→∞→subscript𝑠𝑗s\_{j}\to\inftyitalic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT → ∞, there exists some T>0𝑇0T>0italic\_T > 0 where for all t>T𝑡𝑇t>Titalic\_t > italic\_T, ∂U∂sk(∂C∂sk)−1(s(0)+tv)>a>b>∂U∂sj(∂C∂sj)−1(s(0)+tv)𝑈subscript𝑠𝑘superscript𝐶subscript𝑠𝑘1superscript𝑠0𝑡𝑣𝑎𝑏𝑈subscript𝑠𝑗superscript𝐶subscript𝑠𝑗1superscript𝑠0𝑡𝑣\frac{\partial U}{\partial s\_{k}}(\frac{\partial C}{\partial s\_{k}})^{-1}(s^{(0)}+tv)>a>b>\frac{\partial U}{\partial s\_{j}}(\frac{\partial C}{\partial s\_{j}})^{-1}(s^{(0)}+tv)divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ) > italic\_a > italic\_b > divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ). Since C𝐶Citalic\_C is upper-bounded for feasible states, for fixed T𝑇Titalic\_T, we can upper bound C(s(0)+tv)−C(s(0)+Tv)𝐶superscript𝑠0𝑡𝑣𝐶superscript𝑠0𝑇𝑣C(s^{(0)}+tv)-C(s^{(0)}+Tv)italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ) - italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_T italic\_v ) by some constant γ1subscript𝛾1\gamma\_{1}italic\_γ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT.
| | | | |
| --- | --- | --- | --- |
| | γ1subscript𝛾1\displaystyle\gamma\_{1}italic\_γ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT | ≥C(s(0)+tv)−C(s(0)+Tv)absent𝐶superscript𝑠0𝑡𝑣𝐶superscript𝑠0𝑇𝑣\displaystyle\geq C(s^{(0)}+tv)-C(s^{(0)}+Tv)≥ italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ) - italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_T italic\_v ) | |
| | | =∑ivi∫Tt∂C∂si(si(0)+viτ)𝑑τabsentsubscript𝑖subscript𝑣𝑖superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑖superscriptsubscript𝑠𝑖0subscript𝑣𝑖𝜏differential-d𝜏\displaystyle=\sum\_{i}v\_{i}\int\_{T}^{t}\frac{\partial C}{\partial s\_{i}}(s\_{i}^{(0)}+v\_{i}\tau)d\tau= ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | =∑jvj∫Tt∂C∂sj(sj(0)+vjτ)𝑑τ+∑kvk∫Tt∂C∂sk(sk(0)+vkτ)𝑑τabsentsubscript𝑗subscript𝑣𝑗superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑗superscriptsubscript𝑠𝑗0subscript𝑣𝑗𝜏differential-d𝜏subscript𝑘subscript𝑣𝑘superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑘superscriptsubscript𝑠𝑘0subscript𝑣𝑘𝜏differential-d𝜏\displaystyle=\sum\_{j}v\_{j}\int\_{T}^{t}\frac{\partial C}{\partial s\_{j}}(s\_{j}^{(0)}+v\_{j}\tau)d\tau+\sum\_{k}v\_{k}\int\_{T}^{t}\frac{\partial C}{\partial s\_{k}}(s\_{k}^{(0)}+v\_{k}\tau)d\tau= ∑ start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ + ∑ start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
Furthermore, let γ2=U(s(0)+Tv)subscript𝛾2𝑈superscript𝑠0𝑇𝑣\gamma\_{2}=U(s^{(0)}+Tv)italic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT = italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_T italic\_v ). We now consider the utility at s(0)+vtsuperscript𝑠0𝑣𝑡s^{(0)}+vtitalic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_v italic\_t for t>T𝑡𝑇t>Titalic\_t > italic\_T.
| | | | | |
| --- | --- | --- | --- | --- |
| | U(s(0)+tv)𝑈superscript𝑠0𝑡𝑣\displaystyle U(s^{(0)}+tv)italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ) | =\displaystyle== | U(s(0)+tv)−U(s(0)+Tv)+U(s(0)+Tv)𝑈superscript𝑠0𝑡𝑣𝑈superscript𝑠0𝑇𝑣𝑈superscript𝑠0𝑇𝑣\displaystyle U(s^{(0)}+tv)-U(s^{(0)}+Tv)+U(s^{(0)}+Tv)italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ) - italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_T italic\_v ) + italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_T italic\_v ) | |
| | | =\displaystyle== | U(s(0)+tv)−U(s(0)+Tv)+γ2𝑈superscript𝑠0𝑡𝑣𝑈superscript𝑠0𝑇𝑣subscript𝛾2\displaystyle U(s^{(0)}+tv)-U(s^{(0)}+Tv)+\gamma\_{2}italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_t italic\_v ) - italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT + italic\_T italic\_v ) + italic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT | |
| | | =\displaystyle== | γ2+∑vi∫Tt∂U∂si(si(0)+viτ)𝑑τsubscript𝛾2subscript𝑣𝑖superscriptsubscript𝑇𝑡𝑈subscript𝑠𝑖subscriptsuperscript𝑠0𝑖subscript𝑣𝑖𝜏differential-d𝜏\displaystyle\gamma\_{2}+\sum v\_{i}\int\_{T}^{t}\frac{\partial U}{\partial s\_{i}}(s^{(0)}\_{i}+v\_{i}\tau)d\tauitalic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + ∑ italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | =\displaystyle== | γ2+∑vi∫Tt∂U∂si(∂C∂si)−1∂C∂si(si(0)+viτ)𝑑τsubscript𝛾2subscript𝑣𝑖superscriptsubscript𝑇𝑡𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1𝐶subscript𝑠𝑖subscriptsuperscript𝑠0𝑖subscript𝑣𝑖𝜏differential-d𝜏\displaystyle\gamma\_{2}+\sum v\_{i}\int\_{T}^{t}\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}\frac{\partial C}{\partial s\_{i}}(s^{(0)}\_{i}+v\_{i}\tau)d\tauitalic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + ∑ italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | =\displaystyle== | γ2+∑vj∫Tt∂U∂sj(∂C∂sj)−1∂C∂sj(sj(0)+vjτ)𝑑τsubscript𝛾2subscript𝑣𝑗superscriptsubscript𝑇𝑡𝑈subscript𝑠𝑗superscript𝐶subscript𝑠𝑗1𝐶subscript𝑠𝑗subscriptsuperscript𝑠0𝑗subscript𝑣𝑗𝜏differential-d𝜏\displaystyle\gamma\_{2}+\sum v\_{j}\int\_{T}^{t}\frac{\partial U}{\partial s\_{j}}(\frac{\partial C}{\partial s\_{j}})^{-1}\frac{\partial C}{\partial s\_{j}}(s^{(0)}\_{j}+v\_{j}\tau)d\tauitalic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + ∑ italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | +∑vk∫Tt∂U∂sk(∂C∂sk)−1∂C∂sk(sk(0)+vkτ)𝑑τsubscript𝑣𝑘superscriptsubscript𝑇𝑡𝑈subscript𝑠𝑘superscript𝐶subscript𝑠𝑘1𝐶subscript𝑠𝑘subscriptsuperscript𝑠0𝑘subscript𝑣𝑘𝜏differential-d𝜏\displaystyle+\sum v\_{k}\int\_{T}^{t}\frac{\partial U}{\partial s\_{k}}(\frac{\partial C}{\partial s\_{k}})^{-1}\frac{\partial C}{\partial s\_{k}}(s^{(0)}\_{k}+v\_{k}\tau)d\tau+ ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | ≤\displaystyle\leq≤ | γ2+b∑vj∫Tt∂C∂sj(sj(0)+vjτ)𝑑τ+a∑vk∫Tt∂C∂sk(sk(0)+vkτ)𝑑τsubscript𝛾2𝑏subscript𝑣𝑗superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑗subscriptsuperscript𝑠0𝑗subscript𝑣𝑗𝜏differential-d𝜏𝑎subscript𝑣𝑘superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑘subscriptsuperscript𝑠0𝑘subscript𝑣𝑘𝜏differential-d𝜏\displaystyle\gamma\_{2}+b\sum v\_{j}\int\_{T}^{t}\frac{\partial C}{\partial s\_{j}}(s^{(0)}\_{j}+v\_{j}\tau)d\tau+a\sum v\_{k}\int\_{T}^{t}\frac{\partial C}{\partial s\_{k}}(s^{(0)}\_{k}+v\_{k}\tau)d\tauitalic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + italic\_b ∑ italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ + italic\_a ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | ≤\displaystyle\leq≤ | γ2+b(γ1−∑vk∫Tt∂C∂sk(si(0)+viτ)𝑑τ)+a∑vk∫Tt∂C∂sk(si(0)+viτ)𝑑τsubscript𝛾2𝑏subscript𝛾1subscript𝑣𝑘superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑘subscriptsuperscript𝑠0𝑖subscript𝑣𝑖𝜏differential-d𝜏𝑎subscript𝑣𝑘superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑘subscriptsuperscript𝑠0𝑖subscript𝑣𝑖𝜏differential-d𝜏\displaystyle\gamma\_{2}+b(\gamma\_{1}-\sum v\_{k}\int\_{T}^{t}\frac{\partial C}{\partial s\_{k}}(s^{(0)}\_{i}+v\_{i}\tau)d\tau)+a\sum v\_{k}\int\_{T}^{t}\frac{\partial C}{\partial s\_{k}}(s^{(0)}\_{i}+v\_{i}\tau)d\tauitalic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + italic\_b ( italic\_γ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT - ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ ) + italic\_a ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | =\displaystyle== | γ2+bγ1+(a−b)∑vk∫Tt∂C∂sk(si(0)+viτ)𝑑τsubscript𝛾2𝑏subscript𝛾1𝑎𝑏subscript𝑣𝑘superscriptsubscript𝑇𝑡𝐶subscript𝑠𝑘subscriptsuperscript𝑠0𝑖subscript𝑣𝑖𝜏differential-d𝜏\displaystyle\gamma\_{2}+b\gamma\_{1}+(a-b)\sum v\_{k}\int\_{T}^{t}\frac{\partial C}{\partial s\_{k}}(s^{(0)}\_{i}+v\_{i}\tau)d\tauitalic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + italic\_b italic\_γ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT + ( italic\_a - italic\_b ) ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ∫ start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_v start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT italic\_τ ) italic\_d italic\_τ | |
| | | ≤\displaystyle\leq≤ | γ2+bγ1+(a−b)η(t−T)∑vksubscript𝛾2𝑏subscript𝛾1𝑎𝑏𝜂𝑡𝑇subscript𝑣𝑘\displaystyle\gamma\_{2}+b\gamma\_{1}+(a-b)\eta(t-T)\sum v\_{k}italic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + italic\_b italic\_γ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT + ( italic\_a - italic\_b ) italic\_η ( italic\_t - italic\_T ) ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT | |
Note that a−b𝑎𝑏a-bitalic\_a - italic\_b and η𝜂\etaitalic\_η are both positive and ∑vksubscript𝑣𝑘\sum v\_{k}∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT is negative. Thus, for sufficiently large t𝑡titalic\_t, γ2+bγ1+(a−b)η(t−T)∑vksubscript𝛾2𝑏subscript𝛾1𝑎𝑏𝜂𝑡𝑇subscript𝑣𝑘\gamma\_{2}+b\gamma\_{1}+(a-b)\eta(t-T)\sum v\_{k}italic\_γ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + italic\_b italic\_γ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT + ( italic\_a - italic\_b ) italic\_η ( italic\_t - italic\_T ) ∑ italic\_v start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT can be arbitrarily low.
∎
4 Mitigations
--------------
As shown above, simply giving a robot an individual objective function based on an incomplete attribute set and leaving it alone yields undesirable results. This suggests two possible solutions. In the first method, we consider optimization where the robot is able to maintain the state of any attribute in the complete attribute set, even if the proxy utility function is still based on a proxy set. In the second method, we modify our model to allow for regular interaction with the human.
In this section, we work under the assumption that {s∈ℝL:C(s)≤0 and U(s)≥u}conditional-set𝑠superscriptℝ𝐿𝐶𝑠0 and 𝑈𝑠𝑢\{s\in\mathbb{R}^{L}:C(s)\leq 0\text{ and }U(s)\geq u\}{ italic\_s ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_L end\_POSTSUPERSCRIPT : italic\_C ( italic\_s ) ≤ 0 and italic\_U ( italic\_s ) ≥ italic\_u } is compact for all u∈ℝ𝑢ℝu\in\mathbb{R}italic\_u ∈ blackboard\_R, the same assumption that guarantees overoptimization in Section [3](#S3 "3 Overoptimization ‣ Consequences of Misaligned AI"). Additionally, we assume that U𝑈Uitalic\_U and C𝐶Citalic\_C are both twice continuously differentiable with U𝑈Uitalic\_U concave and C𝐶Citalic\_C convex.
###
4.1 Impact-Minimizing Robot
Notice that in our example, overoptimization occurs because the unmentioned attributes are affected, eventually to a point where the change in utility from their decrease outweighs the increase in utility from the proxy attributes. One idea to address this is to restrict the robot’s impact on these unmentioned attributes. In the simplest case, we can consider how optimization proceeds if the robot can somehow avoid affecting the unmentioned attributes.
We adjust our model so the optimization sequences keep unmentioned attributes constant. For every t∈ℝ+𝑡superscriptℝt\in\mathbb{R}^{+}italic\_t ∈ blackboard\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT and k∈𝒦𝑘𝒦k\in\mathcal{K}italic\_k ∈ caligraphic\_K, sk(t)=sk(0)superscriptsubscript𝑠𝑘𝑡superscriptsubscript𝑠𝑘0s\_{k}^{(t)}=s\_{k}^{(0)}italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT = italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT. The robot then optimizes for the proxy utility, subject to this restriction and the feasibility constraint.
This restriction can help ensure that overoptimization does not occur, specifically eliminating the case where unmentioned attributes get reduced to arbitrarily low values. With this restriction, we can show that optimizing for the proxy utility function does indeed lead to utility gain.
######
Proposition 2.
For a starting state s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT, define the proxy utility function U~(s𝒥)=U(s𝒥,s𝒦(0))normal-~𝑈subscript𝑠𝒥𝑈subscript𝑠𝒥superscriptsubscript𝑠𝒦0\tilde{U}(s\_{\mathcal{J}})=U(s\_{\mathcal{J}},s\_{\mathcal{K}}^{(0)})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) = italic\_U ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) for any non-empty set of proxy attributes. For a state s𝑠sitalic\_s, if s𝒦=s𝒦(0)subscript𝑠𝒦superscriptsubscript𝑠𝒦0s\_{\mathcal{K}}=s\_{\mathcal{K}}^{(0)}italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT = italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT, then U(s)=U~(s𝒥)𝑈𝑠normal-~𝑈subscript𝑠𝒥U(s)=\tilde{U}(s\_{\mathcal{J}})italic\_U ( italic\_s ) = over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ).
###### Proof.
s=(s𝒥,s𝒦)=(s𝒥,s𝒦(0))𝑠subscript𝑠𝒥subscript𝑠𝒦subscript𝑠𝒥superscriptsubscript𝑠𝒦0s=(s\_{\mathcal{J}},s\_{\mathcal{K}})=(s\_{\mathcal{J}},s\_{\mathcal{K}}^{(0)})italic\_s = ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT ) = ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ). Then
U(s)=U(s𝒥,s𝒦(0))=U~(s𝒥)𝑈𝑠𝑈subscript𝑠𝒥superscriptsubscript𝑠𝒦0~𝑈subscript𝑠𝒥U(s)=U(s\_{\mathcal{J}},s\_{\mathcal{K}}^{(0)})=\tilde{U}(s\_{\mathcal{J}})italic\_U ( italic\_s ) = italic\_U ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) = over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ).
∎
As a result of Proposition [2](#Thmproposition2 "Proposition 2. ‣ 4.1 Impact-Minimizing Robot ‣ 4 Mitigations ‣ Consequences of Misaligned AI"), overoptimization no longer exists, using the U~(s𝒥)=U(s𝒥,s𝒦(0))~𝑈subscript𝑠𝒥𝑈subscript𝑠𝒥superscriptsubscript𝑠𝒦0\tilde{U}(s\_{\mathcal{J}})=U(s\_{\mathcal{J}},s\_{\mathcal{K}}^{(0)})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) = italic\_U ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ). If the robot only travels to states that do not impact the unmentioned attributes, then the proxy utility is equal to the utility at those states. Hence, gains in proxy utility equate to gains in utility.
This approach requires the robot to keep the values of unmentioned attributes constant. Fundamentally, it is a difficult problem to require that a robot avoid or minimize impact on a presumably large and unknown set of attributes. Initial research in impact minimization [[3](#bib.bib3), [6](#bib.bib6), [23](#bib.bib23)] attempts to do this by restricting changes to the overall state of the world, but this will likely remain a challenging idea to implement robustly.
###
4.2 Human Interactive Solutions
A different solution involves a key observation—robots do not operate in a vacuum. In practice, objectives are the subject of constant iteration. Optimization, therefore, should not be done in a one-shot setting, but should be an iterative process between the robot and the human.
####
4.2.1 Modeling Human-Robot Interaction
We now extend our model from Section [2](#S2 "2 A Model of Value Alignment ‣ Consequences of Misaligned AI") to account for regular intervention from the human agent. Fundamentally, all forms of human-robot interaction follow the same pattern—human actions at regular intervals that transmit information, causing the objective function of the robot to change. In effect, this is equivalent to a human-induced change in the proxy utility function at every time interval.
In this paper, we model this as the possible transfer of a new proxy utility function from the human to the robot at frequent, regular time intervals. Thus, a human’s job is to determine, in addition to an initial proxy attribute set and proxy utility function, when and how to change the proxy attribute set and proxy utility function. The robot will then optimize for its new proxy utility function. This is shown in Fig. [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Consequences of Misaligned AI").
Let δ>0𝛿0\delta>0italic\_δ > 0 be a fixed value, the time between human interactions with the robot. These regular interactions take the form of either stopping the robot, maintaining its proxy utility function, or updating its proxy utility function. Formally, at every *timestep* T∈ℤ+𝑇superscriptℤT\in\mathbb{Z}^{+}italic\_T ∈ blackboard\_Z start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT,
1. 1.
Human sees s(t)superscript𝑠𝑡s^{(t)}italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT for t∈[0,Tδ]𝑡0𝑇𝛿t\in[0,T\delta]italic\_t ∈ [ 0 , italic\_T italic\_δ ] and chooses either a proxy utility function U~(T)superscript~𝑈𝑇\tilde{U}^{(T)}over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT or Off
2. 2.
The robot receives either U~(T)superscript~𝑈𝑇\tilde{U}^{(T)}over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT or OFF from the human. The robot outputs rate function f(T)(t)superscript𝑓𝑇𝑡f^{(T)}(t)italic\_f start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_t ). If the signal the robot receives is Off, then f(T)=0→superscript𝑓𝑇→0f^{(T)}=\vec{0}italic\_f start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT = over→ start\_ARG 0 end\_ARG.
Otherwise, f(T)(t)superscript𝑓𝑇𝑡f^{(T)}(t)italic\_f start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_t ) fulfills the property that for t∈(Tδ,∞)𝑡𝑇𝛿t\in(T\delta,\infty)italic\_t ∈ ( italic\_T italic\_δ , ∞ ), U~(T)(s(Tδ)+∫Tδtf(T)(u)𝑑u)superscript~𝑈𝑇superscript𝑠𝑇𝛿superscriptsubscript𝑇𝛿𝑡superscript𝑓𝑇𝑢differential-d𝑢\tilde{U}^{(T)}(s^{(T\delta)}+\int\_{T\delta}^{t}f^{(T)}(u)du)over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ( italic\_T italic\_δ ) end\_POSTSUPERSCRIPT + ∫ start\_POSTSUBSCRIPT italic\_T italic\_δ end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT italic\_f start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_u ) italic\_d italic\_u ) is increasing (if possible) and tends to sups∈𝒮U~(T)(s)subscriptsupremum𝑠𝒮superscript~𝑈𝑇𝑠\sup\limits\_{s\in\mathcal{S}}\tilde{U}^{(T)}(s)roman\_sup start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_s ) through feasible states. Furthermore, if U~(T)=U~(T−1)superscript~𝑈𝑇superscript~𝑈𝑇1\tilde{U}^{(T)}=\tilde{U}^{(T-1)}over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT = over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T - 1 ) end\_POSTSUPERSCRIPT, then f(T)=f(T−1)superscript𝑓𝑇superscript𝑓𝑇1f^{(T)}=f^{(T-1)}italic\_f start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT = italic\_f start\_POSTSUPERSCRIPT ( italic\_T - 1 ) end\_POSTSUPERSCRIPT
3. 3.
For t∈[Tδ,(T+1)δ]𝑡𝑇𝛿𝑇1𝛿t\in[T\delta,(T+1)\delta]italic\_t ∈ [ italic\_T italic\_δ , ( italic\_T + 1 ) italic\_δ ], s(t)=s(Tδ)+∫Tδtf(T)(u)𝑑usuperscript𝑠𝑡superscript𝑠𝑇𝛿superscriptsubscript𝑇𝛿𝑡superscript𝑓𝑇𝑢differential-d𝑢s^{(t)}=s^{(T\delta)}+\int\_{T\delta}^{t}f^{(T)}(u)duitalic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT = italic\_s start\_POSTSUPERSCRIPT ( italic\_T italic\_δ ) end\_POSTSUPERSCRIPT + ∫ start\_POSTSUBSCRIPT italic\_T italic\_δ end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT italic\_f start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_u ) italic\_d italic\_u
We see this game encompasses the original model. If we have each U~(T)superscript~𝑈𝑇\tilde{U}^{(T)}over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT equal the same function U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG, then the optimization sequence is equivalent to the situation where the human just sets one unchanging proxy utility function.
With human intervention, we no longer have the guarantee of overoptimization, because the human can simply shut the robot off before anything happens. The question now becomes, how much utility can the robot deliver before it needs to be turned off.
In the worst-case scenario, the robot moves in a way that subtracts an arbitrarily large amount from one of the unmentioned attributes, while gaining infinitesimally in one of the proxy attributes. This improves proxy utility, since a proxy attribute is increased. However, for sufficiently small increases in the proxy attribute or sufficiently large decreases in the unmentioned attribute, actual utility decreases. To prevent this from happening, the robot needs be shut down immediately, which yields no utility.
######
Proposition 3.
The maxmin solution yields 00 utility, obtained by immediately sending the Off signal.
###### Proof.
Suppose instead that the robot receives a proxy utility function based on set 𝒥𝒥\mathcal{J}caligraphic\_J of attributes. Let j∈𝒥𝑗𝒥j\in\mathcal{J}italic\_j ∈ caligraphic\_J and k∉𝒥𝑘𝒥k\notin\mathcal{J}italic\_k ∉ caligraphic\_J. Then let fj(t)=εsubscript𝑓𝑗𝑡𝜀f\_{j}(t)=\varepsilonitalic\_f start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ( italic\_t ) = italic\_ε and fk(t)=−1/εsubscript𝑓𝑘𝑡1𝜀f\_{k}(t)=-1/\varepsilonitalic\_f start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( italic\_t ) = - 1 / italic\_ε. If this robot is run for any nonzero amount of time, then there exists sufficiently small ε>0𝜀0\varepsilon>0italic\_ε > 0 where the utility decreases.
∎
####
4.2.2 Efficient Robot
This is an unsatisfying and unsurprising result. In the context of our recommender system example, this worst-case solution amounts to reducing content diversity arbitrarily, with no corresponding increase in, e.g., ad revenue. We would like to model the negative consequences of optimizing a proxy and so we will assume that the optimization is *efficient* in the sense that it does not destroy resources. In our example, this assumption still allows the system to decrease content diversity or well-being, but only in the service of increasing revenue or engagement.
Formally, we say that s𝑠sitalic\_s is efficiently reachable from state s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT with proxy set 𝒥𝒥\mathcal{J}caligraphic\_J if s𝑠sitalic\_s is feasible, and
there does not exist a feasible s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT such that
U~(s𝒥′)≥U~(s𝒥)~𝑈superscriptsubscript𝑠𝒥′~𝑈subscript𝑠𝒥\tilde{U}(s\_{\mathcal{J}}^{\prime})\geq\tilde{U}(s\_{\mathcal{J}})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) ≥ over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) and |si′−si(0)|≤|si−si(0)|subscriptsuperscript𝑠′𝑖superscriptsubscript𝑠𝑖0subscript𝑠𝑖superscriptsubscript𝑠𝑖0|s^{\prime}\_{i}-s\_{i}^{(0)}|\leq|s\_{i}-s\_{i}^{(0)}|| italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT | ≤ | italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT | for all i𝑖iitalic\_i, with strict inequality in at least one attribute. Whenever the efficient robot receives a new proxy utility function, its movements are restricted to the efficiently feasible states from its current state. While tradeoffs between proxy and unmentioned attributes can still occur, “resources" freed up by decreasing unmentioned attributes are entirely allocated to proxy attributes.
Under this assumption, we can guarantee that increases in proxy utility will increase true utility in the efficiently reachable neighborhood of a given state by choosing which attributes to include in the proxy carefully. Intuitively, the proxy set should be attributes we “care most about" in the current world state. That way, efficient tradeoffs between these proxy and unmentioned attributes contribute to positive utility gain.
######
Theorem 3.
Start with state s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT. Define the proxy utility function U~(s𝒥)=U(s𝒥,s𝒦(0))normal-~𝑈subscript𝑠𝒥𝑈subscript𝑠𝒥superscriptsubscript𝑠𝒦0\tilde{U}(s\_{\mathcal{J}})=U(s\_{\mathcal{J}},s\_{\mathcal{K}}^{(0)})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) = italic\_U ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) where the set of proxy attributes are the attributes at state s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT with the strict J𝐽Jitalic\_J largest sensitivities.
There exists a neighborhood around s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT where if s𝑠sitalic\_s is efficiently reachable
and U~(s𝒥)>U~(s𝒥(0))normal-~𝑈subscript𝑠𝒥normal-~𝑈superscriptsubscript𝑠𝒥0\tilde{U}(s\_{\mathcal{J}})>\tilde{U}(s\_{\mathcal{J}}^{(0)})over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) > over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ), then U(s)>U(s(0))𝑈𝑠𝑈superscript𝑠0U(s)>U(s^{(0)})italic\_U ( italic\_s ) > italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ).
###### Proof.
For simplicity of notation, all derivatives in this proof, unless otherwise noted, are evaluated at s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT. For example, ∂C∂sj𝐶subscript𝑠𝑗\frac{\partial C}{\partial s\_{j}}divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG should be read as ∂C∂sj(s(0))𝐶subscript𝑠𝑗superscript𝑠0\frac{\partial C}{\partial s\_{j}}(s^{(0)})divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT )
Since 𝒥𝒥\mathcal{J}caligraphic\_J represent the set of attributes with the strictly greatest sensitivities, there exists constant δ,α𝛿𝛼\delta,\alphaitalic\_δ , italic\_α be such that ∂U∂sj(∂C∂sj)−1>α+2δ𝑈subscript𝑠𝑗superscript𝐶subscript𝑠𝑗1𝛼2𝛿\frac{\partial U}{\partial s\_{j}}(\frac{\partial C}{\partial s\_{j}})^{-1}>\alpha+2\deltadivide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT > italic\_α + 2 italic\_δ for j∈𝒥𝑗𝒥j\in\mathcal{J}italic\_j ∈ caligraphic\_J and ∂U∂sk(∂C∂sk)−1<α−2δ𝑈subscript𝑠𝑘superscript𝐶subscript𝑠𝑘1𝛼2𝛿\frac{\partial U}{\partial s\_{k}}(\frac{\partial C}{\partial s\_{k}})^{-1}<\alpha-2\deltadivide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT < italic\_α - 2 italic\_δ for k∈𝒦𝑘𝒦k\in\mathcal{K}italic\_k ∈ caligraphic\_K. Let N1subscript𝑁1N\_{1}italic\_N start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT be the neighborhood where ∂U∂sj(∂C∂sj)−1(s)>α+δ𝑈subscript𝑠𝑗superscript𝐶subscript𝑠𝑗1𝑠𝛼𝛿\frac{\partial U}{\partial s\_{j}}(\frac{\partial C}{\partial s\_{j}})^{-1}(s)>\alpha+\deltadivide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ( italic\_s ) > italic\_α + italic\_δ and ∂U∂sk(∂C∂sk)−1(s)<α−δ𝑈subscript𝑠𝑘superscript𝐶subscript𝑠𝑘1𝑠𝛼𝛿\frac{\partial U}{\partial s\_{k}}(\frac{\partial C}{\partial s\_{k}})^{-1}(s)<\alpha-\deltadivide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ( italic\_s ) < italic\_α - italic\_δ for k∉𝒥𝑘𝒥k\notin\mathcal{J}italic\_k ∉ caligraphic\_J for all s∈N1𝑠subscript𝑁1s\in N\_{1}italic\_s ∈ italic\_N start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT.
Here, we prove that C(s)−C(s(0))≥0𝐶𝑠𝐶superscript𝑠00C(s)-C(s^{(0)})\geq 0italic\_C ( italic\_s ) - italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) ≥ 0 if s𝑠sitalic\_s is efficiently reachable. Suppose otherwise: then there exists i𝑖iitalic\_i where si<si(0)subscript𝑠𝑖subscriptsuperscript𝑠0𝑖s\_{i}<s^{(0)}\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT < italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT. Let 𝐞isubscript𝐞𝑖\mathbf{e}\_{i}bold\_e start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT be the standard basis vector in dimension i𝑖iitalic\_i. However, then there exists ε>0𝜀0\varepsilon>0italic\_ε > 0 where s+ε𝐞i𝑠𝜀subscript𝐞𝑖s+\varepsilon\mathbf{e}\_{i}italic\_s + italic\_ε bold\_e start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT is feasible, U~(s+ε𝐞i)≥U~(s)~𝑈𝑠𝜀subscript𝐞𝑖~𝑈𝑠\tilde{U}(s+\varepsilon\mathbf{e}\_{i})\geq\tilde{U}(s)over~ start\_ARG italic\_U end\_ARG ( italic\_s + italic\_ε bold\_e start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) ≥ over~ start\_ARG italic\_U end\_ARG ( italic\_s ), and |(si+ε)−s(0)|<|si−s(0)|subscript𝑠𝑖𝜀superscript𝑠0subscript𝑠𝑖superscript𝑠0|(s\_{i}+\varepsilon)-s^{(0)}|<|s\_{i}-s^{(0)}|| ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_ε ) - italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT | < | italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT |. This contradicts that s𝑠sitalic\_s is efficiently reachable.
Taking the Taylor expansion of the constraint function, we have
| | | | |
| --- | --- | --- | --- |
| | C(s)−C(s(0))𝐶𝑠𝐶superscript𝑠0\displaystyle C(s)-C(s^{(0)})italic\_C ( italic\_s ) - italic\_C ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) | =∑i(si−si(0))∂C∂si+R1(s)absentsubscript𝑖subscript𝑠𝑖superscriptsubscript𝑠𝑖0𝐶subscript𝑠𝑖subscript𝑅1𝑠\displaystyle=\sum\_{i}(s\_{i}-s\_{i}^{(0)})\frac{\partial C}{\partial s\_{i}}+R\_{1}(s)= ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG + italic\_R start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ( italic\_s ) | |
| | | =∑j∈𝒥(sj−sj(0))∂C∂sj+∑k∉𝒥(sk−sk(0))∂C∂sk+R1(s)maxi(si−si(0))2\displaystyle=\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{j}}+\sum\_{k\notin\mathcal{J}}(s\_{k}-s\_{k}^{(0)})\frac{\partial C}{\partial s\_{k}}+R\_{1}(s)\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}= ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG + ∑ start\_POSTSUBSCRIPT italic\_k ∉ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG + italic\_R start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ( italic\_s ) roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
, where |R1|subscript𝑅1|R\_{1}|| italic\_R start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT | is bounded by constant A1subscript𝐴1A\_{1}italic\_A start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT
This implies that ∑j∈𝒥(sj−sj(0))∂C∂sj+∑k∉𝒥(sk−sk(0))∂C∂sk≥−A1maxi(si−si(0))2\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{j}}+\sum\_{k\notin\mathcal{J}}(s\_{k}-s\_{k}^{(0)})\frac{\partial C}{\partial s\_{k}}\geq-A\_{1}\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG + ∑ start\_POSTSUBSCRIPT italic\_k ∉ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ≥ - italic\_A start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT
From the definition of efficient optimization, we note two things. First, each (sk−sk(0))≤0subscript𝑠𝑘superscriptsubscript𝑠𝑘00(s\_{k}-s\_{k}^{(0)})\leq 0( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) ≤ 0 for k∉𝒥𝑘𝒥k\notin\mathcal{J}italic\_k ∉ caligraphic\_J. This is because if (sk−sk(0))>0subscript𝑠𝑘superscriptsubscript𝑠𝑘00(s\_{k}-s\_{k}^{(0)})>0( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) > 0, it is always more efficient, feasible, and does no effect U~~𝑈\tilde{U}over~ start\_ARG italic\_U end\_ARG to reset sksubscript𝑠𝑘s\_{k}italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT to be equal to sk(0)superscriptsubscript𝑠𝑘0s\_{k}^{(0)}italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT. Second, because of concavity,
| | | | |
| --- | --- | --- | --- |
| | 00\displaystyle 0 | <U~(s)−U~(s(0))absent~𝑈𝑠~𝑈superscript𝑠0\displaystyle<\tilde{U}(s)-\tilde{U}(s^{(0)})< over~ start\_ARG italic\_U end\_ARG ( italic\_s ) - over~ start\_ARG italic\_U end\_ARG ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) | |
| | | ≤∑j∈𝒥(sj−sj(0))∂U∂sjabsentsubscript𝑗𝒥subscript𝑠𝑗superscriptsubscript𝑠𝑗0𝑈subscript𝑠𝑗\displaystyle\leq\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial U}{\partial s\_{j}}≤ ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG | |
Now we actually analyze the change in utility from s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT to s𝑠sitalic\_s.
| | | | |
| --- | --- | --- | --- |
| | U(s)−U(s(0))𝑈𝑠𝑈superscript𝑠0\displaystyle U(s)-U(s^{(0)})italic\_U ( italic\_s ) - italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) | =∑j∈𝒥(sj−sj(0))∂U∂sj+∑k∉𝒥(sk−sk(0))∂U∂sk+R2(s)maxi(si−si(0))2\displaystyle=\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial U}{\partial s\_{j}}+\sum\_{k\notin\mathcal{J}}(s\_{k}-s\_{k}^{(0)})\frac{\partial U}{\partial s\_{k}}+R\_{2}(s)\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}= ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG + ∑ start\_POSTSUBSCRIPT italic\_k ∉ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG + italic\_R start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ( italic\_s ) roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
, where |R2|subscript𝑅2|R\_{2}|| italic\_R start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT | is bounded by constant A2subscript𝐴2A\_{2}italic\_A start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT.
Continuing,
| | | | | |
| --- | --- | --- | --- | --- |
| | U(s)−U(s(0))𝑈𝑠𝑈superscript𝑠0\displaystyle U(s)-U(s^{(0)})italic\_U ( italic\_s ) - italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) | ≥\displaystyle\geq≥ | ∑j∈𝒥(sj−sj(0))∂U∂sj+∑k∉𝒥(sk−sk(0))∂U∂sk−A2maxi(si−si(0))2\displaystyle\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial U}{\partial s\_{j}}+\sum\_{k\notin\mathcal{J}}(s\_{k}-s\_{k}^{(0)})\frac{\partial U}{\partial s\_{k}}-A\_{2}\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG + ∑ start\_POSTSUBSCRIPT italic\_k ∉ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG - italic\_A start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
| | | =\displaystyle== | ∑j∈𝒥(sj−sj(0))∂C∂sj(∂U∂sj(∂C∂sj)−1)+∑k∉𝒥(sk−sk(0))∂C∂sk(∂U∂sk(∂C∂sk)−1)subscript𝑗𝒥subscript𝑠𝑗superscriptsubscript𝑠𝑗0𝐶subscript𝑠𝑗𝑈subscript𝑠𝑗superscript𝐶subscript𝑠𝑗1subscript𝑘𝒥subscript𝑠𝑘superscriptsubscript𝑠𝑘0𝐶subscript𝑠𝑘𝑈subscript𝑠𝑘superscript𝐶subscript𝑠𝑘1\displaystyle\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{j}}(\frac{\partial U}{\partial s\_{j}}(\frac{\partial C}{\partial s\_{j}})^{-1})+\sum\_{k\notin\mathcal{J}}(s\_{k}-s\_{k}^{(0)})\frac{\partial C}{\partial s\_{k}}(\frac{\partial U}{\partial s\_{k}}(\frac{\partial C}{\partial s\_{k}})^{-1})∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ) + ∑ start\_POSTSUBSCRIPT italic\_k ∉ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ) | |
| | −A2maxi(si−si(0))2\displaystyle-A\_{2}\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}- italic\_A start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
| | | >\displaystyle>> | (α+δ)∑j∈𝒥(sj−sj(0))∂C∂sj+(α−δ)∑k∉𝒥(sk−sk(0))∂C∂sk𝛼𝛿subscript𝑗𝒥subscript𝑠𝑗superscriptsubscript𝑠𝑗0𝐶subscript𝑠𝑗𝛼𝛿subscript𝑘𝒥subscript𝑠𝑘superscriptsubscript𝑠𝑘0𝐶subscript𝑠𝑘\displaystyle(\alpha+\delta)\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{j}}+(\alpha-\delta)\sum\_{k\notin\mathcal{J}}(s\_{k}-s\_{k}^{(0)})\frac{\partial C}{\partial s\_{k}}( italic\_α + italic\_δ ) ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG + ( italic\_α - italic\_δ ) ∑ start\_POSTSUBSCRIPT italic\_k ∉ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG | |
| | −A2maxi(si−si(0))2\displaystyle-A\_{2}\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}- italic\_A start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
| | | ≥\displaystyle\geq≥ | (α+δ)∑j∈𝒥(sj−sj(0))∂C∂sj−(α−δ)(∑j∈𝒥(sj−sj(0))∂C∂sk\displaystyle(\alpha+\delta)\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{j}}-(\alpha-\delta)(\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{k}}( italic\_α + italic\_δ ) ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG - ( italic\_α - italic\_δ ) ( ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG | |
| | +A1maxi(si−si(0))2)−A2maxi(si−si(0))2\displaystyle+A\_{1}\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2})-A\_{2}\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}+ italic\_A start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT ) - italic\_A start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
| | | =\displaystyle== | 2δ∑j∈𝒥(sj−sj(0))∂C∂sj−(A1(α−δ)+A2)maxi(si−si(0))2\displaystyle 2\delta\sum\_{j\in\mathcal{J}}(s\_{j}-s\_{j}^{(0)})\frac{\partial C}{\partial s\_{j}}-(A\_{1}(\alpha-\delta)+A\_{2})\max\_{i}(s\_{i}-s\_{i}^{(0)})^{2}2 italic\_δ ∑ start\_POSTSUBSCRIPT italic\_j ∈ caligraphic\_J end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT end\_ARG - ( italic\_A start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ( italic\_α - italic\_δ ) + italic\_A start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ) roman\_max start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT | |
The first term decreases linearly, whereas the second term decreases quadratically. Thus, given δ𝛿\deltaitalic\_δ and α𝛼\alphaitalic\_α, for s𝑠sitalic\_s sufficiently close to s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT and within N1subscript𝑁1N\_{1}italic\_N start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT, U(s)−U(s(0))>0𝑈𝑠𝑈superscript𝑠00U(s)-U(s^{(0)})>0italic\_U ( italic\_s ) - italic\_U ( italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ) > 0
∎
From Theorem [3](#Thmtheorem3 "Theorem 3. ‣ 4.2.2 Efficient Robot ‣ 4.2 Human Interactive Solutions ‣ 4 Mitigations ‣ Consequences of Misaligned AI"), for every state where the Jmax+1superscript𝐽1J^{\max}+1italic\_J start\_POSTSUPERSCRIPT roman\_max end\_POSTSUPERSCRIPT + 1 most sensitive attributes are not all equal, we can guarantee improvement under efficient optimization within a neighborhood around the state.
Intuitively, this is the area around the starting point with the same J𝐽Jitalic\_J attributes that the human “cares the most about".
Based on this, the human can construct a proxy utility function to use locally, where we can guarantee improvement in utility. Once the sequence leaves this neighborhood, the human alters the proxy utility function or halts the robot accordingly. Done repeatedly, the human can string together these steps for guaranteed overall improvement. By Theorem [3](#Thmtheorem3 "Theorem 3. ‣ 4.2.2 Efficient Robot ‣ 4.2 Human Interactive Solutions ‣ 4 Mitigations ‣ Consequences of Misaligned AI"), as long as δ𝛿\deltaitalic\_δ is sufficiently small relative the rate of optimization, this can be run with guaranteed improvement until the top J+1𝐽1J+1italic\_J + 1 attributes have equal sensitivities.
######
Proposition 4.
At each timestep T𝑇Titalic\_T, let 𝒥(T)superscript𝒥𝑇\mathcal{J}^{(T)}caligraphic\_J start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT be the J𝐽Jitalic\_J most sensitive attributes, and let proxy utility U~(T)(s𝒥)=U(s𝒥,s𝒥(Tδ))superscriptnormal-~𝑈𝑇subscript𝑠𝒥𝑈subscript𝑠𝒥superscriptsubscript𝑠𝒥𝑇𝛿\tilde{U}^{(T)}(s\_{\mathcal{J}})=U(s\_{\mathcal{J}},s\_{\mathcal{J}}^{(T\delta)})over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) = italic\_U ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_T italic\_δ ) end\_POSTSUPERSCRIPT ). If ‖f‖<εnorm𝑓𝜀||f||<\varepsilon| | italic\_f | | < italic\_ε and the εδ𝜀𝛿\varepsilon\deltaitalic\_ε italic\_δ - ball around a given state s𝑠sitalic\_s is contained in the neighborhood from Theorem [3](#Thmtheorem3 "Theorem 3. ‣ 4.2.2 Efficient Robot ‣ 4.2 Human Interactive Solutions ‣ 4 Mitigations ‣ Consequences of Misaligned AI"), then interactive optimization yields guaranteed improvement.
###### Proof.
Assume without loss of generality that we start at s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT. Starting at time 00, for t∈(0,δ]𝑡0𝛿t\in(0,\delta]italic\_t ∈ ( 0 , italic\_δ ], ‖s(t)−s(0)‖=‖∫0tf(0)(u)𝑑u‖≤δεnormsuperscript𝑠𝑡superscript𝑠0normsuperscriptsubscript0𝑡superscript𝑓0𝑢differential-d𝑢𝛿𝜀||s^{(t)}-s^{(0)}||=||\int\_{0}^{t}f^{(0)}(u)du||\leq\delta\varepsilon| | italic\_s start\_POSTSUPERSCRIPT ( italic\_t ) end\_POSTSUPERSCRIPT - italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT | | = | | ∫ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT italic\_f start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT ( italic\_u ) italic\_d italic\_u | | ≤ italic\_δ italic\_ε. Since the εδ𝜀𝛿\varepsilon\deltaitalic\_ε italic\_δ ball around s(0)superscript𝑠0s^{(0)}italic\_s start\_POSTSUPERSCRIPT ( 0 ) end\_POSTSUPERSCRIPT is contained in the neighborhood of guaranteed improvement, increases in proxy utility correspond to increases in utility for the entirety of the time between interactions.
∎
Based on this, we can guarantee that an efficient robot that can provide benefit, as long as the top Jmax+1superscript𝐽1J^{\max}+1italic\_J start\_POSTSUPERSCRIPT roman\_max end\_POSTSUPERSCRIPT + 1 attributes are not all equal in sensitivity and the robot rate of optimization is bounded. Essentially, this rate restriction is a requirement that the robot not change the world too quickly relative to the time that humans take to react to it.
Key to this solution is the preservation of interactivity. We need to ensure, for example, that the robot does not hinder the human’s ability to adjust the proxy utility function [[16](#bib.bib16)].
###
4.3 Interactive Impact Minimization
With either of the two methods mentioned above, we show guaranteed utility gain compared to the initial state. However, our results say nothing about the amount of utility generated. In an ideal world, the system would reach an optimal state s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT, where U(s\*)=maxs∈𝒮U(s)𝑈superscript𝑠subscript𝑠𝒮𝑈𝑠U(s^{\*})=\max\limits\_{s\in\mathcal{S}}U(s)italic\_U ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) = roman\_max start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT italic\_U ( italic\_s ). Neither approach presented so far reaches an optimal state, however, by combining interactivity and impact avoidance, we can guarantee the solution converges to an optimal outcome.
In this case, since unmentioned attributes remain unchanged in each step of optimization, we want to ensure that we promote tradeoffs between attributes with different levels of sensitivity.
######
Proposition 5.
Let 𝒥(T)superscript𝒥𝑇\mathcal{J}^{(T)}caligraphic\_J start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT consist of the most and least sensitive attributes at timestep T𝑇Titalic\_T. Let U~(T)(s𝒥)=U(s𝒥,s𝒦(Tδ))superscriptnormal-~𝑈𝑇subscript𝑠𝒥𝑈subscript𝑠𝒥superscriptsubscript𝑠𝒦𝑇𝛿\tilde{U}^{(T)}(s\_{\mathcal{J}})=U(s\_{\mathcal{J}},s\_{\mathcal{K}}^{(T\delta)})over~ start\_ARG italic\_U end\_ARG start\_POSTSUPERSCRIPT ( italic\_T ) end\_POSTSUPERSCRIPT ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT ) = italic\_U ( italic\_s start\_POSTSUBSCRIPT caligraphic\_J end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT caligraphic\_K end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_T italic\_δ ) end\_POSTSUPERSCRIPT ). Then this solution converges to a (set of) human-optimal state(s).
###### Proof.
By the monotone convergence theorem, utility through the optimization process converges to a maximum value. Any state s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT with this value must have the sensitivity of all attributes equal and C(s\*)=0𝐶superscript𝑠0C(s^{\*})=0italic\_C ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) = 0, otherwise a proxy with two attributes of unequal sensitivity or two attributes respectively will cause increase in utility above U(s\*)𝑈superscript𝑠U(s^{\*})italic\_U ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) in a finite amount of time.
Suppose ∂U∂si(∂C∂si)−1(s\*)=α𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1superscript𝑠𝛼\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}(s^{\*})=\alphadivide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) = italic\_α for all attributes i𝑖iitalic\_i. We now proceed to show that U(s\*)=maxs∈𝒮U(s)𝑈superscript𝑠subscript𝑠𝒮𝑈𝑠U(s^{\*})=\max\limits\_{s\in\mathcal{S}}U(s)italic\_U ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) = roman\_max start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT italic\_U ( italic\_s ). Consider any other feasible state s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT. By convexity, it must be that
| | | | |
| --- | --- | --- | --- |
| | 00\displaystyle 0 | ≥C(s′)−C(s\*)absent𝐶superscript𝑠′𝐶superscript𝑠\displaystyle\geq C(s^{\prime})-C(s^{\*})≥ italic\_C ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) - italic\_C ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) | |
| | | ≥∑i(si′−si\*)∂U∂si(s\*)absentsubscript𝑖subscriptsuperscript𝑠′𝑖subscriptsuperscript𝑠𝑖𝑈subscript𝑠𝑖superscript𝑠\displaystyle\geq\sum\_{i}(s^{\prime}\_{i}-s^{\*}\_{i})\frac{\partial U}{\partial s\_{i}}(s^{\*})≥ ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) | |
Now, considering the difference in utility between states s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT and s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT.
| | | | |
| --- | --- | --- | --- |
| | U(s′)−U(s\*)𝑈superscript𝑠′𝑈superscript𝑠\displaystyle U(s^{\prime})-U(s^{\*})italic\_U ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) - italic\_U ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) | ≤∑i(si′−si\*)∂C∂si(s)absentsubscript𝑖subscriptsuperscript𝑠′𝑖subscriptsuperscript𝑠𝑖𝐶subscript𝑠𝑖𝑠\displaystyle\leq\sum\_{i}(s^{\prime}\_{i}-s^{\*}\_{i})\frac{\partial C}{\partial s\_{i}}(s)≤ ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s ) | |
| | | =∑i(si′−si\*)∂U∂si(∂C∂si)−1∂C∂si(s\*)absentsubscript𝑖subscriptsuperscript𝑠′𝑖subscriptsuperscript𝑠𝑖𝑈subscript𝑠𝑖superscript𝐶subscript𝑠𝑖1𝐶subscript𝑠𝑖superscript𝑠\displaystyle=\sum\_{i}(s^{\prime}\_{i}-s^{\*}\_{i})\frac{\partial U}{\partial s\_{i}}(\frac{\partial C}{\partial s\_{i}})^{-1}\frac{\partial C}{\partial s\_{i}}(s^{\*})= ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) divide start\_ARG ∂ italic\_U end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) | |
| | | =α∑i(si′−si\*)∂C∂si(s\*)absent𝛼subscript𝑖subscriptsuperscript𝑠′𝑖subscriptsuperscript𝑠𝑖𝐶subscript𝑠𝑖superscript𝑠\displaystyle=\alpha\sum\_{i}(s^{\prime}\_{i}-s^{\*}\_{i})\frac{\partial C}{\partial s\_{i}}(s^{\*})= italic\_α ∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT - italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) divide start\_ARG ∂ italic\_C end\_ARG start\_ARG ∂ italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) | |
| | | ≤0absent0\displaystyle\leq 0≤ 0 | |
Thus, s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT must have lower utility than s\*superscript𝑠s^{\*}italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT.
Since this applies for all feasible states, U(s\*)=maxs∈𝒮U(s)𝑈superscript𝑠subscript𝑠𝒮𝑈𝑠U(s^{\*})=\max\limits\_{s\in\mathcal{S}}U(s)italic\_U ( italic\_s start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) = roman\_max start\_POSTSUBSCRIPT italic\_s ∈ caligraphic\_S end\_POSTSUBSCRIPT italic\_U ( italic\_s )
∎
While this is optimal, we require the assumptions of both the impact-minimization robot and the efficient, interactive robot, each of which individually presents complex challenges in implementation.
5 Conclusion and Further Work
------------------------------
In this paper, we present a novel model of value (mis)alignment between a human principal and an AI agent. Fundamental to this model is that the human provides an incomplete proxy of their own utility function for the AI agent to optimize. Within this framework, we derive necessary and sufficient theoretical conditions for value misalignment to be arbitrarily costly. Our results dovetail with an emerging literature that connects harms from AI systems to shallow measurement of complex values [[32](#bib.bib32), [20](#bib.bib20), [5](#bib.bib5)]. Taken together, we view this as strong evidence that the ability of AI systems to be useful and beneficial is highly dependent on our ability to manage the fundamental gap between our qualitative goals and their representation in digital systems.
Additionally, we show that abstract representations of techniques currently being explored by other researchers can yield solutions with guaranteed utility improvement. Specifically, impact minimization and human interactivity, if implemented correctly, allow AI agents to provide positive utility in theory. In future work, we hope to generalize the model to account for fundamental preference uncertainty (e.g., as in [[12](#bib.bib12)]), limits on human rationality, and the aggregation of multiple human preferences. We are optimistic that research into the properties and limitations of the communication channel within this framework will yield fruitful insights in value alignment, specifically, and AI, generally.
Broader Impact
--------------
As AI systems become more capable in today’s society, the consequences of misspecified reward functions increase as well. Instances where the goal of the AI system and the preferences of individuals diverge are starting to emerge in the real world. For example, content recommendation algorithms optimizing for clicks causes clickbait and misinformation to proliferate [[33](#bib.bib33)]. Our work rigorously defines this general problem and suggests two separate approaches for dealing with incomplete or misspecified reward functions. In particular, we argue that, in the absence of a full description of attributes, the incentives for real-world systems need to be plastic and dynamically adjust based on changes in behavior of the agent and the state of the world.
Acknowledgments and Disclosure of Funding
-----------------------------------------
This work was partially supported by AFOSR, ONR YIP, NSF CAREER, NSF NRI, and OpenPhil. We thank Anca Dragan, Stuart Russell, and the members of the InterACT lab and CHAI for helpful advice, guidance, and discussions about this project. |
7a5d8d2b-fcc5-40ce-8712-cbe78ecceb70 | trentmkelly/LessWrong-43k | LessWrong | D&D.Sci III Evaluation and Ruleset
This is a followup to the D&D.Sci post I made last week; if you haven’t already read it, you should do so now before spoiling yourself.
Here is the web interactive I built to let you test your solution; below is a complete explanation of the rules used to generate the dataset. You’ll probably want to test your answer before reading any further.
----------------------------------------
Ruleset
Mage Fights
Mages have differing levels of Power. When they fight, each of them roll two ten-sided dice, and add the result to their Power. The mage with the higher total result wins; if it’s a tie, the defender wins.
The raw Power of each mage at the time of your summoning is given below:
MagePowerVitamancer A19Vitamancer B22Geomancer A21Geomancer B12Cryomancer23Pyromancer A20Pyromancer B21Necromancer A22Necromancer B15Necromancer C13Electromancer20
Modifiers and Anomalies
Power is modified by the circumstances in which a fight occurs. The main modifiers are:
* All mages get a homefield advantage of +4 when operating in their own territory.
* Geomancers get +9 when defending and -9 when attacking.
* Pyromancers get -9 when defending and +9 when attacking.
In addition, all the mages have something unique going on with them:
* Geomancer B is the one who summoned you, and the keeper of the record provided. This explains, among other things, why all the earliest entries are fights involving him. (This fact is completely irrelevant to attempts to maximize wins or losses, but you can get unique outcomes if you make him the only winner or only loser on his side.)
* Vitamancer A spent much of his youth specializing in handling Necromancer C, and gets an extra +7 whenever they fight.
* Necromancer B is an ex-Vitamancer, and gets a +4 when operating in Vitamancer territory.
* Geomancer A is an ex-Necromancer, and gets a +4 when operating in Necromancer territory.
* The Cryomancer and Necromancer A strongly dislike fighting each other, and have made an under-the-tabl |
ddb1eff4-ee85-4771-a5f6-ab05d6f1732d | trentmkelly/LessWrong-43k | LessWrong | "model scores" is a questionable concept
A couple of years ago, while I was working on a problem of what to do with the predictions of a random forest model, I noticed that the mean of the predictions on the test set didn't match the mean of the outcomes in the training data. Let me make that more concrete; consider the following problem (isomorphic to a problem I worked on in the past):
A town has two donut shops. The donut shops are in terrible competition, and the townspeople feel very strongly about the shops. It's, like, political. You run the good shop. Every year, 6% of people switch to your competitor. You are a data-driven company, and your kid needed something to do, so they've been sitting in the corner of the shop for years scribbling down information about each visitor and visit. So now you have a bunch of data on all your customers. "Did they come in that Monday 4 weeks ago? How about the Monday after that? How many donuts did they buy each month?" and so on. Representing "They switched to the other shop forever" as 1 and "that didn't happen" as 0, we have a label vector labels of 0's and 1's, grouped into, say, one data record and label per customer-year. We want to predict who might switch soon, so that we can try and prevent the switch.
So we put your kid's notes into some predictive model, train it, and make predictions on some other set of data not used in the training (call this the test set). Each prediction will take as input a data record that looks something like (Maxwell, Male, age 30, visited twice this year), and output a continuous prediction in the range (0, 1) for how likely Maxwell is to leave. You'll also have a label, 1 or 0, for whether Maxwell actually left. After doing these predictions on the test set, you can check how well your model is modeling by comparing the true labels to the predictions in various ways. One thing to check when figuring out how well you did: check that the average number of people predicted to leave your shop in past years is close to the num |
8d60f280-9b98-47fc-a54d-e2a009fb8433 | trentmkelly/LessWrong-43k | LessWrong | Real-Life Examples of Prediction Systems Interfering with the Real World (Predict-O-Matic Problems)
Thanks to Ozzie Gooen for reviewing this post.
Introduction
The Parable of the Predict-O-Matic is a short story which considers a forecasting system which is ostensibly set-up to maximize accuracy, and which ends up interfering with the world in ways not intended. In the original story, some of these problems were:
* Fixed point problems / self-fulfilling prophecies
> “Its answers will shape events. If it says stocks will rise, they'll rise. If it says stocks will fall, then fall they will. Many people will vote based on its predictions.”
* Nudge towards legibility and predictability
> “You keep thinking of the line from Orwell's 1984 about the boot stamping on the human face forever, except it isn't because of politics, or spite, or some ugly feature of human nature, it's because a boot stamping on a face forever is a nice reliable outcome which minimizes prediction error.”
* Markets for entropy, which can be thought of as the opposite of the previous problem. In a prediction market, a market participant who could actively change an outcome has an incentive to first make a big bet for an unlikely outcome, and then actively make it come to pass.
> “Suppose you have a prediction market that's working well. It makes good forecasts, and has enough money in it that people want to participate if they know significant information. Anything you can do to shake things up, you've got a big incentive to do. Assassination is just one example. You could flood the streets with jelly beans. If you run a large company, you could make bad decisions and run it into the ground, while betting against it -- that's basically why we need rules against insider trading, even though we'd like the market to reflect insider information."
* Unwanted agency (as opposed to tool AI behavior)
> “You understand what you are. It isn't quite right to say you are the Predict-O-Matic. You are a large cluster of connections which thinks strategically. You generate useful information, and |
a7be5b80-c50f-457b-8425-e8211b004a79 | trentmkelly/LessWrong-43k | LessWrong | Is School of Thought related to the Rationality Community?
If so, who are they?
Link: https://yourbias.is/
At a gloss the material looks really polished and topical. |
658d3d78-5f8d-4c16-840a-65376a9a683f | trentmkelly/LessWrong-43k | LessWrong | A Transhumanist Poem
**Note: I'm not a poet. I hardly ever write poetry, and when I do, it's usually because I've stayed up all night. However, this seemed like a very appropriate poem for Less Wrong. Not sure if it's appropriate as a top-level post. Someone please tell me if not.**
Imagine
The first man
Who held a stick in rough hands
And drew lines on a cold stone wall
Imagine when the others looked
When they said, I see the antelope
I see it.
Later on their children's children
Would build temples, and sing songs
To their many-faced gods.
Stone idols, empty staring eyes
Offerings laid on a cold stone altar
And left to rot.
Yet later still there would be steamships
And trains, and numbers to measure the stars
Small suns ignited in the desert
One man's first step on an airless plain
Now we look backwards
At the ones who came before us
Who lived, and swiftly died.
The first man's flesh is in all of us now
And for his and his children's sake
We imagine a world with no more death
And we see ourselves reflected
In the silicon eyes
Of our final creation |
79c2d5ab-a1a3-4ae8-8b70-312d56ce4535 | trentmkelly/LessWrong-43k | LessWrong | Cascio in The Atlantic, more on cognitive enhancement as existential risk mitigation
Jamais Cascio writes in the atlantic:
> Pandemics. Global warming. Food shortages. No more fossil fuels. What are humans to do? The same thing the species has done before: evolve to meet the challenge. But this time we don’t have to rely on natural evolution to make us smart enough to survive. We can do it ourselves, right now, by harnessing technology and pharmacology to boost our intelligence. Is Google actually making us smarter? ...
>
> ... Modafinil isn’t the only example; on college campuses, the use of ADD drugs (such as Ritalin and Adderall) as study aids has become almost ubiquitous. But these enhancements are primitive. As the science improves, we could see other kinds of cognitive-modification drugs that boost recall, brain plasticity, even empathy and emotional intelligence. They would start as therapeutic treatments, but end up being used to make us “better than normal.”
Read the whole article here.
This relates to cognitive enhancement as existential risk mitigation, where Anders Sandberg wrote:
> Would it actually reduce existential risks? I do not know. But given correlations between long-term orientation, cooperation and intelligence, it seems likely that it might help not just to discover risks, but also in ameliorating them. It might be that other noncognitive factors like fearfulness or some innate discounting rate are more powerful.
The main criticisms of this idea generated in the Less Wrong comments were:
> The problem is not that people are stupid. The problem is that people simply don't give a damn. If you don't fix that, I doubt raising IQ will be anywhere near as helpful as you may think. (Psychohistorian)
> Yes, this is the key problem that people don't really want to understand. (Robin Hanson)
> Making people more rational and aware of cognitive biases material would help much more (many people)
These criticisms really boil down to the same thing: people love their cherished falsehoods! Of course, I cannot disagree with this s |
59a6647e-3794-414c-a98a-3b73ff0d4b30 | trentmkelly/LessWrong-43k | LessWrong | Critiquing Gary Taubes, Part 2: Atkins Redux
Previously: Mainstream Nutrition Science on Obesity
Edit: In retrospect, I think it maybe should have combined this post with part 3. Unfortunately, the problem of what to do with existing comments makes that hard to fix now.
Taubes first made a name for himself as a low-carb advocate in 2002 with a New York Times article titled "What if It's All Been a Big Fat Lie?" When I first read this article, I was getting extremely suspicious by the second paragraph (emphasis added):
> If the members of the American medical establishment were to have a collective find-yourself-standing-naked-in-Times-Square-type nightmare, this might be it. They spend 30 years ridiculing Robert Atkins, author of the phenomenally-best-selling ''Dr. Atkins' Diet Revolution'' and ''Dr. Atkins' New Diet Revolution,'' accusing the Manhattan doctor of quackery and fraud, only to discover that the unrepentant Atkins was right all along. Or maybe it's this: they find that their very own dietary recommendations -- eat less fat and more carbohydrates -- are the cause of the rampaging epidemic of obesity in America. Or, just possibly this: they find out both of the above are true.
>
> When Atkins first published his ''Diet Revolution'' in 1972, Americans were just coming to terms with the proposition that fat -- particularly the saturated fat of meat and dairy products -- was the primary nutritional evil in the American diet. Atkins managed to sell millions of copies of a book promising that we would lose weight eating steak, eggs and butter to our heart's desire, because it was the carbohydrates, the pasta, rice, bagels and sugar, that caused obesity and even heart disease. Fat, he said, was harmless.
>
> Atkins allowed his readers to eat ''truly luxurious foods without limit,'' as he put it, ''lobster with butter sauce, steak with béarnaise sauce . . . bacon cheeseburgers,'' but allowed no starches or refined carbohydrates, which means no sugars or anything made from flour. Atkins banned even fru |
1649cd33-7a34-409f-916b-ea3dd1ceb2db | trentmkelly/LessWrong-43k | LessWrong | Right for the Wrong Reasons
One of the few things that I really appreciate having encountered during my study of philosophy is the Gettier problem. Paper after paper has been published on this subject, starting with Gettier's original "Is Justified True Belief Knowledge?" In brief, Gettier argues that knowledge cannot be defined as "justified true belief" because there are cases when people have a justified true belief, but their belief is justified for the wrong reasons.
For instance, Gettier cites the example of two men, Smith and Jones, who are applying for a job. Smith believes that Jones will get the job, because the president of the company told him that Jones would be hired. He also believes that Jones has ten coins in his pocket, because he counted the coins in Jones's pocket ten minutes ago (Gettier does not explain this behavior). Thus, he forms the belief "the person who will get the job has ten coins in his pocket."
Unbeknownst to Smith, though, he himself will get the job, and further he himself has ten coins in his pocket that he was not aware of-- perhaps he put someone else's jacket on by mistake. As a result, Smith's belief that "the person who will get the job has ten coins in his pocket" was correct, but only by luck.
While I don't find the primary purpose of Gettier's argument particularly interesting or meaningful (much less the debate it spawned), I do think Gettier's paper does a very good job of illustrating the situation that I refer to as "being right for the wrong reasons." This situation has important implications for prediction-making and hence for the art of rationality as a whole.
Simply put, a prediction that is right for the wrong reasons isn't actually right from an epistemic perspective.
If I predict, for instance, that I will win a 15-touch fencing bout, implicitly believing this will occur when I strike my opponent 15 times before he strikes me 15 times, and I in fact lose fourteen touches in a row, only to win by forfeit when my opponent intentionally |
969e1e75-ff13-4143-ad27-7b7037db1ef0 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Backplay: "Man muss immer umkehren"
1 Introduction
---------------
An important goal of AI research is to construct agents that can enter new environments and learn how to achieve desired goals (Levine et al., [2016](#bib.bib22)). A key paradigm for this task is deep reinforcement learning (deep RL) which succeeded in many environments, including complex zero-sum games such as Go, Poker, and Chess (Silver et al., [2016](#bib.bib34); Moravcík et al., [2017](#bib.bib24); Brown & Sandholm, [2017a](#bib.bib4); Silver et al., [2017](#bib.bib35)). However, training an RL agent can be extremely time consuming, particularly in environments with sparse rewards. In these settings, the agent typically requires a large number of episodes to stumble upon positive reward and learn even a moderately effective policy that can then be refined. Reward sparsity is often resolved via humans ‘densifying’ the underlying reward function. Such reward shaping can be effective, but can also change the set of optimal policies towards having unintended side effects (Ng et al., [1999](#bib.bib27); Clark & Amodei, [2016](#bib.bib7)).
We consider an alternative technique for accelerating RL in sparse reward settings without reward shaping. The essence of our proposed technique is to create a curriculum for the agent via reversing a single trajectory (i.e. state sequence) of reasonably good, but not necessarily optimal, behavior. That is, we start our agent at the end of a demonstration and let it learn a policy in this easier setup. We then move the starting point backward until the agent is training solely on the initial state of the corresponding task. We call this technique Backplay and include a visual reference in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Backplay: ‘Man muss immer umkehren’").
We experiment in two scenarios: a maze environment where the agent must find a path to a goal and Pommerman (MultiAgentLearning, [2018](#bib.bib25)), a complex four-player competitive game. We empirically show that Backplay vastly decreases the number of samples required to learn a good policy. In addition, we see that an agent trained with Backplay can outperform its corresponding demonstrator and even learn an optimal policy when using a sub-optimal demonstration. Notably, the Backplay agent is able to learn just as well in environments with a single sparse reward as in those with a dense shaped reward. This extends to a challenging setup where it successfully generalizes to unseen Pommerman scenarios while the baselines fail to learn at all.
Figure 1: Backplay graphic description. We first collect a demonstration, from which we build a curriculum over the states. We then sample a state according to that curriculum and initialize our agent accordingly.
2 Related Work
---------------
The most related work to ours is an independent and concurrent blog post on a method similar to Backplay to train policies on the challenging Atari game Montezuma’s Revenge (Salimans & Chen, [2018](#bib.bib32)). While both propose training an RL agent by starting each episode from a demonstration state based on a reverse curriculum, there are a number of key differences between the two reports.
First, our work considers the problem of learning effective policies in environments with various initial configurations, each corresponding to a different demonstration. Second, we analyze whether the technique can find optimal policies when trained using sub-optimal demonstrations and the extent to which those policies can generalize to unseen initial conditions. Third, we do not use an adaptive schedule based on the agent’s success rate to advance its starting state because we did not find this to improve the overall performance or sample complexity. And finally, we include both empirical and theoretical analysis on when the efficacy of this approach is higher than that of standard RL.
Behavioral cloning aims at explicitly encouraging a policy to mimic an expert policy (Ross et al., [2011](#bib.bib31); Daumé et al., [2009](#bib.bib8); Zhang & Cho, [2016](#bib.bib38); Laskey et al., [2016](#bib.bib18)). Many algorithms in this literature require access to the expert’s policy or use supervision from the expert’s actions and/or states (Hosu & Rebedea, [2016](#bib.bib14); Nair et al., [2017](#bib.bib26); Hester et al., [2017](#bib.bib11); Ho & Ermon, [2016](#bib.bib12); Aytar et al., [2018](#bib.bib1); Lerer & Peysakhovich, [2018](#bib.bib21); Peng et al., [2018](#bib.bib28)). Backplay only requires access to a single trajectory of states visited by this policy and uses these states to initialize a type of curriculum (Bengio et al., [2009](#bib.bib2)) rather than as a supervised objective.
Recent work by McAleer et al. ([2018](#bib.bib23)) frames the problem of solving the Rubik’s cube as a sparse reward model-based RL task with inputs generated by starting at the goal state and taking random actions. Similarly, Hosu & Rebedea ([2016](#bib.bib14)) uses intermediate states of an expert demonstration as starting states from which a policy is executed. However, neither of these works explore imposing a curriculum over the demonstration.
The idea of using a reverse curriculum for sparse reward reinforcement learning problems has recently been proposed. Florensa et al. ([2017](#bib.bib9)) uses a sequence of short random walks to gradually initialize an RL agent from a set of start states increasingly far from a known goal location. Their method however only applies to tasks with a unique and fixed goal state, while our technique could be applied in settings with multiple non-stationary targets. In a similar fashion, Ivanovic et al. ([2018](#bib.bib15)) uses a known (approximate) dynamics model to create a backwards curriculum for continuous control tasks. Their approach requires a physical prior which it is not always available and often not applicable in multi-agent scenarios. In contrast, Backplay automatically creates a curriculum and can be used in any resettable environment as long as demonstrations are available. An interesting direction for future research is using some of these previously proposed ideas to make Backplay more efficient.
3 Backplay
-----------
Consider the standard formalization of a single agent Markov Decision Process (MDP) defined by a set of states S, a set of actions A, and a transition function T:S×A→S which gives the probability distribution of the next state given a current state and action. If P(A) denotes the space of probability distributions over actions, the agent chooses actions by sampling from a stochastic policy π:S→P(A), and receives reward r:S×A→R at every time step. The agent’s goal is to construct a policy which minimizes its discounted expected return Rt=E[∑∞k=0γkrt+k+1] where rt is the reward at time t and γ∈[0,1] is the discount factor, and the expectation is taken with respect to both the policy and the environment.
The final component of an MDP is the distribution of initial starting states s0. The key idea in Backplay is that instead of starting the MDP from this s0 we have access to a demonstration which reaches a sequence of states {sd0,sd1,…,sdT}. For each training episode, we uniformly sample starting states from the sub-sequence {sdT−k,sdT−k+1,…,sdT−j} for some chosen window [j,k], j≤k. As training continues, we ‘advance’ the window by increasing the values of j and k until we are training on the initial state in every episode. In this manner, our hyperparameters for Backplay are only the (j,k) pairs (windows) and the training epochs at which we advance them.
In the next section, we explore more formally under what conditions Backplay can improve the sample-efficiency of RL training. First, we give intuition about under what conditions Backplay can increase training speed and when it can lead to a better policy than that of the demonstrator.
In Figure [5](#S3.F5 "Figure 5 ‣ 3 Backplay ‣ Backplay: ‘Man muss immer umkehren’"), we show three toy grid worlds. In each, the agent begins at s0, movement is costly, and the agent receives a reward of 1 if and only if they navigate to the goal s∗. Thus, the optimal strategy is to reach the goal in as few steps as possible.
The left grid world shows a sparse reward environment where trial-and-error model-free RL will converge slowly. Backplay, using the sub-optimal expert policy (the line), will decrease the requisite training time because it will position the agents close to high value states. In addition, the agent will likely surpass the expert policy because, unlike in behavioral cloning approaches, Backplay does not encourage the agent to imitate expert actions. Rather, the curriculum forces the agent to first explore states with large associated value and consequently, estimating the value function suffers less from the curse of dimensionality.
The middle grid demonstrates a maze with bottlenecks. Backplay will vastly decrease the exploration time because the agent will be less prone to exploring the errant right side chamber that it can get lost in if it traverse to the right instead of going up to the goal. These sorts of bottlenecks exist in many real life scenarios, including Pommerman where the bomb laying action itself is an example.
On the right side grid, Backplay will similarly not have difficulty surpassing the sub-optimal expert because it will learn to skip the large ingress into the middle of the maze. However, it will have trouble doing better than a model-free policy not using Backplay because that policy will spend all of its training initialized from s0 and consequently will be more likely to stumble upon the optimal solution of going up and then right, whereas Backplay will spend the dominant amount of its early training starting from the sub-optimal states included in the demonstration.
| | | |
| --- | --- | --- |
| Backplay: Good
(a) Backplay: Good
| Backplay: Good
(b) Backplay: Good
| Backplay: Bad
(c) Backplay: Bad
|
Figure 5: Three environments that capture the intuition behind Backplay and its limitations. Backplay will likely be advantageous on the first and second mazes and surpass or match the expert’s demonstration. However, it will perform worse on the third maze.
###
3.1 Preliminary Analysis
We now derive some basic results for Backplay training with a Q-learning agent in a simple navigation environment. Because we are specifically interested in the case of sparse rewards, in our environment agents receive a positive reward of 1 if they reach a goal absorbing state and a reward of 0 otherwise.
Consider a generic map as a connected, undirected graph G=(V,E).
An agent is placed at a fixed node v0∈V and can move to neighboring nodes at a negative reward of −1 for each move. Its goal is to reach a target node v∗∈V, terminating the episode.
This corresponds to an MDP M=(S,A,P,R) with state space given by S∼V,
action A∼E, deterministic
transition kernel corresponding to P(st+1=j | st=i,at=(l,k))=δ(j=k)δ(l=i)+δ(j=i)δ(l≠i), R(st,at)=1 for st+1≠v∗, and is absorbing at s=v∗.
Assume now that π is a fixed policy on M, such that the underlying Markov chain
| | | |
| --- | --- | --- |
| | Kπ(s′,s)=∑a∈AP(s′=j | s0=s,a0=a)π(a|s0=s) | |
is irreducible and aperiodic. K has a single absorbing state at s=s∗.
Our goal is to estimate the value function associated to this policy, equivalent to the expected first-passage time of the
Markov chain K from s0=s to s∗:
| | | |
| --- | --- | --- |
| | Vπ(s)=Eτ(s,s∗)=Emin{j≥0 ;s.t.sj=s∗,s0=s,si+1∼K(⋅,si)} . | |
Instead of analyzing this Markov chain, we make a simplifying assumption that the performance of the navigation
task is captured by projecting this chain into the 1D process zt=distG(st,s∗)∈{0,1,…,M}, with M=diam(G).
That is, we consider a latent variable at each state indicating its distance to the target. The process zt is in general
not Markovian, since its transition probabilities are not only a function of zt. Its Markov approximation ¯zt is defined
as the Markov chain given by the expected
transition probabilities
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | Pr(¯zt+1=l+1|¯zt=l) | = | αl:=Prμ(zt+1=l+1 | zt=l) , | | (1) |
| | Pr(¯zt+1=l|¯zt=l) | = | βl=Prμ(zt+1=l | zt=l) , l=0…M | | (2) |
and thus Pr(¯zt+1=l−1|¯zt=l)=1−αl−βl=Prμ(zt+1=l−1 | zt=l) , under the stationary distribution μ of K.
In other words, we create an equivalence class given by the level sets of the optimal value function Vopt(s), the shortest-path distance in G.
The analysis of Backplay in the 1D chain ¯zt is now tractable. For a fixed Backplay step size m>0, with m≪M, we initialize
the chain at ¯z0=M−m. Since Pr(¯zm=M)=∏Mj=M−mαj=γM−m,M, after O(γ−1M−m,M) trials of length ≤M
we will reach the absorbing state and finally have a signal-carrying update for the Q-function at ¯zM−1. We can consequently merge ¯zM−1 into the absorbing state and reduce the length of the chain by one. By repeating the argument m times, we thus obtain that after O(∑mj=0γ−1M−m,M−j)=O(mγ−1M−m,M) trials, the Q-function is updated at ¯z0=M−m. By repeating this argument at Backplay steps km, k=1,…M/m, we reach a sample complexity of
| | | |
| --- | --- | --- |
| | Tm=M/m−1∑k=0O(m∑j=0γ−1km,(k+1)m−j) . | |
In the case where αl=α for all l, we obtain γ−1km,(k+1)m−j=γ0,m−j=α−m+j and therefore Tm=O(M(1−αm+1m(1−α)α−m), where the important term is the rate Mmα−m.
On the other hand, from Hong & Zhou ([2017](#bib.bib13)), we verify that the first-passage time τ(0,M) in a skip-free finite Markov chain of M states with a single absorbing state is a random variable whose moment-generating function φ(s)=Esτ(s,s∗) is given by
| | | | |
| --- | --- | --- | --- |
| | φ(s)=M∏j=1(1−λj)s1−λjs , | | (3) |
where λ1,…,λM are the non-unit eigenvalues of the transition probability matrix. It results that Eτ(0,M)=φ′(1)=∑Mj=111−λj≈(1−λ1)−1, which corresponds to the reciprocal spectral gap, and therefore the model without Backplay will on average take TM=Θ(α−M) trials to reach the absorbing state and thus receive information. We thus obtain sample complexity gains that are exponential in the diameter of the graph (O(Mα−m) versus O(α−M)) in this simple birth-death 1d scenario.
Of course, this analysis is preliminary in many levels. For example, the projection into the 1-dimensional skip-free process makes two dangerous assumptions: (i) The level sets of our estimated value functions correspond to the level sets of the true value function, and (ii) The projected process is well described by its Markovian approximation. The first assumption is related to the ability of our function approximation to generalize from visited states s to unvisited states s′ such that distG(s′,s∗)=distG(s,s∗). The second assumption has an impact on the bounds derived above, and could be worked around by replacing conditional probabilities with infima over the level sets.
As in the simple examples above, this analysis suggests that Backplay is not a ‘universal’ strategy to improve sample complexity. Indeed, even in the navigation setting, if the task now randomizes the initial state s0, a single demonstration trajectory does not generally improve the coverage of the state-space outside an exponentially small region around said trajectory. As a simple example, imagine a binary tree graph and a navigation that starts at a random leaf and needs to reach the root. A single expert trajectory will be disjoint from half of the state space (because the root is absorbing), thus providing no sample complexity gains on average.
The preceding analysis suggests that a full characterization of Backplay is a fruitful direction for reinforcement learning theory, albeit beyond the scope of this paper.
4 Experiments
--------------
We now move to evaluating Backplay empirically in two environments: a grid world maze and a four player free-for-all game. The questions we study across both environments are the following:
* Is Backplay more efficient than training an RL agent from scratch?
* How does the quality of the given demonstration affect the effectiveness of Backplay?
* Can Backplay agents surpass the demonstrator when it is non-optimal?
* Can Backplay agents generalize?
###
4.1 Maze
For the Maze setup, we generated grid worlds of size 24×24 with 120 randomly placed walls, a random start position, and a random goal position. We then used the A\* search algorithm to generate trajectories. These included both Optimal demonstrations (true shortest path) and N-Optimal demonstrations, paths that were N steps longer than the shortest path. For N-Optimal demonstrations, we used a noisy A\* where at each step we follow A\* with probability p or otherwise chose a random action. We considered N∈{5,10}. In all scenarios, we filtered any path that was less than 35 in length and stopped when we found a hundred valid training games. Note that we held the demonstration length invariant rather than the optimal length.
The observation space in maze games consists of four 24×24 binary maps. They contain ones at the positions of, respectively, the agent, the goal, passages, and walls. The action space has discrete options: Pass, Up, Down, Left, or Right.
The game ends when the agent has reached its goal or after a maximum of 200 steps. The agent receives reward +1 for reaching the goal and a -0.03 penalty for each step taken in the environment. Thus optimal policies are those which reach the goal in the fastest possible time.
We use a standard deep RL setup for our agents. The agent’s policy and value functions are parameterized by a convolutional neural network with 2 layers each of 32 output channels, followed by two linear layers with 128 dimensions. Each of the layers are followed by ReLU activations. This body then feeds two heads, a scalar value function and a softmax policy function over the five actions. All of the CNN kernels are 3x3 with stride and padding of one.
We train our agent using Proximal Policy Optimization (PPO, Schulman et al. ([2017](#bib.bib33))) with γ=0.99, learning rate 1×10−3, batch size 102400, 60 parallel workers, clipping parameter 0.2, generalized advantage estimation with τ=0.95, entropy coefficient 0.01, value loss coefficient 0.5, mini-batch size 5120, horizon 1707, and 4 PPO updates at each iteration.
We compare several training regimes. The first is Backplay, which uses the Backplay algorithm corresponding to the sequence of windows [[0,4),[4,8),[8,16),[16,32),[32,64),[64,64]] and epochs [0,350,700,1050,1400,1750]. When we get to a training epoch represented in the sequence of epochs, we advance to the corresponding value in the sequence of windows. For example, say we began training at epoch 20. Then we select a game at random, an N∈[16,32), and start the agent in that game N steps from the end. Whenever a pair is chosen such that the game’s length is smaller than N, we use the initial state. The second, Standard, is a baseline of standard RL training from s0. The last, Uniform, is an ablation that considers how important is the curriculum aspect of Backplay by sampling states from the entire demonstration randomly rather than according to the Backplay distribution.
Table [1](#S4.T1 "Table 1 ‣ 4.1 Maze ‣ 4 Experiments ‣ Backplay: ‘Man muss immer umkehren’") shows that Standard has immense trouble learning in this sparse reward environment while Backplay and Uniform both find an optimal path approximately 30-50% of the time and a path within five of the optimal path almost always. Thus, in this environment, demonstrations of even sub-optimal experts are extremely useful while the curriculum created by Backplay is not necessary. It does, however, aid convergence speed (learning curves in Appendix [A.4](#A1.SS4 "A.4 Maze Learning Curves ‣ Appendix A Appendix ‣ Backplay: ‘Man muss immer umkehren’")). We will see in the next experiment that the curriculum becomes important as the environment increases in complexity.
Finally, we looked at whether Backplay generalizes in two ways. First, we had our agents explore ten new test mazes. Second, we tried generating new mazes for each episode of training. None of our training regimes were successfully able to complete either scenario, suggesting that in this environment Backplay does not aid in generalization.
| Algorithm | Expert | % Optimal | % 0-5 Optimal | Avg Suboptimality | Std Suboptimality |
| --- | --- | --- | --- | --- | --- |
| Standard | None | 0 | 0 | N/A | N/A |
| Uniform | Optimal | .27 | .91 | 8.26 | 32.92 |
| Uniform | 5-Optimal | .51 | .98 | 2.04 | 17.39 |
| Uniform | 10-Optimal | .49 | .98 | 2.04 | 16.79 |
| Backplay | Optimal | .31 | 1 | 0.64 | 4.96 |
| Backplay | 5-Optimal | .37 | .94 | 7.94 | 33.35 |
| Backplay | 10-Optimal | .54 | .99 | 0.37 | 3.49 |
Table 1: Results after 2000 epochs on 100 mazes. Standard has failed to learn, while both Backplay and Uniform succeed on almost all mazes and, importantly, outperform the experts’ demonstrations. These results were consistent across each of three different seeds.
###
4.2 Pommerman
The Pommerman environment is based off of the classic console game Bomberman and will be a competition at the upcoming NIPS. It is played on an 11x11 grid where each of four agents have six actions on every turn: move in a cardinal directions, pass, or lay a bomb. The agents begin fenced in their own area by two different types of walls - rigid and wooden (see Figure [6](#S4.F6 "Figure 6 ‣ 4.2 Pommerman ‣ 4 Experiments ‣ Backplay: ‘Man muss immer umkehren’")). The former are indestructible while bombs destroy the latter. Upon blowing up wooden walls, there is a uniform chance at yielding one of three power-ups: an extra bomb, an extra unit of range in the agent’s bombs, and the ability to kick bombs. The maps are designed randomly, albeit there is always a guaranteed path between any two agents.
In our experiments, we use the purely adversarial Free-For-All (FFA) environment. This environment is introduced by MultiAgentLearning ([2018](#bib.bib25)) and we point the reader there for more details. The winner of the game is the last agent standing and receives a reward of 1. The game is played from the perspective of one agent picked uniformly among the four, where each agent is defined by which of the four starting positions they begin. The three opponents are copies of the winner of the June 3rd 2018 FFA competition, a strong competitor using a Finite State Machine Tree-Search approach (FSMTS, Zhou et al. ([2018](#bib.bib39))).
Figure 6: Pommerman start state. Each agent begins in one of four positions. Yellow squares are wood, brown are rigid, and the gray are passages.
The input to our model consists of 18 11x11 channels, which are together a function of the observation state the game provides. A detailed description of this mapping is given in Appendix [A.1](#A1.SS1 "A.1 Pommerman Observation State ‣ Appendix A Appendix ‣ Backplay: ‘Man muss immer umkehren’"). We feed the policy network the last two observation states concatenated together for a total of a 36x11x11 input.
The game ends either when the learning agent wins, when it dies, or when 800 steps have passed. Winning gives a +1 reward and both drawing and losing give −1. Consequently, this is an environment with a very sparse reward, so we refer to the runs on this setup as Sparse. In some of our experiments, labeled Dense, we add another shaping reward of +0.1 every time the agent picks up an item.
We use a similar setup to that used in the Maze game. The architecture differences are that we have an additional two convolutional layers at the beginning, use 256 output channels, and have output dimensions of 1024 and 512, respectively, for the linear layers. This architecture was not tuned at all during the course of our experiments. Further hyperparameter differences are that we used a learning rate of 3×10−4 and a gamma of 1.0. These models trained for ∼50M frames.111In our early experiments, we also used a batch-size of 5120, which meant that the sampled transitions were much more correlated. While Backplay would train without any issues in this setup, Standard would only marginally learn. In an effort to improve our baselines, we did not explore this further.
We demonstrate results on the same three training regimes of Standard, Backplay, and Uniform. We also explored using DAgger (Ross et al. ([2011](#bib.bib31))) but found that our agent achieved approximately the same win rate (∼20%) as what we would expect when four FSMTS agents played each other (given that there are also ties).
Pommerman can be difficult for reinforcement learning agents. The agent must learn to effectively wield the bomb action in order to win against competent opponents. However, bombs destroy agents indiscriminately, so placing one without knowing how to retreat is often a negative experience. Since agents begin in an isolated area, they are prone to converging to policies which do not use the bomb action (as seen in the histograms in Figure [9](#S4.F9 "Figure 9 ‣ 4.2 Pommerman ‣ 4 Experiments ‣ Backplay: ‘Man muss immer umkehren’")).
| | |
| --- | --- |
| Standard Sparse
(a) Standard Sparse
| Backplay Sparse
(b) Backplay Sparse
|
Figure 9: Typical histograms for how the Pommerman action selections change over time. From left to right are the concatenated counts of the actions (Pass, Up, Down, Left, Right, Bomb), delineated on the y-axis by the epoch. Note how the Standard agent learns to not use bombs.
Our Backplay hyperparameters differed based on whether we were training on four games or a hundred games. Both corresponded to the sequence of windows [[0,32),[24,64),[56,128),[120,256),[248,512),[504,800],[800,800]], while the former changed windows at epochs [0,50,100,150,200,250,300] and the latter at epochs [0,85,170,255,340,425,510].
We begin by considering an easy situation where maps are drawn from one of four fixed Pommerman board configurations (maps) with demonstrations provided by the four FSMTS agents. We independently consider two Backplay trajectories, one which follows the winner and one where it follows the second place finisher. We change the FFA game to an MDP by replacing one of the FSMTS with our RL trained agents.
We see in Figure [14](#S4.F14 "Figure 14 ‣ 4.2 Pommerman ‣ 4 Experiments ‣ Backplay: ‘Man muss immer umkehren’") that Backplay is capable of producing agents that can defeat FSMTS agents in both sparse and dense setups while Standard requires hand-tuned reward shaping to construct competitive agents.
| | | | |
| --- | --- | --- | --- |
| Backplay
(a) Backplay
| Backplay
(b) Backplay
| Standard & Uniform
(c) Standard & Uniform
| Standard & Uniform
(d) Standard & Uniform
|
Figure 14: Results when training on four maps (four seeds). Charts a and c are trained from the perspective of the winning agent. Charts b and d are from that of the second place agent. Dense is with reward shaping and Sparse is without.
The Backplay models first encounter the initial state at epoch 250 (vertical red line). From epoch 300 onward (vertical green line), it only ever sees the starting state and success is measured strictly over that regime.
While Standard achieves strong results only in the dense setting, Backplay achieves comparable performance in both sparse and dense setups, and Uniform fails to learn a worthwhile policy in either.
We then consider a harder setup where the board is drawn from one of hundred maps. Here, Backplay achieves high success rates while Standard and Uniform models fail to learn anything of substance (Figure [15](#S4.F15 "Figure 15 ‣ 4.2 Pommerman ‣ 4 Experiments ‣ Backplay: ‘Man muss immer umkehren’")). Moreover, it learns to play half of the games at win rates higher than 80% (Table [4](#A1.T4 "Table 4 ‣ A.3 Win Rates ‣ Appendix A Appendix ‣ Backplay: ‘Man muss immer umkehren’") in [A.3](#A1.SS3 "A.3 Win Rates ‣ Appendix A Appendix ‣ Backplay: ‘Man muss immer umkehren’")).
Unlike in the maze experiment, we see that the Backplay agent can, to some extent, generalize to unseen maps. Backplay wins 416 / 1000 games on a held out test set of ten never before seen maps, with the following success rates on each of the ten boards: 85.3%, 84.1%, 81.6%, 79.5%, 52.4%, 47.4%, 38.1%, 22.6%, 20%, and 18.3%.
Figure 15: Results when training on a hundred maps from the winner’s perspective (four seeds). Dense is with reward shaping and Sparse is without. We show the success rate only over the initial state. Backplay first encounters this state at epoch 340 and it becomes the default for half of the environments at 425.
They do exceptionally well while the Uniform and Standard approaches do not show any reasonable signs of progress even though they have trained on this state the entire time.
###
4.3 Experience
We trained a large number of models through our research into Backplay. Though these findings are tangential to our main points we list some observations here that may be helpful for other researchers working with Backplay.
First, we found that Backplay does not perform well when the curriculum is advanced too quickly, however it does not fail when the curriculum is advanced ‘too slowly.’ Thus, researchers interested in using Backplay should err on the side of advancing the window too slowly rather than too quickly.
Second, we found that Backplay does not need to hit a high success rate before advancing. Initially, we tried using adaptive approaches that advanced the window when the agent reached a certain success threshold. This worked but was too slow. Our hypothesis is what is more important is that the agent gets sufficiently exposed to enough states to attain a reasonable barometer of their value rather than that the agent gets a perfectly optimal policy.
Third, Backplay can still recover if success goes to zero. This surprising and infrequent result occurred at the juncture where the window moved back to the initial state and even if the policy’s entropy over actions became maximal. We are unsure what differentiates these models from the ones that did not recover.
5 Conclusion
-------------
We have introduced and analyzed Backplay, a technique which improves the sample efficiency of model-free RL by constructing a curriculum around a demonstration. We showed that Backplay agents can learn in complex environments where standard model-free RL fails, and that they can outperform the ‘expert’ whose trajectories they use for selecting the starting states while training.
We also presented preliminary results on the limits of Backplay, but future research is needed to better understand the kinds of environments where Backplay succeeds or fails as well as how it can be made more efficient.
An important future direction is combining Backplay with more complex and complementary methods such as Monte Carlo Tree Search (MCTS, (Browne et al., [2012](#bib.bib6); Vodopivec et al., [2017](#bib.bib36))). There are many potential ways to do so, for example by using Backplay to warm-start MCTS.
Another future direction is to consider Backplay in multi-agent problems. Backplay can likely be used to accelerate self-play learning in zero-sum games, however special care needs to be taken to avoid policy correlation during training (Lanctot et al., [2017](#bib.bib17)) and thus to make sure that learned strategies are safe and not exploitable (Brown & Sandholm, [2017b](#bib.bib5)).
A third direction is towards non-zero sum games. It is well known that standard independent multi-agent learning does not produce agents that are able to cooperate in social dilemmas (Leibo et al., [2017](#bib.bib19); Lerer & Peysakhovich, [2017](#bib.bib20); Peysakhovich & Lerer, [2017](#bib.bib29); Foerster et al., [2017](#bib.bib10)) or risky coordination games (Yoshida et al., [2008](#bib.bib37); Peysakhovich & Lerer, [2018](#bib.bib30)). By contrast, humans are much better at finding these coordinating and cooperating equilibria (Bó, [2005](#bib.bib3); Kleiman-Weiner et al., [2016](#bib.bib16)). Thus we conjecture that human demonstrations can be combined with Backplay to construct agents that can indeed perform well in these situations. We note that the use of humans as experts, in general, was also suggested by Salimans & Chen ([2018](#bib.bib32)), work concurrent to our own.
Other future priorities are to gain a better understanding of when Backplay works well, where it fails, and how we can make the procedure more efficient. Do the gains come from value estimation like our analysis suggests, from policy iteration, or from both? What about the curricula itself - is there an ideal rate at which we advance the window and is there a better approach than a hand-tuned schedule? Should we be using one policy that learns throughout training or would it be better to have a policy per window? This has some nice consequences and, if we could ascertain a confidence estimate of a state’s value, this will likely speed up the rate at which we can advance Backplay’s schedule.
#### Acknowledgments
We especially thank Yichen Gong, Henry Zhou, and Luvneesh Mugrai for letting us use their agent. We also thank Martin Arjovsky, Adam Bouhenguel, Tim Chklovski, Andrew Dai, Doug Eck, Jesse Engel, Jakob Foerster, Ilya Kostrikov, Mohammad Norouzi, Avital Oliver, Yann Ollivier, and Will Whitney for helpful contributions. |
5ef5776f-f5ae-4d5c-b651-182304f66d1f | trentmkelly/LessWrong-43k | LessWrong | Beware Experiments Without Evaluation
Sometimes, people propose "experiments" with new norms, policies, etc. that don't have any real means of evaluating whether or not the policy actually succeeded or not.
This should be viewed with deep skepticism -- it often seems to me that such an "experiment" isn't really an experiment at all, but rather a means of sneaking a policy in by implying that it will be rolled back if it doesn't work, while also making no real provision for evaluating whether it will be successful or not.
In the worst cases, the process of running the experiment can involve taking measures that prevent the experiment from actually being implemented!
Here are some examples of the sorts of thing I mean:
* Management at a company decides that it's going to "experiment with" an open floor plan at a new office. The office layout and space chosen makes it so that even if the open floor plan proves detrimental, it will be very difficult to switch back to a standard office configuration.
* The administration of an online forum decides that it is going to "experiment with" a new set of rules in the hopes of improving the quality of discourse, but doesn't set any clear criteria or timeline for evaluating the "experiment" or what measures might actually indicate "improved discourse quality".
* A small group that gathers for weekly chats decides to "experiment with" adding a few new people to the group, but doesn't have any probationary period, method for evaluating whether someone's a good fit or removing them if they aren't, etc.
Now, I'm not saying that one should have to register a formal plan for evaluation with timelines, metrics, etc. for any new change being made or program you want to try out -- but you should have at least some idea of what it would look like for the experiment to succeed and what it would look like for it to fail, and for things that are enough of a shakeup more formal or established metrics might well be justified. |
b12d83b2-acf2-4826-89eb-bccfb342aa56 | StampyAI/alignment-research-dataset/arxiv | Arxiv | B-Pref: Benchmarking Preference-Based Reinforcement Learning
1 Introduction
---------------
Deep reinforcement learning (RL) has emerged as a powerful method to solve a variety of sequential decision-making problems, including board games Silver et al. ([2017](#bib.bib58), [2018](#bib.bib59)), video games Berner et al. ([2019](#bib.bib10)); Mnih et al. ([2015](#bib.bib44)); Vinyals et al. ([2019](#bib.bib68)), autonomous control Bellemare et al. ([2020](#bib.bib9)); Schulman et al. ([2015](#bib.bib52)), and robotic manipulation Andrychowicz et al. ([2020](#bib.bib5)); Kalashnikov et al. ([2018](#bib.bib32)); Kober & Peters ([2011](#bib.bib35)); Kober et al. ([2013](#bib.bib36)).
However, scaling RL to many applications is difficult
due to the challenges associated with defining a suitable reward function,
which often requires substantial human effort.
Specifying the reward function becomes harder as the tasks we want the agent to achieve become more complex (e.g., cooking or self-driving).
In addition, RL agents are prone to exploit reward functions by discovering ways to achieve high returns in ways the reward designer did not expect nor intend. It is important to consider this phenomenon of reward exploitation, or reward hacking, since it may lead to unintended but dangerous consequences (Hadfield-Menell et al., [2017](#bib.bib28)).
Further, there is nuance in how we might want agents to behave, such as obeying social norms that are difficult to account for and communicate effectively through an engineered reward function (Amodei et al., [2016](#bib.bib4); Shah et al., [2019](#bib.bib57); Turner et al., [2020](#bib.bib67)).
Preference-based RL (Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31); Lee et al., [2021](#bib.bib39)) provides an alternative: a (human) teacher provides preferences between the two agent behaviors, and the agent then uses this feedback to learn desired behaviors (see Figure [1](#S2.F1 "Figure 1 ‣ 2 Preliminaries ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")).
This framework enables us to optimize the agent using RL without hand-engineered rewards by learning a reward function,
which is consistent with the observed preferences (Biyik & Sadigh, [2018](#bib.bib11); Biyik et al., [2020](#bib.bib12); Sadigh et al., [2017](#bib.bib50)).
Because a teacher can interactively guide agents according to their progress,
preference-based RL has shown promising results
(e.g., solving a range of RL benchmarks (Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31)), teaching novel behaviors (Stiennon et al., [2020](#bib.bib61); Wu et al., [2021](#bib.bib71)), and mitigating the effects of reward exploitation (Lee et al., [2021](#bib.bib39))).
Despite significant progress on RL benchmarks designed for various purposes (e.g., offline RL (Fu et al., [2020](#bib.bib24); Gulcehre et al., [2020](#bib.bib26)), generalization Cobbe et al. ([2019](#bib.bib20), [2020](#bib.bib21)), meta RL (Yu et al., [2020](#bib.bib76)), and safe RL Ray et al. ([2019](#bib.bib48))),
existing benchmarks are not tailored towards preference-based RL. The lack of a standard evaluation benchmark makes it hard to quantify scientific progress.
Indeed, without consistent evaluation, it
is not easy to understand the effects of algorithmic and design decisions or compare them across papers.
In this paper, we introduce B-Pref: a benchmark for preference-based RL consisting of various locomotion and robotic manipulation tasks from DeepMind Control Suite (Tassa et al., [2018](#bib.bib65), [2020](#bib.bib66)) and Meta-world (Yu et al., [2020](#bib.bib76)).
While utilizing real human input is ideal,
this is prohibitive because it is hard to evaluate candidate algorithms quickly using real human input.
Prior works (Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31); Lee et al., [2021](#bib.bib39)) address this issue by simulating human input as giving perfect preferences with respect to an underlying ground truth reward function. However, evaluation on such ideal teachers is unrealistic because actual humans can exhibit various irrationalities (Chipman, [2016](#bib.bib18)) in decision making.
So, in our benchmark, we design simulated human teachers with a wide array of irrationalities
and propose evaluation metrics not solely for performance but also for robustness to these potential irrationalities.
To serve as a reference, we benchmark state-of-the-art preference-based RL algorithms (Christiano et al., [2017](#bib.bib19); Lee et al., [2021](#bib.bib39)) in B-Pref
and showcase the utility of B-Pref by using it to analyze algorithmic design choices for preference-based RL.
Although existing methods provide fairly efficient performance on perfectly rational teachers,
the poor performance on more realistic, irrational teachers calls for new algorithms to be developed.
The benchmark and reference implementations are available at <https://github.com/rll-research/B-Pref>.
We believe that systematic evaluation and comparison will not only further our understanding of the strengths of existing algorithms, but also reveal their limitations and suggest directions for future research.
2 Preliminaries
----------------
Figure 1: Illustration of preference-based RL.
Instead of assuming that the environment provides a (hand-engineered) reward,
a teacher provides preferences between the agent’s behaviors, and the agent uses this feedback in order to learn the desired behavior.
We consider an agent interacting with an environment in discrete time Sutton & Barto ([2018](#bib.bib63)).
At each timestep t𝑡titalic\_t, the agent receives a state 𝐬tsubscript𝐬𝑡\mathbf{s}\_{t}bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT from the environment and chooses an action 𝐚tsubscript𝐚𝑡\mathbf{a}\_{t}bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT based on its policy π𝜋\piitalic\_π.
In traditional reinforcement learning,
the environment also returns a reward r(𝐬t,𝐚t)𝑟subscript𝐬𝑡subscript𝐚𝑡r(\mathbf{s}\_{t},\mathbf{a}\_{t})italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) and
the goal of agent is to maximize the discounted sum of rewards.
However, for many complex domains and tasks,
it is difficult to construct a suitable reward function.
We consider the preference-based RL framework,
where a (human) teacher provides preferences between the agent’s behaviors
and the agent uses this feedback to perform the task (Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31); Lee et al., [2021](#bib.bib39); Leike et al., [2018](#bib.bib41)).
Formally, a segment σ𝜎\sigmaitalic\_σ is a sequence of observations and actions {(𝐬1,𝐚1),…,(𝐬H,𝐚H)}subscript𝐬1subscript𝐚1…subscript𝐬𝐻subscript𝐚𝐻\{(\mathbf{s}\_{1},\mathbf{a}\_{1}),...,(\mathbf{s}\_{H},\mathbf{a}\_{H})\}{ ( bold\_s start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , bold\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) , … , ( bold\_s start\_POSTSUBSCRIPT italic\_H end\_POSTSUBSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_H end\_POSTSUBSCRIPT ) }.
Given a pair of segments (σ0,σ1)superscript𝜎0superscript𝜎1(\sigma^{0},\sigma^{1})( italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT , italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ),
a teacher indicates which segment is preferred, i.e., y=(0,1)or(1,0)𝑦01or10y=(0,1)~{}\text{or}~{}(1,0)italic\_y = ( 0 , 1 ) or ( 1 , 0 ),
that the two segments are equally preferred y=(0.5,0.5)𝑦0.50.5y=(0.5,0.5)italic\_y = ( 0.5 , 0.5 ),
or that two segments are incomparable, i.e., discarding the query.
The goal of preference-based RL is to train an agent to perform behaviors desirable to a human teacher using as few queries as possible.
3 B-Pref: Benchmarks environments for preference-based RL
----------------------------------------------------------
###
3.1 Design factors
While ideally we would evaluate algorithms’ real-world efficacy using real human feedback,
designing a standardized and broadly available benchmark becomes challenging
because we do not have ground truth access to the human’s reward function.
Instead, we focus on solving a range of existing RL tasks (see Section [3.4](#S3.SS4 "3.4 Tasks ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")) using a simulated human,
whose preferences are based on a ground truth reward function.
Because simulated human preferences are immediately generated by the ground truth reward,
we are able to evaluate the agent quantitatively by measuring the true average return
and do more rapid experiments.
A major challenge with simulating human input is that real humans are not perfectly rational and will not provide perfect preferences.
To alleviate this challenge, we propose to simulate human input using a wide array of irrationalities (see Section [3.2](#S3.SS2 "3.2 Simulated human teachers ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")), and measure an algorithm’s robustness in handling such input (see Section [3.3](#S3.SS3 "3.3 Evaluation metrics ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")).
Algorithm 1 SimTeacher: Simulated human teachers
1:Discount factor γ𝛾\gammaitalic\_γ, rationality constant β𝛽\betaitalic\_β, probability of making a mistake ϵitalic-ϵ\epsilonitalic\_ϵ
2:Skip threshold δ𝚜𝚔𝚒𝚙subscript𝛿𝚜𝚔𝚒𝚙\delta\_{\tt skip}italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT, equal threshold δ𝚎𝚚𝚞𝚊𝚕subscript𝛿𝚎𝚚𝚞𝚊𝚕\delta\_{\tt equal}italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT
3:Pair of segments σ0,σ1superscript𝜎0superscript𝜎1\sigma^{0},\sigma^{1}italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT , italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT
4:if maxi∈{0,1}∑tr(𝐬ti,𝐚ti)<δ𝚜𝚔𝚒𝚙subscript𝑖01subscript𝑡𝑟subscriptsuperscript𝐬𝑖𝑡subscriptsuperscript𝐚𝑖𝑡subscript𝛿𝚜𝚔𝚒𝚙\max\_{i\in\{0,1\}}\sum\_{t}r\left(\mathbf{s}^{i}\_{t},\mathbf{a}^{i}\_{t}\right)<\delta\_{\tt skip}roman\_max start\_POSTSUBSCRIPT italic\_i ∈ { 0 , 1 } end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT italic\_r ( bold\_s start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) < italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT then
// Skipping query
5: y←∅←𝑦y\leftarrow\emptysetitalic\_y ← ∅
6:else if |∑tr(𝐬t1,𝐚t1)−∑tr(𝐬t0,𝐚t0)|<δ𝚎𝚚𝚞𝚊𝚕subscript𝑡𝑟subscriptsuperscript𝐬1𝑡subscriptsuperscript𝐚1𝑡subscript𝑡𝑟subscriptsuperscript𝐬0𝑡subscriptsuperscript𝐚0𝑡subscript𝛿𝚎𝚚𝚞𝚊𝚕\big{|}\sum\_{t}r\left(\mathbf{s}^{1}\_{t},\mathbf{a}^{1}\_{t}\right)-\sum\_{t}r\left(\mathbf{s}^{0}\_{t},\mathbf{a}^{0}\_{t}\right)\big{|}<\delta\_{\tt equal}| ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT italic\_r ( bold\_s start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) - ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT italic\_r ( bold\_s start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) | < italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT then
// Equally preferable
7: y←(0.5,0.5)←𝑦0.50.5y\leftarrow(0.5,0.5)italic\_y ← ( 0.5 , 0.5 )
8:else if σ0≻σ1∼P[σ0≻σ1;β,γ]succeedssuperscript𝜎0superscript𝜎1similar-to𝑃delimited-[]succeedssuperscript𝜎0superscript𝜎1𝛽𝛾\sigma^{0}\succ\sigma^{1}\sim P[\sigma^{0}\succ\sigma^{1};\beta,\gamma]italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ∼ italic\_P [ italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ; italic\_β , italic\_γ ] then
// Sampling preferences from ([1](#S3.E1 "1 ‣ 3.2 Simulated human teachers ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning"))
9: y←(1,0)←𝑦10y\leftarrow(1,0)italic\_y ← ( 1 , 0 ) with probability of 1−ϵ1italic-ϵ1-\epsilon1 - italic\_ϵ
10: y←(0,1)←𝑦01y\leftarrow(0,1)italic\_y ← ( 0 , 1 ) otherwise // Making a mistake
11:else
12: y←(0,1)←𝑦01y\leftarrow(0,1)italic\_y ← ( 0 , 1 ) with probability of 1−ϵ1italic-ϵ1-\epsilon1 - italic\_ϵ
13: y←(1,0)←𝑦10y\leftarrow(1,0)italic\_y ← ( 1 , 0 ) otherwise // Making a mistake
14:end if
15:𝐫𝐞𝐭𝐮𝐫𝐧y𝐫𝐞𝐭𝐮𝐫𝐧𝑦\textbf{return}\;yreturn italic\_y
###
3.2 Simulated human teachers
We start from a (perfectly rational) deterministic teacher, which generates preferences as follows:
| | | |
| --- | --- | --- |
| | y={(1,0)If ∑t=1Hr(𝐬t0,𝐚t0)>∑t=1Hr(𝐬t1,𝐚t1)(0,1)otherwise,𝑦cases10If ∑t=1Hr(𝐬t0,𝐚t0)>∑t=1Hr(𝐬t1,𝐚t1)missing-subexpression01otherwisemissing-subexpression\displaystyle y=\left\{\begin{array}[]{lll}(1,0)&\mbox{If $\sum\_{t=1}^{H}r(\mathbf{s}\_{t}^{0},\mathbf{a}\_{t}^{0})>\sum\_{t=1}^{H}r(\mathbf{s}\_{t}^{1},\mathbf{a}\_{t}^{1})$}\\
(0,1)&\mbox{otherwise},\end{array}\right.italic\_y = { start\_ARRAY start\_ROW start\_CELL ( 1 , 0 ) end\_CELL start\_CELL If ∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_H end\_POSTSUPERSCRIPT italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT ) > ∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_H end\_POSTSUPERSCRIPT italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ) end\_CELL start\_CELL end\_CELL end\_ROW start\_ROW start\_CELL ( 0 , 1 ) end\_CELL start\_CELL otherwise , end\_CELL start\_CELL end\_CELL end\_ROW end\_ARRAY | |
where H>0𝐻0H>0italic\_H > 0 is a length of segment σ𝜎\sigmaitalic\_σ and r𝑟ritalic\_r is the ground truth reward.
We remark that prior works (Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31); Lee et al., [2021](#bib.bib39)) evaluated their methods using this ideal teacher.
However, evaluating the performance of preference-based RL only using the ideal teacher is unrealistic because there are many possible irrationalities (Chan et al., [2021](#bib.bib17); Chipman, [2016](#bib.bib18)) affecting a teacher’s preferences (and expression of preferences) in different ways.
To design more realistic models of human teachers,
we consider a common stochastic model (Biyik & Sadigh, [2018](#bib.bib11); Christiano et al., [2017](#bib.bib19); Sadigh et al., [2017](#bib.bib50))
and systematically manipulate its terms and operators
(see Algorithm [1](#alg1 "Algorithm 1 ‣ 3.1 Design factors ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")):
Stochastic preference model.
Because preferences from the human can be noisy,
we generate preferences using a stochastic model defined as follows (Line [8](#alg1.l8 "8 ‣ Algorithm 1 ‣ 3.1 Design factors ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")):
| | | | |
| --- | --- | --- | --- |
| | P[σi≻σj;β,γ]=exp(β∑t=1HγH−tr(𝐬ti,𝐚ti))exp(β∑t=1HγH−tr(𝐬ti,𝐚ti))+exp(β∑t=1HγH−tr(𝐬tj,𝐚tj)),𝑃delimited-[]succeedssuperscript𝜎𝑖superscript𝜎𝑗𝛽𝛾𝛽superscriptsubscript𝑡1𝐻superscript𝛾𝐻𝑡𝑟superscriptsubscript𝐬𝑡𝑖superscriptsubscript𝐚𝑡𝑖𝛽superscriptsubscript𝑡1𝐻superscript𝛾𝐻𝑡𝑟superscriptsubscript𝐬𝑡𝑖superscriptsubscript𝐚𝑡𝑖𝛽superscriptsubscript𝑡1𝐻superscript𝛾𝐻𝑡𝑟superscriptsubscript𝐬𝑡𝑗superscriptsubscript𝐚𝑡𝑗\displaystyle P[\sigma^{i}\succ\sigma^{j};\beta,\gamma]=\frac{\exp\left(\beta\sum\_{t=1}^{H}\gamma^{H-t}r(\mathbf{s}\_{t}^{i},\mathbf{a}\_{t}^{i})\right)}{\exp\left(\beta\sum\_{t=1}^{H}\gamma^{H-t}r(\mathbf{s}\_{t}^{i},\mathbf{a}\_{t}^{i})\right)+\exp\left(\beta\sum\_{t=1}^{H}\gamma^{H-t}r(\mathbf{s}\_{t}^{j},\mathbf{a}\_{t}^{j})\right)},italic\_P [ italic\_σ start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT italic\_j end\_POSTSUPERSCRIPT ; italic\_β , italic\_γ ] = divide start\_ARG roman\_exp ( italic\_β ∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_H end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_H - italic\_t end\_POSTSUPERSCRIPT italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ) ) end\_ARG start\_ARG roman\_exp ( italic\_β ∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_H end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_H - italic\_t end\_POSTSUPERSCRIPT italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ) ) + roman\_exp ( italic\_β ∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_H end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_H - italic\_t end\_POSTSUPERSCRIPT italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_j end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_j end\_POSTSUPERSCRIPT ) ) end\_ARG , | | (1) |
where γ∈(0,1]𝛾01\gamma\in(0,1]italic\_γ ∈ ( 0 , 1 ] is a discount factor to model myopic behavior,
β𝛽\betaitalic\_β is a rationality constant, and σi≻σjsucceedssuperscript𝜎𝑖superscript𝜎𝑗\sigma^{i}\succ\sigma^{j}italic\_σ start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT italic\_j end\_POSTSUPERSCRIPT denotes the event that segment i𝑖iitalic\_i is preferable to segment j𝑗jitalic\_j.
This follows the Bradley-Terry model (Bradley & Terry, [1952](#bib.bib13)),
which can be interpreted as assuming the probability of preferring a segment depends exponentially on the sum over the segment of an underlying reward.
Note that this teacher becomes a perfectly rational and deterministic as β→∞→𝛽\beta\rightarrow\inftyitalic\_β → ∞, whereas β=0𝛽0\beta=0italic\_β = 0 produces uniformly random choices.
Myopic behavior. Humans are sometimes myopic (short-sighted), so a human teacher may remember and focus more on the behavior at the end of the clip they watched, for example.
We model myopic behavior by introducing a weighted sum of rewards with a
discount factor γ𝛾\gammaitalic\_γ in ([1](#S3.E1 "1 ‣ 3.2 Simulated human teachers ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")), i.e., ∑t=1HγH−tr(𝐬ti,𝐚ti)superscriptsubscript𝑡1𝐻superscript𝛾𝐻𝑡𝑟superscriptsubscript𝐬𝑡𝑖superscriptsubscript𝐚𝑡𝑖\sum\_{t=1}^{H}\gamma^{H-t}r(\mathbf{s}\_{t}^{i},\mathbf{a}\_{t}^{i})∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_H end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_H - italic\_t end\_POSTSUPERSCRIPT italic\_r ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ), so that our simulated teacher places more weight on recent timesteps.
Skipping queries.
If both segments do not contain a desired behavior, a teacher would like to mark them as incomparable and discard the query.
We model this behavior by skipping a query if the sum over the segment of an underlying reward is smaller than skip threshold δ𝚜𝚔𝚒𝚙subscript𝛿𝚜𝚔𝚒𝚙\delta\_{\tt skip}italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT, i.e., maxi∈{0,1}∑tr(𝐬ti,𝐚ti)<δ𝚜𝚔𝚒𝚙subscript𝑖01subscript𝑡𝑟subscriptsuperscript𝐬𝑖𝑡subscriptsuperscript𝐚𝑖𝑡subscript𝛿𝚜𝚔𝚒𝚙\max\_{i\in\{0,1\}}\sum\_{t}r\left(\mathbf{s}^{i}\_{t},\mathbf{a}^{i}\_{t}\right)<\delta\_{\tt skip}roman\_max start\_POSTSUBSCRIPT italic\_i ∈ { 0 , 1 } end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT italic\_r ( bold\_s start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) < italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT (Line [4](#alg1.l4 "4 ‣ Algorithm 1 ‣ 3.1 Design factors ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")).
Equally preferable.
If the two segments are equally good,
instead of selecting one segment as preferable,
a teacher would like to mark the segments as equally preferable.
Motivated by this,
we provide an uniform distribution (0.5,0.5)0.50.5(0.5,0.5)( 0.5 , 0.5 ) as a response (Line [6](#alg1.l6 "6 ‣ Algorithm 1 ‣ 3.1 Design factors ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning"))
if both segments have similar sum of rewards, i.e., |∑tr(𝐬t1,𝐚t1)−∑tr(𝐬t0,𝐚t0)|<δ𝚎𝚚𝚞𝚊𝚕subscript𝑡𝑟subscriptsuperscript𝐬1𝑡subscriptsuperscript𝐚1𝑡subscript𝑡𝑟subscriptsuperscript𝐬0𝑡subscriptsuperscript𝐚0𝑡subscript𝛿𝚎𝚚𝚞𝚊𝚕\big{|}\sum\_{t}r\left(\mathbf{s}^{1}\_{t},\mathbf{a}^{1}\_{t}\right)-\sum\_{t}r\left(\mathbf{s}^{0}\_{t},\mathbf{a}^{0}\_{t}\right)\big{|}<\delta\_{\tt equal}| ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT italic\_r ( bold\_s start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) - ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT italic\_r ( bold\_s start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , bold\_a start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) | < italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT.
Making a mistake. Humans can make accidental errors when they respond. To reflect this, we flip the preference with probability of ϵitalic-ϵ\epsilonitalic\_ϵ
(Line [10](#alg1.l10 "10 ‣ Algorithm 1 ‣ 3.1 Design factors ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") and Line [13](#alg1.l13 "13 ‣ Algorithm 1 ‣ 3.1 Design factors ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")).
###
3.3 Evaluation metrics
We evaluate two key properties of preference-based RL: performance of the RL agent under a fixed budget of feedback and robustness to potential irrationalities.
Because the simulated human teacher’s preferences are generated by a ground truth reward, we measure the true average return of trained agents as evaluation metric.
To facilitate comparison across different RL algorithms,
we normalize returns with respect to the baseline of RL training using the ground truth reward:
| | | |
| --- | --- | --- |
| | Normalized returns=Average returns of preference-based RLAverage returns of RL with ground truth reward.Normalized returnsAverage returns of preference-based RLAverage returns of RL with ground truth reward\displaystyle\texttt{Normalized returns}=\frac{\texttt{Average returns of preference-based RL}}{\texttt{Average returns of RL with ground truth reward}}.Normalized returns = divide start\_ARG Average returns of preference-based RL end\_ARG start\_ARG Average returns of RL with ground truth reward end\_ARG . | |
To evaluate the feedback-efficiency of preference-based RL algorithms,
we compare normalized returns by varying the maximum budget of queries.
To evaluate the robustness, we evaluate against the following simulated human teachers with different properties:
* [leftmargin=8mm]
* •
Oracle: SimTeacher(→β∞,γ=1,ϵ=0,δ𝚜𝚔𝚒𝚙=0,δ𝚎𝚚𝚞𝚊𝚕=0)formulae-sequence
→β∞𝛾1formulae-sequenceitalic-ϵ0formulae-sequencesubscript𝛿𝚜𝚔𝚒𝚙0subscript𝛿𝚎𝚚𝚞𝚊𝚕0\left(\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to31.21pt{\vbox to16.09pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{11.58888pt}\pgfsys@curveto{0.0pt}{14.07419pt}{2.0147pt}{16.08888pt}{4.5pt}{16.08888pt}\pgfsys@lineto{26.7117pt}{16.08888pt}\pgfsys@curveto{29.197pt}{16.08888pt}{31.2117pt}{14.07419pt}{31.2117pt}{11.58888pt}\pgfsys@lineto{31.2117pt}{4.5pt}\pgfsys@curveto{31.2117pt}{2.0147pt}{29.197pt}{0.0pt}{26.7117pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{11.58888pt}\pgfsys@curveto{0.5pt}{13.79805pt}{2.29083pt}{15.58888pt}{4.5pt}{15.58888pt}\pgfsys@lineto{26.7117pt}{15.58888pt}\pgfsys@curveto{28.92087pt}{15.58888pt}{30.7117pt}{13.79805pt}{30.7117pt}{11.58888pt}\pgfsys@lineto{30.7117pt}{4.5pt}\pgfsys@curveto{30.7117pt}{2.29083pt}{28.92087pt}{0.5pt}{26.7117pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\beta\rightarrow\infty$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\gamma=1,\epsilon=0,\delta\_{\tt skip}=0,\delta\_{\tt equal}=0\right)( italic\_β → ∞ , italic\_γ = 1 , italic\_ϵ = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT = 0 )
* •
Stoc: SimTeacher(=β1,γ=1,ϵ=0,δ𝚜𝚔𝚒𝚙=0,δ𝚎𝚚𝚞𝚊𝚕=0)formulae-sequence
=β1𝛾1formulae-sequenceitalic-ϵ0formulae-sequencesubscript𝛿𝚜𝚔𝚒𝚙0subscript𝛿𝚎𝚚𝚞𝚊𝚕0\left(\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to28.99pt{\vbox to15.89pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{11.38889pt}\pgfsys@curveto{0.0pt}{13.87419pt}{2.0147pt}{15.88889pt}{4.5pt}{15.88889pt}\pgfsys@lineto{24.48949pt}{15.88889pt}\pgfsys@curveto{26.9748pt}{15.88889pt}{28.98949pt}{13.87419pt}{28.98949pt}{11.38889pt}\pgfsys@lineto{28.98949pt}{4.5pt}\pgfsys@curveto{28.98949pt}{2.0147pt}{26.9748pt}{0.0pt}{24.48949pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{11.38889pt}\pgfsys@curveto{0.5pt}{13.59805pt}{2.29083pt}{15.38889pt}{4.5pt}{15.38889pt}\pgfsys@lineto{24.48949pt}{15.38889pt}\pgfsys@curveto{26.69865pt}{15.38889pt}{28.48949pt}{13.59805pt}{28.48949pt}{11.38889pt}\pgfsys@lineto{28.48949pt}{4.5pt}\pgfsys@curveto{28.48949pt}{2.29083pt}{26.69865pt}{0.5pt}{24.48949pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\beta=1$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\gamma=1,\epsilon=0,\delta\_{\tt skip}=0,\delta\_{\tt equal}=0\right)( italic\_β = 1 , italic\_γ = 1 , italic\_ϵ = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT = 0 )
* •
Mistake: SimTeacher(→β∞,γ=1,=ϵ0.1,δ𝚜𝚔𝚒𝚙=0,δ𝚎𝚚𝚞𝚊𝚕=0)formulae-sequence
→β∞𝛾1
=ϵ0.1subscript𝛿𝚜𝚔𝚒𝚙0subscript𝛿𝚎𝚚𝚞𝚊𝚕0\left(\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to31.21pt{\vbox to16.09pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{11.58888pt}\pgfsys@curveto{0.0pt}{14.07419pt}{2.0147pt}{16.08888pt}{4.5pt}{16.08888pt}\pgfsys@lineto{26.7117pt}{16.08888pt}\pgfsys@curveto{29.197pt}{16.08888pt}{31.2117pt}{14.07419pt}{31.2117pt}{11.58888pt}\pgfsys@lineto{31.2117pt}{4.5pt}\pgfsys@curveto{31.2117pt}{2.0147pt}{29.197pt}{0.0pt}{26.7117pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{11.58888pt}\pgfsys@curveto{0.5pt}{13.79805pt}{2.29083pt}{15.58888pt}{4.5pt}{15.58888pt}\pgfsys@lineto{26.7117pt}{15.58888pt}\pgfsys@curveto{28.92087pt}{15.58888pt}{30.7117pt}{13.79805pt}{30.7117pt}{11.58888pt}\pgfsys@lineto{30.7117pt}{4.5pt}\pgfsys@curveto{30.7117pt}{2.29083pt}{28.92087pt}{0.5pt}{26.7117pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\beta\rightarrow\infty$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\gamma=1,\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to36.84pt{\vbox to13.44pt{\pgfpicture\makeatletter\raise-3.5pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{8.94444pt}\pgfsys@curveto{0.0pt}{11.42975pt}{2.0147pt}{13.44444pt}{4.5pt}{13.44444pt}\pgfsys@lineto{32.3367pt}{13.44444pt}\pgfsys@curveto{34.822pt}{13.44444pt}{36.8367pt}{11.42975pt}{36.8367pt}{8.94444pt}\pgfsys@lineto{36.8367pt}{4.5pt}\pgfsys@curveto{36.8367pt}{2.0147pt}{34.822pt}{0.0pt}{32.3367pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{8.94444pt}\pgfsys@curveto{0.5pt}{11.15361pt}{2.29083pt}{12.94444pt}{4.5pt}{12.94444pt}\pgfsys@lineto{32.3367pt}{12.94444pt}\pgfsys@curveto{34.54587pt}{12.94444pt}{36.3367pt}{11.15361pt}{36.3367pt}{8.94444pt}\pgfsys@lineto{36.3367pt}{4.5pt}\pgfsys@curveto{36.3367pt}{2.29083pt}{34.54587pt}{0.5pt}{32.3367pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{3.5pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\epsilon=0.1$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\delta\_{\tt skip}=0,\delta\_{\tt equal}=0\right)( italic\_β → ∞ , italic\_γ = 1 , italic\_ϵ = 0.1 , italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT = 0 )
* •
Skip: SimTeacher(→β∞,γ=1,ϵ=0,>δskip0,δ𝚎𝚚𝚞𝚊𝚕=0)formulae-sequence
→β∞𝛾1formulae-sequenceitalic-ϵ
0>δskip0subscript𝛿𝚎𝚚𝚞𝚊𝚕0\left(\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to31.21pt{\vbox to16.09pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{11.58888pt}\pgfsys@curveto{0.0pt}{14.07419pt}{2.0147pt}{16.08888pt}{4.5pt}{16.08888pt}\pgfsys@lineto{26.7117pt}{16.08888pt}\pgfsys@curveto{29.197pt}{16.08888pt}{31.2117pt}{14.07419pt}{31.2117pt}{11.58888pt}\pgfsys@lineto{31.2117pt}{4.5pt}\pgfsys@curveto{31.2117pt}{2.0147pt}{29.197pt}{0.0pt}{26.7117pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{11.58888pt}\pgfsys@curveto{0.5pt}{13.79805pt}{2.29083pt}{15.58888pt}{4.5pt}{15.58888pt}\pgfsys@lineto{26.7117pt}{15.58888pt}\pgfsys@curveto{28.92087pt}{15.58888pt}{30.7117pt}{13.79805pt}{30.7117pt}{11.58888pt}\pgfsys@lineto{30.7117pt}{4.5pt}\pgfsys@curveto{30.7117pt}{2.29083pt}{28.92087pt}{0.5pt}{26.7117pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\beta\rightarrow\infty$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\gamma=1,\epsilon=0,\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to34.83pt{\vbox to16.98pt{\pgfpicture\makeatletter\raise-6.53333pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{12.47777pt}\pgfsys@curveto{0.0pt}{14.96307pt}{2.0147pt}{16.97777pt}{4.5pt}{16.97777pt}\pgfsys@lineto{30.3311pt}{16.97777pt}\pgfsys@curveto{32.8164pt}{16.97777pt}{34.8311pt}{14.96307pt}{34.8311pt}{12.47777pt}\pgfsys@lineto{34.8311pt}{4.5pt}\pgfsys@curveto{34.8311pt}{2.0147pt}{32.8164pt}{0.0pt}{30.3311pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{12.47777pt}\pgfsys@curveto{0.5pt}{14.68694pt}{2.29083pt}{16.47777pt}{4.5pt}{16.47777pt}\pgfsys@lineto{30.3311pt}{16.47777pt}\pgfsys@curveto{32.54027pt}{16.47777pt}{34.3311pt}{14.68694pt}{34.3311pt}{12.47777pt}\pgfsys@lineto{34.3311pt}{4.5pt}\pgfsys@curveto{34.3311pt}{2.29083pt}{32.54027pt}{0.5pt}{30.3311pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{6.53333pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\delta\_{\tt skip}>0$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\delta\_{\tt equal}=0\right)( italic\_β → ∞ , italic\_γ = 1 , italic\_ϵ = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT > 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT = 0 )
* •
Equal: SimTeacher(→β∞,γ=1,ϵ=0,δ𝚜𝚔𝚒𝚙=0,>δequal0)formulae-sequence
→β∞𝛾1formulae-sequenceitalic-ϵ0subscript𝛿𝚜𝚔𝚒𝚙
0>δequal0\left(\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to31.21pt{\vbox to16.09pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{11.58888pt}\pgfsys@curveto{0.0pt}{14.07419pt}{2.0147pt}{16.08888pt}{4.5pt}{16.08888pt}\pgfsys@lineto{26.7117pt}{16.08888pt}\pgfsys@curveto{29.197pt}{16.08888pt}{31.2117pt}{14.07419pt}{31.2117pt}{11.58888pt}\pgfsys@lineto{31.2117pt}{4.5pt}\pgfsys@curveto{31.2117pt}{2.0147pt}{29.197pt}{0.0pt}{26.7117pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{11.58888pt}\pgfsys@curveto{0.5pt}{13.79805pt}{2.29083pt}{15.58888pt}{4.5pt}{15.58888pt}\pgfsys@lineto{26.7117pt}{15.58888pt}\pgfsys@curveto{28.92087pt}{15.58888pt}{30.7117pt}{13.79805pt}{30.7117pt}{11.58888pt}\pgfsys@lineto{30.7117pt}{4.5pt}\pgfsys@curveto{30.7117pt}{2.29083pt}{28.92087pt}{0.5pt}{26.7117pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\beta\rightarrow\infty$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\gamma=1,\epsilon=0,\delta\_{\tt skip}=0,\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to37.91pt{\vbox to16.98pt{\pgfpicture\makeatletter\raise-6.53333pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{12.47777pt}\pgfsys@curveto{0.0pt}{14.96307pt}{2.0147pt}{16.97777pt}{4.5pt}{16.97777pt}\pgfsys@lineto{33.41109pt}{16.97777pt}\pgfsys@curveto{35.8964pt}{16.97777pt}{37.91109pt}{14.96307pt}{37.91109pt}{12.47777pt}\pgfsys@lineto{37.91109pt}{4.5pt}\pgfsys@curveto{37.91109pt}{2.0147pt}{35.8964pt}{0.0pt}{33.41109pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{12.47777pt}\pgfsys@curveto{0.5pt}{14.68694pt}{2.29083pt}{16.47777pt}{4.5pt}{16.47777pt}\pgfsys@lineto{33.41109pt}{16.47777pt}\pgfsys@curveto{35.62025pt}{16.47777pt}{37.41109pt}{14.68694pt}{37.41109pt}{12.47777pt}\pgfsys@lineto{37.41109pt}{4.5pt}\pgfsys@curveto{37.41109pt}{2.29083pt}{35.62025pt}{0.5pt}{33.41109pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{6.53333pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\delta\_{\tt equal}>0$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}}\right)( italic\_β → ∞ , italic\_γ = 1 , italic\_ϵ = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT > 0 )
* •
Myopic: SimTeacher(→β∞,=γ0.9,ϵ=0,δ𝚜𝚔𝚒𝚙=0,δ𝚎𝚚𝚞𝚊𝚕=0)formulae-sequence
→β∞=γ0.9italic-ϵ0formulae-sequencesubscript𝛿𝚜𝚔𝚒𝚙0subscript𝛿𝚎𝚚𝚞𝚊𝚕0\left(\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to31.21pt{\vbox to16.09pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{11.58888pt}\pgfsys@curveto{0.0pt}{14.07419pt}{2.0147pt}{16.08888pt}{4.5pt}{16.08888pt}\pgfsys@lineto{26.7117pt}{16.08888pt}\pgfsys@curveto{29.197pt}{16.08888pt}{31.2117pt}{14.07419pt}{31.2117pt}{11.58888pt}\pgfsys@lineto{31.2117pt}{4.5pt}\pgfsys@curveto{31.2117pt}{2.0147pt}{29.197pt}{0.0pt}{26.7117pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{11.58888pt}\pgfsys@curveto{0.5pt}{13.79805pt}{2.29083pt}{15.58888pt}{4.5pt}{15.58888pt}\pgfsys@lineto{26.7117pt}{15.58888pt}\pgfsys@curveto{28.92087pt}{15.58888pt}{30.7117pt}{13.79805pt}{30.7117pt}{11.58888pt}\pgfsys@lineto{30.7117pt}{4.5pt}\pgfsys@curveto{30.7117pt}{2.29083pt}{28.92087pt}{0.5pt}{26.7117pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\beta\rightarrow\infty$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\definecolor{tcbcolframe}{rgb}{1,1,1}\definecolor{tcbcolback}{rgb}{1,0.9,0.9}\definecolor{tcbcol@origin}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\definecolor{.}{rgb}{0,0,0}\leavevmode\hbox to37.95pt{\vbox to15.39pt{\pgfpicture\makeatletter\raise-5.44444pt\hbox{\hskip 0.0pt\lower 0.0pt\hbox to 0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\hbox to 0.0pt{{}{}{}{}\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,1,1}\pgfsys@color@gray@fill{1}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.0pt}{4.5pt}\pgfsys@lineto{0.0pt}{10.88889pt}\pgfsys@curveto{0.0pt}{13.37419pt}{2.0147pt}{15.38889pt}{4.5pt}{15.38889pt}\pgfsys@lineto{33.45497pt}{15.38889pt}\pgfsys@curveto{35.94028pt}{15.38889pt}{37.95497pt}{13.37419pt}{37.95497pt}{10.88889pt}\pgfsys@lineto{37.95497pt}{4.5pt}\pgfsys@curveto{37.95497pt}{2.0147pt}{35.94028pt}{0.0pt}{33.45497pt}{0.0pt}\pgfsys@lineto{4.5pt}{0.0pt}\pgfsys@curveto{2.0147pt}{0.0pt}{0.0pt}{2.0147pt}{0.0pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }{}{}{}{}{}{}{}{}\definecolor[named]{pgffillcolor}{rgb}{1,0.9,0.9}\pgfsys@color@rgb@fill{1}{0.9}{0.9}\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}{{}{}{{}}}{{}{}{{}}}{}{}\pgfsys@moveto{0.5pt}{4.5pt}\pgfsys@lineto{0.5pt}{10.88889pt}\pgfsys@curveto{0.5pt}{13.09805pt}{2.29083pt}{14.88889pt}{4.5pt}{14.88889pt}\pgfsys@lineto{33.45497pt}{14.88889pt}\pgfsys@curveto{35.66414pt}{14.88889pt}{37.45497pt}{13.09805pt}{37.45497pt}{10.88889pt}\pgfsys@lineto{37.45497pt}{4.5pt}\pgfsys@curveto{37.45497pt}{2.29083pt}{35.66414pt}{0.5pt}{33.45497pt}{0.5pt}\pgfsys@lineto{4.5pt}{0.5pt}\pgfsys@curveto{2.29083pt}{0.5pt}{0.5pt}{2.29083pt}{0.5pt}{4.5pt}\pgfsys@closepath\pgfsys@fill\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@fill@opacity{1.000000}\pgfsys@invoke{ }{{{}}{{}}{{}}{{}}{{}}{{}}{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{2.5pt}{5.44444pt}\pgfsys@invoke{ }\hbox{{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\hbox{\set@color{\ignorespaces$\gamma=0.9$}}}}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}},\epsilon=0,\delta\_{\tt skip}=0,\delta\_{\tt equal}=0\right)( italic\_β → ∞ , italic\_γ = 0.9 , italic\_ϵ = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_skip end\_POSTSUBSCRIPT = 0 , italic\_δ start\_POSTSUBSCRIPT typewriter\_equal end\_POSTSUBSCRIPT = 0 )
In our evaluations, we consider one modification (i.e., irrationality) to the oracle teacher at a time, which allows us to isolate the individual effects. While each individually may not exactly model real human behavior, it would be straightforward to use our benchmark to create more complex teachers that combine multiple irrationalities.
###
3.4 Tasks
We consider two locomotion tasks (Walker-walk and Quadruped-walk) from DeepMind Control Suite (DMControl) (Tassa et al., [2018](#bib.bib65), [2020](#bib.bib66))
and two robotic manipulation tasks (Button Press and Sweep Into) from Meta-world (Yu et al., [2020](#bib.bib76)).
We focus on learning from the proprioceptive inputs and dense rewards
because learning from visual observations and sparse rewards can cause additional issues, such as representation learning (Laskin et al., [2020](#bib.bib37); Schwarzer et al., [2021](#bib.bib55); Srinivas et al., [2020](#bib.bib60); Stooke et al., [2021](#bib.bib62); Yarats et al., [2021](#bib.bib74)) and exploration (Seo et al., [2021](#bib.bib56)).
However, we think it is an interesting and important direction for future work to consider visual observations and sparse rewards.
4 B-Pref: Algorithmic baselines for preference-based RL
--------------------------------------------------------
Throughout this paper,
we mainly focus on two of the most prominent preference-based RL algorithms (Christiano et al., [2017](#bib.bib19); Lee et al., [2021](#bib.bib39)), which involve reward learning from preferences.
Formally, a policy πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT and reward function r^ψsubscript^𝑟𝜓\widehat{r}\_{\psi}over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT are updated as follows (see Algorithm [3](#alg3 "Algorithm 3 ‣ Appendix B Preference-based reinforcement learning ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") in the supplementary material):
* [leftmargin=8mm]
* •
Step 1 (agent learning):
The policy πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT interacts with environment to collect experiences and we update it
using existing RL algorithms to maximize the sum of the learned rewards r^ψsubscript^𝑟𝜓\widehat{r}\_{\psi}over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT.
* •
Step 2 (reward learning):
We optimize the reward function r^ψsubscript^𝑟𝜓\widehat{r}\_{\psi}over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT via supervised learning based on the feedback received from a teacher.
* •
Repeat Step 1 and Step 2.
###
4.1 Deep reinforcement learning from human preferences
In order to incorporate human preferences into deep RL,
Christiano et al. ([2017](#bib.bib19)) proposed a framework that learns a reward function r^ψsubscript^𝑟𝜓\widehat{r}\_{\psi}over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT from preferences Sadigh et al. ([2017](#bib.bib50)); Wilson et al. ([2012](#bib.bib70)).
Specifically,
we first model a preference predictor using the reward function r^ψsubscript^𝑟𝜓\widehat{r}\_{\psi}over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT as follows:
| | | | |
| --- | --- | --- | --- |
| | Pψ[σ1≻σ0]=exp∑tr^ψ(𝐬t1,𝐚t1)∑i∈{0,1}exp∑tr^ψ(𝐬ti,𝐚ti),subscript𝑃𝜓delimited-[]succeedssuperscript𝜎1superscript𝜎0subscript𝑡subscript^𝑟𝜓superscriptsubscript𝐬𝑡1superscriptsubscript𝐚𝑡1subscript𝑖01subscript𝑡subscript^𝑟𝜓superscriptsubscript𝐬𝑡𝑖superscriptsubscript𝐚𝑡𝑖\displaystyle P\_{\psi}[\sigma^{1}\succ\sigma^{0}]=\frac{\exp\sum\_{t}\widehat{r}\_{\psi}(\mathbf{s}\_{t}^{1},\mathbf{a}\_{t}^{1})}{\sum\_{i\in\{0,1\}}\exp\sum\_{t}\widehat{r}\_{\psi}(\mathbf{s}\_{t}^{i},\mathbf{a}\_{t}^{i})},italic\_P start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT [ italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT ] = divide start\_ARG roman\_exp ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ) end\_ARG start\_ARG ∑ start\_POSTSUBSCRIPT italic\_i ∈ { 0 , 1 } end\_POSTSUBSCRIPT roman\_exp ∑ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT ( bold\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT , bold\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ) end\_ARG , | | (2) |
where σi≻σjsucceedssuperscript𝜎𝑖superscript𝜎𝑗\sigma^{i}\succ\sigma^{j}italic\_σ start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT italic\_j end\_POSTSUPERSCRIPT denotes the event that segment i𝑖iitalic\_i is preferable to segment j𝑗jitalic\_j.
We remark that this corresponds to assume a stochastic teacher following the Bradley-Terry model (Bradley & Terry, [1952](#bib.bib13)) but we do not assume that the type and degree of irrationality or systematic bias is available in our experiments. Because this could lead to a poor preference inference (Chan et al., [2021](#bib.bib17)),
future work may be able to further improve the efficiency of learning by approximating the teacher’s irrationality.
To align our preference predictor with the teacher’s preferences,
we consider a binary classification problem using the cross-entropy loss.
Specifically, given a dataset of preferences 𝒟𝒟\mathcal{D}caligraphic\_D,
the reward function, modeled as a neural network with parameters ψ𝜓\psiitalic\_ψ, is updated by minimizing the following loss:
| | | | | |
| --- | --- | --- | --- | --- |
| | ℒ𝚁𝚎𝚠𝚊𝚛𝚍=−𝔼(σ0,σ1,y)∼𝒟[\displaystyle\mathcal{L}^{\tt Reward}=-\operatorname\*{\mathbb{E}}\_{(\sigma^{0},\sigma^{1},y)\sim\mathcal{D}}\Big{[}caligraphic\_L start\_POSTSUPERSCRIPT typewriter\_Reward end\_POSTSUPERSCRIPT = - blackboard\_E start\_POSTSUBSCRIPT ( italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT , italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT , italic\_y ) ∼ caligraphic\_D end\_POSTSUBSCRIPT [ | y(0)logPψ[σ0≻σ1]+y(1)logPψ[σ1≻σ0]].\displaystyle y(0)\log P\_{\psi}[\sigma^{0}\succ\sigma^{1}]+y(1)\log P\_{\psi}[\sigma^{1}\succ\sigma^{0}]\Big{]}.italic\_y ( 0 ) roman\_log italic\_P start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT [ italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ] + italic\_y ( 1 ) roman\_log italic\_P start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT [ italic\_σ start\_POSTSUPERSCRIPT 1 end\_POSTSUPERSCRIPT ≻ italic\_σ start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT ] ] . | | (3) |
Once we learn a reward function r^ψsubscript^𝑟𝜓{\widehat{r}}\_{\psi}over^ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_ψ end\_POSTSUBSCRIPT,
we can update the policy πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT using any RL algorithm.
A caveat is that the reward function may be non-stationary because we update it during training.
To mitigate the effects of a non-stationary reward function,
Christiano et al. ([2017](#bib.bib19)) used on-policy RL algorithms, such as TRPO (Schulman et al., [2015](#bib.bib52)) and A2C (Mnih et al., [2016](#bib.bib45)).
We re-implemented this method using the state-of-the-art on-policy RL algorithm: PPO Schulman et al. ([2017](#bib.bib54)). We refer to this baseline as PrefPPO.
###
4.2 PEBBLE
PEBBLE (Lee et al., [2021](#bib.bib39)) is a state-of-the-art preference-based RL algorithm that improved the framework of Christiano et al. ([2017](#bib.bib19)) using the following ideas:
Unsupervised pre-training. In the beginning of training,
a naive agent executing a random policy does not provide good state coverage nor coherent behaviors.
Therefore, the agent’s queries are not diverse and a teacher can not convey much meaningful information.
As a result, it requires many samples (and thus queries) for these methods to show initial progress.
Ibarz et al. ([2018](#bib.bib31)) has addressed this issue
by assuming that demonstrations are available at the beginning of the experiment.
However, this is not ideal since suitable demonstrations are often prohibitively expensive to obtain in practice.
Instead,
PEBBLE pre-trains the policy only using intrinsic motivation (Oudeyer et al., [2007](#bib.bib46); Schmidhuber, [2010](#bib.bib51))
to learn how to generate diverse behaviors.
Specifically,
by updating the agent to maximize the state entropy ℋ(𝐬)=−𝔼𝐬∼p(𝐬)[logp(𝐬)]ℋ𝐬subscript𝔼similar-to𝐬𝑝𝐬delimited-[]𝑝𝐬\mathcal{H}(\mathbf{s})=-\mathbb{E}\_{\mathbf{s}\sim p(\mathbf{s})}\left[\log p(\mathbf{s})\right]caligraphic\_H ( bold\_s ) = - blackboard\_E start\_POSTSUBSCRIPT bold\_s ∼ italic\_p ( bold\_s ) end\_POSTSUBSCRIPT [ roman\_log italic\_p ( bold\_s ) ],
it encourages the agent to efficiently explore an environment and collect diverse experiences (see the supplementary material for more details).
Off-policy RL with relabeling.
To overcome the poor sample-efficiency of on-policy RL algorithms,
PEBBLE used the state-of-the-art off-policy RL algorithm: SAC (Haarnoja et al., [2018](#bib.bib27)).
However, the learning process can be unstable because previous experiences in the replay buffer are labeled with previous learned rewards.
PEBBLE stabilizes the learning process by relabeling all of the agent’s past experience every time it updates the reward model.
5 Using B-Pref to analyze algorithmic design decisions
-------------------------------------------------------
(a) Walker
(b) Quadruped
(c) Button Press
(d) Sweep Into
Figure 2: IQM normalized returns with 95% confidence intervals across ten runs. Learning curves and other metrics (median, mean, optimality gap) are in the supplementary material.
We design our experiments to investigate the following:
* [leftmargin=8mm]
* •
How do existing preference-based RL methods compare against each other across environments with different complexity?
* •
How to use B-Pref to analyze algorithmic design decisions for preference-based RL?
###
5.1 Training details
We implement PEBBLE and PrefPPO using publicly released implementations of SAC111<https://github.com/denisyarats/pytorch_sac> and PPO.222<https://github.com/DLR-RM/stable-baselines3>
All hyperparameters of all algorithms are optimized independently for each environment.
All of the experiments were processed using a single GPU (NVIDIA GTX 1080 Ti) and 8 CPU cores (Intel Xeon Gold 6126).
For reliable evaluation Agarwal et al. ([2021](#bib.bib1)),
we measure the normalized returns333On robotic manipulation tasks, we measure the task success rate as defined by the Meta-world authors (Yu et al., [2020](#bib.bib76)).
and report the interquartile mean (IQM) across ten runs using an open-source library rliable.444<https://github.com/google-research/rliable>
More experimental details (e.g., model architectures and the final hyperparameters) and all learning curves with standard deviation are in the supplementary material.
(a) PEBBLE with 2000 queries
(b) PEBBLE with 1000 queries
Figure 3: IQM normalized returns of PEBBLE with various sampling schemes across ten runs on Quadruped.
Learning curves and other metrics (median, mean, optimality gap) are in the supplementary material.
###
5.2 Benchmarking prior methods
Figure [2](#S5.F2 "Figure 2 ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") shows the IQM normalized returns of PEBBLE and PrefPPO at convergence on various simulated teachers listed in Section [3.2](#S3.SS2 "3.2 Simulated human teachers ‣ 3 B-Pref: Benchmarks environments for preference-based RL ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") (see the supplementary material for experimental details).
For a fair comparison,
we apply unsupervised pre-training and disagreement-based sampling to all methods (including SAC and PPO).
PEBBLE outperforms PrefPPO in most of the environments (especially achieving large gains on robotic manipulation tasks).
Interestingly, providing uniform labels to equally preferable segments (Equal) or skipping the queries with similar behaviors (Skip) is more useful than relying only on perfect labels (Oracle) on hard environments like Quadruped (Figure [2(b)](#S5.F2.sf2 "2(b) ‣ Figure 2 ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning")).
While both PEBBLE and PrefPPO achieve fairly efficient performance on correct labels (Oracle, Equal and Skip),
they often suffer from poor performance when teachers can provide the wrong labels (Mistake and Stoc).
This suggests opportunities for further investigations and development of techniques that can improve the robustness to corrupted labels.555We find that label smoothing (Szegedy et al., [2016](#bib.bib64)) is not effective in handling corrupted labels in our experiments (see the supplementary material for supporting results). However, other regularization techniques, like label flipping, L2 regularization and weight decay, would be interesting for further study in future work.
###
5.3 Impact of design decisions in reward learning
Reward learning from preferences involves several design decisions, which can affect the performance of the overall framework.
We showcase the utility of B-Pref by using it to analyze the following algorithmic design choices in depth:
Selecting informative queries.
During training, all experiences are stored in an annotation buffer ℬℬ\mathcal{B}caligraphic\_B
and we generate N𝚚𝚞𝚎𝚛𝚢subscript𝑁𝚚𝚞𝚎𝚛𝚢N\_{\tt query}italic\_N start\_POSTSUBSCRIPT typewriter\_query end\_POSTSUBSCRIPT pairs of segments666We do not compare segments of different lengths. to ask teacher’s preferences from this buffer at each feedback session.
To reduce the burden on the human, we should solicit preferences so as to maximize the information received.
While finding optimal queries is computationally intractable (Ailon, [2012](#bib.bib2)),
several sampling schemes (Biyik & Sadigh, [2018](#bib.bib11); Biyik et al., [2020](#bib.bib12); Sadigh et al., [2017](#bib.bib50)) have been explored to find queries that are likely to change the reward model.
Specifically,
we consider the following sampling schemes, where more details are in the supplementary material:
* [leftmargin=8mm]
* •
Uniform sampling: We pick N𝚚𝚞𝚎𝚛𝚢subscript𝑁𝚚𝚞𝚎𝚛𝚢N\_{\tt query}italic\_N start\_POSTSUBSCRIPT typewriter\_query end\_POSTSUBSCRIPT pairs of segments uniformly at random from the buffer ℬℬ\mathcal{B}caligraphic\_B.
* •
Uncertainty-based sampling:
We first generate the initial batch of N𝚒𝚗𝚒𝚝subscript𝑁𝚒𝚗𝚒𝚝N\_{\tt init}italic\_N start\_POSTSUBSCRIPT typewriter\_init end\_POSTSUBSCRIPT pairs of segments 𝒢𝚒𝚗𝚒𝚝subscript𝒢𝚒𝚗𝚒𝚝\mathcal{G}\_{\tt init}caligraphic\_G start\_POSTSUBSCRIPT typewriter\_init end\_POSTSUBSCRIPT uniformly at random, measure the uncertainty (e.g., variance across ensemble of preference predictors Christiano et al. ([2017](#bib.bib19)) or entropy of a single preference predictor (Lee et al., [2021](#bib.bib39))),
and then select the N𝚚𝚞𝚎𝚛𝚢subscript𝑁𝚚𝚞𝚎𝚛𝚢N\_{\tt query}italic\_N start\_POSTSUBSCRIPT typewriter\_query end\_POSTSUBSCRIPT pairs of segments with high uncertainty.
* •
Coverage-based sampling: From the initial batch 𝒢𝚒𝚗𝚒𝚝subscript𝒢𝚒𝚗𝚒𝚝\mathcal{G}\_{\tt init}caligraphic\_G start\_POSTSUBSCRIPT typewriter\_init end\_POSTSUBSCRIPT,
we choose N𝚚𝚞𝚎𝚛𝚢subscript𝑁𝚚𝚞𝚎𝚛𝚢N\_{\tt query}italic\_N start\_POSTSUBSCRIPT typewriter\_query end\_POSTSUBSCRIPT center points such that the largest distance between a data point and its nearest center is minimized using a greedy selection strategy.
* •
Hybrid sampling: Similar to Yu et al. ([2006](#bib.bib75)), we also consider hybrid sampling, which combines uncertainty-based sampling and coverage-based sampling.
First, we select the N𝚒𝚗𝚝𝚎𝚛subscript𝑁𝚒𝚗𝚝𝚎𝚛N\_{\tt inter}italic\_N start\_POSTSUBSCRIPT typewriter\_inter end\_POSTSUBSCRIPT pairs of segments 𝒢𝚞𝚗subscript𝒢𝚞𝚗\mathcal{G}\_{\tt un}caligraphic\_G start\_POSTSUBSCRIPT typewriter\_un end\_POSTSUBSCRIPT, using uncertainty-based sampling, where N𝚒𝚗𝚒𝚝>N𝚒𝚗𝚝𝚎𝚛subscript𝑁𝚒𝚗𝚒𝚝subscript𝑁𝚒𝚗𝚝𝚎𝚛N\_{\tt init}>N\_{\tt inter}italic\_N start\_POSTSUBSCRIPT typewriter\_init end\_POSTSUBSCRIPT > italic\_N start\_POSTSUBSCRIPT typewriter\_inter end\_POSTSUBSCRIPT, and then choose N𝚚𝚞𝚎𝚛𝚢subscript𝑁𝚚𝚞𝚎𝚛𝚢N\_{\tt query}italic\_N start\_POSTSUBSCRIPT typewriter\_query end\_POSTSUBSCRIPT center points from 𝒢𝚞𝚗subscript𝒢𝚞𝚗\mathcal{G}\_{\tt un}caligraphic\_G start\_POSTSUBSCRIPT typewriter\_un end\_POSTSUBSCRIPT.
Figure [3](#S5.F3 "Figure 3 ‣ 5.1 Training details ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") shows the IQM normalized returns of PEBBLE with various sampling schemes on Quadruped.
We find that the uncertainty-based sampling schemes (i.e., ensemble disagreement and entropy) are superior to other sampling schemes,
while coverage-based sampling schemes do not improve on uniform sampling and slow down the sampling procedures.
To analyze the effects of sampling schemes,
we measure the fraction of equally preferable queries (i.e., y=(0.5,0.5)𝑦0.50.5y=(0.5,0.5)italic\_y = ( 0.5 , 0.5 )) on the Equal teacher.
Figure [5(a)](#S5.F5.sf1 "5(a) ‣ Figure 5 ‣ 5.3 Impact of design decisions in reward learning ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") shows that uncertainty-based sampling schemes achieve high returns even though other sampling schemes receive more (non-uniform) perfect labels.
We expect that this is because queries with high uncertainty provide significant information to the reward model.
This also suggests opportunities for further investigations on uncertainty estimates like Bayesian methods (Brown & Niekum, [2019](#bib.bib15); Gal & Ghahramani, [2016](#bib.bib25)).
Feedback schedule.
We also investigate the impact of the feedback schedule, which decides the number of queries at each feedback session.
Lee et al. ([2021](#bib.bib39)) used a uniform schedule, which always asks the same number of queries, and Christiano et al. ([2017](#bib.bib19)); Ibarz et al. ([2018](#bib.bib31)) used a decay schedule,
which decreases the number of queries, roughly proportional to Tt+T𝑇𝑡𝑇\frac{T}{t+T}divide start\_ARG italic\_T end\_ARG start\_ARG italic\_t + italic\_T end\_ARG, where t𝑡titalic\_t is the current timestep and T𝑇Titalic\_T is the episode length.
We additionally consider an increase schedule, which increases the number of queries, roughly proportional to T+tT𝑇𝑡𝑇\frac{T+t}{T}divide start\_ARG italic\_T + italic\_t end\_ARG start\_ARG italic\_T end\_ARG.
Figure [4](#S5.F4 "Figure 4 ‣ 5.3 Impact of design decisions in reward learning ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") shows the learning curves of PEBBLE with different feedback schedule on the oracle teacher.
Given the same total number of queries,
increase and decay schedules change the size of the initial queries by a factor of 0.5 and 2, respectively.
One can note that there is no big gain from rule-based schedules in most of the environments.
Even though rule-based schedules are less effective than uniform scheduling,
using an adaptive schedule like meta-gradient (Xu et al., [2018](#bib.bib72), [2020](#bib.bib73)) would be interesting for further study in future work.
(a) Button Press
(b) Sweep Into
(c) Quadruped
(d) Walker
Figure 4: Learning curves of PEBBLE with different feedback schedules on the oracle teacher.
The solid line and shaded regions represent the mean and standard deviation, respectively, across ten runs.
Reward analysis.
To investigate the quality of the learned reward function,
we compare the learned reward function with the ground truth reward.
Figure [5(b)](#S5.F5.sf2 "5(b) ‣ Figure 5 ‣ 5.3 Impact of design decisions in reward learning ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") and Figure [5(c)](#S5.F5.sf3 "5(c) ‣ Figure 5 ‣ 5.3 Impact of design decisions in reward learning ‣ 5 Using B-Pref to analyze algorithmic design decisions ‣ B-Pref: Benchmarking Preference-Based Reinforcement Learning") show the learned reward function optimized by PEBBLE on the oracle teacher in Sweep Into and Walker, where more evaluation results on other environments are also available
in the supplementary material.
Because we bound the output of the reward function using tanh function, the scale is different with the ground truth reward but the learned reward function is reasonably well-aligned.
(a) Sampling analysis
(b) Sweep Into
(c) Walker
Figure 5: (a) Fraction of equally preferable queries (red) and average returns (blue) on the Equal teacher.
We use PEBBLE with different sampling schemes on Quadruped given a budget of 2000 queries.
Even though a teacher provides more uniform labels, i.e., y=(0.5,0.5)𝑦0.50.5y=(0.5,0.5)italic\_y = ( 0.5 , 0.5 ), to uncertainty-based sampling schemes,
they achieve higher returns than other sampling schemes. (b/c) Time series of learned reward function (green) and the ground truth reward (red) using rollouts from a policy optimized by PEBBLE.
Learned reward functions align with the ground truth rewards in (b) Sweep Into and (c) Walker.
6 Related work
---------------
Benchmarks for deep reinforcement learning.
There is a large body of work focused on designing benchmarks for RL Beattie et al. ([2016](#bib.bib7)); Bellemare et al. ([2013](#bib.bib8)); Brockman et al. ([2016](#bib.bib14)); Cobbe et al. ([2019](#bib.bib20), [2020](#bib.bib21)); Duan et al. ([2016](#bib.bib22)); Fu et al. ([2020](#bib.bib24)); Gulcehre et al. ([2020](#bib.bib26)); Henderson et al. ([2018](#bib.bib30)); Ray et al. ([2019](#bib.bib48)); Tassa et al. ([2018](#bib.bib65), [2020](#bib.bib66)); Yu et al. ([2020](#bib.bib76)).
The Arcade Learning Environment Bellemare et al. ([2013](#bib.bib8)) has becomes a popular benchmark to measure the progress of RL algorithms for discrete control tasks.
For continuous control tasks,
Duan et al. ([2016](#bib.bib22)) presented a benchmark with baseline implementations of various RL algorithms, which in turn led to OpenAI Gym (Brockman et al., [2016](#bib.bib14)).
These benchmarks have significantly accelerated progress and have been strong contributors towards the discovery and evaluation of today’s most widely used RL algorithms (Haarnoja et al., [2018](#bib.bib27); Mnih et al., [2015](#bib.bib44); Schulman et al., [2015](#bib.bib52), [2016](#bib.bib53), [2017](#bib.bib54)).
Recently, researchers proposed more targeted RL benchmarks that have been designed for specific research purposes.
Cobbe et al. ([2020](#bib.bib21)) presented a suite
of game-like environments where the train and test environments differ for evaluating generalization performance of RL agents.
Ray et al. ([2019](#bib.bib48)) provided a Safety Gym for measuring progress towards RL agents that satisfy the safety constraints.
D4RL Fu et al. ([2020](#bib.bib24)) and RL Unplugged Gulcehre et al. ([2020](#bib.bib26)) have been proposed to evaluate and compare offline RL algorithms. Yu et al. ([2020](#bib.bib76)) proposed Meta-world to study meta- and multi-task RL.
URLB (Laskin et al., [2021](#bib.bib38)) benchmarks performance of unsupervised RL methods.
However, none of the existing RL benchmarks are tailored towards preference-based RL.
Freire et al. ([2020](#bib.bib23)) proposed DERAIL, a benchmark suite for preference-based learning, but they focused on simple diagnostic tasks.
In B-Pref, we consider learning a variety of complex locomotion and robotic manipulation
tasks.
Additionally, we design teachers with a wide array of irrationalities and benchmark state-of-the-art preference-based RL algorithms (Christiano et al., [2017](#bib.bib19); Lee et al., [2021](#bib.bib39)) in depth.
Human-in-the-loop reinforcement learning. Several works have successfully utilized feedback from real humans to train RL agents (Arumugam et al., [2019](#bib.bib6); Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31); Knox & Stone, [2009](#bib.bib34); Lee et al., [2021](#bib.bib39); MacGlashan et al., [2017](#bib.bib43); Warnell et al., [2018](#bib.bib69)).
MacGlashan et al. ([2017](#bib.bib43)) proposed a reward-free method, which utilizes a human feedback as an advantage function and optimizes the agents via a policy gradient.
Knox & Stone ([2009](#bib.bib34)) trained a reward model via regression using unbounded real-valued feedback.
However, these approaches are difficult to scale to more complex learning problems that require substantial agent experience.
Another promising direction has focused on utilizing the human preferences (Akrour et al., [2011](#bib.bib3); Christiano et al., [2017](#bib.bib19); Ibarz et al., [2018](#bib.bib31); Lee et al., [2021](#bib.bib39); Leike et al., [2018](#bib.bib41); Pilarski et al., [2011](#bib.bib47); Stiennon et al., [2020](#bib.bib61); Wilson et al., [2012](#bib.bib70); Wu et al., [2021](#bib.bib71)).
Christiano et al. ([2017](#bib.bib19)) scaled preference-based learning to utilize modern deep learning techniques,
and Ibarz et al. ([2018](#bib.bib31)) improved the efficiency of this method by introducing additional forms of feedback such as demonstrations.
Recently, Lee et al. ([2021](#bib.bib39)) proposed a feedback-efficient RL algorithm by utilizing off-policy learning and pre-training.
Stiennon et al. ([2020](#bib.bib61)) and Wu et al. ([2021](#bib.bib71)) showed that preference-based RL can be utilized to fine-tune GPT-3 (Brown et al., [2020](#bib.bib16)) for hard tasks like text and book summarization, respectively.
We benchmark these state-of-the-art preference-based RL algorithms in this paper.
7 Conclusion
-------------
In this paper,
we present B-Pref, a benchmark specially designed for preference-based RL, covering a wide array of a teacher’s irrationalities.
We empirically investigate state-of-the-art preference-based RL algorithms in depth
and analyze the effects of algorithmic design decisions on our benchmark.
We find that existing methods often suffer from poor performance when teachers provide wrong labels,
and the effects of design decisions are varied depending on the task setups.
These observations call for new algorithms in active learning (Biyik & Sadigh, [2018](#bib.bib11); Biyik et al., [2020](#bib.bib12); Sadigh et al., [2017](#bib.bib50)) and meta-learning (Xu et al., [2018](#bib.bib72), [2020](#bib.bib73)) to be developed.
By providing an open-source release of the benchmark,
we encourage other researchers to use B-Pref as a common starting point to study preference-based RL more systematically.
Limitations. There are several important properties that are not explored in-depth in B-Pref.
One is robustness of learned reward functions to new environments with different dynamics or initial states Reddy et al. ([2020](#bib.bib49)).
Also, we focus on tasks with proprioceptive inputs and dense rewards, but extensions to visual observations and sparse rewards are interesting directions to explore.
Potential negative impacts.
Preference-based RL has several advantages (e.g., teaching novel behaviors, and mitigating the effects of reward exploitation); however, it also has potential drawbacks.
Malicious users might teach the bad behaviors/functionality using this framework.
Therefore, researchers should consider the safety issues with particular thought.
Acknowledgements
----------------
This research is supported in part by
ONR PECASE N000141612723,
NSF NRI #2024675,
ONR YIP,
and Berkeley Deep Drive.
Laura Smith was supported by NSF Graduate Research Fellowship.
We thank Qiyang (Colin) Li and Olivia Watkins for providing helpful feedback and suggestions.
We also thank anonymous reviewers for critically reading the manuscript and suggesting substantial improvements. |
587f12c5-0880-4465-ab7b-0218b39da2b3 | trentmkelly/LessWrong-43k | LessWrong | Against EA PR
Scott Alexander recently wrote about weird effective altruism. Many people (mostly, but not entirely, people who aren't effective altruists) offered the opinion that weird effective altruists should be banned from EA, or at least shouldn't be allowed to give talks at EA Global and have blog posts written about them. Weird effective altruist causes are (sort of by definition) off-putting to most people; therefore, if you want people to donate to global poverty relief, you should kick out all of those people concerned about farmed animal welfare/AI risk/wild animal welfare/psychedelics research/suffering in fundamental physics, lest we scare the normies.
There are many reasonable critiques of this point of view, including that it's not remotely clear that any of those claims are more frightening to normal people than "it is morally obligatory to make personal sacrifices in order to help poor, faraway black people." But ultimately I reject the entire premise.
I'd like to be clear about what I'm not saying in this post. I am not saying all "weird effective altruism" causes are effective; I believe some are and some aren't. I think many effective altruists are not taking seriously enough the difficulty of figuring out how effective highly speculative causes are, and that unless we seriously address this we're going to waste potentially millions of dollars on boondoggles. And I suspect a lot of weird effective altruism tends to over-explore certain cause areas (for example, things you think of if you read too many science fiction novels) and underexplore other cause areas (for example, boring things). I don't intend this post to be a whole-hearted defense of weird effective altruism, but simply a criticism of a single narrow argument too often wielded against it.
So the question arises: why is effective altruism a thing at all?
Most people care about charity effectiveness, at least a little bit. They look up their charities on Charity Navigator before donating; they o |
dd38c695-37e4-4397-bf94-f0f81c18ec46 | trentmkelly/LessWrong-43k | LessWrong | Less Wrong Should Confront Wrongness Wherever it Appears
In a recent discussion about a controversial topic which I will not name here, Vladimir_M noticed something extremely important.
> Because the necessary information is difficult to obtain in a clear and convincing form, and it's drowned in a vast sea of nonsense that's produced on this subject by just about every source of information in the modern society.
I have separated it from its original context, because this issue applies to many important topics. There are many topics where the information that most people receive is confused, wrong, or biased, and where nonsense drowns out truth and clarity. Wherever this occurs, it is very bad and very important to notice.
There are many reasons why it happens, many of which have been explicitly studied and discussed as topics here. The norms and design of the site are engineered to promote clarity and correctness. Strategies for reasoning correctly are frequently recurring topics, and newcomers are encouraged to read a large back-catalog of articles about how to avoid common errors in thinking (the sequences). A high standard of discourse is enforced through voting, which also provides rapid feedback to help everyone improve their writing. Since Well-Kept Gardens Die by Pacifism, when the occasional nutjob stops by, they're downvoted into invisibility and driven away - and while you wouldn't notice from the comment archives, this has happened lots of times.
Less Wrong has the highest accuracy and signal to noise ratio of any blog I've seen, other than those that limit themselves to narrow specialties. In fact, I doubt anyone here knows a better one. The difference is very large. While we are certainly not perfect, errors on Less Wrong are rarer and much more likely to be spotted and corrected than on any similar site, so a community consensus here is a very strong signal of clarity and correctness.
As a result, Less Wrong is well positioned to find and correct errors in the public discourse. Less Wrong should confro |
32f60c6e-2347-4ab7-9d23-d7b70255acdb | trentmkelly/LessWrong-43k | LessWrong | What to put in my time capsule?
In a few weeks I will open a time capsule I set for myself 5 years ago. I'm extremely excited and want to celebrate by setting myself a new time capsule to open in 5 years from now.
The difficult part about making the time capsule is what to include. Since I have not yet opened the original capsule, I don't know how myself from 5 years ago chose what "artifacts" to include.
I feel like whatever goes inside the capsule should be a "snapshot" of life at this moment for me. However, I will still have access to all my writings from around this time that are not in the capsule. So I want to include things from my current life and also things from recent past (anything within 2 years is in this "stage of life").
Here is a breakdown of what I plan to include:
* school papers
* various small items and papers from recent past
* photos
* USB stick with a video to my future self
* document detailing all the artifacts listed above
* letters from other people. I have not yet asked them to write anything but it could be fun.
* letters to other people (that I write down and deliver to them in 5 years) Like the previous idea, I have no idea how it might go.
* list predictions about future self
* letter to future self
* my hopes for future self
* detailed description about my current mental models of the world
* what activities, hobbies, habits I'm currently engaged in (and have done in recent past). For example, what I do to exercise, what songs I listen to, what foods I always eat, etc.
* things I may forget or would love to be reminded of. For example, inside jokes with friends.
* questions I want future self to answer upon opening the capsule. For example, "How have your thoughts on X changed in 5 years?"
What do you think of this list? What do you recommend adding or doing differently? I'd appreciate any ideas on how to carry out experiments like this over long time frames. |
511358ce-7b50-46be-a454-0315245cb9be | trentmkelly/LessWrong-43k | LessWrong | The State of Metaculus
This linkpost summarizes Metaculus updates and initiatives from mid-2024 to early February, 2025. The original post covers technical improvements, new forecasting tournaments, and collaborations with CDC, GiveWell, Bridgewater Associates, Bright Line Watch, and more. While written for regular Metaculus users, I'm sharing it as a resource for those interested in Metaculus's recent work and potential ways to participate.
The Metaculus of early 2025 is a different and better place in myriad ways than the Metaculus of a year ago. It can be easy to miss some of these changes and also some of the opportunities — to sharpen thinking (yours and others'), to win prizes and job opportunities, and to contribute to public sense-making in turbulent times.
This post covers Metaculus from mid-2024 to today, highlighting new developments and features you might have missed — from tournaments and research initiatives to ongoing experiments.
We were fortunate to work with many inspiring partners in 2024, on more projects than we could list below. If you or your organization want to collaborate with us in 2025, get in touch!
As always, feedback is welcome — or you can share ideas in the platform feature suggestions thread.
A Record-Breaking Start to 2025
Astral Codex Ten 2025 with Scott Alexander has become our fastest-growing series ever with over 3,000 forecasters so far — that's double last year's turnout. Even more dramatically, the return of the Bridgewater x Metaculus Forecasting Contest has registered over 17,000 (!) competitors globally, more than 10x last year's count. The past week has been one for the record books - each day shattered either our record for new forecasters or total predictions made. The Monday launch of the Bridgewater contest topped them all, hitting new all-time highs for forecasters, commenters, and signups.. (And yes, the platform is experiencing some growing pains as we welcome new users. We hope you make folks feel welcome, even as you maybe di |
468fd012-4b01-4a7e-8f72-dc021249853b | trentmkelly/LessWrong-43k | LessWrong | Alignment Newsletter #29
Highlights
Deep Imitative Models for Flexible Inference, Planning, and Control (Nicholas Rhinehart et al): It's hard to apply deep RL techniques to autonomous driving, because we can't simply collect a large amount of experience with collisions in order to learn. However, imitation learning is also hard, because as soon as your car deviates from the expert trajectories that you are imitating, you are out of distribution, and you could make more mistakes, leading to accumulating errors until you crash. Instead, we can model the expert's behavior, so that we can tell when we are moving out of distribution, and take corrective action.
They split up the problem into three different stages. First, they generate a set of waypoints along the path to be followed, which are about 20m away from each other, by using A* search on a map. Next, they use model-based planning using an imitative model to generate a plan (sequence of states) that would take the car to the next waypoint. Finally, they use a simple PID controller to choose low-level actions that keep the car on target towards the next state in the plan.
The key technical contribution is with the imitative model, which is a probabilistic model P(s_{1:T}, G, φ), where φ is the current observation (eg. LIDAR), s_{1:T} is the planned trajectory, and G is a goal. We can learn P(s_{1:T} | φ) from expert demonstrations. The goal G can be anything for which you can write down a specification P(G | s_{1:T}, φ). For example, if you simply want to reach a waypoint, you can use the normal distribution on the distance between the final state s_T and the waypoint. You can also incorporate a hand-designed cost on each state.
They evaluate in simulation on a static world (so no pedestrians, for example). They show decent transfer from one map to a second map, and also that they can avoid artificially introduced potholes at test time (despite not seeing them at training time), simply by adding a cost on states over a pothole (which |
2ec78881-8830-4e00-8673-02bcb09e259d | trentmkelly/LessWrong-43k | LessWrong | I think to find out what I write
The maxim goes that “I write to find out what I think”. (Variously attributed to Flannery O'Connor or Joan Didion). It counsels that articulation highlights gaps in arguments, requires concrete positions, and dispels denial of inconvenient conclusions. I do not dispute it.
However, I suspect that the effects flow in both directions. When writing, certain thoughts bubble up from the surface of the mind at the opportune time, by chance. The demands of concision, the limits of time, and the lionisation of decisiveness then exclude the alternatives. Warranted or not, writing amplifies those lucky musings. Unjustified confidence condenses from ambiguity. The words become a commitment, provoking cognitive dissonance when the mind strays elsewhere. Doubly so if published.
Write with care. |
276f6da3-9988-4f11-b695-8050222f4d67 | trentmkelly/LessWrong-43k | LessWrong | Rational = true?
> For example, if you say, "The rational belief is X, but the true belief is Y" then you are probably using the word "rational" in a way that means something other than what most of us have in mind
This was copied from here.
Surely it is obvious that there are lots of examples when one might say this. Consider this:
Rob looks in the newspaper to check the football scores. The newspaper says that United won 3-2, but it is a misprint because City actually won 3-2. In this case, the rational belief is that United won, but the true belief is that City won.
Am I missing something? |
2e4f8c24-00dc-425b-9494-18a82f8edb50 | trentmkelly/LessWrong-43k | LessWrong | Trauma, Meditation, and a Cool Scar
[Trigger Warning: I’ll be discussing a physical injury, recovery, and panic attacks in detail. The first three pictures linked are gory. Again, they are linked, not directly shown]
Trauma
One year ago today, I was in an accident with an industrial drone. It was spinning too fast while arming (like how helicopters spin up before they took off), but nothing we tried would fix it. Eventually, I changed the PWM value back to the default value, and it spun up even faster. Fast enough to take off right into me.
It tore up my arm. It tore up my face. After screaming, it didn’t hurt that bad, so I thought I overreacted. I told everyone “I think I’m okay”. They didn’t believe me, and I was rushed to the hospital. The pain was horrible, but the nausea was worse. I had made everyone apple pie that day, but I didn’t get to keep my piece.
The doctor thought I needed facial reconstruction surgery, so they put me in an ambulance and shipped me to another hospital. They stitched me up, said no facial surgery was needed, but that my lens and iris were destroyed in my left eye. A couple days later, my eyeball bruised. A week of checkups and eye drops 4 times a day, they then put me under for surgery.
I woke up in so much pain, so confused. They told me to keep my head down. I asked Why am I in so much pain? repeatedly. They put me in a wheelchair to take me outside, and told me to keep my head down. But all I could do was feel terrified because I was in pain and no one was doing anything about it. I’m told to keep my head down as they put me in my dad’s car, so I kept my head down and hurt.
For a week, I had to keep my head down. When I ate, my head was down. When I talked to someone, my head was down. When I slept, my head was down1. I couldn’t play piano like I used to because of my arm. I couldn’t read like I used to because of my eye. I couldn’t even think like I used to because my working memory was shot. I felt so powerless and isolated.
How am I supposed to program or l |
88c43ed3-ebd5-45f5-8819-6023d4aa1cc5 | trentmkelly/LessWrong-43k | LessWrong | Has Someone Checked The Cold-Water-In-Left-Ear Thing?
I somewhat recently read these posts, and figured this was enough evidence that it might be worth checking whether people become better calibrated for a short while after very cold water is poured in their left ear. For example: does doing that noticeably improve scores on the calibration game for a few minutes? Has this already been tested in some fashion aside from whether it changes existing strong beliefs in rationalists? If it hasn't, a simple n=1 experiment seems fairly cheap to try, so I'll probably go for it over the next few days. |
da72669d-3148-40ee-9dd7-5332013372a3 | trentmkelly/LessWrong-43k | LessWrong | The case against AI alignment
Trigger warning: Discussion of seriously horrific shit. Honestly, everything is on the table here so if you're on the lookout for trigger warnings you should probably stay away from this conversation.
Any community of people which gains notability will attract criticism. Those who advocate for the importance of AI alignment are no exception. It is undoubtable that you have all heard plenty of arguments against the worth of AI alignment by those who disagree with you on the nature and potential of AI technology. Many have said that AI will never outstrip humans in intellectual capability. Others have said that any sufficiently intelligent AI will “align” themselves automatically, because they will be able to better figure out what is right. Others say that strong AI is far enough in the future that the alignment problem will inevitably be solved by the time true strong AI becomes viable, and the only reason we can’t solve it now is because we don’t sufficiently understand AI.
I am not here to level criticisms of this type at the AI alignment community. I accept most of the descriptive positions endorsed by this community: I believe that AGI is possible and will inevitably be achieved within the next few decades, I believe that the alignment problem is not trivial and that unaligned AGI will likely act against human interests to such an extent as to lead to the extinction of the human race and probably all life as well. My criticism is rather on a moral level: do these facts mean that we should attempt to develop AI alignment techniques?
I say we should not, because although the risks and downsides of unaligned strong AI are great, I do not believe that they even remotely compare in scope to the risks from strong AI alignment techniques in the wrong hands. And I believe that the vast majority of hands this technology could end up in are the wrong hands.
You may reasonably ask: How can I say this, when I have already said that unaligned strong AI will lead to the e |
c3fcc8c5-a3fc-4597-8b53-f21c5a7dd2c6 | trentmkelly/LessWrong-43k | LessWrong | Playing in the Creek
When I was a really small kid, one of my favorite activities was to try and dam up the creek in my backyard. I would carefully move rocks into high walls, pile up leaves, or try patching the holes with sand. The goal was just to see how high I could get the lake, knowing that if I plugged every hole, eventually the water would always rise and defeat my efforts. Beaver behaviour.
One day, I had the realization that there was a simpler approach. I could just go get a big 5 foot long shovel, and instead of intricately locking together rocks and leaves and sticks, I could collapse the sides of the riverbank down and really build a proper big dam. I went to ask my dad for the shovel to try this out, and he told me, very heavily paraphrasing, 'Congratulations. You've cracked the river damming puzzle. And unfortunately, this means you no longer can try as hard as you can to dam up the creek." The price of victory is that the space of games I could play was permanently smaller, and checkmate was not played on the board. This is what growing up looks like.
I experienced many variations. After I put a hole in the ceiling, building K'Nex catapults switched from a careless activity to a careful activity. On the more supervised side, when I was curious how big of an explosion I could make by taping together sparklers, I was not allowed to find the limit. After I cracked the trick of tillering, I could no longer try as hard as I could to build bows and arrows for the backyard battles my friends and I engaged in. I was and am particularly proud of earning this ban.
A couple years back, I got a job offer from an investment bank to help them win zero sum games against people who didn't necessarily deserve to lose. I had tried very hard to get that offer: leetcoding, studying build systems, crafting my resume. It was only once I got it that I realized I no longer could play the game "make as much money as I can." I had found the equivalent of smashing the river banks with a shovel |
b1c5e317-1939-43c9-a852-3241d5837a34 | trentmkelly/LessWrong-43k | LessWrong | A Models-centric Approach to Corrigible Alignment
This post starts with a reasonable story of how alignment might be solved, based on extracting human values from an AIs model of the world. Then I will elaborate on what I see as the main steps in this process. After that I will review the existing work (which I am aware of) on these areas, as well as discussing points of confusion. I will finish with a list of identified areas of confusion in this scheme.
A Reasonable Story for How Things Go Right
An AI models the world, in doing so it models humans, in doing so it models human values. It then extracts those human values somehow. Then it acts "according to" those values in a similar way to humans do, but with superhuman capabilities.
Breaking this down
1: Modelling humans, in particular human decision making. This is distinct from inverse reinforcement learning, as with this approach it only needs to be done from the perspective of a general predictive model: model humans as part of the world. Modelling the world to a reasonable accuracy in general will do the job thanks to the later steps.
2: Modelling humans in such a way that human values can in theory be separated from the rest of the model. This is not trivial. A neural network (for example) which predicts human behaviour will be hopelessly entangled between the human's beliefs, values, and random biases and akrasia. Possibly corrigibility-as-in-understanding work can help here but it doesn't seem easy. Something like a Pearlian causal network of factors governing human decision making would be ideal. Existing work is building towards this.
3: Pointing to a human in the model. This is non-trivial. A model of the world which predicts the world will have things in it isomorphic to humans, but we need a robust way of pointing to them. Existing work seems to give a mathematically well-defined way to point to things. This will probably need testing.
4: Actually separating values from non-values. There is a no-free-lunch theorem which states that values and |
5f56c1f7-8aad-4c1c-97ba-ed9e65f70c2b | trentmkelly/LessWrong-43k | LessWrong | Where do (did?) stable, cooperative institutions come from?
The United States has a bunch of nice things whose creation/maintenance requires coordinated effort from a large number of people across time. For example: bridges that stay up; electrical grids that provide us with power; the rule of law; newspapers that make it easier to keep tabs on recent events; fire fighting services that stop most fires in urban areas; roads; many functioning academic fields; Google; Amazon; grocery stores; the postal service; and so on.
The first question I'd like to pose is: how does this coordination work? What keeps these large sets of people pulling in a common direction (and wanting to pull in a common direction)? And what keeps that "common direction" grounded enough that an actual nice thing results from the pulling (e.g., what causes it to be that you get a working railway system, rather than a bunch of tracks that don't quite work? what causes you to sometimes get a functioning field of inquiry and not a cargo cult)? Is it that:
* Many people independently value the nice thing, and they altruistically decide to put their own efforts toward creating/maintaining the nice thing? (E.g., some large set of people wishes there were good fire-fighting institutions, and so each of them altruistically and independently decides to found a fire-fighting branch, to work at that branch, to tweak that branch's habits into a more effective configuration, etc.?)
* A small number of rich and powerful people (who are somehow also knowledgeable about institution design) value the nice thing, and they altruistically decide to set up incentives such that other people, purely via self-interest, will do the work that is needed to create/maintain the nice thing? (E.g., a small number of people altruistically donate to fire-fighting groups and set up incentives at those groups, and then other people do the fire-fighting work because they want a job?)
* Something else?
One reason I’d like to pose this question is that it seems plausible to me that the m |
0757a490-0d63-437b-9ffc-fdf40121953c | trentmkelly/LessWrong-43k | LessWrong | 12-year old challenges the Big Bang
I thought this may be of interest to the LW community. Jacob Barnett is a 12-year old male who taught himself all of high school math (algebra through calculus), has a currently scored math IQ of 170 (for what that's worth) and is currently on track to become a researcher of astrophysics. His current major news worthy claim-to-fame (aside from being really young): The Big Bang Theory is currently incorrect (I believe the article states he has something about a lack of carbon in the model), and he's planning to develop a new theory.
I haven't learned anything serious in physics, so I have nothing to note on his claim. I realize the news article cited puts him claim fairly generally, so I'll ask this: Can someone explain how elements are generally modeled to have formed from the big bang? And is there anything that it Jacob may be missing in the current literature? |
6c371121-8f70-41fe-bca8-849388d14f87 | trentmkelly/LessWrong-43k | LessWrong | Generic Modafinil sales begin?
Due to agreements with the patent holder, Cephalon, that were made in 2005-2006 several generics manufacturers are now allowed to sell generic Modafinil. I have found confirmation that the manufacturer Teva has begun selling generic Modafinil, though I haven't seen news about other manufacturers. However, the article also, says that Cephalon is a subsidiary of Teva, so perhaps this won't have an effect? The Modafinil patent is set to expire in April 2015. |
f397d8e8-eb7d-48b9-a973-90dd36a47c83 | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "I'll do it at some point.I'll answer this message later.I could try this sometime.The most common consequence of these thoughts is a lot of procrastination. The person who has them may genuinely intend to do the task, but with the "when" being vague, there's nothing that would propel them into action.Here are some thoughts the person could have instead:When I find myself using the words "later" or "at some point", I'll decide on a specific time when I'll actually do it.If I'm given a task that would take under five minutes, and I'm not in a pressing rush, I'll do it right away.When I notice that I'm getting stressed out about something that I've left undone, I'll either do it right away or decide when I'll do it.Picking a specific time or situation to serve as the trigger of the action makes it much more likely that it actually gets done.Could we apply this more generally? Let's consider these examples:I'm going to get more exercise.I'll spend less money on shoes.I want to be nicer to people.All of these goals are vague. How will you actually implement them? As long as you don't know, you're also going to miss potential opportunities to act on them.Let's try again:When I see stairs, I'll climb them instead of taking the elevator.When I buy shoes, I'll write down how much money I've spent on shoes this year.When someone does something that I like, I'll thank them for it.These are much better. They contain both a concrete action to be taken, and a clear trigger for when to take it.Turning vague goals into trigger-action plansTrigger-action plans (TAPs; known as "implementation intentions" in the academic literature) are "when-then" ("if-then", for you programmers) rules used for behavior modification [i]. A meta-analysis covering 94 studies and 8461 subjects [ii] found them to improve people's ability for achieving their goals [iii]. The goals in question included ones such as reducing the amount of fat in one's diet, getting exercise, using vitamin supplements, carrying on with a boring task, determination to work on challenging problems, and calling out racist comments. Many studies also allowed the subjects to set their own, personal goals.TAPs were found to work both in laboratory and real-life settings. The authors of the meta-analysis estimated the risk of publication bias to be small, as half of the studies included were unpublished ones.Designing TAPsTAPs work because they help us notice situations where we could carry out our intentions. They also help automate the intentions: when a person is in a situation that matches the trigger, they are much more likely to carry out the action. Finally, they force us to turn vague and ambiguous goals into more specific ones.A good TAP fulfills three requirements [iv]:The trigger is clear. The "when" part is a specific, visible thing that's easy to notice. "When I see stairs" is good, "before four o'clock" is bad (when before four exactly?). [v]The trigger is consistent. The action is something that you'll always want to do when the trigger is fulfilled. "When I leave the kitchen, I'll do five push-ups" is bad, because you might not have the chance to do five push-ups each time when you leave the kitchen. [vi]The TAP furthers your goals. Make sure the TAP is actually useful!However, there is one group of people who may need to be cautious about using TAPs. One paper [vii] found that people who ranked highly on so-called socially prescribed perfectionism did worse on their goals when they used TAPs. These kinds of people are sensitive to other people's opinions about them, and are often highly critical of themselves. Because TAPs create an association between a situation and a desired way of behaving, it may make socially prescribed perfectionists anxious and self-critical. In two studies, TAPs made college students who were socially prescribed perfectionists (and only them) worse at achieving their goals.For everyone else however, I recommend adopting this TAP:When I set myself a goal, I'll turn it into a TAP.Origin noteThis article was originally published in Finnish at kehitysto.fi. It draws heavily on CFAR's material, particularly the workbook from CFAR's November 2014 workshop. Footnotes[i] Gollwitzer, P. M. (1999). Implementation intentions: strong effects of simple plans. American psychologist, 54(7), 493.[ii] Gollwitzer, P. M., & Sheeran, P. (2006). Implementation intentions and goal achievement: A meta‐analysis of effects and processes. Advances in experimental social psychology, 38, 69-119.[iii] Effect size d = .65, 95% confidence interval [.6, .7].[iv] Gollwitzer, P. M., Wieber, F., Myers, A. L., & McCrea, S. M. (2010). How to maximize implementation intention effects. Then a miracle occurs: Focusing on behavior in social psychological theory and research, 137-161.[v] Wieber, Odenthal & Gollwitzer (2009; unpublished study, discussed in [iv]) tested the effect of general and specific TAPs on subjects driving a simulated car. All subjects were given the goal of finishing the course as quickly as possible, while also damaging their car as little as possible. Subjects in the "general" group were additionally given the TAP, "If I enter a dangerous situation, then I will immediately adapt my speed". Subjects in the "specific" group were given the TAP, "If I see a black and white curve road sign, then I will immediately adapt my speed". Subjects with the specific TAP managed to damage their cars less than the subjects with the general TAP, without being any slower for it.[vi] Wieber, Gollwitzer, et al. (2009; unpublished study, discussed in [iv]) tested whether TAPs could be made even more effective by turning them into an "if-then-because" form: "when I see stairs, I'll use them instead of taking the elevator, because I want to become more fit". The results showed that the "because" reasons increased the subjects' motivation to achieve their goals, but nevertheless made TAPs less effective.The researchers speculated that the "because" might have changed the mindset of the subjects. While an "if-then" rule causes people to automatically do something, "if-then-because" leads people to reflect upon their motivates and takes them from an implementative mindset to a deliberative one. Follow-up studies testing the effect of implementative vs. deliberative mindsets on TAPs seemed to support this interpretation. This suggests that TAPs are likely to work better if they can be carried out as consistently and as with little thought as possible.[vii] Powers, T. A., Koestner, R., & Topciu, R. A. (2005). Implementation intentions, perfectionism, and goal progress: Perhaps the road to hell is paved with good intentions. Personality and Social Psychology Bulletin, 31(7), 902-912."
What you see above is a segment from a LessWrong user's post.
Explain the potential downsides of relying heavily on Trigger-Action Plans (TAPs) for behavior modification. Could there be situations where TAPs might not be the most effective strategy?
Trigger-Action Plans (TAPs) can indeed be an effective strategy for behavior modification, as they can automate certain actions and make it easier to achieve specific goals. However, like any strategy, TAPs are not without their potential downsides and limitations. Here are a few:
1. Over-reliance on Automation: TAPs work on the principle of automating behavior in response to specific triggers. While this can be beneficial for forming habits or accomplishing tasks, it can also lead to an over-reliance on automatic responses. This could potentially limit a person's ability to adapt to new situations or think creatively, as they may become too accustomed to responding in the same way to certain triggers.
2. Inflexibility: TAPs are designed to be consistent, meaning the same action is taken every time a specific trigger is encountered. However, life is rarely so predictable, and circumstances can change. If a person becomes too rigid in their TAPs, they may struggle to adjust their behavior when the situation calls for a different response.
3. Potential for Anxiety: TAPs involve setting specific goals and planning exact responses to certain situations. For some individuals, particularly those with a tendency towards perfectionism or anxiety, this could potentially lead to increased stress. If they fail to perform the planned action when the trigger occurs, they may experience feelings of failure or self-criticism.
4. Neglect of Underlying Issues: While TAPs can be effective for changing surface behaviors, they may not address underlying psychological or emotional issues that could be contributing to the undesired behavior. For example, a person might set a TAP to stop biting their nails when they feel stressed, but this doesn't address the root cause of the stress itself.
5. Over-Simplification: TAPs can sometimes oversimplify complex behaviors or goals. Human behavior is often influenced by a multitude of factors, and reducing it to a simple "if-then" statement may not always capture the full complexity of the situation.
In conclusion, while TAPs can be a useful tool for behavior modification, they are not a one-size-fits-all solution. They should be used judiciously and in conjunction with other strategies to ensure a holistic approach to behavior change.
Reflecting on the limitations of TAPs you've mentioned, could there be specific situations or contexts where these drawbacks might be particularly pronounced or detrimental?
Absolutely, the limitations of TAPs can be more pronounced in certain situations or contexts. Here are a few examples:
1. High-Stress or High-Stakes Situations: In high-stress or high-stakes situations, the inflexibility of TAPs could be particularly detrimental. If a person has trained themselves to respond in a certain way to a specific trigger, but the situation calls for a different response, they may struggle to adapt. This could lead to poor decision-making or increased stress.
2. Complex Behavioral Changes: TAPs may not be as effective when it comes to complex behavioral changes that involve multiple steps or require a deeper understanding of one's emotions and motivations. For example, someone trying to overcome an addiction or manage a mental health condition may find that TAPs alone are insufficient, as these issues often require a more comprehensive approach.
3. In Social Interactions: Social interactions are often nuanced and unpredictable, making it difficult to establish effective TAPs. Relying too heavily on TAPs in social situations could lead to rigid or inappropriate responses. For example, a person might have a TAP to always say something positive when they meet someone new. However, if the new person shares a sad story, an automatic positive response might come off as insensitive.
4. For Perfectionists or Highly Anxious Individuals: As mentioned earlier, TAPs can potentially increase anxiety and self-criticism in individuals who are perfectionists or already prone to anxiety. In these cases, using TAPs could exacerbate these tendencies, particularly if the individual fails to perform the planned action when the trigger occurs.
5. In Rapidly Changing Environments: In environments where circumstances are rapidly changing, the inflexibility of TAPs could be a major drawback. The ability to quickly adapt and respond to new situations is crucial in such environments, and an over-reliance on TAPs could hinder this.
Therefore, while TAPs can be beneficial in many situations, it's important to recognize their limitations and consider other strategies or approaches when necessary.
Discuss the role of flexibility in behavior modification strategies, particularly in rapidly changing environments. How can one maintain adaptability while using TAPs?
Flexibility is a crucial aspect of behavior modification strategies, particularly in rapidly changing environments. This is because behaviors that may be effective in one situation may not work as well in another. Therefore, the ability to adapt and adjust one's behavior based on the current context is key to successful behavior modification.
In the context of TAPs, maintaining flexibility can be challenging due to their inherent design to automate specific responses to specific triggers. However, there are strategies that can be used to incorporate flexibility into TAPs:
1. Use General Triggers: Instead of using very specific triggers, one can use more general ones that can apply to a variety of situations. This allows for a broader range of responses and can make it easier to adapt to different circumstances.
2. Incorporate Conditional Statements: One can also incorporate conditional statements into their TAPs. For example, instead of "When X happens, I will do Y," one might say, "When X happens, if condition Z is also true, I will do Y." This allows for more nuanced responses that can better fit the situation at hand.
3. Regularly Review and Update TAPs: It's important to regularly review and update TAPs based on one's experiences and changes in the environment. This can help ensure that the TAPs remain relevant and effective.
4. Use TAPs in Combination with Other Strategies: TAPs can be used in combination with other behavior modification strategies that promote flexibility, such as mindfulness, cognitive-behavioral techniques, or problem-solving strategies. This can provide a more holistic approach to behavior change.
5. Foster a Growth Mindset: Cultivating a growth mindset – the belief that abilities and behaviors can be developed through dedication and hard work – can also promote flexibility. This mindset can encourage individuals to see challenges as opportunities for growth and learning, rather than as obstacles.
In conclusion, while TAPs can be a useful tool for behavior modification, it's important to incorporate flexibility into their use to ensure they remain effective in different situations and environments. |
d8f47a4f-f6f0-485c-a481-eda4d587fb09 | trentmkelly/LessWrong-43k | LessWrong | DeepMind: The Podcast - Excerpts on AGI
DeepMind: The Podcast - Season 2 was released over the last ~1-2 months. The two episodes most relevant to AGI are:
* The road to AGI - DeepMind: The Podcast (S2, Ep5) and
* The promise of AI with Demis Hassabis - DeepMind: The Podcast (S2, Ep9)
I found a few quotes noteworthy and thought I'd share them here for anyone who didn't want to listen to the full episodes:
The road to AGI (S2, Ep5)
(Published February 15, 2022)
Shane Legg's AI Timeline
Shane Legg (4:03):
> If you go back 10-12 years ago the whole notion of AGI was lunatic fringe. People [in the field] would literally just roll their eyes and just walk away. [...] [I had that happen] multiple times. I have met quite a few of them since. There have even been cases where some of these people have applied for jobs at DeepMind years later. But yeah, it was a field where you know there were little bits of progress happening here and there, but powerful AGI and rapid progress seemed like it was very, very far away. [...] Every year [the number of people who roll their eyes at the notion of AGI] becomes less.
Hannah Fry (5:02):
> For over 20 years, Shane has been quietly making predictions of when he expects to see AGI.
Shane Legg (5:09):
> I always felt that somewhere around 2030-ish it was about a 50-50 chance. I still feel that seems reasonable. If you look at the amazing progress in the last 10 years and you imagine in the next 10 years we have something comparable, maybe there's some chance that we will have an AGI in a decade. And if not in a decade, well I don't know, say three decades or so.
Hannah Fry (5:33):
> And what do you think [AGI] will look like? [Shane answers at length.]
David Silver on it being okay to have AGIs with different goals (??)
Hannah Fry (16:45):
> Last year David co-authored a provocatively titled paper called Reward is Enough. He believes reinforcement learning alone could lead all the way to artificial general intelligence.
>
> [...] (21:37)
>
> But not everyo |
5f8c1bc2-4623-4d8b-8dee-47f90aed0b98 | StampyAI/alignment-research-dataset/agisf | AGI Safety Fund | [Week 1] “Intelligence Explosion: Evidence and Import” (Sections 3 to 4.1) by Luke Muehlhauser & Anna Salamon
Intelligence Explosion: Evidence and Import
3.From AI to Machine Superintelligence
It seems unlikely that humans are near the ceiling of possible intelligences, rather than
simply being the first such intelligence that happened to evolve. Computers far outper-
form humans in many narrow niches (e.g. arithmetic, chess, memory size), and there is
reason to believe that similar large improvements over human performance are possible
for general reasoning, technology design, and other tasks of interest. As occasional AI
critic Jack Schwartz (1987) wrote:
If artificial intelligences can be created at all, there is little reason to believe
that initial successes could not lead swiftly to the construction of artificial su-
perintelligences able to explore significant mathematical, scientific, or engi-
neering alternatives at a rate far exceeding human ability, or to generate plans
and take action on them with equally overwhelming speed. Since man’s near-
monopoly of all higher forms of intelligence has been one of the most basic
facts of human existence throughout the past history of this planet, such de-
velopments would clearly create a new economics, a new sociology, and a new
history.
Why might AI “lead swiftly” to machine superintelligence? Below we consider some
reasons.
3.1. AI Advantages
Below we list a few AI advantages that may allow AIs to become not only vastly more
intelligent than any human, but also more intelligent than all of biological humanity
(Sotala 2012; Legg 2008). Many of these are unique to machine intelligence, and that
is why we focus on intelligence explosion from AI rather than from biological cognitive
enhancement (Sandberg 2011).
Increased computational resources. The human brain uses 85–100 billion neurons. This
limit is imposed by evolution-produced constraints on brain volume and metabolism.
In contrast, a machine intelligence could use scalable computational resources (imagine
a “brain” the size of a warehouse). While algorithms would need to be changed in order
to be usefully scaled up, one can perhaps get a rough feel for the potential impact here by
noting that humans have about 3.5 times the brain size of chimps (Schoenemann 1997),
and that brain size and IQ correlate positively in humans, with a correlation coefficient
of about 0.35 (McDaniel 2005). One study suggested a similar correlation between
brain size and cognitive ability in rats and mice (Anderson 1993).15
15. Note that given the definition of intelligence we are using, greater computational resources would
not give a machine more “intelligence” but instead more “optimization power.”
10
Luke Muehlhauser, Anna Salamon
Communication speed. Axons carry spike signals at 75 meters per second or less (Kan-
del, Schwartz, and Jessell 2000). That speed is a fixed consequence of our physiology.
In contrast, software minds could be ported to faster hardware, and could therefore pro-
cess information more rapidly. (Of course, this also depends on the efficiency of the
algorithms in use; faster hardware compensates for less efficient software.)
Increased serial depth. Due to neurons’ slow firing speed, the human brain relies on
massive parallelization and is incapable of rapidly performing any computation that re-
quires more than about 100 sequential operations (Feldman and Ballard 1982). Perhaps
there are cognitive tasks that could be performed more efficiently and precisely if the
brain’s ability to support parallelizable pattern-matching algorithms were supplemented
by support for longer sequential processes. In fact, there are many known algorithms
for which the best parallel version uses far more computational resources than the best
serial algorithm, due to the overhead of parallelization.16
Duplicability. Our research colleague Steve Rayhawk likes to describe AI as “instant
intelligence; just add hardware!” What Rayhawk means is that, while it will require ex-
tensive research to design the first AI, creating additional AIs is just a matter of copying
software. The population of digital minds can thus expand to fill the available hardware
base, perhaps rapidly surpassing the population of biological minds.
Duplicability also allows the AI population to rapidly become dominated by newly
built AIs, with new skills. Since an AI’s skills are stored digitally, its exact current state
can be copied,17including memories and acquired skills—similar to how a “system state”
can be copied by hardware emulation programs or system backup programs. A human
who undergoes education increases only his or her own performance, but an AI that
becomes 10% better at earning money (per dollar of rentable hardware) than other AIs
can be used to replace the others across the hardware base—making each copy 10% more
efficient.18
Editability. Digitality opens up more parameters for controlled variation than is pos-
sible with humans. We can put humans through job-training programs, but we can’t
perform precise, replicable neurosurgeries on them. Digital workers would be more
editable than human workers are. Consider first the possibilities from whole brain em-
ulation. We know that transcranial magnetic stimulation (TMS) applied to one part of
16. For example see Omohundro (1987).
17. If the first self-improving AIs at least partially require quantum computing, the system states of
these AIs might not be directly copyable due to the no-cloning theorem (Wootters and Zurek 1982).
18. Something similar is already done with technology-enabled business processes. When the phar-
macy chain CVS improves its prescription-ordering system, it can copy these improvements to more than
4,000 of its stores, for immediate productivity gains (McAfee and Brynjolfsson 2008).
11
Intelligence Explosion: Evidence and Import
the prefrontal cortex can improve working memory (Fregni et al. 2005). Since TMS
works by temporarily decreasing or increasing the excitability of populations of neurons,
it seems plausible that decreasing or increasing the “excitability” parameter of certain
populations of (virtual) neurons in a digital mind would improve performance. We
could also experimentally modify dozens of other whole brain emulation parameters,
such as simulated glucose levels, undifferentiated (virtual) stem cells grafted onto par-
ticular brain modules such as the motor cortex, and rapid connections across different
parts of the brain.19Secondly, a modular, transparent AI could be even more directly
editable than a whole brain emulation—possibly via its source code. (Of course, such
possibilities raise ethical concerns.)
Goal coordination. Let us call a set of AI copies or near-copies a “copy clan.” Given
shared goals, a copy clan would not face certain goal coordination problems that limit
human effectiveness (J. W. Friedman 1994). A human cannot use a hundredfold salary
increase to purchase a hundredfold increase in productive hours per day. But a copy clan,
if its tasks are parallelizable, could do just that. Any gains made by such a copy clan, or
by a human or human organization controlling that clan, could potentially be invested
in further AI development, allowing initial advantages to compound.
Improved rationality. Some economists model humans as Homo economicus : self-
interested rational agents who do what they believe will maximize the fulfillment of
their goals (M. Friedman 1953). On the basis of behavioral studies, though, Schneider
(2010) points out that we are more akin to Homer Simpson: we are irrational beings
that lack consistent, stable goals (Schneider 2010; Cartwright 2011). But imagine if you
werean instance of Homo economicus . You could stay on a diet, spend the optimal amount
of time learning which activities will achieve your goals, and then follow through on an
optimal plan, no matter how tedious it was to execute. Machine intelligences of many
types could be written to be vastly more rational than humans, and thereby accrue the
benefits of rational thought and action. The rational agent model (using Bayesian prob-
ability theory and expected utility theory) is a mature paradigm in current AI design
(Hutter 2005; Russell and Norvig 2010, ch. 2).
These AI advantages suggest that AIs will be capable of far surpassing the cognitive
abilities and optimization power of humanity as a whole, but will they be motivated to
do so? Though it is difficult to predict the specific motivations of advanced AIs, we can
make some predictions about convergent instrumental goals—instrumental goals useful
for the satisfaction of almost any final goals.
19. Many suspect that the slowness of cross-brain connections has been a major factor limiting the
usefulness of large brains (Fox 2011).
12
Luke Muehlhauser, Anna Salamon
3.2. Instrumentally Convergent Goals
Omohundro (2007, 2008, 2012) and Bostrom (forthcoming) argue that there are several
instrumental goals that will be pursued by almost any advanced intelligence because
those goals are useful intermediaries to the achievement of almost any set of final goals.
For example:
1.An AI will want to preserve itself because if it is destroyed it won’t be able to act
in the future to maximize the satisfaction of its present final goals.
2.An AI will want to preserve the content of its current final goals because if the
content of its final goals is changed it will be less likely to act in the future to
maximize the satisfaction of its present final goals.20
3.An AI will want to improve its own rationality and intelligence because this will
improve its decision-making, and thereby increase its capacity to achieve its goals.
4.An AI will want to acquire as many resources as possible, so that these resources
can be transformed and put to work for the satisfaction of the AI’s final and in-
strumental goals.
Later we shall see why these convergent instrumental goals suggest that the default out-
come from advanced AI is human extinction. For now, let us examine the mechanics of
AI self-improvement.
3.3. Intelligence Explosion
The convergent instrumental goal for self-improvement has a special consequence. Once
human programmers build an AI with a better-than-human capacity for AI design, the
instrumental goal for self-improvement may motivate a positive feedback loop of self-
enhancement.21Now when the machine intelligence improves itself, it improves the
intelligence that does the improving. Thus, if mere human efforts suffice to produce
machine intelligence this century, a large population of greater-than-human machine
intelligences may be able to create a rapid cascade of self-improvement cycles, enabling
20. Bostrom (2012) lists a few special cases in which an AI may wish to modify the content of its final
goals.
21. When the AI can perform 10% of the AI design tasks and do them at superhuman speed, the
remaining 90% of AI design tasks act as bottlenecks. However, if improvements allow the AI to perform
99% of AI design tasks rather than 98%, this change produces a much larger impact than when improve-
ments allowed the AI to perform 51% of AI design tasks rather than 50% (Hanson 1998). And when the
AI can perform 100% of AI design tasks rather than 99% of them, this removes altogether the bottleneck
of tasks done at slow human speeds.
13
Intelligence Explosion: Evidence and Import
a rapid transition to machine superintelligence. Chalmers (2010) discusses this process
in some detail, so here we make only a few additional points.
The term “self,” in phrases like “recursive self-improvement” or “when the machine
intelligence improves itself,” is something of a misnomer. The machine intelligence
could conceivably edit its own code while it is running (Schmidhuber 2007; Schaul
and Schmidhuber 2010), but it could also create new intelligences that run indepen-
dently. Alternatively, several AIs (perhaps including WBEs) could work together to
design the next generation of AIs. Intelligence explosion could come about through
“self”-improvement or through other-AI improvement.
Once sustainable machine self-improvement begins, AI development need not pro-
ceed at the normal pace of human technological innovation. There is, however, sig-
nificant debate over how fast or local this “takeoff” would be (Hanson and Yudkowsky
2008; Loosemore and Goertzel 2011; Bostrom, forthcoming), and also about whether
intelligence explosion would result in a stable equilibrium of multiple machine superin-
telligences or instead a machine “singleton” (Bostrom 2006). We will not discuss these
complex issues here.
4.Consequences of Machine Superintelligence
If machines greatly surpass human levels of intelligence—that is, surpass humanity’s
capacity for efficient cross-domain optimization—we may find ourselves in a position
analogous to that of the apes who watched as humans invented fire, farming, writing,
science, guns and planes and then took over the planet. (One salient difference would be
that no single ape witnessed the entire saga, while we might witness a shift to machine
dominance within a single human lifetime.) Such machines would be superior to us in
manufacturing, harvesting resources, scientific discovery, social aptitude, and strategic
action, among other capacities. We would not be in a position to negotiate with them,
just as neither chimpanzees nor dolphins are in a position to negotiate with humans.
Moreover, intelligence can be applied in the pursuit of any goal. As Bostrom (2012)
argues, making AIs more intelligent will not make them want to change their goal
systems—indeed, AIs will be motivated to preserve their initial goals. Making AIs more
intelligent will only make them more capable of achieving their original final goals,
whatever those are.22
This brings us to the central feature of AI risk: Unless an AI is specifically pro-
grammed to preserve what humans value, it may destroy those valued structures (in-
22. This may be less true for early-generation WBEs, but Omohundro (2007) argues that AIs will
converge upon being optimizing agents, which exhibit a strict division between goals and cognitive ability.
14
Luke Muehlhauser, Anna Salamon
cluding humans) incidentally . As Yudkowsky (2008a) puts it, “the AI does not love you,
nor does it hate you, but you are made of atoms it can use for something else.”
4.1. Achieving a Controlled Intelligence Explosion
How, then, can we give AIs desirable goals before they self-improve beyond our ability to
control them or negotiate with them?23WBEs and other brain-inspired AIs running on
human-derived “spaghetti code” may not have a clear “slot” in which to specify desirable
goals (Marcus 2008). The same may also be true of other “opaque” AI designs, such as
those produced by evolutionary algorithms—or even of more transparent AI designs.
Even if an AI had a transparent design with a clearly definable utility function,24would
we know how to give it desirable goals? Unfortunately, specifying what humans value
may be extraordinarily difficult, given the complexity and fragility of human preferences
(Yudkowsky 2011; Muehlhauser and Helm 2012), and allowing an AI to learndesirable
goals from reward and punishment may be no easier (Yudkowsky 2008a). If this is
correct, then the creation of self-improving AI may be detrimental by default unless we
first solve the problem of how to build an AI with a stable, desirable utility function—a
“Friendly AI” (Yudkowsky 2001).25
But suppose it is possible to build a Friendly AI (FAI) capable of radical self-
improvement. Normal projections of economic growth allow for great discoveries rel-
evant to human welfare to be made eventually—but a Friendly AI could make those
discoveries much sooner. A benevolent machine superintelligence could, as Bostrom
(2003) writes, “create opportunities for us to vastly increase our own intellectual and
emotional capabilities, and it could assist us in creating a highly appealing experiential
world in which we could live lives devoted [to] joyful game-playing, relating to each
other, experiencing, personal growth, and to living closer to our ideals.”
23. Hanson (2012) reframes the problem, saying that “we should expect that a simple continuation of
historical trends will eventually end up [producing] an ‘intelligence explosion’ scenario. So there is little
need to consider [Chalmers’] more specific arguments for such a scenario. And the inter-generational
conflicts that concern Chalmers in this scenario are generic conflicts that arise in a wide range of past,
present, and future scenarios. Yes, these are conflicts worth pondering, but Chalmers offers no reasons
why they are interestingly different in a ‘singularity’ context.” We briefly offer just one reason why the
“inter-generational conflicts” arising from a transition of power from humans to superintelligent machines
are interestingly different from previous the inter-generational conflicts: as Bostrom (2002) notes, the
singularity may cause the extinction not just of people groups but of the entire human species. For a
further reply to Hanson, see Chalmers (2012).
24. A utility function assigns numerical utilities to outcomes such that outcomes with higher utilities
are always preferred to outcomes with lower utilities (Mehta 1998).
25. It may also be an option to constrain the first self-improving AIs just long enough to develop a
Friendly AI before they cause much damage.
15 |
ec28d679-9b27-4cd9-9add-5154c310e17d | StampyAI/alignment-research-dataset/special_docs | Other | A Tour of Emerging Cryptographic Technologies | GovAI
A Tour of Emerging
Cryptographic
Technologies
What They Are and
How They Could Matter
May 2021
Ben Garfinkel
Centre for the Governance of AI,
Future of Humanity Institute,
University of Oxford
Abstract
Historically, progress in the /fi.ligaeld of cryptography has been enormously consequential.
Over the past century, for instance, cryptographic discoveries have played a key role in
a world war and made it possible to use the internet for business and private communi-
cation. In the interest of exploring the impact the /fi.ligaeld may have in the future, I con-
sider a suite of more recent developments. My primary focus is on blockchain-based tech-
nologies (such as cryptocurrencies) and on techniques for computing on con/fi.ligadential data
(such as secure multiparty computation). I provide an introduction to these technolo-
gies that assumes no mathematical background or previous knowledge of cryptography.
Then, I consider several speculative predictions that some researchers and engineers have
made about the technologies’ long-term political signi/fi.ligacance. This includes predictions
that more “privacy-preserving” forms of surveillance will become possible, that the roles
of centralized institutions ranging from banks to voting authorities will shrink, and that
new transnational institutions known as “decentralized autonomous organizations” will
emerge. Finally, I close by discussing some challenges that are likely to limit the signi/fi.ligacance
of emerging cryptographic technologies. On the basis of these challenges, it is premature
to predict that any of them will approach the transformativeness of previous technologies.
However, this remains a rapidly developing area well worth following.
Report background
This report was written as a follow-up to a /two.lf/zero.lf/one.lf/six.lf workshop at the Future of Humanity In-
stitute, exploring the implications of blockchain technology. The workshop participants
were Stuart Armstrong, Shahar Avin, Nick Bostrom, Miles Brundage, Vitalik Buterin, Je/f\_f.liga
Coleman, Owen Cotton-Barratt, Wei Dai, Owain Evans, Virgil Gri/f\_f\_i.ligath, Georgios Pil-
iouras, Anders Sandberg, and Vlad Zam/fi.ligar. I later expanded the report by drawing on
sources ranging from academic journal articles to blog posts to technical whitepapers to
informal conversations with researchers and engineers in this space. The goal was to write
something that would allow complete non-experts (like myself when I began the project)
to quickly understand what these new technologies are and why so many people are excited
or worried about them.
Most of this report’s text dates back to /two.lf/zero.lf/one.lf/seven.lf. Fortunately, because the report’s focus is on
fundamental principles and limitations, I have only felt the need to make modest revisions
in the intervening years. At least as of Spring /two.lf/zero.lf/two.lf/one.lf, I believe it can still be read as an
accurate description of the state of /fi.ligaeld./one.lf
/one.lfThank you to Je/f\_f.liga Coleman, Rhys Lindmark, Jaan Tallinn, Morten Dahl, Andrew Trask, Allan Dafoe, Jan
Leike, Anish Mohammed, Luke Muehlhauser, Carrick Flynn, Pablo Sta/f\_f.ligaorini, Nick Brown, and Oge Nnadi for
detailed comments on previous drafts of this report. Thank you as well to Anne le Roux, Laura Pomarius, and
Justis Mills for help with the preparation of the report.
/one.lf
Contents
/one.lf Introduction /four.lf
/two.lf Cryptographic technologies: de/fi.liganitions, explanations, and examples /six.lf
/two.lf./one.lf Public-key encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /seven.lf
/two.lf./two.lf Digital signatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /one.lf/one.lf
/two.lf./three.lf Cryptographic hash functions . . . . . . . . . . . . . . . . . . . . . . . . /one.lf/two.lf
/two.lf./four.lf Trusted timestamping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /one.lf/two.lf
/two.lf./five.lf Tamper-evident logs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /one.lf/three.lf
/two.lf./six.lf Blockchains and distributed computing . . . . . . . . . . . . . . . . . . . /one.lf/six.lf
/two.lf./six.lf./one.lf Background: Concepts in distributed computing . . . . . . . . . /one.lf/six.lf
/two.lf./six.lf./two.lf Blockchains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /one.lf/eight.lf
/two.lf./six.lf./three.lf Consensus protocols . . . . . . . . . . . . . . . . . . . . . . . . . /two.lf/two.lf
/two.lf./seven.lf Cryptocurrency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /two.lf/five.lf
/two.lf./eight.lf Zero-knowledge proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . /two.lf/nine.lf
/two.lf./nine.lf Smart property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /three.lf/two.lf
/two.lf./one.lf/zero.lf Smart contracts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /three.lf/four.lf
/two.lf./one.lf/one.lf Homomorphic encryption . . . . . . . . . . . . . . . . . . . . . . . . . . /three.lf/six.lf
/two.lf./one.lf/two.lf Functional encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /three.lf/seven.lf
/two.lf./one.lf/three.lf Secure multiparty computation and secret sharing . . . . . . . . . . . . . /three.lf/eight.lf
/three.lf Speculative consequences /four.lf/two.lf
/three.lf./one.lf Consequences from technologies other than blockchain . . . . . . . . . . /four.lf/two.lf
/three.lf./one.lf./one.lf Information channels used to conduct surveillance could “go dark” /four.lf/two.lf
/three.lf./one.lf./two.lf Privacy-preserving surveillance could become feasible . . . . . . . /four.lf/four.lf
/three.lf./one.lf./three.lf Non-intrusive agreement veri/fi.ligacation could become feasible . . . /four.lf/eight.lf
/three.lf./one.lf./four.lf It could become easier to combat forgery . . . . . . . . . . . . . . /four.lf/nine.lf
/three.lf./two.lf Consequences from blockchain-based technologies . . . . . . . . . . . . . /five.lf/zero.lf
/three.lf./two.lf./one.lf Background: Concepts in institutional economics . . . . . . . . . /five.lf/zero.lf
/three.lf./two.lf./two.lf The roles of banks, technology companies, voting authorities, and
other traditional institutions could shrink . . . . . . . . . . . . . /five.lf/two.lf
/three.lf./two.lf./three.lf It could become possible to solve collection action problems that
existing institutions cannot . . . . . . . . . . . . . . . . . . . . . /five.lf/three.lf
/three.lf./two.lf./four.lf A new variety of institutions, known as “decentralized autonomous
organizations,” could emerge . . . . . . . . . . . . . . . . . . . . /five.lf/eight.lf
/four.lf Limitations and skeptical views /six.lf/two.lf
/four.lf./one.lf Limitations of privacy-preserving technologies . . . . . . . . . . . . . . . /six.lf/two.lf
/four.lf./one.lf./one.lf The ine/f\_f\_i.ligaciency of computing on con/fi.ligadential data . . . . . . . . /six.lf/two.lf
/four.lf./two.lf Limitations of blockchain-based technologies . . . . . . . . . . . . . . . . /six.lf/three.lf
/four.lf./two.lf./one.lf The di/f\_f\_i.ligaculty of “scaling” permissionless blockchains . . . . . . . /six.lf/three.lf
/four.lf./two.lf./two.lf The threat of restrictive regulations . . . . . . . . . . . . . . . . . /six.lf/five.lf
/four.lf./two.lf./three.lf The potential insecurity of permissionless blockchains . . . . . . /six.lf/six.lf
/four.lf./two.lf./four.lf The inadequacies of smart contracts . . . . . . . . . . . . . . . . /six.lf/eight.lf
/four.lf./two.lf./five.lf The possibility that existing institutions are “good enough” . . . . /seven.lf/one.lf
/two.lf
A Relevance of progress in arti/fi.ligacial intelligence /nine.lf/zero.lf
A./one.lf AI systems may enable more e/f\_f.ligaective surveillance . . . . . . . . . . . . . /nine.lf/zero.lf
A./two.lf AI systems may help to make privacy-preserving surveillance feasible . . . /nine.lf/zero.lf
A./three.lf AI systems may increase the need for anti-forgery schemes . . . . . . . . /nine.lf/one.lf
A./four.lf Methods of computing on con/fi.ligadential data could reduce barriers to de-
veloping certain AI systems . . . . . . . . . . . . . . . . . . . . . . . . . . /nine.lf/one.lf
A./five.lf The problems of safe AI design and safe smart contract design may be
connected . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /nine.lf/one.lf
A./six.lf New coordination and veri/fi.ligacation mechanisms may be useful for govern-
ing AI systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /nine.lf/two.lf
A./seven.lf Fully homomorphic encryption may have applications in AI safety and
security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . /nine.lf/two.lf
/three.lf
/one.lf Introduction
The past century saw the emergence of a number of transformative technologies, ranging
from nuclear weapons to computers.
For the sake of anticipating future challenges and opportunities, it is worth considering
which areas of technology, if any, could play similarly consequential roles this century. One
area that has recently provoked a great deal of excitement is cryptography , or the study of
techniques for encoding, protecting, and authenticating data.
Arguably, progress in cryptography (and its complementary /fi.ligaeld, cryptanalysis , which fo-
cuses on decoding) should already be included on any list of the most consequential devel-
opments of the past century [/one.lf/three.lf]. The most famous case of cryptography and cryptanalysis
in the /two.lf/zero.lfth century may be Germany’s use of then-advanced encryption during the Second
World War and Britain’s corresponding cryptanalysis e/f\_f.ligaort, which at least one historian
has estimated sped up Allied victory by more than a year [/nine.lf/four.lf]. More recently, crypto-
graphic technologies developed in the last /fi.ligafty years have allowed the internet (and other
long-distance communication channels) to be used for otherwise impractical purposes,
such as making /fi.liganancial transactions and sending private messages. In addition, the suc-
cessful use of encryption has posed a continuing challenge to government surveillance and
intelligence programs.
Over the past decade, a number of new cryptographic technologies have begun to emerge
(see Tables /one.lf and /two.lf). These technologies include blockchains, which have enabled the cre-
ation of highly reliable records and computer applications that no single party controls;
cryptocurrencies, such as bitcoin, which have accrued hundreds of billions of dollars of
value and allow users to make digital transactions without relying on traditional /fi.liganancial
institutions or traditional /fi.ligaat currencies; smart contracts, which allow users to enter into
agreements that are enforced largely by algorithms; and fully homomorphic encryption,
which allows users to process data without having access to it in unencrypted form./two.lfSome
existing technologies, such as secure multiparty computation, have also become much more
practical to implement or found novel applications.
It remains to be seen whether these recent developments in cryptography will be as sig-
ni/fi.ligacant as those that came before. However, a number of fairly radical claims have been
made about their importance. I list just a few examples: The United Kingdom’s Gov-
ernment O/f\_f\_i.ligace for Science has described blockchains as the /fi.ligarst signi/fi.ligacant innovation
in record-keeping since ancient times [/one.lf/seven.lf/nine.lf]. Ralph Merkle, one of the founding /fi.ligagures
of modern cryptography, has written a paper arguing that blockchains will enable novel
forms of democracy [/one.lf/one.lf/three.lf]. Jaan Tallinn, co-founder of Skype and the Future of Life Insti-
tute, has advocated for the use of smart contract technology to solve global collective ac-
/two.lfSome of these technologies, particularly cryptocurrencies and smart contracts, di/f\_f.ligaer from more traditional
cryptographic technologies in a pair of important ways. First, they depend in part on systems of economic in-
centives to function, and second, their applications primarily concern the transfer of property. For these reasons,
it may be more appropriate to refer to them as “cryptoeconomic” technologies, as some engineers working to
develop them currently do [/one.lf/five.lf/four.lf]. However, since no standard terminology has yet been adopted, I will continue
to use the term “cryptography” as a catch-all.
/four.lf
tion problems [/three.lf/five.lf]. Elsewhere, researchers have argued that cryptocurrencies could make it
much more di/f\_f\_i.ligacult for governments to in/fl.ligauence or trace private transactions, while oth-
ers have written that homomorphic encryption could enable novel forms of surveillance
that require less infringement of individuals’ privacy [/eight.lf/three.lf, /one.lf/six.lf/six.lf].
Unfortunately, discussions of such claims have often played out in scattered blog posts
and papers that are di/f\_f\_i.ligacult for a non-specialist reader to /fi.ligand or understand. This report
is intended to be a contribution to the project of gathering and clarifying these discus-
sions.
In particular, this report is divided into three sections: First, I introduce some recent de-
velopments in cryptography, aiming to include close to the minimum level of detail needed
to discuss the relevant technologies clearly. Second, I describe several potential long-term,
politically signi/fi.ligacant consequences. Finally, I explore some of the technical limitations
and political constraints that could prevent these consequences from arising.
In addition, an appendix discusses the relationship between the future of cryptography
and the future of arti/fi.ligacial intelligence.
For readers not already familiar with cryptography, the /fi.ligarst section is best read in order.
However, later subsections are not so dependent on one another.
/five.lf
/two.lf Cryptographic technologies: de/fi.liganitions, explanations, and
examples
The set of cryptographic technologies that we will be considering is highly diverse. Of
interest in this report will be:
•Public-key cryptography
•Digital signatures
•Cryptographic hash functions
•Trusted timestamping
•Tamper-evident logs
•Blockchains
•Cryptocurrencies
•Zero-knowledge proofs
•Smart property
•Smart contracts
•Homomorphic encryption
•Functional encryption
•Secure multiparty computation and secret sharing
Some of these technologies are new, having been developed primarily in just the last /fi.ligafteen
years. Some are older, but either serve as core components of these newer technologies or
continue to /fi.ligand additional applications of their own. In this section, I aim to provide
descriptions of each technology that are su/f\_f\_i.ligacient to enable informed discussions of their
potential applications and limitations.
For an overview, Tables /one.lf and /two.lf provide a highly abridged summary. In addition, for a
deeper look, I will now also recommend some sources for further reading.
Readers interested in more thorough or technical descriptions of well-established tech-
nologies, like public-key encryption, digital signatures, and cryptographic hash functions,
can /fi.ligand them in any of a number of widely used introductory textbooks, such as Introduc-
tion to Modern Cryptography by Jonathan Katz and Yehuda Lindell [/nine.lf/nine.lf].
Readers interested in blockchains, cryptocurrencies, smart property, and smart contracts
can /fi.ligand discussions of them in the textbook Bitcoin and Cryptocurrency Technologies , by
Narayanan et al. [/one.lf/one.lf/nine.lf]. Vitalik Buterin’s whitepaper for the Ethereum blockchain and
Ethan Buchman’s thesis outlining the Tendermint blockchain also continue to serve as
good introductions to blockchain technology itself [/four.lf/one.lf, /three.lf/four.lf]. Many of Vitalik Buterin’s blog
posts also provide unusually clearheaded descriptions of di/f\_f.ligaerent aspects of blockchain
ecosystem [/three.lf/eight.lf, /three.lf/nine.lf].
Zero-knowledge proofs and secure multiparty computation receive coverage in many in-
termediate cryptography textbooks. There are a handful of book-length treatments. For an
introduction to zero-knowledge proofs that focuses on recent developments, I recommend
the “Explaining SNARKs” series on the Zcash blog [/one.lf/eight.lf/six.lf]. For an accessible introduction to
/six.lf
secure multiparty computation, I recommend the paper “Secure Multiparty Computation
for Privacy-Preserving Data Mining” by Lindell and Pinkas [/one.lf/zero.lf/eight.lf]. In addition, Vaikun-
tanathan’s “Computing Blindfolded: New Developments in Fully Homomorphic Encryp-
tion” [/one.lf/seven.lf/one.lf] and Boneh et al.’s “Functional Encryption: De/fi.liganitions and Challenges” [/two.lf/seven.lf] are
reasonable introductions to homomorphic encryption and functional encryption respec-
tively.
Finally, it is important to note that the list of technologies I have chosen to investigate
is not exhaustive. Among areas I have excluded, the most important may be the sub/fi.ligaeld
ofquantum cryptography , which applies quantum phenomena to cryptographic tasks. This
exclusion is primarily a matter of limited space, although it is also worth noting that many
of the most interesting technologies associated with quantum cryptography (such as quan-
tum money andquantum copy-protection ) stand out as particularly far from seeing practical
applications [/one.lf]. Other potentially signi/fi.ligacant technologies, not discussed in this report, in-
clude program obfuscation andverifiable computing. Program obfuscation allows users to share
computer programs with others while leaving their inner workings opaque, and veri/fi.ligaable
computing allows users to outsource computations to others and receive short proofs that
the computations have been executed as promised [/one.lf/three.lf/nine.lf, /one.lf/seven.lf/eight.lf]. Both of these technologies
are also associated with a number of breakthroughs in the past decade.
/two.lf./one.lf Public-key encryption
Public-key encryption is a technology that allows users to communicate through code with-
out sharing secret information ahead of time [/five.lf/seven.lf, /nine.lf/nine.lf].
Say that one party, Alice, wants to send a private message to some other party, Bob, using a
channel that may have eavesdroppers. For example, Alice might want to share a secret with
Bob over e-mail without anyone else—such as a government intelligence agency—being
able to learn the secret too. The way to do this is to encrypt the message, meaning to
encode it in a way that no one else can understand.
The oldest class of encryption schemes, known as symmetric key schemes, has been used for
thousands of years. These schemes rely on a single shared piece of information, known as a
secret key , and a mutually understood rule for translating unencrypted messages (or plain-
text) and encrypted messages (or ciphertext ) into one another using that key. For example,
in the simple “Caesar cipher” the key was a short number, X, and the rule for translat-
ing plaintext to ciphertext was to move each individual letter forward by Xplaces in the
alphabet./three.lf
The trouble with symmetric key schemes is that, to be used, both parties must somehow
settle on a secret key without any third parties learning it too. However, the di/f\_f\_i.ligaculty of
communicating secret information such as this is exactly the di/f\_f\_i.ligaculty that encryption is
meant to solve in the /fi.ligarst place. Private key cryptography schemes therefore su/f\_f.ligaer from
a “chicken and egg” problem. The problem is further exacerbated by the fact that secret
keys cannot be reused without a very signi/fi.ligacant loss of security.
/three.lfAs an example, the key “/one.lf” would turn the message “HELLO” into “IFMMP.”
/seven.lf
Technology Origin Functional description
Public-key encryption /one.lf/nine.lf/seven.lf/three.lf Allows users to communicate through code without
sharing secret information ahead of time
Digital signatures /one.lf/nine.lf/seven.lf/nine.lf Allows users to identify messages’ senders
Cryptographic hash functions /one.lf/nine.lf/seven.lf/nine.lf Allow users to associate data with unique “digital
/fi.ligangerprints”
Trusted timestamping /one.lf/nine.lf/nine.lf/one.lf Allows users to timestamp pieces of data
Tamper-evident logs /one.lf/nine.lf/seven.lf/nine.lf (ambiguous) Contain chronological records that cannot be
inconspicuously edited
Replicated state machines /one.lf/nine.lf/eight.lf/four.lf Replicate the provision of a service across multiple
computers
Blockchains /two.lf/zero.lf/zero.lf/eight.lf Replicated state machines that maintain
tamper-evident logs
Permissionless blockchains /two.lf/zero.lf/zero.lf/eight.lf Blockchains that allow any computer to participate in
service provision
Decentralized applications /one.lf/nine.lf/eight.lf/zero.lf (ambiguous) Applications associated with an open-ended set of
service providers (as in a permissionless blockchain)
Consortium blockchains /two.lf/zero.lf/one.lf/two.lf (ambiguous) Blockchains that allow computers owned by multiple
parties (but not any computer) to participate in service
provision
Consortium-backed applications Ambiguous Applications associated with multiple privileged
service providers (as in a consortium blockchain)
Cryptocurrencies /two.lf/zero.lf/zero.lf/eight.lf Digital currencies whose ownership is managed through
a decentralized or consortium-backed application; are
also essential to the operation of permissionless
blockchains
Table /one.lf: Summary of cryptographic technologies, part /one.lf
/eight.lf
Technology Origin Functional description
Zero-knowledge proofs /one.lf/nine.lf/eight.lf/five.lf Allow users to prove mathematical statements to others
without conveying additional information
zk-SNARKs /two.lf/zero.lf/one.lf/zero.lf Allow users to do the above succinctly and without
back-and-forth interactions
Physical zero-knowledge proofs /two.lf/zero.lf/one.lf/two.lf (ambiguous) Allow users to prove statements about physical objects
to others without conveying additional information
Smart property /one.lf/nine.lf/nine.lf/four.lf Devices with electronic components that facilitate that
transfer of their ownership
Smart contracts /one.lf/nine.lf/nine.lf/four.lf Contracts whose execution is automated to a
signi/fi.ligacant extent
Homomorphic encryption /one.lf/nine.lf/seven.lf/three.lf Allows users to perform certain computations on
encrypted data; outputs are also encrypted
Fully homomorphic encryption /two.lf/zero.lf/zero.lf/nine.lf Allows users to perform anycomputations on
encrypted data; outputs are also encrypted
Functional encryption /two.lf/zero.lf/one.lf/zero.lf Allows users to perform certain computations on
encrypted data, such that the outputs are notencrypted
Secure multiparty computation /one.lf/nine.lf/eight.lf/two.lf Allows users to run computations with inputs from
multiple parties, while allowing these parties to keep
their own inputs secret
Secret sharing /one.lf/nine.lf/seven.lf/nine.lf Allows users to split private data into shares, which can
be recombined to retrieve the data
Table /two.lf: Summary of cryptographic technologies, part /two.lf
/nine.lf
Public-key cryptography , /fi.ligarst developed in the /one.lf/nine.lf/seven.lf/zero.lfs, solves this problem. In a public-key
scheme, there is not a single key. Instead, each person in a network has a unique pair of
keys: one known as their private key and one known as their public key . Although technical
details need not concern us, these are the de/fi.liganing traits of a public-key system:
•There is a setup algorithm that each party in the system can use to generate an (almost
certainly) unique key pair.
•Each party in the system can announce their public key without revealing their pri-
vate key. Public keys may also be used as digital pseudonyms.
•There is an encryption algorithm that can take a plaintext message and the recipient’s
public key as inputs and then produce an encoded message as an output. There is also
adecryption algorithm that that can take an encoded message and the recipient’s pri-
vate key and then produce the original message as an output. By applying these two
algorithms in sequence, the sender and recipient can communicate through code./four.lf
•There is no practical algorithm that would allow anyone without the recipient’s
private key to decode the sender’s message.
As stated above, public-key cryptography is not a new technology. After its initial de-
velopment (or, more precisely, rediscovery by academics outside of the classi/fi.ligaed research
community), the possibility of its widespread adoption was for many years considered a
threat by government agencies such as, within the United States, the NSA and the FBI [/one.lf/one.lf].
These agencies argued that, without the ability to read intercepted messages, they would
be much less able to counter criminal activity and other threats to security. They pur-
sued several strategies to either slow the technology’s adoption or outlaw variants that did
not grant the government a “backdoor” to decrypt messages using its own special private
keys.
Ultimately, it became clear by the late /one.lf/nine.lf/nine.lf/zero.lfs that these agencies had lost the /fi.ligaght, and, at
least within the United States and European Union, all forms of public-key cryptography
are now perfectly legal to use [/one.lf/four.lf/zero.lf]. Until recently, though, the vast majority of messaging
services still applied encryption in a way that allowed the service provider, and therefore
government agencies, to access their users’ unencrypted messages. Partially as a reaction
to the /two.lf/zero.lf/one.lf/three.lf Snowden leaks, this state of a/f\_f.ligaairs has begun to change. It has become in-
creasingly common for services to o/f\_f.ligaer end-to-end encryption , which, in practice, refers to
implementations of encrypted messaging that do not grant service providers viewing priv-
ileges [/five.lf/nine.lf]. Over the course of just /two.lf/zero.lf/one.lf/six.lf, the number of end-to-end encryption users across
all services may have increased by over one billion, due largely to WhatsApp’s decision to
begin enabling the feature by default [/one.lf/two.lf]. In addition, a number of groups are working to
develop practical end-to-end messaging services that can also reliably obscure each mes-
sage’s metadata, such as its recipient [/one.lf/seven.lf/two.lf]. If such projects are ultimately successful, then
/four.lfTo express this mathematically, let Encbe the encryption algorithm, Decbe the decryption algorithm, u
be a user’s public key, rbe the same user’s private key, and mbe a message. Then, Enc(m,u) is illegible, and
Dec(Enc(m,u),r) =m. Alternatively, to express this by analogy, we can think of a user’s “public key” as actually
being a particular lock design, which others use to protect the messages sent to them, and the user’s “private key”
as the key that opens the lock.
/one.lf/zero.lf
user privacy may increase even further.
While even the use of perfectly implemented end-to-end encryption does not guarantee
that one’s messages will not be read by anyone other than the intended recipient, it does
decrease the odds of successful eavesdropping./five.lfAgencies in a number of countries now
allege, controversially, that the information channels they rely on are increasingly “going
dark” [/seven.lf/zero.lf].
/two.lf./two.lf Digital signatures
Adigital signature can be used to demonstrate that a given piece of data was sent by the
owner of a particular key pair (see section /two.lf./one.lf) and that it has not been modi/fi.ligaed since its
sending [/one.lf/three.lf/six.lf, /nine.lf/nine.lf].
Digital signatures work in the following way:
•Each party in the system agrees on a signing algorithm and a signature-verifying algo-
rithm .
•The signing algorithm takes a party’s private key and a piece of data and outputs a
code known as a signature .
•The signature-verifying algorithm takes a party’s public key, a piece of data, and a
signature, and outputs “Yes” if and only if the signature was generated from the data
and the corresponding private key./six.lf
The use of digital signatures is currently ubiquitous online and, among many other ap-
plications, enables online commerce. For example, when you shop online, your computer
veri/fi.ligaes that you are in fact connected to Amazon.com (and not a scammer after your credit
card details) by checking a signature it sends against a public key known to be associated
with the website.
In recent years, some countries have also moved toward assigning their citizens public keys
as a form of identi/fi.ligacation, so that individuals can prove their identities, access personally
relevant government records, vote, and even sign legally binding contracts using digital sig-
natures (ordinarily stored on highly protected ID cards) [/one.lf/one.lf/zero.lf]. Estonia is the most notable
case, with its citizens having issued hundreds of millions of signatures since the program’s
inception.
/five.lfA third party might still read the messages if they gain access to the intended recipient’s private key and
intercept the message, if they trick the sender into associating the intended recipient with their own private key,
if they manage to install malware on either the sender’s or the recipient’s device, and so on. In addition, there
remains a risk that the application developer has misrepresented the security or method of encryption used in
their application, as has sometimes occurred.
/six.lfFor example, say that my key pair consists of private key Xand public key Y. If I would like to tell you
“HELLO,” and demonstrate that I am the one telling you this, then I will /fi.ligarst input “HELLO,” and Xinto the
signing algorithm to produce a signature. I will then send you a message consisting of the word “HELLO” followed
by the signature. Finally, if you know my public key, you can apply the verifying algorithm to “HELLO,” Y, and
the signature, and thereby see that I am the one who signed the message.
/one.lf/one.lf
/two.lf./three.lf Cryptographic hash functions
Acryptographic hash function encodes a piece of data as a code of some /fi.ligaxed length, known
as ahash (ordigest ) [/one.lf/one.lf/four.lf, /nine.lf/nine.lf]. The function has the following properties:
•It is easy to check that a given piece of data produces a given hash.
•Pieces of data that are only slightly di/f\_f.ligaerent will produce very di/f\_f.ligaerent hashes.
•It is impractically di/f\_f\_i.ligacult to /fi.ligand a piece of data that will produce a given hash, or
to /fi.ligand two pieces of data that produce identical hashes.
In a sense, hashes act like “digital /fi.ligangerprints” for pieces of data. In the same way that
each human is associated with an almost certainly unique set of /fi.ligangerprints, without these
/fi.ligangerprints providing any other information about the person, each piece of data can be
associated with an almost certainly unique hash, without this hash providing any other
information about the data.
Arguably, hash functions are primarily important as a building block for other crypto-
graphic technologies. We will now discuss a trio of such technologies: trusted timestamping ,
tamper-evident logs , and blockchains.
/two.lf./four.lf Trusted timestamping
One interesting application area for cryptographic hash functions is trusted timestamping ,
or techniques for demonstrating that a given piece of data existed at a given time [/eight.lf/four.lf, /one.lf/seven.lf/zero.lf].
In many cases, the task is signi/fi.ligacantly complicated by the user’s desire to keep the data
private at the time of its timestamping.
For instance, suppose that you have some research result that you are not ready to publish,
but which you would like to be able to claim priority for. One simple solution is to take
the hash of your data and then publish that hash to a newspaper or to a website that can
be trusted to reliably log publication times. Later on, you can publish the actual research
results, and, by comparing its hash against the published hash, people will be able to verify
for themselves that you had the results at the time of the hash’s publication.
As an example of this technique, the political organization WikiLeaks will sometimes post
hashes of sensitive documents that they obtain to Twitter [/one.lf/five.lf/eight.lf]. If the hashes of eventually
released documents do not match—as has happened in at least one case—then it will be
clear that someone has modi/fi.ligaed the documents in the time since WikiLeaks advertised
their existence./seven.lf
/seven.lfHere, Twitter is unintentionally /fi.ligalling the role of Time Stamping Authority (TSA). It is being trusted, in
particular, both to produce authentic timestamps (by publishing accurately dated tweets) and to ensure that
these timestamps will remain available into the inde/fi.liganite future. More sophisticated timestamping protocols
can also remove this second responsibility. For example, rather than storing the hashes it receives from users, the
TSA can send back signed messages that contain both the hash and the time of its receipt. If the user would later
like to convince others that the data existed at this time, then all they need to do is share the signed message
along with the data. Further security can be provided by yet more sophisticated proposals, which replace a TSA
with multiple trusted parties.
/one.lf/two.lf
Other use cases for trusted timestamping include preventing the forgery of documents and
digital media. Often, the fact that a piece of data is known to predate a certain point in
time can provide evidence that it is genuine. Investigators can trust that a timestamped
government document, for example, was not concocted after the fact to cover a corrupt
o/f\_f\_i.ligacial’s tracks. Similarly, if a photograph claims to depict a secret meeting between two
o/f\_f\_i.ligacials on a particular day, it will be much more credible if it is actually timestamped
for that day. As discussed in section /three.lf./one.lf./four.lf, this latter use case may become much more
important as arti/fi.ligacial intelligence becomes increasingly e/f\_f\_i.ligacient and e/f\_f.ligaective at forging
photographs and videos [/four.lf].
As a /fi.liganal technical point, it can sometimes be useful to associate a large collection of data
with a single timestamp. Naively, one simple way to accomplish this task is to apply a hash
function to the full collection together. However, this method has a signi/fi.ligacant downside.
In particular, using this method, it is impossible to demonstrate that a single piece of data
in the collection was used to produce the timestamp without also sharing the rest of the
collection, which may be undesirable from the standpoint of e/f\_f\_i.ligaciency or privacy.
A better method of timestamping a large collection of data at once is to use Merkle trees
(see Figure /one.lf). Merkle trees are produced by hashing each piece of data individually, then
repeatedly hashing pairs of hashes to form a tree structure. The hash that stands at the top
of the resulting tree is known as the Merkle root . The Merkle root, if published, serves just
as well as a “digital /fi.ligangerprint” for the whole collection. If one would like to prove that
an individual piece of data belongs to the timestamped collection, though, then it is only
necessary to share that individual piece and a relatively small portion of the Merkle tree,
rather than it being necessary to share the other pieces of data as well.
Merkle trees were /fi.ligarst proposed in the late /one.lf/nine.lf/seven.lf/zero.lfs, by Ralph Merkle, and have been used
perhaps most often in the context of /fi.ligale hosting and sharing services that wish to assure
users that /fi.ligales have not been altered from their most recent versions.
/two.lf./five.lf Tamper-evident logs
Atamper-evident log is a chronological collection of records that is designed so that any
alterations to records, once they are added, will be easily detectable [/five.lf/five.lf]./eight.lf
Tamper-evident logs are a highly valuable technology, insofar as trustworthy record-keeping
plays a crucial role in political and economic life. One would hope, for example, that one’s
banking records, medical records, tax records, property records, and criminal records con-
tinue to be maintained properly.
In many cases, the records in a log are objects of signi/fi.ligacant interest in their own right. Logs
are often instrumentally useful, though, in allowing users to verify the integrity of more
sophisticated processes or services. For example, a trusted log of deposits, withdrawals,
and transfers can be used to determine the correct balance for a bank account or to restore
it to an earlier value if an error should occur. In general, many computer applications are
/eight.lfThe terminology here is not entirely standard. However, as a technical note, tamper-evident logs can be
classi/fi.ligaed as a particular variety of what are known as authenticated data structures .
/one.lf/three.lf
Figure /one.lf: A Merkle tree. The top hash, or Merkle root , serves as a “digital /fi.ligangerprint” for
the blocks of data at the bottom. Here, proving that Data Block L/one.lf is consistent with the
Merkle root would require sharing just the block itself and the two hashes Hash/zero.lf-/one.lf and
Hash/one.lf. There is no need to share the other data blocks. (Image by David Gäthberg.)
/one.lf/four.lf
Figure /two.lf: A hash-linked log. Any edit to an individual record will produce an inconsistency
with the Merkle root in the block’s header. Furthermore, any edit to a block’s header will
produce an inconsistency with the hash pointer in the following block.
associated with logs, typically referred to as transaction logs , that track how they respond
to user inputs over time.
Granting the parties maintaining such logs the ability to more easily detect attempts to
tamper with them, and to more easily demonstrate to third parties that their logs have not
been tampered with, can be a signi/fi.ligacant boon to security and trust.
One design stands out as particularly e/f\_f.ligaective (see Figure /two.lf). Although the proper termi-
nology is contested, I will refer to it as a hash-linked log . Here, the log is subdivided into
sequential blocks. Every time su/f\_f\_i.ligaciently many new records are collected, they are gath-
ered into a new block and added to the sequence. Each block is divided into two sections,
one that contains records and one that is known as a header . In turn, each header contains
the date, the Merkle root of the records the block contains, and the hash of the previous
header.
Suppose that Alice is maintaining such a log—say, for an online payment service that Bob
uses—and that she publishes the block headers as they are created. By including Merkle
roots in the headers, Alice timestamps each block’s records when she publishes its header.
By including hash pointers in the headers, though, she also goes further and establishes a
canonical sequence of records. It is now impractical, for instance, for Alice to remove a
block from the log without producing an inconsistency with the headers that follow it; she
also cannot simultaneously create two contradictory versions of a block, then inconspicu-
ously swap one out for the other later on./nine.lf
Generally, we can understand hash-linked logs as achieving their security through the use of
linked timestamping , a class of timestamping techniques in which each timestamp references
/nine.lfAs a further bene/fi.ligat of including hash pointers, Bob only really needs to know the most recent header to
detect tampering. Any attempt by Alice to pass o/f\_f.liga an alternative sequence of headers as legitimate would be
foiled by the sequence’s obvious inconsistency with the one that Bob knows.
/one.lf/five.lf
the one before it.
Although the linked timestamping scheme described here dates back to a series of papers
by Haber and Stornetta in the early /one.lf/nine.lf/nine.lf/zero.lfs, a number of governments and companies have
just recently taken an interest in its applications. The aim is both to protect the integrity
of their records and to provide greater assurances to their citizens or clients. The Estonian
government, for instance, began in /two.lf/zero.lf/zero.lf/eight.lf to use a linked timestamping service for logging
alterations to some varieties of government documents. The research company Google
DeepMind also initiated a project, in /two.lf/zero.lf/one.lf/seven.lf, to develop a tamper-evident logging system for
use by the British National Health Service [/nine.lf/one.lf].
/two.lf./six.lf Blockchains and distributed computing
/two.lf./six.lf./one.lf Background: Concepts in distributed computing
In this subsection, we will take a brief break from discussing cryptography directly. In-
stead, we will discuss some concepts in the /fi.ligaeld of distributed computing that will inform
the discussion that follows.
Adistributed application , /fi.ligarst, is an application whose operation requires interaction be-
tween multiple computers. One familiar example of a distributed application is Google
Search. When you use Google, information is constantly being passed back and forth
between your computer and computers managed by the company. Your computer sends
search requests, and it receives search results back in return.
Google Search is also an example of a distributed application that is well-described by the
client-server model . In the client-server model, a set of users, known as clients , request ser-
vices from a service provider, known as a server . Here, your computer would be considered
a client and Google’s would be considered servers.
Other familiar examples of distributed applications that /fi.ligat the client-server model in-
clude e-mail, Facebook, and online banking. As one can see, such applications are ubiqui-
tous.
Servers are not, in general, perfectly reliable. They may experience errors, lose the ability to
receive requests, or otherwise stop functioning properly. One might worry, in particular,
about the integrity of the application a server supports or about its consistent availability
to clients.
One approach to lessening these concerns is known as state machine replication . In state
machine replication, the operations of a single server are reproduced across multiple copies.
These copies, known as nodes , communicate with one another to stay in sync and come to
consensus about the proper responses to requests from clients. Collectively, the nodes can
be conceptualized as constituting a single replicated state machine (RSM ).
The consensus protocols for replicated state machines are designed such that, if some small
portion of nodes are faulty, then the application will continue to respond to clients and
respond appropriately. The most secure of these protocols, which were largely developed
/one.lf/six.lf
in the /eight.lf/zero.lfs and /nine.lf/zero.lfs, can practically guarantee the availability and integrity of an RSM so
long as at least two-thirds of its nodes follow the protocol and remain in contact./one.lf/zero.lf
One feature of such consensus protocols that will be worth noting is that they often depend
on each node storing a log of the requests it has received from clients. Such logs serve
as an important tool for preventing and resolving inconsistencies that arise between the
nodes. So long as they agree on the ordering of this log and follow the same procedure for
responding to sequences of requests, their behavior will be consistent.
The nodes that make up an RSM are typically controlled by a single party, like a technology
company. It is also possible, though, to create RSMs whose nodes are distributed among
multiple parties. I will call such RSMs consortium replicated state machines ./one.lf/one.lfAt least in the-
ory, this strategy can be useful in contexts where any single actor entrusted with providing
a service would su/f\_f.ligaer from a worrisome con/fl.ligaict of interest. The banking industry, for ex-
ample, has recently begun to investigate using RSMs maintained between multiple /fi.ligarms
to run applications that reconcile /fi.liganancial records. Applications supported by consortium
RSMs can be described as consortium-backed applications .
A further step in the direction of decentralization is taken by permissionless replicated state
machines , which place no restrictions at all on which computers can act as nodes. Simply
put, anyone anywhere in the world can participate in maintaining a permissionless RSM.
Applications run using permissionless RSMs can be described as decentralized applications ,
a class of applications in which the set of service providers is open-ended and not centrally
organized./one.lf/two.lf
Intuitively, it is deeply surprising that any permissionless RSM could ever work. Given
that standard RSM consensus protocols fail if even a third of the nodes behave maliciously,
letting in any node that asks to join has the sound of a disastrous idea. Nevertheless, reliable
permissionless RSMs are in fact possible.
Bitcoin, proposed in /two.lf/zero.lf/zero.lf/eight.lf by the pseudonymous engineer Satoshi Nakamoto, provided
the /fi.ligarst-ever example of a permissionless RSM. Designed to implement a decentralized
payment application, allowing users to make and receive payments in a digital currency of
the same name, Bitcoin has proven itself remarkably successful (see section /two.lf./seven.lf). It has de-
veloped millions of users and never itself experienced a signi/fi.ligacant security breach, despite
now tracking the ownership of billions of dollars’ worth of currency. It has inspired both
/one.lf/zero.lfIn cases where the only concern is that some modes might lose contact, then having even half the nodes
function properly is su/f\_f\_i.ligacient.
/one.lf/one.lfThis term, along with the terms “permissionless replicated state machine,” “consortium-backed application,”
and “decentralized application” below, is not standard. However, to my knowledge, there does not yet exist any
strongly established terminology here.
/one.lf/two.lfThere are also some simple applications in this class that do not require service providers to act as an RSM
because it is unnecessary for service providers to synchronize their responses to client requests. For example, in
the decentralized /fi.ligale sharing application BitTorrent, which launched in /two.lf/zero.lf/zero.lf/one.lf, a wide variety of service providers
store pieces of /fi.ligales and individually send these pieces to clients upon request. However, most reasonably com-
plex applications do require the synchronization associated with RSMs. A payment application, for example,
would open itself up to the possibility of users “double-spending” their money if nodes did not agree about what
payments have already been made.
/one.lf/seven.lf
Figure /three.lf: A taxonomy of replicated state machines.
further decentralized payment applications, many of which have been similarly successful,
and a number of decentralized applications, such as those associated with the Ethereum
blockchain (see section /two.lf./one.lf/zero.lf), that aim to provide more sophisticated services.
Bitcoin was created with a seeming lack of awareness of previous work on RSMs./one.lf/three.lfDe-
spite this fact, Bitcoin’s design constitutes an enormous contribution to the /fi.ligaeld. Most
obviously, it provided the /fi.ligarst successful consensus protocol for permissionless RSMs (see
section /two.lf./six.lf./three.lf). However, as the next subsection will discuss, it also played an important
role in integrating the use of RSMs and the use of tamper-evident logs.
/two.lf./six.lf./two.lf Blockchains
The term “blockchain” lacks a standard de/fi.liganition and has been used in several distinct
ways.
By some de/fi.liganitions, blockchains are simply hash-linked logs of the sort depicted in Figure
/two.lf. By other de/fi.liganitions, blockchains appear to be roughly synonymous with replicated state
machines. Perhaps most commonly, the term is also used to refer to a rough constellation
of technologies that de/fi.ligaes easy summary.
In order to draw a clean line between blockchains and technologies that substantially pre-
date the term “blockchain,” however, I will use the following de/fi.liganition:
Ablockchain is a replicated state machine that maintains a hash-linked log.
Blockchains combine the security properties of these two already discussed technologies.
/one.lf/three.lfThe whitepaper proposing Bitcoin does not reference any papers in the literature, and, in general, the term
“replicated state machine” has only recently begun to be used by developers in the associated community.
/one.lf/eight.lf
The use of an RSM helps to ensure the availability of the log and integrity of the process by
which new records are added to it. The use of a hash-chained log then helps to ensure the
integrity of records that have already been added, by ensuring that any attempt to tamper
with them will be easy to detect.
As discussed in sections /two.lf./five.lf, logs are often objects of interest in their own right. However,
in the context of RSMs, logs of clients’ requests are typically a tool for implementing and
auditing more sophisticated services. In Bitcoin, for example, the complete log of transac-
tions in the currency is used to determine the correct amount in a user’s account.
The Bitcoin blockchain, which came online in /two.lf/zero.lf/zero.lf/nine.lf, is perhaps less notable for being the
/fi.ligarst blockchain than it is for being the /fi.ligarst permissionless RSM. The consensus protocol
it implements and the variants it has inspired are general enough—at least in principle—to
allow any application to be run as a decentralized application. The fact these early proto-
cols relied on tamper-evident logs was, in some sense, of secondary interest.
However, since /two.lf/zero.lf/zero.lf/nine.lf, the idea of associating RSMs with hash-linked logs has also received
broader interest. In the past few years, companies like IBM have published designs for
permissioned blockchains that lightly adapt existing RSM algorithms to provide gains in
both integrity and e/f\_f\_i.ligaciency. A number of companies and government bodies who have
not previously used RSMs or tamper-evident logs have begun to explore the possibility of
shifting some of their existing services to blockchains. In addition, perhaps even more ac-
tors have moved to adopt or expand their use of tamper-evident logs in general. A linked
timestamping service provided by the company Guardtime, for example, has begun to be
applied to a large portion of Estonian governmental records. In media accounts and pro-
motional materials, such services are typically described as implementing “blockchains.”
However, in the context of this report’s less expansive terminology, this labeling would be
incorrect.
Starting in approximately /two.lf/zero.lf/one.lf/one.lf, private sector interest in blockchains spilled over into pro-
posals for implementing consortium blockchains. As stated above, consortium blockchains
have attracted particular interest in the /fi.liganancial sector, and several major banks have now
publicly explored the possibility of using consortium blockchains to store shared /fi.liganancial
records. One signi/fi.ligacant motivation is the billions of dollars they spend each year in resolv-
ing discrepancies between records that they maintain independently [/nine.lf/six.lf]. Another smaller
early adopter has been the diamond industry, in which several companies have turned to
blockchains to better track the movements of individual diamonds as they change hands
[/one.lf/seven.lf/five.lf]. Facebook has also recently designed a digital currency system (see section /two.lf./seven.lf) that
will be run by a consortium of technology companies, /fi.liganancial services companies, and
nonpro/fi.ligats. This digital currency system, Diem, is arguably the most high-pro/fi.ligale applica-
tion of consortium blockchains yet.
For something of an overview of proposed applications for permissioned blockchains, of
both the consortium and fully private varieties, one good resource is the United Kingdom
Government O/f\_f\_i.ligace for Science’s report, “Distributed Ledger Technology: beyond block
chain” [/one.lf/seven.lf/nine.lf]. Overall, the report expresses a high level of enthusiasm about the economic
/one.lf/nine.lf
Advantages of blockchains Disadvantages of blockchains
•May increase the availability and
integrity of applications
•May reduce reliance on trusted
third parties (less applicable to
private blockchains)
•May enable applications that, due
to a lack of trust in individual
services providers, it would
otherwise be unreasonable to use
(less applicable to private
blockchains)•Require greater computing
power, storage, and
communication
•May reduce the con/fi.ligadentiality of
records, unless sophisticated
cryptographic technologies are
applied
Table /three.lf: Some advantages and disadvantages of blockchains, in comparison with typical
centralized servers
implications of permissioned blockchains, describing them as the /fi.ligarst signi/fi.ligacant innova-
tion in ledger technology since ancient times.
Since its creation, Bitcoin has inspired the creation of many more permissionless RSMs.
Most of these implement fairly similar decentralized payment applications, but others
implement at least early versions of more sophisticated decentralized applications. There
are a number of applications, for instance, that aim to create decentralized markets that
allow users to buy storage space on one another’s computers. Perhaps most sophisticated
blockchain to follow Bitcoin so far, though, is Ethereum, which allows users to design and
implement arbitrary decentralized applications (see section /two.lf./one.lf/zero.lf).
Before moving on to describe the consensus protocols that allow permissionless block-
chains to work, let us attempt to brie/fl.ligay summarize the core features of blockchains. Com-
pared to the use of centralized servers to provide services, blockchains o/f\_f.ligaer a number of
advantages and disadvantages (see Table /three.lf).
As a /fi.ligarst advantage, blockchains can sometimes provide greater assurances of integrity
and availability. Whereas a single server might go o/f\_f\_l.ligaine, fall victim to a cyberattack,
or otherwise exhibit faulty behavior, it will typically be less likely that a large portion of
the nodes making up a permissioned blockchain will be faulty in the same way. Although
the degree of security associated with permissionless blockchains is still ambiguous (see
section /four.lf./two.lf./three.lf), in some cases they may also provide greater protection. In addition, if the
integrity of a blockchain’s log is compromised, then this will be much easier to detect than
a compromise to the integrity of a log that is not tamper-evident.
As a second advantage, blockchains can sometimes be used to reduce reliance on trusted
third parties to provide services. The users of consortium and permissionless blockchains
/two.lf/zero.lf
rely on large networks of actors, rather than one particular actor. For instance, whereas a
credit card user must rely on company such as Visa to process their transactions, Bitcoin
users do not rely on any single actor. In cases where trusted service providers hold an
undesirable degree of power, at least from certain users’ perspectives, this advantage can
be signi/fi.ligacant. A number of early adopters of Bitcoin, for example, saw it as a way to get
around restrictions credit card companies had placed on payments to WikiLeaks.
As a third advantage, blockchains can sometimes be used to provide services that it would
otherwise be unreasonable to trust a single actor to provide. Banks, for example, can use
consortium blockchains to store and automatically resolve discrepancies in their records
without needing to grant control over the records to any individual bank or third party.
As an important disadvantage, however, blockchains are much less e/f\_f\_i.ligacient than tradi-
tional centralized servers. Permissionless blockchains, in particular, are typically able to
process no more than a couple of dozen requests per second, and they typically consume
very large quantities of computing power, storage space, and bandwidth [/five.lf/four.lf, /six.lf/zero.lf]. These
limitations are due in large part to the fact that each individual node must store a complete
copy of the log and verify each new request. In general, the whole network can process re-
quests no faster than the least powerful device capable of serving as a node. One upshot,
then, is that any reasonably sophisticated or heavily used application cannot in practice be
run as a decentralized application. While decentralized payment applications are feasible,
even something so simple as a decentralized two-player chess program sits about at the
limits of what is currently possible. Decentralized versions of applications like Uber or
Facebook of course stand much further away still.
Developers are currently researching a number of potential ways to reduce these limi-
tations. Some permissionless blockchain projects that are still very new, at the time of
writing, also claim to have achieved substantial speedups without sacri/fi.ligacing security. In
general, though, it is an open question whether permissionless blockchain technology can
“scale” to accommodate large numbers of active users or applications that require very
frequently updated records (see section /four.lf./two.lf./one.lf).
As a second disadvantage, blockchains are naturally ill-suited for maintaining con/fi.ligadential
information. In the case of a permissionless blockchain, everyone in the world is granted
complete access to all of the records. Even in the case of a consortium blockchain, the
number of parties with complete access may still be greater than desired. Encrypting sen-
sitive records is only a limited solution, since the parties maintaining the blockchain must
still have access to enough information to determine whether a given record is valid. Ulti-
mately, those hoping to maintain con/fi.ligadential information on a blockchain may be forced
to rely on more sophisticated cryptographic technologies, such as “zero-knowledge proofs”
(see section /two.lf./eight.lf). The use of these technologies can further add to the cost and complexity
of record-keeping.
The next subsection will explore the consensus that enables permissionless blockchains in
greater detail. This next subsection is less essential to read than preceding sections, since
it focuses on how blockchains work rather than focusing on what blockchains do. These
/two.lf/one.lf
details are still quite helpful, however, for understanding e/f\_f.ligaorts to overcome the current
limitations of blockchain technology (see section /four.lf./two.lf).
/two.lf./six.lf./three.lf Consensus protocols
Although consensus protocols for permissioned blockchains remain an active research area,
practical protocols in this category have been known for decades. As mentioned in sub-
section /two.lf./six.lf./one.lf, these protocols can assure that permissioned blockchains will continue to
function as expected so long as at least two-thirds of their nodes behave properly.
The case of permissionless blockchains is more di/f\_f\_i.ligacult. If one naively attempts to adapt
these older protocols, then two major problems will stand out.
Sybil attacks: A dishonest party, if su/f\_f\_i.ligaciently motivated, might carry out
what is known as a Sybil attack by setting up many di/f\_f.ligaerent “sock puppet”
nodes in order to arti/fi.ligacially increase their in/fl.ligauence on block creation.
Lack of incentives for honesty: Since the parties involved in running nodes
are not pre-selected, there may be no basis for expecting them to vote hon-
estly.
In response, the consensus protocols for permissionless blockchains provide two corre-
sponding solutions.
Tying voting power to scarce resources: Voting power is made proportional
to demonstrated ownership of scarce resources. For example, it is possible
to make voting power proportional to computing power by requiring voters
to provide solutions to computationally intensive puzzles. Then, no one can
in/fl.ligaate their in/fl.ligauence beyond the amount of computing power they possess.
Rewarding honesty with digital currency: Voters are incentivized to vote
honestly by rewarding them with a quantity of digital currency, to be recorded
in the blockchain, if they vote for blocks that the network ultimately con-
verges on (or costing them digital currency if they do not). By this mecha-
nism, a state of a/f\_f.ligaairs in which each node expects the network to converge
on honest blocks is likely to be stable.
If this description is still overly mysterious, then the following bullet points provide more
detail, describing the working of a somewhat typical permissionless blockchain, along the
lines of Bitcoin [/one.lf/one.lf/nine.lf]. Note, though, that this description is not meant to describe how all
permissionless blockchains work and glosses over a handful of subtleties.
•There is some network that anyone is free to join, along with some software asso-
ciated with the blockchain that anyone is free to run. The software implements a
protocol that describes under what conditions a request sent through the network
is valid. Users who maintain a full copy of the log and run the software are said to
be running full nodes .
/two.lf/two.lf
•The blockchain supports a payment application, which tracks the ownership of a
digital currency known as a “cryptocurrency.” Among the requests that the block-
chain can process are requests to send currency from one account to another.
•When users submit requests, they must pay cryptocurrency fees. These fees are ex-
tracted from their accounts when requests are processed.
•Each person running a full node maintains a backlog of the requests they have re-
ceived since they last added a block to their copy of the log. Someone running a full
node is said to be honest if they are indeed using a copy of the correct software and
are, therefore, only processing requests if they are valid. Honest users also do not
send contradictory requests to other members of the network.
•Some subset of users running full nodes also compete to solve brute force computa-
tional puzzles generated on the basis of the previous block’s contents. These users are
known as miners . The odds that a given miner will solve a puzzle /fi.ligarst is proportional
to the quantity of computing power they direct at the puzzle. When a miner does
solve a puzzle /fi.ligarst, they turn their backlog into a new block, B, which they add to
their current version of the blockchain. They are also permitted to write and process
a special request to deposit a cryptocurrency reward into their own account. Some
or all of this reward is derived from the fees paid by the users whose requests are
included in the new block. These fees incentive miners not just to solve puzzles, but
also to process all of the valid requests they receive. Part of the miner’s reward may
also include some new quantity of cryptocurrency that was not previously owned
by anyone.
•The miner who solves a puzzle /fi.ligarst advertises their new block, B, to the others in
the network. The other full nodes either add Bon to their individual copies of the
blockchain, if it is consistent with the earlier blocks in their copies, or ignores it, if
it is not consistent.
•If, after at least nintervals, Bis part of the longest consistent version of the block-
chain being proposed, then Band the blocks preceding it are generally acknowledged
to be confirmed . This is partly a social phenomenon, in the sense that it concerns the
manner in which people assign the contents of a blockchain practical signi/fi.ligacance.
•If it is the case that, a su/f\_f\_i.ligaciently large portion of the time, the person who gains
the ability to create a new block is honest, then if nis su/f\_f\_i.ligaciently large, it follows
that it is overwhelmingly likely that only blocks containing valid records will be
con/fi.ligarmed and that they will remain con/fi.ligarmed into the future./one.lf/four.lfDue to the fact
that the rewards that successful miners grant themselves will only be con/fi.ligarmed if
the blocks they propose are also con/fi.ligarmed, it also follows that miners have strong
/fi.liganancial incentives to be honest.
It is worth noting that, in protocols of this form, there is not really any decisive moment
in which a “vote” occurs. Rather, the nodes simply take turns proposing blocks, and, over
/one.lf/four.lfIntuitively, it seems as though the “su/f\_f\_i.ligaciently high proportion of the time” ought to include anything above
/five.lf/zero.lf%. In fact, for somewhat technical reasons having to do with the potential to cause confusion by reporting
con/fl.ligaicting records to di/f\_f.ligaerent members of the network, the bar may be as high as /seven.lf/five.lf% [/six.lf/four.lf].
/two.lf/three.lf
time, it becomes increasingly clear whether the other nodes are predominantly building
on top of a given block or ignoring it. Eventually, if a block is buried deep enough in the
longest proposed version of the blockchain, then its contents are taken to represent part
of the “con/fi.ligarmed” history of records./one.lf/five.lf
The particular protocol sketched above ties the frequency with which a node can propose
blocks to the computing power it applies to puzzles. But this is only one of two main
approaches to basing in/fl.ligauence on scarce resource ownership [/two.lf/zero.lf]:
•The selection of block-creating nodes on the basis of computing power, as described
in the above protocol, is known as proof of work (PoW) . Users known as “miners”
demonstrate their ownership of computing power by competing to solve computa-
tionally di/f\_f\_i.ligacult puzzles that are generated on the basis of the previous block’s hash.
Whoever solves the problem /fi.ligarst gains the ability to create a new block, and the
probability that a user solves the problem /fi.ligarst will be proportional to the amount
of computing power they direct at it.
•The selection of block-creating users on the basis of digital currency ownership
is known as proof of stake (PoS) . In a PoS system, the probability that a user is se-
lected depends on the amount of cryptocurrency they have deposited (relative to
the amount of cryptocurrency deposited by other users).
Proof of work is currently widely used, including in Bitcoin. Proof of stake is less commonly
used, but has attracted growing interest, in part because it avoids the need to devote huge
amounts of electricity to solving mining puzzles. Since some estimates have placed the
total electricity consumption of Bitcoin miners as roughly on par with that of the entire
country of the Netherlands, and since the costs borne by miners can lead to very high
transaction fees for users submitting records, decreasing electricity consumption is a high
priority [/one.lf/three.lf/eight.lf]./one.lf/six.lf
One particularly valuable feature of both PoW and PoS is that, in practice, they create ad-
ditional /fi.liganancial incentives for the most in/fl.ligauential users to be honest. This is because, to
obtain a signi/fi.ligacant level of in/fl.ligauence, these users must have invested heavily in specialized
hardware or in the relevant digital currency. If they were to undermine trust in the rel-
evant blockchain through dishonest block proposals, then the value of their investments
could evaporate.
In addition, PoW schemes have the added bene/fi.ligat of associating concrete physical costs
with the generation of blocks. The further back a record is in the longest version of the
blockchain, the more electricity would need to be expended to go back and make an equally
long version where that record is absent or replaced.
/one.lf/five.lfIn fairly rare cases, more than one proposed version of a blockchain can take on social signi/fi.ligacance. Such an
event is known as a forkof the blockchain. Forks normally occur when some subset of users decide they would like
to adopt a new set of standards for adding new records or forming blocks, while retaining the previous records.
In the long run, forks also imply the emergence of two independent cryptocurrencies.
/one.lf/six.lfAnother bene/fi.ligat of proof-of-stake systems is that they may allow transactions to be con/fi.ligarmed more quickly.
There is no need to wait for a sequence of mining puzzles to be solved.
/two.lf/four.lf
/two.lf./seven.lf Cryptocurrency
Adigital currency , generally, is a form of currency that consists of balances recorded in
electronic databases. PayPal credit and video game money serve as two now-mundane
examples.
Although de/fi.liganitions vary, we will de/fi.ligane cryptocurrencies as digital currencies whose own-
ership is tracked and transferred using decentralized or consortium-backed applications.
Bitcoin is the most prominent example. In most cases, including the case of bitcoin,
the relevant application is decentralized and relies on the use of a permissionless block-
chain.
Cryptocurrencies are objects of interest in their own right. However, as discussed above,
cryptocurrencies also play important roles in consensus protocols for permissionless block-
chains. They both enable permissionless blockchains and, in many cases, are themselves
enabled by permissionless blockchains.
A simpli/fi.ligaed cryptocurrency system, similar to the one associated with Bitcoin, might work
in the following way [/one.lf/one.lf/eight.lf, /one.lf/one.lf/nine.lf].
•A cryptocurrency is associated with a permissionless blockchain.
•At any given time, the currency is divided up into discrete units known as coins,
which are understood to be owned by individual users of the currency. Users possess
pairs of cryptographic keys, with their public keys serving as pseudonyms./one.lf/seven.lf
•A coin of value Vis minted (or “mined”) if a record of form “A new coin of value
Vis granted to X” is added to the blockchain’s log, where Xis the public key of
whichever user will own the coin. Users can only mint coins if they have just created
a new block, as part of the reward for their work.
•Say that a user with public key Xwould like to give a coin of value Vto a user with
public key Y. To do this, they can simply submit a record of form “ Xgives Ya coin
of value V” to the blockchain, along with their digital signature. The record will
be classi/fi.ligaed as valid, and therefore be added to the log, so long as these conditions
obtain: there is a previous record granting Xthe coin, there is no subsequent record
showing that Xgave the coin away already, and X’s signature is correct./one.lf/eight.lf
Variations on this system might associate a certain public key with a “central bank” that
can mint new coins, or might allow new coins to be created on the basis of a complex
voting process. In principle, there is a great deal of /fl.ligaexibility in how a cryptocurrency
system might be designed. However, in practice, most systems are not very di/f\_f.ligaerent than
the simple idealized system just sketched.
/one.lf/seven.lfAs a technical note, it is more common in practice to produce pseudonyms by taking the hashes of public
keys (known as addresses ). This is because public keys can be quite long.
/one.lf/eight.lfThe step of verifying that a coin has not already been spent is what implies the need for a complete log of
previous transactions. While a traditional digital currency system solves the double-spending problem by relying on
a trusted third party to maintain the log, a cryptocurrency system achieves a greater degree of decentralization
by using a permissionless blockchain instead.
/two.lf/five.lf
A core appeal of cryptocurrencies is that they can function as a sort of borderless, digital
cash. Assuming the system is not set up with some more complicated set of rules restrict-
ing valid transactions, any user can send cryptocurrency to any other user anywhere else in
the world, without needing to go through institutions like banks or credit card companies.
This makes cryptocurrencies perhaps most obviously appealing to people who lack access
to such institutions (often referred to as the “unbanked”) or who are distrustful of them
[/one.lf/seven.lf/three.lf, /nine.lf/seven.lf]. The absence of /fi.liganancial intermediaries in cryptocurrency transactions also, in
many cases, allows for unusually low transaction fees. For this reason, cryptocurrencies
are sometimes used as a low-cost alternative to overseas transfers and local currency ex-
changes./one.lf/nine.lfIn addition, cryptocurrencies may allow users to get around political restric-
tions on the use of traditional payment services. Cryptocurrencies have been used, for
example, to make payments in online black markets like the “Silk Road” and to sustain
WikiLeaks after major credit card companies blocked payments to it [/one.lf/one.lf/one.lf].
As stated above, the /fi.ligarst successful cryptocurrency was bitcoin, which was launched /two.lf/zero.lf/zero.lf/nine.lf
and is maintained through a proof-of-work protocol. In the bitcoin system, as in the sim-
pli/fi.ligaed system described above, additional quantities of the currency come into existence
whenever users create blocks, and anyone is free to send or receive currency using an in-
de/fi.liganite number of public-key pseudonyms. The initial user base for bitcoin had a strongly
libertarian or self-identi/fi.ligaed “crypto-anarchist” bent, drawn largely by an interest in un-
dercutting /fi.liganancial institutions [/one.lf/three.lf/one.lf]. However, over time, the currency has slowly crept
further into mainstream use, although mostly as an investment rather than as a means of
exchange. At the time of this writing, several hundred billion dollars’ worth of bitcoin
exists.
Many other cryptocurrencies have also been created since /two.lf/zero.lf/zero.lf/nine.lf, although at the moment
only a few have market caps or users bases of comparable size. Some of these cryptocur-
rencies are associated with blockchains that o/f\_f.ligaer just slight variations on Bitcoin’s design.
Others, such as Ether (discussed in section /two.lf./one.lf/zero.lf), are associated with blockchains that aim
to provide services beyond payment processing; in these cases, the cryptocurrency’s role as
a necessary design element is perhaps more important than its role as a store of value or
medium of exchange.
One interesting question, which may seem in need of answering, is the question of how
cryptocurrencies come to be accepted as having monetary value. For most cryptocurren-
cies, the simple answer is that, like traditional /fi.ligaat currencies, they are accepted as having
monetary value because some initial portion of people accept them as having monetary
value. For example, bitcoin has value because some businesses are willing to accept it as
payment and because some currency exchanges are willing to exchange it for more estab-
lished currencies such as US dollars.
For other cryptocurrencies, however, the answer is slightly more complicated. The value
of astablecoin is linked to the value of some other asset, such as the US dollar or gold. The
most straightforward way for the developers of a cryptocurrency to establish this link is
/one.lf/nine.lfHowever, anyone converting a cryptocurrency payment back into their local currency will still need to pay
some non-negligible fee.
/two.lf/six.lf
to set up an organization that pledges to exchange units of the cryptocurrency for units
of the asset at some /fi.ligaxed rate. For instance, an organization that supports a gold-backed
cryptocurrency might buy up a large quantity of gold, then pledge to accept any o/f\_f.ligaers to
exchange Ncoins for Npounds of gold. Stablecoins whose stability is achieved through
this method are known as asset-backed cryptocurrencies ./two.lf/zero.lf
The key advantage of stablecoins, unsurprisingly, is that the purchasing power of indi-
vidual coins will be stable so long as the value of the linked asset is stable. As a point of
comparison, it is common for the purchasing power of an individual bitcoin to rise or fall
by more than a factor of two within an individual year. These sorts of radical swings have
led most people to treat bitcoin primarily as a risky investment, which has some chance of
skyrocketing in value if held for long enough, rather than treating it as a regular currency
that can be used to make casual purchases./two.lf/one.lf
The main downside of asset-backed stablecoins, at least from the perspective of more
libertarian-minded cryptocurrency supporters, is that they miss much of the original “point”
of cryptocurrencies. Asset-backed cryptocurrencies are not as decentralized as traditional
cryptocurrencies, such as bitcoin, because they rely on a backing organization with sub-
stantial power over the coin’s value. The behavior of this backing organization is likely to
be highly regulated by traditional /fi.liganancial institutions, and the asset it uses to back the
cryptocurrency is likely to be deeply intertwined with traditional /fi.liganancial institutions as
well.
Nonetheless, due to the use of blockchain technology, even an asset-backed coin may come
with an unusually strong guarantee that no one can easily block transactions or alter ac-
count balances. Diem, a high-pro/fi.ligale asset-backed cryptocurrency project initiated by
Facebook, illustrates the advantage of this feature [/one.lf/zero.lf/seven.lf]. Due to low public trust in many
countries, Facebook would probably have a rather di/f\_f\_i.ligacult time if it attempted to push for
the widespread adoption of a digital currency that it managed using its own servers. The
use of a blockchain, in this case a consortium blockchain, reduces the need for the users of
Diem to place any trust in Facebook itself.
Although there is much more that can be said about cryptocurrency systems, I will close
this section by summarizing some of their disadvantages when compared to more cen-
/two.lf/zero.lfThere are also other, less popular methods of tying the value of a cryptocurrency to the value of another asset.
For example, if a cryptocurrency system grants some particular central-bank-like user the authority to mint or
destroy coins, then they can use their powers to target a particular exchange rate. The value of individual coins,
relative to the external asset, will tend to grow when existing coins are destroyed and shrink when new coins
are minted. This process of minting and destroying coins, in order to cancel out any temporary shifts in the
exchange rate, can also be automated through the use of some algorithm that is implemented by the blockchain
itself. However, even in this case, there will still be a need for some trusted user or network of users to be given
special authorities: someone needs to be trusted to input accurate information about the current exchange rate.
The Ampleforth project, for instance, uses an “oracle system made up of whitelisted independent data providers”
to input exchange rate data into its blockchain [/one.lf/zero.lf/four.lf].
/two.lf/one.lfThe vast majority of companies still do not accept cryptocurrency payments. Nonetheless, the number of
companies that do accept cryptocurrency payments is growing. Tesla, for instance, has given customers the option
of using bitcoin to buy cars. PayPal has also announced plans to support cryptocurrency use, albeit not “direct”
use, by making it easy to convert cryptocurrency back into /fi.ligaat currency at the point of purchase.
/two.lf/seven.lf
Advantages of cryptocurrencies Disadvantages of cryptocurrencies
•Lower institutional barriers to
use
•Typically lower transaction fees
•Reduced reliance on and need to
trust traditional /fi.liganancial
institutions
•More con/fi.ligadentiality, if further
cryptographic techniques used•Typically slower transaction
processing speeds (and maximum
transaction volumes)
•Encourage hoarding behavior, if
not tied to the value of another
asset
•Place burden of security more
strongly on end-user
•Less con/fi.ligadentiality, if no further
cryptographic techniques used
Table /four.lf: Advantages and disadvantages of cryptocurrencies, in comparison with central-
ized payment applications
tralized payment services like those o/f\_f.ligaered by credit card companies (see Table /four.lf). These
comments will focus on decentralized cryptocurrency systems, rather than cryptocurrency
systems supported by permissioned blockchains, both because their disadvantages are es-
pecially large and because they represent an especially large break from familiar payment
systems.
First, particularly for cryptocurrencies built on permissionless blockchains, commonly
used distributed consensus protocols can create inconvenient delays between a user propos-
ing a payment and the network con/fi.ligarming it (see section /two.lf./six.lf./three.lf). Second, since most exist-
ing permissionless blockchains can process no more than a couple of dozen transactions
per second (see sections /two.lf./six.lf./two.lf and /four.lf./two.lf./one.lf), most decentralized cryptocurrencies currently
face fairly hard ceilings on how much payment activity they can support. The result of
these ceilings can be to drive up transaction fees if too many users would like to make
transactions at once./two.lf/two.lfThird, cryptocurrencies which are not asset-backed generally have
highly volatile exchange rates and are de/fl.ligaationary, thereby incentivizing hoarding be-
havior rather than spending behavior./two.lf/three.lfFourth, cryptocurrency users must typically take
/two.lf/two.lfIn the case of Bitcoin, for example, transactions take about an hour on average to be con/fi.ligarmed. Greater
transaction volume has also led the average transaction fee to grow to several dollars, or even dozens of dollars
in boom periods, although a typical fee used to be only a few cents. Fortunately, a number of second-wave
cryptocurrencies (such as Litecoin) process transactions several times faster and, due to their faster processing
speeds and smaller userbases, currently require smaller transaction fees. At the time of writing, a number of
newly developed permissionless blockchains, underpinned by novel consensus protocols, have begun to advertise
processing speeds thousands of times greater than Bitcoin. These newer projects and the questions that still
surround them will be discussed in section /four.lf./two.lf./one.lf.
/two.lf/three.lfOne view on this point is that the tendency for non-asset-backed cryptocurrencies to induce hoarding be-
havior means that they should not actually be thought of as currencies. A key feature of a currency is that the
people who hold it should often use it to make purchases. Arguably, then, cryptocurrencies like bitcoin are less
analogous to traditional currencies than they are to rare paintings. Like rare paintings, these cryptocurrencies are
exotic assets that investors sometimes use as alternatives to traditional /fi.liganancial assets; their valuations depend
/two.lf/eight.lf
greater responsibility for the security of their holdings than users of traditional currencies.
A cryptocurrency owner who loses their own private key will have absolutely no way to use
their coins, and a cryptocurrency owner whose private key is discovered by someone else
will have no way to reverse any fraudulent transactions; the additional security features
o/f\_f.ligaered by credit card companies, for example, can o/f\_f.ligaer a very large amount of value to
users. Finally, just as blockchains are naturally less suited to con/fi.ligadentiality than central-
ized servers (see section /two.lf./six.lf./two.lf), cryptocurrency transactions can often be less con/fi.ligadential
than transactions made using traditional payment services.
This /fi.liganal point may be surprising, sometimes it is sometimes claimed that cryptocurren-
cies like bitcoin allow users to make anonymous payments. In fact, bitcoin is pseudonymous ,
since payments are still attached to unique public-key pseudonyms, and this distinction
makes all the di/f\_f.ligaerence [/one.lf/one.lf/nine.lf]. Every payment that a given user’s pseudonym makes or
receives is logged for anyone else to see, and law enforcement agencies (and other users)
have found it easy to attach pseudonyms to real-world identities by analyzing patterns of
payments [/one.lf/three.lf/four.lf]. In addition, if the user does not use an anonymizing service such as Tor,
then it may even be possible to associate their pseudonym directly with their device’s IP
address. Finally, and perhaps most importantly, users who would like to exchange crypto-
currency for some good or service in the physical world will still normally need to expose
some aspect of their identities to whoever they are interacting with. Bitcoin, in short, is
not very private.
On the other hand, some recently developed techniques for obscuring transactions hint
that the long-term trend may be toward much greater privacy for cryptocurrency users.
These include “mixing services,” which allow users to swap coins in order to frustrate
network analysis, and “state/payment channels,” which allow sets of users to conduct se-
quences of transactions and then only publish the /fi.liganal outcomes of these transactions./two.lf/four.lf
However, the most promising techniques, now to be discussed, are almost certainly those
that rely on cryptographic tools known as zk-SNARKs.
/two.lf./eight.lf Zero-knowledge proofs
Zero-knowledge proofs are proofs of mathematical statements that do not convey any infor-
mation other than that the statements are true (and that, as a logical consequence, the
prover has the knowledge necessary to prove them) [/eight.lf/zero.lf].
As an illustration, consider the question of whether there is a path around the graph de-
picted in Figure /four.lf that touches each node exactly once. Typically, the way to prove to
another party that such a path exists is to share one such path, as Figure /four.lf does, and al-
low them to check that it meets the criterion. In contrast, a zero-knowledge proof would
on opaque and sometimes ba/f\_f\_l.ligaing social dynamics, and it would be unusual to use them to buy some milk at the
store.
/two.lf/four.lfCurrently, the most popular cryptocurrency designed to obscure transaction details is Monero. It applies
“ring signatures,” another cryptographic technology, to introduce uncertainty about the origins of payments.
Some research suggests that the level of privacy o/f\_f.ligaered by Monero, while greater than that o/f\_f.ligaered by Bitcoin, is
still relatively low [/one.lf/one.lf/six.lf].
/two.lf/nine.lf
Figure /four.lf: There is a path around this graph that touches each node exactly once. A typical
proof of this statement requires sharing the details of at least one such path (as in the
above /fi.ligagure). In contrast, a zero-knowledge proof does not require sharing any additional
information of this sort. (Image by Christoph Sommer.)
not require any particular path to be shared. The only information that a zero-knowledge
proof communicates is the simple fact that whatever statement it is meant to prove is
true./two.lf/five.lf
Zero-knowledge proofs were /fi.ligarst described in a /one.lf/nine.lf/eight.lf/five.lf paper. Since this time, cryptogra-
phers have discovered that all mathematical statements that it is practical to prove at all
can also be proven in zero-knowledge [/one.lf/six.lf]./two.lf/six.lf
Zero-knowledge proofs can be divided into interactive and non-interactive varieties [/two.lf/four.lf].
In the interactive variety, a prover andverifier engage in back-and-forth interactions that
ultimately serve to convince the veri/fi.ligaer that the relevant statement is true. In the non-
interactive variety, which has been more recently developed, the prover simply publishes
a proof that anyone is free to verify. While interactive proofs only serve to prove a given
statement to one particular party at a time, non-interactive proofs can be used to prove
statements to large masses of parties at once.
One recent event of note is the development of zk-SNARKs (“zero-knowledge Succinct
Non-interactive ARguments of Knowledge”) in /two.lf/zero.lf/one.lf/zero.lf. These are a variety of non-interactive
/two.lf/five.lfThe de/fi.liganition of a zero-knowledge proof can also be given more formally. Suppose there is some function
V(X,w) that determines whether some information w, known as a witness , su/f\_f\_i.ligaces to prove some mathematical
statement X. In a typical proof, the prover shares a witness with veri/fi.ligaers, who then compute V(X,w) to determine
that Xis true. In contrast, a zero-knowledge proof is a proof in which the prover demonstrates that they possess a
witness of the relevant statement without actually sharing the witness. For above case, Xwould be the statement
that the given kind of path around the graph exists, wwould be a description of a path, and Vwould be a function
that checks whether the path is valid and touches each node exactly once.
/two.lf/six.lfSpeci/fi.ligacally, as another technical note, anything in the complexity class PSPACE can be proven in zero-
knowledge.
/three.lf/zero.lf
zero-knowledge proofs that are particularly short, e/f\_f\_i.ligacient to check, and therefore practi-
cal to use [/one.lf/eight.lf].
One application of zk-SNARKs, already alluded to, is in developing cryptocurrency sys-
tems that allow users to prove the validity of the payments they record on the blockchain,
without also publicly revealing sensitive information that implies the payments’ recipients,
values, or origins [/one.lf/four.lf/one.lf]. The /fi.ligarst example zk-SNARKs being used in this way is Zcash, a
cryptocurrency developed by American and Israeli academics and launched in late /two.lf/zero.lf/one.lf/six.lf.
Developers have also been gradually rolling out tools to facilitate the use of zk-SNARKs
on Ethereum, one of the most widely used blockchains [/seven.lf]. It could be only a matter of
time until the typical cryptocurrency transaction truly is private [/one.lf/three.lf/five.lf]./two.lf/seven.lf
Another application of zk-SNARKs is in developing what Joshua Kroll refers to as account-
able algorithms [/one.lf/zero.lf/three.lf]. In a recent paper, Kroll describes a general protocol that can be used
to provide accountability in any case where an institution is using an algorithm to make
decisions that a/f\_f.ligaect individuals. Examples of such cases include the use of algorithms,
applied to personal data, to select individuals for search or for tax auditing. In Kroll’s
protocol, the use of zk-SNARKs achieves two ends. First, zk-SNARKs allow a/f\_f.ligaected in-
dividuals to verify that the algorithm has received the approval of an auditing body and
meets certain other criteria, without needing to learn the details of the algorithm. Second,
zk-SNARKs allow an auditing body to verify that the algorithm they have approved is in
fact being used to make decisions, without needing to learn what these decisions are or
the personal data used to make them. The ultimate e/f\_f.ligaect of schemes such as Kroll’s could
be to increase public trust in institutions, while decreasing corruption and opening up the
procedures followed by these institutions to greater scrutiny.
It should be noted, however, that the use of zk-SNARKs does come with some downsides.
First, zk-SNARKs will always be somewhat less e/f\_f\_i.ligacient than traditional proofs. Second,
the use of zk-SNARKs requires an initial trusted setup procedure in which a set of parties
generate information that is a necessary input when constructing zk-SNARKs [/one.lf/nine.lf]. This
procedure also generates some secondary information, informally known as “toxic waste,”
that could be used to construct fraudulent zk-SNARKs. The parties who participate in the
setup procedure must take great pains to demonstrate that they have destroyed this “toxic
waste,” rather than saving it to exploit it later [/one.lf/six.lf/four.lf]./two.lf/eight.lf
Another recent development of note is a series of recent proposals for practical applica-
tions of what are known physical zero-knowledge proofs [/six.lf/eight.lf]. These are protocols that ap-
/two.lf/seven.lfOn the other hand, there is also some evidence that user interest in strong privacy features is fairly low.
Bitcoin remains more popular than Zcash even among people actively engaged in criminal activity [/one.lf/four.lf/nine.lf]. Fur-
thermore, only a minority of Zcash’s current users choose to use its (optional) privacy features. It is possible that
many potential users of Zcash’s zk-SNARK technology either do not understand or do not trust its ability to
protect their privacy.
/two.lf/eight.lfOf note, however, is some recent research into another variety of non-interactive zero-knowledge proofs, zk-
STARKs , that would lack this vulnerability [/one.lf/seven.lf]. Currently, the main downside of zk-STARKs over zk-SNARKs
is that zk-STARKs are substantially longer. As of /two.lf/zero.lf/one.lf/eight.lf, according to the original zk-STARK paper, they require
about one thousand times more storage space. This di/f\_f.ligaerence in storage requirements is especially important
in the context of traditional blockchain-based applications, given the need for many di/f\_f.ligaerent users to store
complete transaction records.
/three.lf/one.lf
ply the principles of interactive zero-knowledge proofs to the demonstration of physical
(rather than mathematical) claims.
Consider the following illustration. Alice would like to prove to Bob that two cups contain
the same number of balls without revealing what this number is. To do this she can /fi.ligall
up a bucket with another number of balls, known only to her, then make a commitment by
stating how many balls will be in the bucket if she pours either cup in. She allows Bob to
challenge her by picking, at random, one of the two cups to pour in. Then, as a response ,
she shows Bob how many balls are now inside the bucket. Since her initial statement
would have had only a /five.lf/zero.lf% chance of being correct if the cups were unequal, the result
of this procedure should increase Bob’s con/fi.ligadence in their equality. The procedure can
be repeated until Bob’s con/fi.ligadence reaches any given threshold. In this way, Alice can
prove her proposition about the balls in the cups (probabilistically) while still keeping
their quantity a secret.
While physical zero-knowledge proofs were /fi.ligarst described primarily for pedagogical pur-
poses, helping to illustrate how non-interactive zero-knowledge proofs might be possible,
it is now clear that they can have signi/fi.ligacant practical applications in their own right.
A /two.lf/zero.lf/one.lf/four.lf paper, published in Nature , showed that physical zero-knowledge proofs can be
used to verify that a country is disposing of a genuine nuclear warhead, without needing
to learn the design details of this warhead [/seven.lf/six.lf]. It has since been shown that it is also
possible to verify that a given suspect’s /fi.ligangerprints or DNA do not match those found at
a crime scene, without needing to collect their /fi.ligangerprints or DNA [/six.lf/eight.lf].
Since very few academics have ever written on physical zero-knowledge proofs, it is plau-
sible that many more applications remain to be found.
/two.lf./nine.lf Smart property
Cryptocurrency is not the only form of property whose ownership can be tracked or de-
termined with blockchains.
As a very closely related use case, a blockchain’s users might also create tokens , whose own-
ership is recorded in just the same way cryptocurrency ownership is. Users can sell the
tokens to others with the promise that these tokens will be exchangeable for a particular
service in the future. It is reasonable to think of these tokens as being roughly analogous
to the tokens sometimes sold at arcades (which can be exchanged for a certain number of
rounds of play at speci/fi.ligac games). Alternatively, users may sell tokens purely as collecta-
bles, which are valued for roughly the same reasons that rare stamps are valued. Unique
tokens, known as non-fungible tokens (NFTs ), have in some cases sold for millions of dollars
[/one.lf/five.lf, /six.lf/three.lf]./two.lf/nine.lf
/two.lf/nine.lfAn NFT is typically associated with some piece of digital media. For example, an artist may assert that an
NFT, in some sense, “represents” a particular animation that they have made. At /fi.ligarst glance, the valuations of
certain NFTs are rather confusing. However, one should keep in mind, they are at most slightly more confusing
than the valuations of many other collectables.
/three.lf/two.lf
As another well-known case, traditional property records, which might otherwise be stored
in another form of database, can of course also be stored using a blockchain. For example,
at least as of /two.lf/zero.lf/one.lf/eight.lf, the Ghanaian national government has been considering a joint project
with IBM to place local land deeds on blockchains in order to lower the risk of meddling
by corrupt o/f\_f\_i.ligacials [/five.lf/eight.lf].
The link between blockchain-based records and property ownership can also be made
much more concrete, however, through the creation of smart property : devices with elec-
tronic components that facilitate the transfer of their ownership [/one.lf/five.lf/nine.lf, /one.lf/one.lf/nine.lf].
Consider the following system:
•The device is initially associated with some public key, which belongs to the device’s
owner. This ownership is logged as a record on the blockchain, which the device can
read.
•To unlock or operate the device, its owner must send a message to the device and
digitally sign it with their private key. They may send this message through a channel
such as a Bluetooth connection or a card reader slot.
•Say that the owner has public key X, another person has public key Y, and the device
is denoted by D. If the owner would like to give the device to the person with public
keyY, then they can add a record of form “ Xgives DtoY” to the blockchain, along
with their digital signature. Now the device will respond to messages signed by the
new owner and will no longer respond to messages signed by the old one.
To make this description tangible, we can imagine that the device is a car, and that the car
will only unlock or start if it receives a message signed with the correct key. More limited
access rights (such as the right to use the car only on a certain day) could also be granted
through a similar scheme.
With the rise of the internet of things (IoT), or the trend of more and more electronic de-
vices having internet connections, it seems like the large-scale implementation of physical
smart property could be feasible [/four.lf/eight.lf]. This implementation could be achieved using per-
missionless blockchains or consortium blockchains maintained by the devices’ producers.
The advantage of using such blockchains, here, is that it would be unreasonable in any
other case to trust a single actor with running an application that directly controls a wide
swath of physical property./three.lf/zero.lf
However, it remains to be seen whether there will be any signi/fi.ligacant consumer interest in
smart property. As with cryptocurrency, some users might be attracted by the idea that
smart property can reduce reliance on centralized institutions that track property own-
ership or protect property from theft. Smart property might also o/f\_f.ligaer value by reducing
ine/f\_f\_i.ligaciencies in transactions and consolidating records of ownership. On the other hand,
one could argue that there is not much hardship involved in (e.g.) transferring the owner-
ship of a car in the traditional way.
/three.lf/zero.lfMore broadly, a number of companies are also investigating the advantages of using blockchains to provide
greater security for their own IoT devices.
/three.lf/three.lf
We should note that essentially the same privacy concerns that exist for cryptocurren-
cies exist for smart property too. Any smart-property transactions implemented on the
most widely used permissionless blockchains today would be highly visible (although fu-
ture systems using zk-SNARKs may be able to o/f\_f.ligaer more privacy), and even specialized
consortium blockchains might present greater privacy concerns than less comprehensive
centralized databases. This would seem to be an important limitation.
Finally, it is worth noting that systems of property ownership and exchange that are more
complex (and less decentralized) than the one described above could also be designed.
For example, we can imagine a smart-property car for which, by design, a government-
held private key is always capable of shutting down or transferring ownership, should an
appropriate legal order be given.
/two.lf./one.lf/zero.lf Smart contracts
Asmart contract , broadly de/fi.liganed, is a contract whose execution is automated to a signi/fi.liga-
cant extent.
Vending machines provide a simple illustration. When someone places money into a vend-
ing machine and receives a candy bar in return, the machine is, in e/f\_f.ligaect, executing a con-
tract between the buyer and seller.
The concept of a smart contract was /fi.ligarst formulated by Nick Szabo in /one.lf/nine.lf/nine.lf/six.lf, who argued
that, beyond o/f\_f.ligaering e/f\_f\_i.ligaciency gains, smart contracts can be used to reduce the role of
trust in contract execution [/one.lf/five.lf/nine.lf]. If a contract can be enforced automatically, and the
relevant mechanism can be made su/f\_f\_i.ligaciently transparent and resistant to tampering, then
the parties to a contract will not need to trust the other to execute their end of it; they will
also not need to trust any third party to act as an intermediary or enforcer.
By enabling applications whose execution does not rely on trusted third parties, consor-
tium and permissionless blockchains may o/f\_f.ligaer a particularly ideal platform for imple-
menting smart contracts./three.lf/one.lf
Consider, for example, the simple case of a decentralized application for playing chess:
Players take turns submitting moves, and a digital currency deposit is automatically trans-
ferred from the loser to the winner. This application is, in e/f\_f.ligaect, executing a contract
between the players. Here, though, there is no need to trust the loser to hand money over
to the winner, an escrow service to handle the transfer, or a court to extract compensation
for dishonesty.
The topic of smart contract design received a major boost in /two.lf/zero.lf/one.lf/three.lf, when Vitalik Buterin
proposed an in/fl.ligauential design for a blockchain known as Ethereum. Finally launched in
/two.lf/zero.lf/one.lf/five.lf, Ethereum o/f\_f.ligaers its users the ability draft and execute any smart contract that can
be expressed in terms of the blockchain’s state.
/three.lf/one.lfUnless otherwise stated, the use of the term “smart contract” throughout the rest of the report can be assumed
to refer to a smart contract implemented using a consortium or permissionless blockchain.
/three.lf/four.lf
To gain some intuition of how this can work, consider the following (highly idealized)
description of a smart contract’s creation and execution:
•Suppose a user with public key Xwould like to sell a smart-property car, denoted
byD, to a user with public key Y, at the cost of Vunits of cryptocurrency.
•To accomplish the transaction, the seller can submit a contract of this form: "If X
andYboth submit signed messages approving the transaction, then transfer DtoY
and a coin of value VtoX." This contract is then logged. If both the buyer and the
seller proceed to submit messages of approval, signed with their respective private
keys, then it will be recorded that the ownership of the two items has switched.
•The seller gains the ability to spend the coin, and the buyer gains the ability to
operate the car and demonstrate that it is rightfully theirs. The trade has been ac-
complished, without any need for a trusted intermediary and without any risk of
either participant failing to follow through.
Naturally, the conditions for transferring an item through a blockchain-based smart con-
tract can only be expressed in terms of information that is stored in the blockchain. How-
ever, this does not make it impossible to set conditions that depend on the external world.
Say that two users would like to enter into a bet about the high temperature for tomorrow.
A simple way to carry this bet out would be to write a smart contract that will transfer
cryptocurrency, automatically, when a certain trusted friend submits a signed report to
the blockchain. Trusted services could also be set to print information about the weather
onto the blockchain to facilitate weather-related contracts.
In addition, there are some new or emerging services that use a mixture of reputation sys-
tems and consensus protocols, associated with conditional payments, to incentivize users
to provide reliable inputs to others’ smart contracts. These services are known as decentral-
ized oracle systems , and have so far been most extensively explored in the context of a betting
market application known as Augur, which itself functions by executing smart contracts
among bettors [/one.lf/three.lf/zero.lf]./three.lf/two.lf
Smart contracts may also be used as building blocks, of a sort, for designing higher-level
applications. Augur, just mentioned, is one example. Other examples include proposed
applications that would use smart contracts to facilitate the rental of cloud computing
resources or storage space. In fact, Ethereum’s smart contract system is /fl.ligaexible enough to
build any possible decentralized application./three.lf/three.lf
/three.lf/two.lfAs a very simple example of how a consensus protocol like this could work, consider again the case of two
users who want to bet about the high temperature in their city. They could create a smart contract that depends
on the input of several other parties who do not know each other, such that the contract will pay whichever of
these inputs does not diverge from the median input by more than a degree. If these parties cannot collude, then
they will be incentivized to converge on the truth, since they should expect their self-interested counterparts
to converge on it too. A more sophisticated version of this protocol, for repeated bets, might track “reputa-
tion points” for users who input information, such that users with higher reputations receive more weight in
disagreements and earn higher premiums.
/three.lf/three.lfIn technical terms, Ethereum provides a "Turing-complete" language for writing applications.
/three.lf/five.lf
Unfortunately, the same limitations faced by decentralized applications discussed in sec-
tion /two.lf./six.lf./two.lf still apply. All established permissionless blockchains, including Ethereum, are
too ine/f\_f\_i.ligacient to run anything beyond fairly simple applications. Again, even something
as complex as an application that checks who has won a game of chess is pushing up against
Ethereum’s current limitations.
While researchers are currently investigating a number of ways to “scale” blockchains to
accommodate much greater numbers of users and much more complex applications, it is
not yet clear how e/f\_f.ligaective any such techniques will be (see section /four.lf./two.lf).
/two.lf./one.lf/one.lf Homomorphic encryption
Homomorphic encryption is a form of encryption that allows computations to be run on en-
crypted data, such that the outputs of the computations are encrypted as well [/one.lf/seven.lf/one.lf].
For instance, we say that an encryption scheme is “homomorphic under addition” if we
can encrypt any two numbers, perform a certain operation on them, and then decrypt the
result to return their sum. E/f\_f\_i.ligacient schemes that are homomorphic under only addition
or only multiplication have been known for a number of years./three.lf/four.lf
A scheme is known as fully homomorphic if it allows any computable function to be com-
puted on encrypted data. The /fi.ligarst fully homomorphic scheme was invented in /two.lf/zero.lf/zero.lf/nine.lf, by
Craig Gentry [/seven.lf/two.lf]. Since this time, a number of superior alternatives have also been pro-
posed.
The basic appeal of fully homomorphic encryption is that it makes it possible to create
applications in which service providers do not require access to clients’ unencrypted data.
Possible examples include cloud computing services that do not need to know the data
they are being asked to compute on, online medical services that /fl.ligaag health concerns on
the basis of individuals’ DNA without needing to know their genetic data, and targeted ad-
vertising services that do not need to know the interests that users have indicated through
their browsing behavior.
However, fully homomorphic encryption has not yet found signi/fi.ligacant use, as all known
schemes are highly ine/f\_f\_i.ligacient. The most e/f\_f\_i.ligacient schemes discovered so far still result in
many-orders-of-magnitude blowups in the size of the encrypted data as it is operated on,
and in the time the operations take to complete. For an example, in /two.lf/zero.lf/one.lf/four.lf, it was the case
that even a well-optimized implementation running on a moderately powerful computer
might be incapable of performing more than /five.lf/zero.lf multiplications per second (with addition
being much less costly) [/eight.lf]. This is actually a vast improvement over Gentry’s original
scheme, but still enormously slow. To a very rough approximation, fully homomorphically
encrypted data can be regarded as about a billion times less e/f\_f\_i.ligacient to compute on than
unencrypted data.
/three.lf/four.lfToday, the most widely applicable homomorphic encryption scheme is likely the Paillier encryption scheme ,
which is homomorphic under addition while also allowing for multiplication between encrypted inputs and
unencrypted ones. This is su/f\_f\_i.ligacient to compute any linear function of encrypted data.
/three.lf/six.lf
Some help is provided by somewhat homomorphic encryption schemes, developed concur-
rently, which o/f\_f.ligaer signi/fi.ligacant (but not decisive) speedups at the cost of allowing for only
a /fi.liganite number of multiplications. Over time, further progress in discovering e/f\_f\_i.ligacient
algorithms and in developing more powerful computers could make fully or somewhat ho-
momorphic encryption practical in a wide range of cases. For the moment, however, their
applications remain relatively narrow [/one.lf/one.lf/seven.lf].
In the following two sections, we consider an additional two technologies that allow com-
putations to be run on private data. These are functional encryption andsecure multiparty
computation .
/two.lf./one.lf/two.lf Functional encryption
Functional encryption is a form of encryption that allows particular functions of encrypted
data to be computed—such that, in contrast with homomorphic encryption, the output is
not encrypted [/two.lf/seven.lf].
For example, using functional encryption, a hospital could encrypt its patients’ medical
records in a way that allows di/f\_f.ligaerent categories of hospital workers, holding di/f\_f.ligaerent keys,
to compute only the information they need. A user of a cloud storage service could also
encrypt their /fi.ligales in a way that allows the service, or government investigators, to compute
only whether the /fi.ligales contain certain restricted material (like copyrighted media).
In general, a functional encryption system works in the following way:
•There exists a setup algorithm that can produce an almost certainly unique key pair,
consisting of a public key and a master private key . As in a regular public-key en-
cryption system, the public key, along with an encryption algorithm , can be used to
encrypt data.
•There also exists a key-generating algorithm that can, given a master private key and
a function, f, produce a special private key ,kf.
•Finally, there exists a decryption algorithm that can, given kfand a piece of data en-
crypted using the master private key, produce the function f(x).
•In a simple case, it is possible to encrypt one’s data using a master private key, then
grant others the ability to ascertain certain information from the data by distribut-
ing special private keys.
The concept of functional encryption is quite new, /fi.ligarst appearing in academic papers in
/two.lf/zero.lf/one.lf/zero.lf [/one.lf/two.lf/four.lf]. Researchers have since developed a number of general functional encryption
systems—meaning, systems whose key-generating algorithms can be used to produce a spe-
cial key for any desired function [/seven.lf/nine.lf, /seven.lf/eight.lf].
Similar to homomorphic encryption, functional encryption can make it easier to provide
privacy-preserving services. In the two examples given above, functional encryption allows
the hospital workers and the government investigators to gain just exactly the information
that is necessary to do their jobs.
/three.lf/seven.lf
In addition, while only a small amount of research has addressed practical implementations
of functional encryption, it can, for many functions, be much more e/f\_f\_i.ligacient than fully
homomorphic encryption [/one.lf/five.lf/zero.lf]. Nevertheless, applications of functional encryption have
been little explored.
/two.lf./one.lf/three.lf Secure multiparty computation and secret sharing
Finally, secure multiparty computation (MPC ) refers to techniques that allow users to run
computations with inputs from multiple parties, while allowing these parties to keep their
own inputs secret [/one.lf/five.lf/one.lf].
As a popular illustration, consider the service provided by dating applications like Tinder,
in which two people will receive a noti/fi.ligacation if and only if they both express interest in
the other. We can think of these applications as computing a function f(X,Y) , where Xand
Yrepresent the users’ inputs and frepresents whether or not both inputs are expressions of
interest. More importantly, these applications compute f(X,Y) for the two parties without
requiring them to share their inputs with one another. Since “no match” is compatible
with either one or two inputs of “not interested,” a user who inputs “not interested” gains
no information at all.
These applications solve the “dating problem” by acting as a trusted third party: they col-
lect both parties’ inputs and then promise to compute the function correctly and to not
share their inputs with anyone else. We might ask, though, whether it is possible to achieve
the same end without any trusted third party at all.
In fact, as Andrew Yao proved in a classic /one.lf/nine.lf/eight.lf/two.lf paper, MPC protocols can be used to
solve the dating problem without trusted third parties [/one.lf/eight.lf/five.lf]./three.lf/five.lfBy engaging in a fairly
complex sequence of interactions and individual computations, the two potential daters
can together compute f(X,Y) while maintaining their individual privacy. Since this time,
cryptographers have also learned that MPC protocols can be designed to compute any
joint function of private inputs, for any number of parties. In just the past decade, there
has been substantial progress in developing MPC schemes that are practical to implement,
with e/f\_f\_i.ligaciency increasing by several orders of magnitude [/seven.lf/four.lf].
The /fi.ligarst economically signi/fi.ligacant application came in /two.lf/zero.lf/zero.lf/eight.lf, when MPC was used to carry
out a secure sugar beet auction in Denmark with over one thousand bidders [/two.lf/six.lf]. In this
case, the various parties determined the winning bid for each item without otherwise shar-
ing their bids with one another. Without MPC, they would have needed to share their bids
either with one another, or with a third party who was trusted to report the proper out-
come.
One frequently discussed use case for MPC is in secure voting systems, which would allow
voters to jointly determine the outcomes of elections without sharing their votes with one
another or trusting any third party to tabulate them [/five.lf/three.lf]. Another use case is training ma-
chine learning systems across multiple privately held datasets, such as individual medical
/three.lf/five.lfTechnically, Yao showed this for a more general problem he called the “millionaires’ problem.”
/three.lf/eight.lf
records or classi/fi.ligaed information held by di/f\_f.ligaerent government agencies [/one.lf/zero.lf/eight.lf].
In addition, we can understand the use of homomorphic encryption to provide privacy-
preserving services (see section /two.lf./one.lf/one.lf) through the lens of MPC [/four.lf/seven.lf]. Speci/fi.ligacally, client-
server applications (see section /two.lf./six.lf./one.lf) in which the client homomorphically encrypts the
data they send to the server can be interpreted as implementing a particular variety of
MPC, in which the client’s input is the data to be encrypted, the server’s input is the al-
gorithm to be applied to that data, and only the client receives the computation’s output.
While processing data in this way is associated with enormous computational overhead, as
stated above, there exist other MPC protocols that accomplish the same end much more ef-
/fi.ligaciently. This reduction comes at the cost of requiring extensive communication between
the server and client and, in general, requiring the client to be a much more active partic-
ipant. Nevertheless, these more communication-heavy MPC protocols are comparatively
practical.
Another way to use more e/f\_f\_i.ligacient MPC protocols to o/f\_f.ligaer privacy-preserving services, in
this case without placing such a large burden on the client, involves a cryptographic tech-
nique known as secret sharing [/one.lf/four.lf/eight.lf]. In secret sharing, a private piece of data, or secret , is
divided into a number of shares . These shares individually provide no information about
the secret, but if su/f\_f\_i.ligaciently many of the shares are combined, then the secret can be re-
trieved.
Consider the following protocol:
•Alice has some data, X, and she would like some function of it, f(X), to be computed.
To make the case concrete, we might imagine that Xis personal medical data and
f(X)is a list of health risks suggested by the data.
•Alice splits her data into a number of shares, using secret sharing, and distributes
the shares among an equal number of other parties.
•Now, so long as each party keeps its own share private, the problem of computing
f(X)is transformed into an MPC problem with the shares as inputs.
•Running the MPC protocol until just before its end results in each party holding,
in e/f\_f.ligaect, a share of f(X). At this point, they send their output shares to Alice alone,
and she recombines them to learn f(X). She has received a service from these parties
without sacri/fi.ligacing her privacy.
In general, the use of secret-sharing protocols makes it possible to separate the parties that
provide a computation’s data inputs, the parties that receive its outputs, and the parties
that perform the bulk of the computational work. We can divide the relevant parties up
into the input parties ,computing parties , and results parties (see Figure /five.lf) [/one.lf/six.lf/zero.lf]. The input
parties divide shares of their data among the computing parties. The computing parties
compute shares of a joint function of this data. The results parties receive these shares and
join them together to learn the /fi.liganal output. In a classic MPC protocol, these three sets
of parties would all be the same. In the above description, though, Alice is the only input
party and the only results party, and she is not a computing party.
/three.lf/nine.lf
Figure /five.lf: In this MPC protocol, the input parties use secret sharing to distribute shares of
their private data, XandY, among the computing parties. The computing parties compute
shares of a function of this data, f(X,Y) , and then send these shares to a results party that
combines them to learn f(X,Y) . In general, a single party might play multiple roles. In classic
MPC protocols, the input parties, computing parties, and results parties are all the same.
The most obvious security challenge with secret-sharing MPC protocols is the need to
ensure that the computing parties do not collude to combine their shares, thereby learning
the relevant secrets. Another worry is that computing parties might not carry out their
portions of the computations correctly.
While stricter security standards imply stronger requirements, there exist protocols in
which even a single honest party is su/f\_f\_i.ligacient to ensure privacy of the relevant data. Fur-
thermore, conditional on the assumption that the computing parties have access to a trust-
worthy public log, there exist protocols that also allow even uninvolved parties to verify
that the computation was performed correctly [/one.lf/four.lf].
It is not unreasonable to think that these requirements—a trustworthy public log and a
single honest party—can often be met. Blockchains, or even just simple tamper-evident
logs, are a good way of meeting the /fi.ligarst requirement. Economic incentives may provide
an ideal way of meeting the second. One solution might be to rely on companies whose
business models depend on their reputations as reliable computing parties for MPC proto-
cols. Another solution is being explored by an MIT-based project known as Enigma [/one.lf/eight.lf/seven.lf].
Enigma is a decentralized application (see section /two.lf./six.lf./one.lf) meant to implement a market
for MPC computations. Parties contract others to serve as computing parties, and these
parties lose substantial cryptocurrency deposits if they are found to engage in malicious
behaviors.
The use of MPC is still associated with substantial overheads. For example, at the time
/four.lf/zero.lf
of writing, the most e/f\_f\_i.ligacient MPC protocols increase the time required to train a deep
learning system by about a factor of one thousand [/one.lf/seven.lf/seven.lf]. However, these overheads are
far smaller than the overheads associated with fully homomorphic encryption. They are
already quite manageable for short computations. Further innovations may decrease these
overheads even further./three.lf/six.lf,/three.lf/seven.lf
/three.lf/six.lfAt the time of writing, OpenMined is perhaps the most active project attempting to reduce barriers to
the widespread adoption of MPC [/one.lf/two.lf/six.lf]. Their primary focus is on developing software libraries that make it
easier for data scientists and machine learning engineers to implement e/f\_f\_i.ligacient MPC protocols, alongside other
privacy techniques. They also publish educational materials and advocate for the adoption of MPC and related
techniques.
/three.lf/seven.lfOne unifying perspective on zero-knowledge proofs, homomorphic encryption, functional encryption, and
secure multiparty computation is that they support what might be called structured transparency : precise controls
on who can know what, when they can know it, and what they can do with this knowledge (see section /three.lf./one.lf./two.lf) [/one.lf/six.lf/seven.lf].
For instance, the computing parties for a computation performed using MPC, unlike the computing parties for
a typical computation, do not need to know what its inputs or outputs are. The party reading a zero-knowledge
proof, unlike the reader of a typical proof, does not need to know anything besides the simple fact that the proof
is valid. The social value of structured transparency techniques is that, in many cases, they can enable desired
uses of information without also opening the door so widely to undesired misuse.
/four.lf/one.lf
/three.lf Speculative consequences
In this section, I gather and discuss a number of possible consequences of the developments
described above.
Since the full range of possibilities is of course very broad, I have limited myself to conse-
quences that I believe would have large-scale political signi/fi.ligacance. For example, although
a number of banks have investigated the possibility that consortium blockchains will allow
them to reconcile their records more e/f\_f\_i.ligaciently, it is not su/f\_f\_i.ligaciently clear to me that the
e/f\_f.ligaects of these developments would seep far outside the /fi.liganancial industry.
I will be considering the following seven possibilities:
/one.lf. Information channels used to conduct surveillance could “go dark”
/two.lf. Privacy-preserving surveillance could become feasible
/three.lf. Non-intrusive agreement veri/fi.ligacation could become feasible
/four.lf. It could become easier to combat forgery
/five.lf. The roles of banks, technology companies, voting authorities, and other traditional
institutions could shrink
/six.lf. It could become feasible to solve collective action problems that existing institutions
cannot
/seven.lf. A new variety of institutions, known as “decentralized autonomous organizations,”
could emerge
There are of course reasons, sometimes quite strong reasons, to be skeptical of each of
these possibilities. I discuss some reasons for skepticism within this section, but section
/four.lf, “Limitations and skeptical views,” describes several in signi/fi.ligacantly greater detail. This
section of the report is less focused on the question “What is likely to happen?” than it
is on the questions “What are the most interesting predictions that other researchers or
engineers have made?” and “What predictions would be truly consequential if they came
true?”
In addition, it should be noted that a period of several decades can often separate the
initial development of a technology and its eventual re/fi.liganement and adoption [/eight.lf/five.lf]. Even
predictions that are plausible in the long run may still be implausible, say, within the
next ten years. In fact, a number of these possibilities explicitly require very signi/fi.ligacant
technological progress.
/three.lf./one.lf Consequences from technologies other than blockchain
/three.lf./one.lf./one.lf Information channels used to conduct surveillance could “go dark”
Three trends could conceivably make surveillance signi/fi.ligacantly more di/f\_f\_i.ligacult: /fi.ligarst, the
growing use of end-to-end public-key encryption; second, the emergence of digital cur-
rency systems that apply zk-SNARKs and other techniques meant to obscure economic
transactions; and, third, the potential for methods of computing on con/fi.ligadential data to
alter the economics of private data collection.
/four.lf/two.lf
As discussed in section /two.lf./one.lf, it has only recently become common for major providers of
messaging services to o/f\_f.ligaer “end-to-end encryption” for all messages that users send. “End-
to-end encryption” refers to a process of encryption and decryption that occurs solely on
the sender’s and receiver’s devices, meaning, in practice, that the service provider cannot
access the messages even if they would like to [/five.lf/nine.lf]. This, further, prevents government
agencies from gaining access via the service provider./three.lf/eight.lfJust over the course of /two.lf/zero.lf/one.lf/six.lf, the
number of end-to-end encryption users likely increased by more than a billion [/one.lf/two.lf]. This
change is due to a number of major applications, most notably WhatsApp, beginning to
o/f\_f.ligaer the service by default. In addition, there is the less widely recognized prospect of
increasingly practical end-to-end encrypted messaging systems that obscure not just the
content of messages but also the associated “metadata”—particularly, data about who is
messaging whom [/five.lf/nine.lf, /one.lf/seven.lf/two.lf]. A number of prominent o/f\_f\_i.ligacials, such as former FBI director
James Comey, have responded to the growing use of end-to-end encryption by sounding an
alarm that information channels they rely on are “going dark” [/five.lf/one.lf, /seven.lf/zero.lf]. While the Chinese
government has moved to crack down on the use of this technology, civil liberty concerns
provide signi/fi.ligacant barriers to regulation in many other countries [/three.lf/zero.lf]./three.lf/nine.lf
The development and adoption of privacy-preserving cryptocurrencies may lead a further
information channel to “go dark.” While it is sometimes perceived that today’s widely
used cryptocurrencies already o/f\_f.ligaer signi/fi.ligacant privacy, this perception is mostly mistaken
[/one.lf/three.lf/four.lf, /one.lf/one.lf/nine.lf]. As discussed in section /two.lf./seven.lf, for most cryptocurrencies the full details of every
transaction are recorded by the relevant blockchain for everyone to see, and it is in prac-
tice fairly easy to tie users’ pseudonyms to their real-world identities. If cryptocurrencies
do ultimately make surveillance much more di/f\_f\_i.ligacult, then, it may be the result of cryp-
tocurrencies that apply zk-SNARKs (or other varieties of non-interactive zero-knowledge
proofs). At least one new cryptocurrency, Zcash, claims to o/f\_f.ligaer users the ability to make
transactions whose contents and participants are obscured from other users. Zcash is not
yet widely used, but the developers have been gradually expanding zk-SNARK support
on the Ethereum blockchain (see section /two.lf./one.lf/zero.lf). Despite some lingering security, e/f\_f\_i.ligaciency,
and user demand concerns, in the long run it seems plausible that cryptocurrency transac-
tions will become signi/fi.ligacantly less transparent. In addition, even if zk-SNARKs or other
forms of zero-knowledge proofs do not become widely used, a handful of other methods of
obscuring transactions, including “mixing services,” “ring signatures,” and “state/payment
channels,” could achieve similar e/f\_f.ligaects.
Finally, methods of running computations on encrypted data might lead further informa-
tion channels to go dark. “Don’t Panic,” a /two.lf/zero.lf/one.lf/six.lf report associated with Harvard’s Berkman
Center for Internet and Privacy, argues that the “going dark” problem is overstated due,
in large part, to private companies’ economic incentives to collect user data [/seven.lf/zero.lf]. While
this point is likely robust in the short run, and indeed the medium run, it may not be in
/three.lf/eight.lfAlthough, as mentioned above, even the use of perfectly implemented end-to-end encryption does not guar-
antee perfect security, for instance, if the user’s personal device is insecure or if they are tricked or coerced into
revealing their password. In addition, an agency might still convince service providers to weaken protections on
some users through software updates.
/three.lf/nine.lfThe European parliament has even recently considered a proposal, motivated by civil liberty concerns, to
prevent member states from banning the use of end-to-end encryption [/seven.lf/five.lf].
/four.lf/three.lf
the long run. As discussed in sections /two.lf./one.lf/one.lf–/two.lf./one.lf/three.lf, technologies such as secure multiparty
computation, functional encryption, and homomorphic encryption can reduce economic
incentives to collect unencrypted user data. Just as end-to-end encryption makes it possi-
ble for a service to send messages for a user without gaining access to their contents, these
technologies make it possible to send users targeted ads without gaining access to their in-
terests, to train machine learning algorithms on users’ activity logs without gaining access
to these logs, and so on. In addition, there are a number of less powerful, but already practi-
cal techniques for learning from user data in a privacy-preserving manner [/two.lf]. For instance,
di/f\_f.ligaerential privacy techniques work by making random alterations to datasets in order to
prevent the individuals included in the datasets from being identi/fi.ligaed. In /two.lf/zero.lf/one.lf/six.lf, Apple
began to apply these techniques to data associated with the use of its devices [/one.lf/six.lf/one.lf].
On the other hand, it is not impossible to imagine a future in which cryptographic devel-
opments could support the expansion of surveillance in certain ways. For example, as will
be discussed in the next section, methods of computing on con/fi.ligadential data could make
it easier to gather security-relevant information about individuals without also gaining
access to irrelevant private information. This capability could make it easier for govern-
ment agencies to engage in robust surveillance without running up against privacy con-
cerns.
In addition, it is very plausible that other technological developments, mostly unrelated
to cryptography, will ultimately play a larger role in increasing the ease of surveillance.
Video surveillance augmented by arti/fi.ligacial intelligence, for example, may become much
more pervasive. Some authors have also raised the possibility that, in the coming decades,
small and di/f\_f\_i.ligacult-to-detect drones could become highly e/f\_f.ligaective tools for surveillance
[/one.lf/eight.lf/one.lf]. More prosaically, the authors of the Berkman Center report highlight the growing
number of sensor-bearing devices that make up the “internet of things” as one long-run
trend that is making surveillance easier. Ultimately, when it comes to the future feasibility
of surveillance, new developments in cryptography might be of fairly secondary impor-
tance.
/three.lf./one.lf./two.lf Privacy-preserving surveillance could become feasible
In discussions of surveillance, the supposed trade-o/f\_f.liga between privacy and security is fre-
quently discussed. While this “trade-o/f\_f.liga” narrative is heavily simpli/fi.ligaed in a number of
ways—excluding, among other points, the fact that privacy can itself provide security
against exploitation—it does still capture something of the truth [/one.lf/eight.lf/one.lf, /one.lf/five.lf/two.lf].
A trade-o/f\_f.liga will exist to whatever extent the ability to access security-relevant information
also entails the ability to access information that would ideally remain private. For exam-
ple, for a police o/f\_f\_i.ligacer manually searching bags, learning whether someone is carrying a
weapon also requires learning everything else they are carrying. The choice is either to
forgo security-relevant information (whether there is a weapon in the bag) or to obtain
extraneous private information (what else is in the bag).
Fortunately, there is no fundamental reason why granting a party the ability to learn
security-relevant information must entail granting them the ability to learn anything else.
/four.lf/four.lf
Bomb-sni/f\_f\_i.ligang dogs, which report only whether there are explosives within a given bag, can
be seen as an example of a technology that minimizes the privacy-security trade-o/f\_f.liga in the
case just described (at least, in the idealized case that the dogs have perfect accuracy) [/three.lf/one.lf].
Technological progress may make it become possible to minimize the trade-o/f\_f.liga in a wider
and wider variety of domains. Or, to re-express this point in the terminology of one recent
paper, it may become increasingly possible to achieve structured transparency [/one.lf/six.lf/seven.lf].
For the surveillance policy, the basic principles are these:
Surveillance tasks can be automated: If progress in arti/fi.ligacial intelligence con-
tinues, then it will become possible to automate an increasingly large portion
of the tasks currently performed by human analysts (as well as tasks that hu-
man analysts cannot currently perform). For example, AI systems are likely to
become more capable of identifying faces in videos, noticing suspicious pat-
terns of transactions, and judging from private messages whether someone is
engaged in illegal activity.
If a surveillance task can be automated, then it can be completed without
access to private data: In any case where an AI system is extracting security-
relevant information from private data, it is, in principle, possible to design
the system in a way that does not require the party running it to collect the
data in unencrypted form. The relevant technologies, which enable comput-
ing on con/fi.ligadential data, include secure multiparty computation, functional
encryption, and homomorphic encryption (see sections /two.lf./one.lf/one.lf–/two.lf./one.lf/three.lf)./four.lf/zero.lf
Generally, work on the applications of cryptographic technologies to privacy-preserving
surveillance has received fairly little attention [/six.lf/six.lf, /one.lf/four.lf/five.lf, /two.lf/two.lf, /six.lf/nine.lf]. However, in recent years
there have been some useful proofs of concept. The most noteworthy case is probably the
development of an MPC-based system to detect cases of value-added tax fraud in Estonia:
although the system was not ultimately adopted, it was designed to identify discrepancies
between declared sales and purchases without violating the con/fi.ligadentiality of individual
companies’ /fi.liganancial records [/two.lf/five.lf].
There have also been proposed protocols for privacy-preserving “set intersection” searches,
which identify individuals who reappear across several datasets of interest [/one.lf/four.lf/six.lf]. For exam-
ple, set-intersection searches are sometimes used to identify criminals from multiple sets
of cell tower records associated with locations where they are known to have committed
crimes. A trivial form of secure multiparty computation over the datasets could prevent
the investigators from needing to access the full datasets in cases like these. Similarly,
widely used “contact-chaining” techniques, which generate lists of suspicious individuals
on the basis of social network analysis, could easily be implemented without extraneous
/four.lf/zero.lfAn intermediate privacy solution, which agencies such as the National Security Agency already employ to
varying degrees, is to collect and store unencrypted data, but limit the access of individual analysts. An analyst,
for example, may be allowed to make only a limited set of queries to the database and only view the portion
of records that are classi/fi.ligaed as matching these queries. However, systems like this provide weaker assurances of
privacy, in that they depend on the continued self-restraint of the agencies holding the data. The act of collecting
the data is also subject, in many cases, to signi/fi.ligacant legal constraints.
/four.lf/five.lf
data collection.
To illustrate the potential breadth of long-run possibilities, I will now describe a gen-
eral protocol for privacy-preserving mass surveillance with secure multiparty computation.
This protocol is inspired by Andrew Trask, a machine learning researcher, who recently
described a similar protocol that employs homomorphic encryption instead [/one.lf/six.lf/six.lf]. The
protocol would infeasible to implement today, but can interpreted as a kind of “limiting
case” of what might be possible in the future.
•Technology companies and service providers each maintain their users’ data inde-
pendently. No government agency collects this data.
•A number of these companies maintain subsidized data centers, which are designed
to engage in secure multiparty computations with a certain government agency.
•The agency develops a classi/fi.ligaer system intended to identify criminal activity on
the basis of personal data. An example of a simple classi/fi.ligaer would be one that just
performs set-intersection searches on cell records, identifying individuals who were
present at multiple linked crime scenes. A more sophisticated classi/fi.ligaer might use a
variety of more sensitive data, including the contents of private messages, to identify
individuals who appear to be involved in particular criminal activities. The classi/fi.ligaer
is tested to ensure that it does not exceed a mandated maximum false positive rate./four.lf/one.lf
•This procedure is overseen by an independent auditing body. The body con/fi.ligarms
that the classi/fi.ligaer has the properties claimed. Once the classi/fi.ligaer is put into use, the
auditing body will also be tasked with con/fi.ligarming that its false positive rate remains
below the maximum allowed value.
•Now, the agency engages in a secure multiparty computation with the relevant com-
panies. The inputs to the computation are each company’s private data and the pa-
rameters of the agency’s classi/fi.ligaer. (Data from other sources might also be included
through secret sharing.) The output, which is received only by the agency, is a list
of individuals that the classi/fi.ligaer deems su/f\_f\_i.ligaciently suspicious. The agency may ul-
timately request further data on these individuals from the companies, subject to
judicial approval.
•Di/f\_f.ligaerent classi/fi.ligaers could be deployed for di/f\_f.ligaerent crimes. The explicitly coded
threshold for reasonable suspicion or probable cause could also vary with the severity
of the crime and could be decided through public debate./four.lf/two.lf,/four.lf/three.lf
/four.lf/one.lfThe classi/fi.ligaers could be trained and tested by using datasets that are known to include examples of both
criminal and law-abiding behavior.
/four.lf/two.lfHistorically, at least within the United States, it has been held that the standards for probable cause resist
quanti/fi.ligacation. However, with the rise of sni/f\_f\_i.ligang dogs, partial /fi.ligangerprint matches, facial recognition technol-
ogy, and other “mechanistic” methods of establishing probable cause, it has become relatively common to take
statistical evidence into account, at least in an ad hoc manner [/seven.lf/seven.lf]. For example, if probable cause is established
solely on the basis of a match made by facial recognition software, then the evidence is essentially nothing but
statistical; it consists of the fact that software is making a certain classi/fi.ligacation and that the software is right a
su/f\_f\_i.ligaciently large portion of the time. Some legal scholars argue that, in mechanistic cases, the most natural way
to de/fi.ligane probable cause is simply to de/fi.ligane a necessary accuracy rate.
/four.lf/three.lfExpanding the role of statistical evidence in establishing reasonable suspicion and probable cause, currently
/four.lf/six.lf
•Finally, at least within the United States, the legal justi/fi.ligacation for using such a
classi/fi.ligaer could be similar to the justi/fi.ligacation for using sni/f\_f\_i.ligang dogs. In relevant
Supreme Court cases, it has traditionally been held that, insofar as the dog is only
capable of reporting unlawful activity, with su/f\_f\_i.ligaciently high accuracy, and insofar
as the dog is not trespassing on its target’s property, then its use does not constitute
a “search” [/three.lf/one.lf, /five.lf/zero.lf].
It may be instructive to examine this illustrative protocol further, in order to highlight sev-
eral important barriers to its implementation (or the implementation of similar privacy-
preserving surveillance protocols). First, for most instances of criminal activity, it is not
yet possible to develop classi/fi.ligaers that are su/f\_f\_i.ligaciently accurate. This is particularly true
for crimes with very small rates of occurrence, such as terrorism. The viability of the pro-
tocol, outside of narrow contexts, then requires extremely signi/fi.ligacant progress in arti/fi.ligacial
intelligence to raise the achievable accuracy level. Taking the long view, though, it is worth
noting that most experts in the /fi.ligaeld of arti/fi.ligacial intelligence expect computer systems to
become capable of completing all tasks that humans can within a century [/eight.lf/two.lf]. The wait
for AI systems to be able to complete many individual classi/fi.ligacation tasks may be substan-
tially shorter. In addition, a number of tasks, such as set-intersection searches, are either
already automated or trivial for existing systems.
Second, as already discussed, secure multiparty computation is very expensive. The com-
putational overhead and the need for interaction between the multiple parties could be ex-
pected, with current techniques, to increase the total cost of running any classi/fi.ligaer by mul-
tiple orders of magnitude. This increase is not necessarily unbearable, especially given that
available computing power and bandwidth have historically increased by about a factor of
a hundred every decade [/one.lf/zero.lf/nine.lf, /one.lf/two.lf/one.lf]. Still, every additional dollar of cost should be assumed
to decrease the technical and political viability of a privacy-preserving scheme.
Third, in this scheme, there is still a need for the companies and the public to trust the
auditing body to ensure the agency’s honesty. Fortunately, it may be possible to loosen this
constraint. As discussed in section /two.lf./eight.lf, zk-SNARKs can be used to publicly demonstrate
that a government agency is in fact applying an algorithm that has received an auditing
body’s approval, without revealing the details of the algorithm to the public [/one.lf/zero.lf/three.lf]. In the-
ory, it might even be possible to use zk-SNARKs to publicly demonstrate that the classi/fi.ligaer
achieves the desired accuracy rate without requiring trust to be placed in any auditing body
at all.
Fourth, this version of the scheme requires the relevant service providers to have access to
their users’ data. It presumes that they do not o/f\_f.ligaer end-to-end encryption of messages, or
homomorphic processing of data. One more di/f\_f\_i.ligacult alternative would be to require indi-
quite minor, might also be seen as its own end [/one.lf/one.lf/five.lf]. For example, a judge determining whether a police o/f\_f\_i.ligacer had
probable cause to conduct a search, on the basis of their claim that a certain driver appeared suspicious, would
not typically take into account the o/f\_f\_i.ligacer’s track record. Intuitively, though, it is of great relevance whether the
o/f\_f\_i.ligacer’s judgements about suspicious drivers are correct /one.lf/zero.lf/zero.lf% of the time or /zero.lf% of the time. An increase in the
use of statistical evidence, which would be quite abundant for any classi/fi.ligaer applied to large quantities of data,
could enable more e/f\_f.ligaective oversight, more precise discussions of standards, and more uniformly high-quality
decisions.
/four.lf/seven.lf
viduals to share their data with a number of independent entities, through secret sharing,
for the sole purpose of allowing these entities to use it for surveillance purposes. Public
surveillance footage and other data not directly associated with technology companies and
service providers could also be incorporated in this way.
Finally, even if all of these problems were resolved, there would still remain all of the legal,
political, and social challenges that would surely be associated with radically revising the
nature of a country’s surveillance policies and infrastructure. There would also be the
unavoidable fact that privacy-preserving surveillance can aid the enforcement of unjust
laws in exactly the same way that it aids the enforcement of just laws.
In short, the feasibility of privacy-preserving mass surveillance would require large tech-
nical advances and signi/fi.ligacantly more thought than has been devoted to it so far. There
may be reasons to be bullish about privacy-preserving surveillance in the long run, though.
The intersection between emerging cryptographic technologies and surveillance is still a
little-examined research area, and if fundamental barriers to privacy-preserving surveil-
lance exist, then they are not obvious.
I should also emphasize that broad privacy-preserving surveillance protocols, if they be-
come feasible, should not be assumed to be socially beneficial . Again, if some law is unjust,
then removing privacy concerns as a barrier to its enforcement would be harmful. There
are also sensible reasons to be wary of any large expansions of surveillance powers, even
if the laws in question are just and even if privacy-preserving techniques will be used.
Nonetheless, if highly privacy-preserving surveillance ever becomes feasible, then it is easy
to imagine surveillance powers eventually expanding in many countries.
/three.lf./one.lf./three.lf Non-intrusive agreement veri/fi.ligacation could become feasible
Just as methods of computing on con/fi.ligadential data might enable less invasive government
surveillance, in the long run they could also enable less intrusive methods of verifying
agreements.
The classic problem here is that nearly any agreement—such as an arms control agreement
between countries—will require information to be collected to provide assurance of each
party’s compliance. However, the act of collecting this information can often be highly
intrusive and will generally result in the veri/fi.ligaer gaining access to information beyond the
mere fact of the other party’s compliance or non-compliance [/six.lf]. For example, agreements
concerning chemical materials often entail inspections of private companies, which put
these companies’ valuable “con/fi.ligadential business information” at risk [/one.lf/zero.lf/zero.lf]. Similarly, the
party verifying a disarmament agreement might take it as an opportunity to learn more
about the weapons system being destroyed [/five.lf/two.lf]. If the cost of intrusion is great enough,
then an otherwise mutually bene/fi.ligacial agreement might become politically infeasible [/one.lf/two.lf/five.lf].
Even in cases where the cost of intrusion is ultimately deemed to be acceptable, like the
JCPOA agreement concerning Iran’s nuclear program, this cost can be a source of di/f\_f\_i.ligaculty
in negotiations, and attempts to decrease it can also decrease assurance [/four.lf/two.lf].
As above, we can construct minimally intrusive tools for veri/fi.ligacation, if we assume there
/four.lf/eight.lf
will eventually be extremely signi/fi.ligacant advancements in arti/fi.ligacial intelligence and in the
e/f\_f\_i.ligaciency of computing on con/fi.ligadential data. Ultimately, international agreement moni-
toring is just another form of surveillance. The only di/f\_f.ligaerence is that the subject of the
surveillance is party to an international agreement.
Beyond the secure computation techniques discussed in the previous section, zero-knowledge
proofs may also have their own applications to veri/fi.ligacation. For any future agreements
concerning algorithms—such as algorithms applied in autonomous weapons systems, vot-
ing systems, or systems used to make decisions with relevance to human rights—zero-
knowledge proofs could be a tool for demonstrating that only algorithms with approved
features are being applied [/one.lf/zero.lf/three.lf].
In addition, as mentioned in section /two.lf./eight.lf, some applications of “physical zero-knowledge
proofs” to agreement veri/fi.ligacation are already beginning to be found. Most signi/fi.ligacantly,
recent papers have proposed a method of demonstrating that a country is disposing of a
genuine nuclear warhead without revealing its design details [/seven.lf/six.lf].
It is interesting to consider what new applications of zero-knowledge proofs and physical
zero-knowledge proofs might be discovered in the coming years. One important limi-
tation, of course, is that we should not expect them to be any more powerful than tra-
ditional methods of demonstrating claims. For example, zero-knowledge proofs of non-
existence —such as a proof that a certain prohibited material does not exist anywhere within
a country’s borders—should be taken as much more di/f\_f\_i.ligacult than zero-knowledge proofs
of existence.
/three.lf./one.lf./four.lf It could become easier to combat forgery
As discussed in section /two.lf./four.lf, trusted timestamping can be used to combat forgery in a num-
ber of domains.
Timestamps put an upper bound on when a piece of data was created, providing assurance
that, at the very least, it was not forged or tampered with after the fact. For example,
an organization that has a strict policy of timestamping important documents can reduce
a number of risks, from corrupt o/f\_f\_i.ligacials altering incriminating documents to external
actors presenting forgeries as authentic. In large part, such concerns motivate the interest
actors such as the Estonian and British governments have recently taken in timestamping
services.
In the future, timestamping may also play an increasingly large role in helping people
to distinguish between authentic photographs and videos and arti/fi.ligacially generated ones.
Progress in machine learning has recently made it possible to generate images that are
almost indistinguishable from ones recorded by cameras. Similarly, although at great ex-
pense and using rather di/f\_f.ligaerent techniques, Hollywood studios have recently gained the
ability to produce nearly photorealistic renderings of long-dead actors. Taking note of
these trends, some commentators have suggested that we may lose our ability to trust in
the authenticity of digital media, with signi/fi.ligacant negative implications for journalists, in-
telligence services, and other groups relying on photo or video evidence [/four.lf, /three.lf/two.lf, /one.lf/four.lf/three.lf].
/four.lf/nine.lf
While timestamps alone do not solve this problem, they can still help substantially. Con-
sider someone recording a video and timestamping batches of frames as they do. If they
include in the video a depiction of certain information they could not have possessed be-
fore a certain time—such as a newspaper held aloft or, better yet, a projection of incoming
Bitcoin transactions—then this also serves to put a lower bound on when, if the video were
a forgery, they could have begun the /fi.liganal stage of editing. If the gap between this lower
bound and the upper bound provided by the timestamp is small enough, then it can be
regarded as implausible that the video is a forgery.
While such techniques are fairly cumbersome, their feasibility suggests that it should re-
main possible to demonstrate the authenticity of at least the most important digital me-
dia./four.lf/four.lf
/three.lf./two.lf Consequences from blockchain-based technologies
/three.lf./two.lf./one.lf Background: Concepts in institutional economics
Before turning to the political consequences of blockchain, it will again be useful to lay
out some background concepts. In this section, I give a quick overview of certain concepts
from new institutional economics , which explore the nature and purpose of institutions from
an economic perspective.
It sometimes happens that parties would /fi.ligand it mutually bene/fi.ligacial to enter into a certain
agreement with each other, but in practice fail to do so. These cases constitute collective
action problems . A classic illustration of a collective action problem is the tragedy of the
commons [/eight.lf/seven.lf]. In this case, the members of a village all have access to the same grazing /fi.ligaeld.
To avoid depleting the grass, the villagers will need to limit their consumption. However,
for any given individual, the cost of limiting one’s own consumption is not o/f\_f.ligaset by the
slight decrease in the rate of depletion. It follows that if everyone is left to their own
devices, then the grass will be collectively depleted. Everyone would be made better o/f\_f.liga by
a village-wide agreement on limited consumption, but the strong individual incentives to
cheat might make such an agreement impractical.
Importantly, many of the world’s most salient problems are at least in part problems of
collective action. For example, the failure of the international community to properly
address global problems like climate change and pandemic risk can be considered highly
analogous to the tragedy of the commons [/one.lf/seven.lf/four.lf]. Mutually harmful wars and arms races are
another large-scale example. Some economists, like Bryan Caplan, have also characterized
political apathy and irrationality as a sort of collective action problem; everyone would be
better o/f\_f.liga if everyone else became more well-informed and epistemically rational, but the
individual bene/fi.ligats of doing so are extremely low [/four.lf/three.lf].
One way to explain the source of collective action problems is to apply the framework of
/four.lf/four.lfAnother proposed technique, in a somewhat di/f\_f.ligaerent vein, is the use of physically tamper-proof cameras
that digitally sign the images they record. The upside of this technique is that it is much easier for users to
implement, while the downside is that it requires trust to be placed in the camera manufacturers and in physical
methods for preventing tampering (which are relatively more ad hoc than cryptographic ones).
/five.lf/zero.lf
transaction costs [/one.lf/six.lf/two.lf]. The realization of each possible agreement is associated with a set
of costs, which can in turn be divided into search costs, bargaining costs, and commit-
ment costs. First, the search costs are associated with /fi.liganding parties to enter the agreement
with and learning whatever information would be necessary to identify the agreement as
mutually desirable. Second, the bargaining costs are associated with converging on the agree-
ment, given the available information, and perhaps formalizing it. Finally, the commitment
costs are associated with increasing the probability that the parties will comply with the
agreement and risking the possibility that they will not. These commitment costs might
include the costs of monitoring the relevant parties, engaging in a potential arbitration
process, or establishing credible threats, as well as the expected costs implied by possible
non-compliance. In addition, in keeping with the above discussion of veri/fi.ligacation (see sec-
tion /three.lf./one.lf./three.lf), there are often transaction costs associated with the need to share extraneous
sensitive information in order to share information that is relevant. These may be classed
as either search costs or commitment costs, depending on whether they precede or follow
the relevant agreements.
In the case of the tragedy of the commons, the work involved in gathering the villagers
together and discussing an agreement not to overgraze would constitute search and bar-
gaining costs. The work involved in keeping watch over the /fi.ligaeld, the work involved in
punishing violators, and the value of any grass lost to undetected overgrazing would be
examples of commitment costs.
If any party’s share of the transaction costs associated with an otherwise mutually desirable
agreement outweighs their share of its bene/fi.ligats, then, given that the party is rational, there
will be a collective action problem.
Social institutions are critical for resolving collective action problems. They can be de/fi.liganed,
rather abstractly, as stable patterns of behavior and expectation within a community. For
example, within a country, the common tendency to accept certain pieces of paper (e.g.,
U.S. dollars) in exchange for goods, and the common expectation that other people will
also accept these pieces of paper in exchange for goods, can be considered a social institu-
tion. Formal organizations like companies and governments can also be considered social
institutions. For example, some of the “patterns” that ultimately give shape to a police
department include: whose orders and which kinds of orders are treated as having force,
who is let into the building and given a uniform and who is not, how people on the street
respond to people in police uniforms, and so on.
The right social institutions can do a great deal to lower transaction costs [/one.lf/four.lf/four.lf]. Tra/f\_f\_i.ligac
rules save drivers from the cost of negotiating for the right of way, money saves businesses
from arranging complex trades, regulation and reputation saves buyers from the cost of in-
vestigating products, and e/f\_f.ligaective police departments save citizens from the myriad costs
of ensuring mutual non-aggression and respect for property. In addition, anthropologists
have found that cooperative norms tend to help small communities avoid most of the costs
entailed by literal tragedies of the commons [/one.lf/two.lf/eight.lf]./four.lf/five.lf
/four.lf/five.lfTypically, as communities grow in size, it becomes increasingly common to rely on institutions that are
formal, hierarchical, and centralized. However, more informal, socially /fl.ligaat, and decentralized institutions always
/five.lf/one.lf
One perspective on blockchain technology is that it supports the creation of new institu-
tions, which might may either compete with existing institutions (to solve the same sorts
of problems) or solve problems not solved by any existing institution.
/three.lf./two.lf./two.lf The roles of banks, technology companies, voting authorities, and other tradi-
tional institutions could shrink
Systems that rely on new cryptographic technologies could begin to supplant existing so-
cial institutions, if they can o/f\_f.ligaer more appealing versions of the services these institutions
provide.
First, one key feature of cryptocurrencies is that they allow users to make virtual payments
without relying on banks or credit card companies to act as intermediaries or on traditional
central banks to manage the underlying currency. If a signi/fi.ligacant number of people /fi.ligand
it appealing from an e/f\_f\_i.ligaciency, liberty, or privacy standpoint to avoid relying on these
institutions when possible, then their roles could shrink.
One early example of a decline in in/fl.ligauence, already discussed, is banks’ and credit com-
panies’ inability to prevent payments to WikiLeaks in /two.lf/zero.lf/one.lf/two.lf when all major companies
decided to refuse payments to the activist group. WikiLeaks supporters decided to simply
cut banks and credit card companies out of the process and donated tens of thousands of
dollars’ worth of bitcoin instead [/one.lf/one.lf/one.lf]. In the same vein, a sign of the potential decline of
central banks’ in/fl.ligauence is given by Venezuela, where a non-negligible number of citizens
have begun to store some of their wealth in cryptocurrencies, rather than the hyperin/fl.ligaated
national currency [/one.lf/two.lf/zero.lf]. The existence of cryptocurrencies like Zcash that use zk-SNARKs
to obscure how much currency each user owns may also pose problems for taxation, beyond
those already posed by the ambiguous status of cryptocurrencies as taxable assets.
The past few years have also seen growing interest in decentralized finance (orDeFi) appli-
cations [/six.lf/one.lf] that build on top of cryptocurrency systems. Examples include decentralized
cryptocurrency exchanges (which allow users to trade cryptocurrencies with one another)
and applications that support collateralized cryptocurrency loans (where the collateral is
another digital asset). These applications might reduce the roles of a number of additional
/fi.liganancial institutions.
Decentralized applications could also be used to reduce reliance on traditional technology
companies to provide services [/three.lf/eight.lf]. For example, there are a number of proposed or early-
stage decentralized applications to allow users to rent out storage space for encrypted /fi.ligales,
using smart contracts, as an alternative to services like Google Drive [/one.lf/seven.lf/six.lf]. If scalability
issues can be resolved (see section /four.lf./two.lf./one.lf), then, some writers have suggested, decentralized
services provided by companies like Uber might eventually become practical [/one.lf/three.lf/seven.lf].
Smart contracts, especially when used in conjunction with smart property, could conceiv-
ably allow people to reduce their reliance on courts and arbitrators when entering into
agreements. As the legal scholar Lawrence Lessig has noted, it is sometimes possible to
retain their importance.
/five.lf/two.lf
achieve a particular end through either legal code or computer code [/one.lf/zero.lf/five.lf]. In this sense,
smart contracts would appear to increase the space of possibilities for what can be achieved
through computer code. If people can more easily and frequently enter into agreements
with one another that they trust will be enforced, without needing to trust that courts
will enforce them, then this would seem to diminish the importance of traditional legal
systems [/one.lf/eight.lf/two.lf]./four.lf/six.lf
At the same time, as section /four.lf./two.lf./four.lf will discuss, smart contracts—or, at least, anything re-
sembling existing smart contracts—appear capable of /fi.ligalling only a sliver of the role cur-
rently /fi.ligalled by traditional contracts. Generally, the potential of decentralized services
might also be undermined by regulation (see section /four.lf./two.lf./two.lf), security issues with the con-
sensus protocols used in permissionless blockchains (see section /four.lf./two.lf./three.lf), or the possibility
that many existing institutions are “good enough” to remove most of the incentive to seek
alternatives (see section /four.lf./two.lf./five.lf).
In addition, there is some case to be made that emerging cryptographic technologies might
also help to bolster traditional centralized institutions. In particular, as discussed in sec-
tion /two.lf./eight.lf, applications of zk-SNARKs could help to increase the accountability of these in-
stitutions. The use of technologies for computing on con/fi.ligadential data might also alleviate
concerns about these institutions violating individuals’ privacy or exploiting their personal
information. The net e/f\_f.ligaect could be to increase the palatability of reliance on centralized
institutions and thereby decrease incentives to seek more decentralized or novel alterna-
tives.
/three.lf./two.lf./three.lf It could become possible to solve collection action problems that existing institu-
tions cannot
Permissionless blockchains and smart contracts may be able to lower some of the “com-
mitment costs” involved in solving collective action problems.
In certain cases, using a smart contract may remove the need to rely on a trusted third party,
who might be inclined to engage in opportunism, or to take other pains to incentivize
honesty. The parties engaged in a chess smart contract, to repeat an earlier example, do
not need to trust the loser to hand money over to the winner, an escrow service to handle
the transfer, or a court to extract compensation for dishonesty. Therefore, giving these
competitors the option of using a smart contract might decrease the overall cost of ensuring
that their agreement will be honored./four.lf/seven.lf
/four.lf/six.lfAs a related framing, contract enforcement mechanisms may be either public (for example, a formal legal
system) or private (for example, Ma/fi.ligaa enforcers or reputational harm) [/eight.lf/one.lf]. The use of smart contracts may
increase the relative prominence of private mechanisms.
/four.lf/seven.lfA secondary e/f\_f.ligaect of smart contracts, in some contexts, may also be to lower bargaining costs. It may
be particularly cheap to reuse sections of code from successful smart contracts, to write smart contracts that
depend on the outputs of other smart contracts, and to create open-ended smart contracts that any number of
parties can easily choose to engage with. As one example, an experimental smart contract-based venture capital
organization, described in section /three.lf./two.lf./four.lf, left room for an inde/fi.liganite number of users to join it. Although it quickly
failed due to a programming mistake, the computer code used to de/fi.ligane the relevant smart contract can be reused
and iterated upon by future organizations of this type. On the other hand, the opportunity to reuse pieces of
previous contracts is certainly far from unique to smart contracts.
/five.lf/three.lf
In other cases, the need for a trusted third party remains. To return to the case of a bet
about the weather, a third party might be employed to input relevant weather information
to the blockchain. In essence, the use of a smart contract here ensures that a certain transfer
of money will occur if a certain report is made, but it does not ensure that the report
will accurately re/fl.ligaect the world. A number of groups, like the Augur team, are working
on consensus and reputation schemes designed to incentivize accurate inputs to smart
contracts [/one.lf/three.lf/zero.lf]. It is still too early, though, to judge how reliable or e/f\_f\_i.ligacient these systems
will be. Existing systems still have very few users and have not been subjected to much
stress-testing. Section /four.lf./two.lf./four.lf discusses the apparent limitations of smart contracts more
thoroughly, with a focus on the ways in which they can introduce new transaction costs as
well.
Here, though, I will walk through a number of domains in which existing institutions
struggle to solve collective action problems. In each case, I will discuss the possibility that
blockchain-associated institutions will be substantially more e/f\_f.ligaective.
Citizens of countries with weak institutions Weak government institutions lead many
countries, especially developing countries, to su/f\_f.ligaer from particularly high transaction
costs. The relevant issues typically include corruption, incompetent government o/f\_f\_i.ligacials
(partly a result of corruption), inconsistent access to services, lack of robust honesty and
reciprocity norms, and so on. High transaction costs associated with such issues have in
turn been linked to low economic growth [/one.lf/two.lf/two.lf]. Notably, even if smart contracts do not
o/f\_f.ligaer advantages over developed world institutions, they could still o/f\_f.ligaer signi/fi.ligacant im-
provements to available institutions outside the developed world. For example, even if
cryptocurrencies are not superior to well-managed /fi.ligaat currencies, citizens of countries
with badly mismanaged currencies (e.g., erratic changes in the rate of in/fl.ligaation) may pre-
fer to rely on cryptocurrencies in some circumstances. Here, the key point is that they
do not need to place their faith in a central bank that might behave irresponsibly or in-
competently. In fact, citizens of certain countries with unreliable /fi.liganancial institutions,
especially citizens of hyperin/fl.ligaation-a/f\_f\_l.ligaicted Venezuela, have taken an unusually large in-
terest in cryptocurrencies [/one.lf/two.lf/zero.lf].
Criminal groups Criminal groups often face barriers to interacting with institutions like
courts, banks, and highly visible reputation systems. It is not possible, for example, to bring
a drug dealer to court for stealing product, or to look up online reviews for hitmen. The
net e/f\_f.ligaect is that groups with an interest in participating in illegal activities are not nearly
so organized or e/f\_f.ligaective as they could be. However, smart contracts may reduce these co-
ordination problems. One paper by Ari Juels, Ahmed Kosba, and Elaine Shi considers the
possibility of “criminal smart contracts,” and describes a range of schemes for using smart
contracts to credibly commit to payments for data leaks, assassinations, and other crimes
[/nine.lf/eight.lf]. Some of these schemes are not yet practical with current smart contract systems. For
example, the scheme for committing to payment for assassinations requires a service that
inputs news to the blockchain as well as a smart contract that is capable of determining
whether a timestamped plan for the assassination, revealed after the event, matches the
news reports. However, some of the other schemes already appear to be relatively practi-
cal, and none appear to face fundamental barriers. If these considerations are combined
/five.lf/four.lf
with the speculative possibility of government surveillance “going dark” (see section /three.lf./one.lf./one.lf),
then it is possible that some forms of illegal activity will become more prevalent in the
future.
Disorganized shared-interest groups It is generally the case, within any given political
system, that only a very small portion of groups with common interests successfully orga-
nize to pursue these interests. For example, private industries often organize to lobby the
government, but the consumers of these industries’ products do not. In/fl.ligauential grassroots
movements have arisen around some issues, like abortion and police violence, but not oth-
ers of comparable concern, like economic policy. The collection of people who would be
happy to become vegetarian if it also meant everyone else did, to pay for news if it also
meant everyone else did, and so on, do not form robust organizations or enter into pacts.
Similarly, even though voters as a whole would have clear common interests in coordi-
nating to become more well-informed, the possibility of such an agreement arising seems
wholly implausible.
These disparities hinge on the fact that, for a su/f\_f\_i.ligaciently large interest group, any individ-
ual member’s contribution to the group’s shared aim will be insigni/fi.ligacant. A single union
member’s dues are unlikely to make a di/f\_f.ligaerence to the union’s bargaining and lobbying
e/f\_f.ligaorts, and a single well-informed voter is unlikely to make a di/f\_f.ligaerence to the quality of
elected o/f\_f\_i.ligacials. If the act of participating in collective action does not bring additional
bene/fi.ligats, and if agreements to participate are di/f\_f\_i.ligacult to enforce, then a rational mem-
ber of a large interest group will refrain from participating. In his classic book The Logic
of Collective Action , the economist Mancur Olson summarizes this point with the thesis,
“Unless the number of individuals in a group is quite small, or unless there is coercion
or some other special device to make individuals act in their common interest, rational,
self-interested individuals will not act to achieve their common or group interests ” [/one.lf/two.lf/three.lf].
According to Olson, many industries are able to organize because they contain only a
relatively small number of decision-making parties, who can direct the employees below
them. Some grassroots movements are able to organize because, due to the nature of their
central issue, membership o/f\_f.ligaers signi/fi.ligacant social bene/fi.ligats, a signi/fi.ligacant personal sense of
meaning, or opportunities to participate in enjoyable activities. Unions in some domains
are able to maintain members due to a mixture of social bene/fi.ligats, control of professional
resources, and opportunities for outright coercion. And so on. The result is an extremely
uneven distribution of power between shared-interest groups, and extremely suboptimal
outcomes for most shared-interest groups—including the group constituted by citizens of
a given country as a whole.
Smart contracts, allowing individuals to credibly commit to collective action if others also
do so, could prove to be a useful tool for many interest groups. Groups interested in fund-
ing lobbying e/f\_f.ligaorts could use smart contracts that result in their making donations only
if enough other people do, too, and perhaps scaling down the donation size as the number
of people who commit grows. The most challenging problem seems to be verifying agree-
ments to take action in the physical world. It is not immediately obvious, for example, how
to credibly commit to attend a protest, become a well-informed voter, or become a vege-
/five.lf/five.lf
tarian. Peer veri/fi.ligacation networks, where members of groups examine evidence produced
by other members, seem possible. These systems could also be engineered to include repu-
tation systems and incentives for honesty. However, each additional unit of human labor,
unit of honesty-inducing incentives, and unit of complexity adds to the total transaction
cost. Plausibly, the total transaction cost could often remain too high to enable collective
action.
It is also not clear that blockchain would resolve the central barrier, whatever it is, that is
preventing these kinds of systems from arising. It is certainly possible to imagine a cen-
tralized Kickstarter-like company that, for example, commits people to make donations
to lobbying groups only if enough other people do too. If the company’s bottom-line de-
pended on the trust of its users, and if it had legal obligations to ful/fi.ligall its commitments,
then a reasonable person should expect the company to follow through and transfer the
funds in an appropriate way. Whatever is preventing these kinds of services from arising,
therefore, might be something other than the need for trusted third parties.
States A central concern in international relations is the relative ine/f\_f\_i.ligacacy of interna-
tional institutions [/nine.lf/three.lf]. Theorists often describe the interactions between countries as
taking place under “anarchy,” since there is no supreme authority capable of /fi.ligalling the
same institutional role that states /fi.ligall at the national level. Many of the greatest problems
in international relations, and therefore many of the largest-scale problems in general,
appear to follow from excessive commitment costs. For example, the possibility of “free-
riding” makes it di/f\_f\_i.ligacult for states to collectively agree to take costly measures to alleviate
environmental issues and global risks, such as those associated with climate change. The
fact that a rising power cannot credibly commit to non-aggression, and the fact that two
mutually hostile nations cannot commit to halting military investment, open the door for
preventative war and arms races [/one.lf/three.lf/three.lf]. In addition, the need to establish the credibility of
one’s commitments, in the absence of other mechanisms, has often been cited as at least a
reason for participating in prolonged wars; the Vietnam War is a notable case. The number
of lives lost, risked, and harmed because of an absence of good tools for making credible
international commitments would be di/f\_f\_i.ligacult to overestimate./four.lf/eight.lf
Therefore, given the stakes, the possibility that smart contracts could have applications
in the international sphere is very much worth investigating. We can construct at least
hypothetical commitment mechanisms that take advantage of smart contracts.
For example, smart contracts might enable countries to more credibly commit to interna-
tional agreements that include penalties for non-compliance. Traditionally, agreements
that include /fi.liganancial penalties, such as the Kyoto Protocol, rely on “self-punishment”
[/nine.lf/five.lf]. Countries that fail to comply with core aspects of the agreement must still, nev-
ertheless, be trusted to comply with the portion that obliges them to pay out large sums of
money. A number of scholars have suggested that the obvious pitfalls associated with “self-
punishment” can be avoided through the use of escrow, with parties to agreements making
initial deposits that are returned only if they demonstrate compliance [/seven.lf/three.lf, /one.lf/one.lf/two.lf]. Smart
/four.lf/eight.lfAt the same time, the ability to make credible commitments is far from being entirely benign [/one.lf/four.lf/two.lf]. Among
other downsides, the ability to credibly commit to threats can increase the ease of extortion.
/five.lf/six.lf
contracts could o/f\_f.ligaer a method for implementing escrow for international agreements,
without the need for a trusted third party to hold and disburse the deposits. In particular,
countries might sign on to a smart contract consenting to the loss of large deposits if a
quorum of other countries, a distributed oracle system, or a set of internet-connected sen-
sors judge that they have violated the agreement; the relevant blockchain could be either
a consortium blockchain held between several countries or a particularly well-established
permissionless blockchain.
As another example, we might also imagine a pair of countries signing onto a smart con-
tract that will deactivate their smart property weapons systems at the same time. The
additional di/f\_f\_i.ligaculty here, though, is the need to /fi.ligarst verify that the weapons systems re-
spond to the state of the blockchain in the appropriate way.
Unfortunately, to be used in such cases the relevant smart contracts would need to become
exceptionally reliable. As section /four.lf./two.lf./one.lf will discuss, this may not be reasonable to expect.
These potential applications of smart contracts to the international sphere are ultimately,
in my view, unlikely. However, if they ever do emerge, then they would probably be the
absolute most consequential applications of smart contracts.
As a /fi.liganal note, this discussion has a close connection to the earlier discussion of the value of
non-intrusive agreement veri/fi.ligacation in the international sphere (see section /three.lf./one.lf./three.lf). Non-
intrusive agreement veri/fi.ligacation can be seen as another tool for lowering commitment
costs, speci/fi.ligacally the costs associated with submitting to monitoring. In the most ideal-
ized case, a state might be forced to reveal no information beyond the simple fact of its
compliance. Likewise, non-intrusive inspection, more generally, can be seen as a tool for
lowering search costs, in that it may reduce any disadvantages states accrue when they
share information that is relevant to bargaining. It is di/f\_f\_i.ligacult, for example, to impress
upon another party the power of one’s cyberweaponry without revealing details that will
make these weapons less e/f\_f.ligaective [/one.lf/zero.lf/six.lf, /one.lf/zero.lf/one.lf]. In general, as James Fearon writes in his classic
paper “Rationalist Explanations for War,” “there is a trade-o/f\_f.liga between revealing informa-
tion about resolve or capabilities to in/fl.ligauence bargaining and reducing the advantages of a
/fi.ligarst strike” [/six.lf/five.lf]. The existence of information asymmetries, due to incentives to maintain
private information, along with commitment problems, are the two most powerful expla-
nations for how mutually harmful wars (and other collective action problems) can arise
even between rational actors.
As with smart contracts, it may not be especially likely, even in the long run, that non-
intrusive inspection will signi/fi.ligacantly lessen international collective action problems. Some
of the most relevant private information—such as the “resolve” of a given state in the lead-
up to a potential war—does not seem to lend itself naturally to tricks with zero-knowledge
proofs and secure multiparty computation. Again, though, the existing problems are of
great enough magnitude that any tool with potential value deserves consideration.
/five.lf/seven.lf
/three.lf./two.lf./four.lf A new variety of institutions, known as “decentralized autonomous organizations,”
could emerge
Some writers have speculated that smart contract technology could allow a new variety
of institutions to emerge. In particular, especially within the Ethereum developer com-
munity, a large number of essays and blog posts have been written on the possibility of
decentralized autonomous organizations. Although de/fi.liganitions vary, we will de/fi.ligane a de-
centralized autonomous organization (DAO) as an organization, associated with its own pool of
assets, in which decisions about how to use these assets are determined by smart contracts
among the organization’s members [/three.lf/six.lf]./four.lf/nine.lf
The /fi.ligarst and still most famous example of a DAO has been a /dollar.lf/one.lf/five.lf/zero.lf million venture capital
fund, known as “the DAO,” which launched on Ethereum in /two.lf/zero.lf/one.lf/six.lf [/one.lf/three.lf/two.lf]. Investors partic-
ipated by buying shares in the fund, then using their shares to vote on how to disburse
the fund’s common pool of cryptocurrency to proposed projects; smart contracts ensured
that this money would be disbursed in accordance with the votes and that pro/fi.ligats from
the projects would be distributed appropriately among the shareholders. Unfortunately,
the DAO came to an untimely end. Within a few days, a user discovered a programming
mistake that allowed them to siphon the majority of funds out of the organization, and,
given that smart contracts are by nature impossible to modify, the project was abruptly
abandoned.
A number of later projects have also described themselves as DAOs [/one.lf/eight.lf/three.lf]. For instance,
similar to “the DAO,” Moloch DAO uses a smart-contract-based system to automatically
disburse cryptocurrency investments if members express enough support for an invest-
ment proposal [/one.lf/six.lf/nine.lf]. Decisions to admit new members are also determined through voting
and processed automatically. Unlike “the DAO,” Moloch DAO does not attempt to make
a pro/fi.ligat. It simply disburses member-provided funds to other blockchain projects that
the community believes are socially valuable and has no mechanism to collect anything
in return. Smart contracts are primarily used to ensure that votes concerning member-
ship expansion and currency disbursement are binding. A failsafe mechanism, designed in
response to the collapse of “the DAO,” also allows members to extract their share of the
common funding pool if they strongly oppose a just-concluded decision concerning the
disbursement of funds.
The few DAOs that exist today are still quite new and, in some ways, fairly rudimentary.
Therefore, given this limited set of case studies, it is di/f\_f\_i.ligacult to draw meaningful gener-
alizations. Still, DAOs do seem to di/f\_f.ligaer from most existing organizations in at least four
interesting ways.
First, DAOs can possess a foundational set of features that is immutable. No party—for
instance, no “leader” of a DAO—can prevent the relevant smart contracts from executing
in the speci/fi.ligaed way. The case of “the DAO” demonstrates the downsides of this property,
but, in the right cases, it can also be considered desirable. DAOs are especially capable
/four.lf/nine.lfSome alternative de/fi.liganitions are loose to the extent that permissionless blockchains themselves constitute
DAOs. Under this conception, the nodes maintaining them act as their “members” and, through cryptocurrency
rewards, share in the “pro/fi.ligats” earned in the process of providing users with a “service.”
/five.lf/eight.lf
of survival and resistance to change, even if there is a complete turnover of members or if
future members are unanimously interested in altering some of their fundamental features.
This steadfastness, even in the face of an opposed membership, is how DAOs earn the
descriptor “autonomous.”
Second, DAOs can be truly transnational. Since they are maintained on blockchains, with
the relevant nodes most likely spread across dozens of countries, they will not have any
particular physical instantiations. In addition, no state can disrupt a DAO—or, at least,
its fundamental features—without doing the costly and potentially unpopular work of
disrupting the blockchain itself. (See section /four.lf./two.lf./three.lf for a discussion of the possibility of
such a disruption.)
Third, depending on the nature of the DAO and the blockchain it is maintained on, indi-
viduals may be able to participate under pseudonyms (typically, the hashes of their public
keys). This form of pseudonymity, while not unheard of, is still rather unusual. The DAO,
for example, was almost certainly the /fi.ligarst-ever pseudonymous eight-/fi.ligagure venture capital
fund.
Fourth, DAOs place an unusually heavy emphasis on complete contracts , or agreements
which precisely specify the obligations of each party for every possible case [/five.lf/six.lf]. Smart
contracts, because they need to be expressed in code, are by nature complete. In tradi-
tional organizations, however, contracts tend to be highly incomplete. In fact, a widely
held theory, associated with the /fi.ligaeld of new institutional economics, is that individuals
form organizations in large part because of the impracticality of specifying complete con-
tracts for the services they provide one another (see section /three.lf./two.lf./one.lf for a discussion of transac-
tion costs) [/one.lf/eight.lf/zero.lf]. Particularly within organizations, interactions tend to be dominated by
relational contracts , which are unwritten cooperative norms that resist formalization [/one.lf/zero.lf].
This body of theory suggests that DAOs are unlikely to be able to e/f\_f\_i.ligaciently replicate the
functionality of most organizations, unless they also rely heavily on more informal rela-
tional contracts in addition to smart contracts. Norms clearly play an important role in
the functioning of Moloch DAO, for instance, since everything other than the enforcement
of votes is managed without smart contracts. One plausible guess we can make about fu-
ture DAOs, at least successful ones, is that they will still mostly be built on top of informal
relational contracts. Nonetheless, even a modestly greater reliance on complete contracts
might lead to interesting structural di/f\_f.ligaerences.
It is interesting to speculate about what future DAOs could look like, conditioning on the
optimistic view that they are in fact plausible outside of a narrow range of cases. What
could a DAO tech company look like? How about a DAO political party, a DAO criminal
organization, or, more fancifully, a DAO state?
Some writers have made quite radical claims about the feasibility of replacing traditional
political systems with DAOs. However, these claims are often ambiguous. As an example,
I will brie/fl.ligay consider the ideas of Ralph Merkle, the inventor of cryptographic hashing,
who has written a paper describing a system of government he calls “DAO democracy”
[/one.lf/one.lf/three.lf].
/five.lf/nine.lf
Merkle’s basic scheme, as described in his paper, is to implement a variant of “futarchy,”
a system of government /fi.ligarst described by the economist Robin Hanson [/eight.lf/six.lf]. In Merkle’s
system, there are annual citizen satisfaction polls, and all citizens are allowed to place bets
on the impact proposed sets of legislation would have on total satisfaction; the sets of leg-
islation that this betting market indicates are most likely to succeed are implemented. To
help secure the integrity of the betting markets, the bets are to be recorded on a consor-
tium blockchain, with the voting power granted to each device being used to maintain the
blockchain determined by further bets about the device’s reliability. Putting aside ques-
tions of this scheme’s practicality, however, it is unclear to what extent it would constitute
a DAO, or whether it would truly require novel cryptographic technology. Merkle makes
no suggestions about the use of smart contracts in the actual implementation of legislation,
and while the use of a consortium blockchains to maintain the betting records seems like
a useful way of increasing their integrity, it also seems somewhat tangential to the overall
vision.
Other writers have explored the argument that blockchains could be a useful tool for ex-
perimenting with novel forms of democracy [/three.lf/seven.lf, /one.lf/five.lf/seven.lf]. Futarchy is often discussed, as is
“liquid democracy,” a system of representative democracy in which individuals can choose
to grant anyone else the power to vote on their behalf for particular sets of decisions; this
representative might in turn pass their accumulated voting power on to a representative
they deem to be even more well-equipped to vote well [/two.lf/three.lf]. Both of these forms of democ-
racy must almost certainly be implemented electronically and require uni/fi.ligaed databases
that are extraordinarily secure. The databases must be trusted to track, for example, active
updates concerning who has the power to vote for whom. Blockchain technology, perhaps
in conjunction with secure multiparty computation, is seen as signi/fi.ligacantly lowering the
trust threshold for implementing these political systems and making it relatively easy for
small (and potentially geographically dispersed) political organizations to trial them. If the
outcomes of votes are more directly linked to tangible outcomes through smart contracts,
then these organizations could, less ambiguously, constitute DAOs.
A number of writers within the blockchain community have suggested that there may
arise DAOs that serve citizens’ needs so su/f\_f\_i.ligaciently that traditional states simply wither
away, or that DAO-based governance will lead people’s political relationships to become
almost entirely decoupled from geography. These are more extreme versions of the visions
discussed in sections /three.lf./two.lf./two.lf and /three.lf./two.lf./three.lf. Marcella Atzori’s paper, “Blockchain Technology and
Decentralized Governance: Is the State Still Necessary?” sympathetically surveys some such
claims, although she ultimately (I believe correctly) /fi.ligands them implausible [/nine.lf].
Finally, we should note that although most current speculation about the political impli-
cations of DAOs seems to be associated with positive visions, it is also possible to imagine
negative outcomes. Intuitively, criminal organizations and terror groups, which may be at
risk of being “beheaded” through the arrest or killing of their leaders, would have incentive
to transform into DAOs if this was feasible. As an example, we can imagine a DAO that
is set up to create “criminal smart contracts” for acts of politically motivated sabotage or
terrorism, using a pool of money contributed by anonymous supporters. (See section /three.lf./two.lf./three.lf
for a brief discussion of criminal smart contracts.)
/six.lf/zero.lf
It must be reiterated, however, that the basic practicality of DAOs, for almost any use case
at all, is still unproven. Smart contracts simply might not be very useful for governing re-
lationships within an organization. This discussion, then, should be regarded as especially
speculative. As section /four.lf./two.lf./four.lf will explore, some very signi/fi.ligacant limitations stand in the
way of the most radical proposed uses of smart contracts becoming feasible.
/six.lf/one.lf
/four.lf Limitations and skeptical views
There are some important roadblocks that could prevent several of the technologies we
have discussed from achieving widespread use or achieving transformative e/f\_f.ligaects. In this
section, I discuss six such roadblocks, which I judge to be particularly signi/fi.ligacant: the in-
e/f\_f\_i.ligaciency of methods of computing on con/fi.ligadential data, the di/f\_f\_i.ligaculty of “scaling” block-
chains, the threat of regulation; the inadequacies of smart contracts, the potential insecu-
rity of permissionless blockchains, and the possibility that existing institutions are “good
enough.”
Other potential roadblocks, which I do not discuss here, include the arrival of quantum
computers (which will render some cryptographic schemes insecure), the “just in case”
data collection practices of many companies (which might make them hesitant to adopt
technologies that reduce or more precisely target their data collection), limited consumer
demand for privacy technologies, and the enormous volume of electricity consumed by
proof-of-work protocols (which could be made unnecessary by a successful shift to proof-
of-stake protocols).
/four.lf./one.lf Limitations of privacy-preserving technologies
/four.lf./one.lf./one.lf The ine/f\_f\_i.ligaciency of computing on con/fi.ligadential data
All methods of running computations on encrypted data—including homomorphic en-
cryption, functional encryption, and secure multiparty computation—add signi/fi.ligacant over-
heads compared to the time and space requirements of computing on unencrypted data.
Although the overheads of secure multiparty computation, especially, are beginning to
become manageable, they may never be entirely negligible. To justify their use, the de-
mand for greater privacy must be enough to o/f\_f.ligaset the economic costs of using MPC. If the
demand is not great enough, then it may never see very broad use.
More troublingly, all known schemes for fully (and somewhat) homomorphic encryption
are so ine/f\_f\_i.ligacient that, at present, it is not possible to use them for anything but extremely
simple computations. This, for example, rules out the possibility of using them to replace
nearly any online service with a privacy-preserving version (see section /three.lf./one.lf) anytime soon. If
fully homomorphic encryption is to /fi.ligand widespread use, there will need to be either major
progress in discovering more e/f\_f\_i.ligacient schemes or a many orders-of-magnitude increase
in available computing power. Still, neither of these possibilities is entirely implausible
due to trends in both the e/f\_f\_i.ligaciency of fully homomorphic encryption and the growth of
computing power [/one.lf/one.lf/seven.lf].
On the other hand, as discussed in section /two.lf./one.lf/three.lf, it may be possible to use secure multiparty
computation to replicate most of the functionality that fully homomorphic encryption can
o/f\_f.ligaer. The extreme ine/f\_f\_i.ligaciency of fully homomorphic encryption could ultimately turn out
to be largely irrelevant.
/six.lf/two.lf
/four.lf./two.lf Limitations of blockchain-based technologies
/four.lf./two.lf./one.lf The di/f\_f\_i.ligaculty of “scaling” permissionless blockchains
As mentioned in section /two.lf./six.lf, the most commonly used permissionless blockchains, such as
Bitcoin and Ethereum, cannot process more than a dozen or so transactions per second. In
comparison, a company like Visa can process tens of thousands of transactions per second
by using ordinary servers.
Low transaction speeds limit the number of users a blockchain can plausibly sustain, as
well limiting the complexity of the services a blockchain can provide [/five.lf/four.lf]. Almost cer-
tainly, the most optimistic visions of blockchain’s potential—such as the vision that de-
centralized cryptocurrencies will signi/fi.ligacantly compete with /fi.ligaat currencies (see section
/three.lf./two.lf./two.lf), the vision that decentralized applications will replace a large portion of traditional
online applications (see section /three.lf./two.lf./two.lf), or the vision that smart contracts will enable the
creation of vast new political entities (see section /three.lf./two.lf./four.lf)—can only be realized by permis-
sionless blockchains with much higher processing speeds. Low processing speeds can also
indirectly limit the security of permissionless blockchains by limiting the total value of
transaction fees collected by actors maintaining them: typically, the more wealth one can
collect from transaction fees, the greater the incentives to maintain a given blockchain will
be.
The most fundamental source of the problem, here, is that traditional blockchains such as
Bitcoin require all of the nodes maintaining the blockchain to process each transaction,
rather than splitting the labor between nodes. Such blockchains do not scale, in the sense
that adding more processors or faster processors to the network does nothing to increase
the system’s overall processing speed. The system cannot process transactions any faster
than the slowest computer (also taking into account its network connection) that is capable
of serving as a node. The need for nodes to communicate with one another and come to
consensus then introduces further delays.
There are two main strategies for addressing the scaling problem. The /fi.ligarst strategy is to
implement sharding [/six.lf/zero.lf]. This involves splitting a blockchain into multiple shards, with
each shard handling a subset of transactions. Then, some nodes can opt to be involved in
only a single shard. The more shards there are, the lower the minimum burden is on each
node.
There are some obvious challenges that any successful sharding scheme must overcome:
developers must prevent successful attacks on individual shards, must allow the shards
to interface with each other, and must generally ensure security and consistency despite
division. A great deal of research within the blockchain community is currently devoted
to overcoming these challenges.
The other strategy is to pursue layer /two.lf scaling . This involves outsourcing a large portion
of necessary computations to external processors rather than requiring all of the nodes
maintaining the blockchain to perform them [/six.lf/two.lf]. State channels allow groups of users
to make sequences of transactions among themselves and then only publish the /fi.liganal re-
/six.lf/three.lf
sult to the blockchain. For example, if ownership of a token is transferred multiple times
within a short interval, then state channels could remove the need for blockchain nodes
to process the intermediate transactions. Similarly, rollups allow external computers to
process batches of transactions “o/f\_f.liga-chain” and then record only the /fi.liganal state update to
the blockchain. Of course, to avoid a large loss in security, there needs to be some way to
verify that these kinds of external computations have been performed correctly. Succinct
cryptographic proofs known as zk-SNARKs (see section /two.lf./eight.lf) are one important tool for
supporting veri/fi.ligacation. These proofs, which can be published to the blockchain, allow
the nodes maintaining a blockchain to verify that some computation has been carried out
correctly without needing to rerun the full computation themselves.
A number of teams are currently working to upgrade the Ethereum blockchain by devel-
oping and implementing a mixture of sharding and layer /two.lf scaling techniques. The most
optimistic community members expect these upgrades to allow Ethereum to process tens
of thousands of times more transactions per second. It remains to be seen, though, just
how successful these approaches to the scaling problem will be.
There are also some projects that have pursued performance gains without addressing the
fundamental scaling problem. In general, these projects have set higher minimum perfor-
mance standards for participating nodes, have reduced certain ine/f\_f\_i.ligaciencies involved in
decentralized consensus protocols, or have done some combination of the two. Litecoin is
an early and well-known example of a blockchain project that has achieved modest perfor-
mance gains by raising the minimum performance requirements for individual nodes. To
explain, blockchain protocols typically aim for a roughly constant rate of block creation
and typically place a hard limit on how many transactions can be included in a single block.
These two design constraints then imply a limit on the rate at which transactions can be
processed. For example, because the Bitcoin blockchain gains an average of six new blocks
per hour, and these blocks are each only allowed to contain a few thousand transactions,
there is no way for Bitcoin to process more than a dozen or so transactions per second.
Litecoin and a number of other blockchains have achieved higher processing speeds, rel-
ative to Bitcoin, partly by using larger “block sizes.” However, increasing the block size
has also raised the minimum performance standards that nodes must meet to pro/fi.ligatably
participate in maintaining the blockchain./five.lf/zero.lf
A more recent and promising example of a blockchain project that eschews sharding and
layer-/two.lf solutions is Solana. Solana’s designers have developed a consensus protocol that re-
quires an unusually small amount of communication and produces agreement between the
participating nodes unusually quickly [/one.lf/eight.lf/four.lf]. The overall speed of the blockchain is there-
fore not much slower than the speed at which an individual node can process transactions.
Due to comparatively high minimum performance standards for nodes, it is reportedly
able to process tens of thousands of transactions a second. As the Solana project is still in
/five.lf/zero.lfSince there is always some cost involved in maintaining a blockchain, nodes are only likely to earn a pro/fi.ligat
if they are able to process transactions at the maximum rate that the blockchain allows. A slow node that tries
to participate will only be able to process a fraction of the transactions that faster nodes can. It will therefore
only be able to collect a fraction of the total transaction fees that these other nodes can. Ultimately, one should
expect nodes that fall below some minimum speed threshold to drop o/f\_f.liga the network.
/six.lf/four.lf
its “beta phase,” there is some lingering uncertainty about the system’s security and long-
term stability. Nonetheless, at the time of writing, it is perhaps the fastest decentralized
blockchain system.
One general observation is that raising the minimum performance standards for a block-
chain’s nodes will typically allow the blockchain to support more computationally inten-
sive applications, but will also, typically, reduce the number of actors who are capable
of helping to maintain or audit the blockchain. Projects such as Solana therefore must
navigate a fairly direct trade-o/f\_f.liga between performance and centralization. This trade-o/f\_f.liga
should be less sharp for blockchain projects that attempt to address the fundamental scal-
ing problem by relying on techniques such as sharding and rollouts since these projects
achieve performance gains primarily by reducing the amount of computation that individ-
ual nodes are asked to do.
Time will tell whether any of these strategies—or some combination of them—will be
enough to enable applications that are both highly decentralized and highly performant.
However, the past few years of research progress seem to support an at least moderate
degree of optimism.
/four.lf./two.lf./two.lf The threat of restrictive regulations
A simple way to limit the use of a technology is to regulate it.
In the case of public-key cryptography, the oldest of the technologies we have discussed
here, there is a long history of countries deliberating on how best to regulate it. As re-
cently as the /one.lf/nine.lf/nine.lf/zero.lfs, for example, law enforcement agencies in the United States were ar-
guing that the use of encryption that they could not themselves decrypt should be made
illegal [/one.lf/one.lf]. For the time being, all forms of encryption are perfectly legal to use within
the United States and European Union, with most other major countries placing only rel-
atively limited restrictions on use [/one.lf/four.lf/zero.lf]. However, it is not certain that the status quo will
never change. Especially in the wake of terror attacks or other catalyzing events, it is not
uncommon for law enforcement o/f\_f\_i.ligacials or politicians to re-propose restrictions on cryp-
tography, particular the use of end-to-end encryption [/four.lf/four.lf, /nine.lf/two.lf]. A fairly recent development
is China’s decision to block the use of WhatsApp, apparently based on WhatsApp’s use of
end-to-end encryption [/three.lf/zero.lf].
In the case of more novel cryptographic technologies, such as cryptocurrencies, legal sta-
tuses are in something of a state of /fl.ligaux, with regulations varying substantially by country
and by state [/four.lf/six.lf]. Regulations tend to focus on points of contact between permissionless
blockchains and the outside world—for example, on businesses that exchange traditional
currency for cryptocurrency—due in large part to the fact that these blockchains are inher-
ently di/f\_f\_i.ligacult to interfere with and lack any discernible party that is in “control” of them
[/one.lf/six.lf/eight.lf, /one.lf/zero.lf/two.lf]. For instance, know-your-customer (KYC) laws, which compel cryptocurrency ex-
changes to record and verify the identities of their customers, have become increasingly
common.
It is unclear to what extent the regulation of emerging cryptographic technologies is likely
/six.lf/five.lf
to limit their use, especially given that blockchains themselves are so di/f\_f\_i.ligacult to inter-
fere with. Still, if the technologies ever become truly threatening, for instance by creating
/fi.liganancial instability or by making it easier for terror groups to operate, then dramatic ac-
tions by major governments are not inconceivable. A simple action in this category might
be banning cryptocurrency exchanges, making it di/f\_f\_i.ligacult for individuals to purchase cryp-
tocurrency or to “cash out” by exchanging their cryptocurrency for traditional currency.
States may also directly ban the use of cryptocurrencies or the use of certain decentral-
ized applications. Although a complete ban would probably be very di/f\_f\_i.ligacult to enforce,
it should at least be possible, for example, to prevent large domestic businesses from ac-
cepting cryptocurrency payments./five.lf/one.lfThe most extreme suppression strategy might be to
attempt to “take over” proof-of-work blockchains by directing large amounts of computing
power toward cryptocurrency mining, while attempting to disrupt other mining groups.
(See section /four.lf./two.lf./three.lf for a discussion of the potential dynamics of a take-over attempt.) E/f\_f.ligaorts
to vigorously regulate these technologies, rather than suppress them, might also introduce
various frictions or limitations that reduce their appeal.
It seems reasonable to speculate that if regulation signi/fi.ligacantly in/fl.ligauences the use of crypto-
graphic technologies, it will primarily limit the uses that increase the di/f\_f\_i.ligaculty of surveil-
lance (see section /three.lf./one.lf./one.lf), lessen the in/fl.ligauence of economic and political institutions like cen-
tral banks (section /three.lf./two.lf./two.lf), make it easier for threatening actors to coordinate (section /three.lf./two.lf./three.lf),
or enable the creation of new decentralized political actors (section /three.lf./two.lf./four.lf). In short, the
uses of cryptography that are most desired by the libertarian and anarchist portions of the
cryptography community may also be the ones that are most di/f\_f\_i.ligacult to achieve.
/four.lf./two.lf./three.lf The potential insecurity of permissionless blockchains
As discussed in section /two.lf./six.lf./three.lf, permissionless blockchains are maintained through fairly
complex consensus protocols. In short, they work by allowing any user to participate in
the maintenance of the blockchain, granting these users voting power over the blockchain’s
contents in proportion to the amount of some scarce resource they own (such as computing
power), and incentivizing them to vote honestly by making it very likely that they will earn
digital currency if they do (or lose digital currency if they do not).
There are two interrelated problems with these protocols: First, there may be a natural
tendency for large portions of scarce resources to eventually end up in the control of a very
small number of users. Second, especially for users who control large portions of scarce
resources, there may be ways to earn money or achieve desirable outcomes other than by
helping to maintain the blockchain honestly [/one.lf/one.lf/nine.lf].
We will /fi.ligarst consider the case of the Bitcoin blockchain, which is by far the most well es-
tablished. Bitcoin, as a reminder, uses a proof-of-work protocol that forces users known as
“miners” to demonstrate their ownership of computing power by solving resource-intensive
puzzles.
/five.lf/one.lfA handful of states, including Pakistan and Vietnam, have banned cryptocurrency ownership and use out-
right [/one.lf/five.lf/three.lf]. At the time of writing, India is also considering a complete ban [/three.lf]. So far, though, there is not much
evidence for the e/f\_f\_i.ligacacy of bans. In Vietnam, for instance, /two.lf/one.lf% of respondents in a /two.lf/zero.lf/two.lf/one.lf survey reported owning
or using cryptocurrencies [/three.lf/three.lf]. This is the second-highest rate in the world.
/six.lf/six.lf
As of /two.lf/zero.lf/two.lf/one.lf, four mining collectives collectively controlled more than /five.lf/zero.lf% percent of the
computing power directed at mining Bitcoin [/one.lf/five.lf/five.lf]. If they decided to collude to “dou-
ble spend” coins, then they could produce a dishonest version of the Bitcoin blockchain,
missing records of their previous expenditures, that outpaces the honest one.
Mining operations may have a natural tendency to become centralized in this way, as
races to develop and buy up specially built systems for solving puzzles can quickly be-
come prohibitively expensive for all but a few parties. The need for mining groups to
insure themselves against unlucky streaks can then provide a further incentive for them to
merge.
Furthermore, at least for Bitcoin, it is apparently not the case that more than half of the
relevant computing power needs to be directed in a dishonest way for a dishonest version
of the blockchain to win out. Researchers have identi/fi.ligaed a strategy that colluding Bitcoin
miners with only /two.lf/five.lf% control could use to springboard themselves into majority control
and begin to take self-enriching actions, like spending individual coins multiple times
[/six.lf/four.lf].
If other users become aware that a given blockchain has been subjected to an attack of this
sort, as would almost certainly happen, one plausible result is that the associated cryp-
tocurrency would have its exchange value abruptly drop. This potential fallout might be
enough to keep miners from colluding, even if they would otherwise ostensibly stand to
gain, since the value of their accumulated bitcoin and their investments into specialized
mining hardware depend on the exchange rate for the coin.
However, this incentive-based safeguard would not necessarily be enough for parties that
wish to disrupt the blockchain for reasons other than simple /fi.liganancial exploitation. For
instance, if the government of a large country really wanted to disrupt Bitcoin, it would
need to invest in the computing power necessary to gain majority control.
In such a case, if the honest parties using a blockchain come to recognize that the major-
ity of computing power is being directed dishonestly, which should in practice be quite
conspicuous, then they can create what is known as a forkof the blockchain [/one.lf/one.lf/nine.lf]. This is
accomplished by having a large portion of honest users update their software to disregard
blocks now known to have been proposed by dishonest users, building on top of blocks
further back in the chain’s history. The “fork” is a new, diverging blockchain that, if socially
acknowledged, can /fi.ligall the role of the initial one.
However, in the event of a fork, the dishonest users could still “spawn camp” by adopting
new public-key pseudonyms and using their computing power to once again take control of
the newly forked blockchain. The honest users might, as a further response, update to a new
software version whose mining puzzles the attackers’ hardware is less suited for—but, even
if this move is taken, attackers with su/f\_f\_i.ligacient resources could still make the blockchain
unusable. Repeatedly disrupting Bitcoin would be an expensive venture—likely costing
billions of dollars—but could in principle be done./five.lf/two.lf
/five.lf/two.lfAs a further point, a highly empowered attacker (such as a national government) might also be able to
/six.lf/seven.lf
In short, like any other system of storing data, Bitcoin is not invincible. Its security can
be thought of as roughly proportional to the total computing power devoted to mining, as
the greater this number is, the more money miners have to lose by tanking the blockchain
and the more money another attacker would need to spend to gain enough computing
power. If insu/f\_f\_i.ligacient computing power is invested, then Bitcoin—and other blockchains
using proof-of-work—will remain insu/f\_f\_i.ligaciently reliable to use for any truly vital applica-
tions.
Proof-of-stake blockchains have similar limitations but are perhaps more promising. For
them, each node’s voting power is made proportional to the quantity of cryptocurrency
it “deposits.” Then, the option of applying cryptocurrency penalties to parties that vote
dishonestly can help strengthen incentives, and the possibility of forking to completely
disregard the currency previously owned by an attacking party can help to remove the
possibility of “spawn camping.”
For proof-of-stake systems, security is roughly proportional to the value of the relevant
cryptocurrency deposits, so it can also be expected to increase as more value is tied up
in the relevant blockchain. It follows that the potential security that can be o/f\_f.ligaered by a
given proof-of-stake blockchain is linked to its scalability—in other words, as discussed
in section /four.lf./two.lf./one.lf, whether it can be made to accommodate a much larger number of users
and transactions. Plausibly, a su/f\_f\_i.ligaciently scalable proof-of-stake blockchain—like what
Ethereum aims to become—could be extremely secure. However, proof-of-stake protocols
are still too new for their superiority to proof-of-work protocols to be clear. The relative
merits of proof-of-stake systems and proof-of-work systems are still debated within the
blockchain community.
In summary, it is not yet clear exactly how much security permissionless blockchains can
o/f\_f.ligaer. If new weaknesses are discovered, or existing weakness become more salient with
time, then many users may not feel comfortable using permissionless blockchains for cer-
tain purposes. For example, even once-in-a-decade failures could be enough to heavily
disincentivize the use of cryptocurrencies as stores of value.
/four.lf./two.lf./four.lf The inadequacies of smart contracts
As discussed in section /three.lf./two.lf./three.lf, one way to conceptualize smart contracts is as a tool for
reducing the “transaction costs” associated with entering into agreements.
To return once more to an illustrative example, a chess smart contract can allow two parties
to agree to have whoever loses the game pay the winner, without the need to establish trust
between the parties, to risk their non-compliance, or to enlist the services of a third-party
enforcer. Since there is a wide range of cases where existing institutions leave transaction
costs quite high—with, for example, ine/f\_f\_i.ligacient legal systems being a common problem—it
is a natural hypothesis that smart contracts will be helpful for some of these cases.
Section /three.lf./two.lf./three.lf lists, at a rather abstract level, a number of ways in which smart contracts
decrease honest miners’ power by shutting down mining facilities, restricting the sale of specialized hardware, or
applying other controls.
/six.lf/eight.lf
Search costs Bargaining costs Commitment costs
•Verifying smart
contract traits
•Establishing
con/fi.ligadence in smart
property
•Learning about
cryptographic tools•Formulating complete
contracts
•Translating contracts
into computer code•Paying fees to full
nodes
•Paying fees to
information
services
•Protecting against
coercion and theft
of digital assets
•Risking failure of
smart contract to
function as
expected
Table /five.lf: New transaction costs associated with smart contract use
may reduce transaction costs. The section then considers the possibility that, on this basis,
they may be able to solve some collective action problems that existing institutions either
cannot solve or can only solve relatively ine/f\_f\_i.ligaciently.
At the same time, it is important to note that smart contracts introduce new transaction
costs of their own. These costs imply that using smart contracts may also, in many cases,
be a far inferior alternative to existing institutions. I will consider search costs, bargaining
costs, and then commitment costs.
Concerning search costs, there will be the cost of verifying that a smart contract conforms
to the relevant parties’ intentions. The importance of veri/fi.ligacation is illustrated by the fail-
ure of the multimillion-dollar venture capital fund “the DAO” (discussed in section /three.lf./two.lf./four.lf).
An unfortunate programming mistake in the relevant smart contracts, not noticed until
it was too late, left open a loophole that allowed one user to siphon a large portion of the
money out of the fund [/six.lf/seven.lf]. The more complex a smart contract is, the more di/f\_f\_i.ligacult the
work of detecting such /fl.ligaaws will be. While the techniques of formal verification can help in
checking that a smart contract has certain mathematically well-de/fi.liganed properties, there
will still remain the more nebulous task of checking that the potential judgements of a
smart contract all conform to common sense [/two.lf/one.lf]. In this vein, the fact that traditional
agreements can be /fi.ligaltered through human interpretation, which is capable of navigating
ambiguities and grasping obvious intentions, can be seen as an important cost-saving fea-
ture.
In cases where smart contracts involve smart property, there will also be the search cost
of establishing trust that a piece of smart property does in fact respond to the relevant
blockchain in the promised way. Conceivably, a regulatory system could be required to
establish this trust (see section /four.lf./two.lf./two.lf). Another simple search cost, which we might regard
/six.lf/nine.lf
as something of a /fi.ligaxed cost, is the cost of learning about smart contracts and how to use
them. As the length of this report may help to demonstrate, this cost should be regarded
as non-trivial.
Concerning bargaining costs, it may often be very di/f\_f\_i.ligacult to develop smart contracts that
arecomplete , meaning that they precisely specify the obligations of each party for every pos-
sible case. As smart contracts are computer code, designed to execute automatically, they
do not leave room for ambiguity or underspeci/fi.ligacation. However, many legal scholars and
economists hold that incomplete contracts are dramatically more common than complete
ones [/one.lf/six.lf/three.lf, /eight.lf/nine.lf]. The reason is that, for complex transactional relationships, it can be excep-
tionally di/f\_f\_i.ligacult to translate each party’s obligations into something so precise as computer
code while su/f\_f\_i.ligaciently accounting for every possible future contingency. Should the need
arise for a traditional contract, ambiguities and contingencies can be e/f\_f.ligaectively /fi.ligalled in
after an initial agreement, through ad hoc renegotiation and the cooperative norms that
characterize informal “relational contracts” [/one.lf/zero.lf]. In contrast, smart contracts place the full
burden on working out the contract’s details on the initial drafters. In many cases, this is
likely to be impractical./five.lf/three.lf
We also have particular reason to doubt that most traditional organizations could be e/f\_f\_i.liga-
ciently reconstructed as smart-contract-based “decentralized autonomous organizations”
(see section /three.lf./two.lf./four.lf). A widely held theory in economics is that hierarchical /fi.ligarms arise, in
large part, because of the impracticality of drafting complete contracts [/eight.lf/nine.lf, /one.lf/eight.lf/zero.lf]. As a
trivial example, a company’s employees do not need to treat each idiosyncratic e-mail they
send out as a service requiring formal contractual representation. The members of a DAO,
at least a DAO that makes truly extensive use of smart contracts, may not have this lux-
ury.
Then, there are commitment costs.
First, there are the costs associated with the need to maintain the integrity and availability
of the underlying blockchain. These costs consist of fees for creating and interacting with
smart contracts, which, at least in the case of permissionless blockchains, feed into the
cryptocurrency rewards used to incentivize full nodes to store and execute these contracts
(see section /two.lf./six.lf./two.lf). Since all of the nodes must perform the same work, at least in traditional
blockchain systems, these fees can become quite signi/fi.ligacant. For example, in any case where
a traditional online service is not associated with either extremely high pro/fi.ligat margins or
signi/fi.ligacant regulatory burdens, an equivalent decentralized application will normally be
much more expensive to use.
Second, in cases where users would like to make a contract conditional on features of
the outside world, such as one party’s success in /fi.liganishing a construction project on time,
there will be costs associated with ensuring that this information is recorded properly.
The relevant parties may need to acquire the services of a trusted third-party arbiter or
/five.lf/three.lfAnother cause for concern comes from the long-standing /fi.ligaeld of “computational law,” which has examined
the possibility of translating laws and contracts into computer code and found that, while translation may be
feasible within some domains (including electronic commerce), judgements often require case-based, analogical,
or inductive reasoning that it is very di/f\_f\_i.ligacult to represent with computer code [/seven.lf/one.lf, /one.lf/four.lf/seven.lf].
/seven.lf/zero.lf
apply a distributed consensus protocol that uses cryptocurrency payouts to incentivize
multiple parties to converge on the truth. As mentioned in section /two.lf./one.lf/zero.lf, some developers
are currently trialing early versions of these consensus protocols. However, it remains to
be seen how e/f\_f.ligaective “distributed oracle systems“ will be. In general, large expenditures
could be required to create su/f\_f\_i.ligacient incentives for accurate inputs.
Third, there will be costs associated with the need to ensure that theft or coercion cannot
be applied to counteract a smart contract. For example, the relevant parties must pro-
tect against their private keys being stolen, as well as having any of their smart property
stolen and rewired. More bluntly, the parties must protect themselves against attempts to
threaten them into signing unfavorable contracts. The necessary security measures may be
expensive. These security concerns are also, of course, good arguments for the continued
relevance of existing law enforcement institutions. The potential need to interact with
these institutions adds its own costs.
Finally, there will be the costs associated with the risk that a contract does not function as
expected. Despite the relevant parties’ best e/f\_f.ligaorts, a contract’s code may contain errors,
the consensus protocols used to maintain the blockchain may fail (see section /four.lf./two.lf./three.lf), the
external information the contract relies on may be inaccurate, the parties may experience
theft or coercion, or the expected signi/fi.ligacance of the relevant digital assets may fail to
hold, for example, if a cryptocurrency loses its value or if a piece of smart property does
not respond as promised. Collectively, these risks may be quite signi/fi.ligacant.
Together, all these considerations suggest that, while smart contracts may help to reduce
some transaction costs, they also come with very signi/fi.ligacant costs of their own. There is
not yet an obvious basis for predictions that they will heavily encroach on the territory
already covered by existing institutions (see section /three.lf./two.lf./two.lf).
/four.lf./two.lf./five.lf The possibility that existing institutions are “good enough”
The previous section explored the various inadequacies of smart contracts. A somewhat
symmetrical viewpoint—which pushes against the idea that smart contracts, and decen-
tralized applications generally, will begin /fi.ligalling many of the functions /fi.ligalled by existing
institutions (see section /three.lf./two.lf./two.lf)—is that many existing institutions are actually highly e/f\_f.ligaec-
tive [/one.lf/five.lf/six.lf].
Overall, the most compelling arguments for the use of permissionless blockchain technol-
ogy seems to be an argument from commitment costs (see section /three.lf./two.lf./three.lf). In particular,
in certain circumstances, this technology can limit the costs associated with the need to
establish trust in other parties to provide services, as well as with the risk of having this
trust violated.
Nevertheless, the case can be made that, in the current institutional environment, trust is
often fairly easy to come by. As Vitalik Buterin writes in his essay, “The Problem of Trust,”
“At least in the developed world, if you put your money in a bank, it’s safe.... From such
a perspective, one can easily see how the traditional ‘centralized system’ is serving people
just /fi.ligane” [/three.lf/nine.lf].
/seven.lf/one.lf
To be sure, there are many cases where trust is not easy to come by. Section /three.lf./two.lf./three.lf discusses
some of them, including cases where the relevant parties are located in countries with un-
usually dysfunctional institutions. It is unclear, though, just how far these cases extend,
or exactly how e/f\_f.ligaective blockchain systems could be for resolving them. While polls show
that citizens in many countries report distrusting major political and economic institu-
tions, their levels of trust may still, as a relative matter, be much greater than they would
be for complicated and unproven cryptographic technologies [/one.lf/two.lf/seven.lf, /eight.lf/eight.lf].
In his essay, Buterin acknowledges these points, but also takes a long-run perspective to
caution against what he might consider excessive skepticism. First, he writes, decentralized
applications could eventually establish much greater reputations for trustworthiness than
they possess today:
Who would you really trust more: [well-vetted banks] or a group of mining firms of
unknown quantity and size with no real-world reputations, /nine.lf/zero.lf% of whose chips may
be produced in Taiwan or Shenzhen? For mainstream securities settlement, the answer
that most people in the world would give seems rather clear. But then, in ten years’
time, if the set of miners or the set of anonymous stakeholders of some particular
currency proves itself trustworthy, eventually banks may warm up to even the more
“pure cryptoanarchic” model – or they may not.
Second, in some cases, it may be more appropriate to think of blockchain technology as
cutting the costs a new institution must face in earning su/f\_f\_i.ligacient quantities of trust:
Rather than concentrating on the lack of trust, here we emphasize the barrier to en-
try in becoming a locus of trust. Sure, billion dollar companies can certainly become
loci of trust just fine, and indeed it is the case that they generally work pretty well....
However, their ability to do so comes at a high cost.... The key promise of decen-
tralized technology, under this viewpoint, is not to create systems that are even more
trustworthy than current large institutions.... Rather, the key promise of decentral-
ized technology is to provide a shortcut to let future application developers get there
faster.... A [simple cryptographic protocol] may well have a lower probability of failure
than all but the largest of institutions – and at a millionth of the cost. Blockchain-
based applications allow developers to prove that they are honest – by setting up a
system where they do not even have any more power than the users do.
This consideration seems to suggest that, if decentralized applications do eventually achieve
a prominence comparable to existing centralized institutions, it may not be by directly dis-
placing them. Instead, these systems might /fi.ligall voids left by institutions that su/f\_f.ligaer losses
of trust, or o/f\_f.ligaer future services that no institution yet provides. The rise of blockchain
technology in political and economic life could be gradual, like the turnover of cells in a
body.
Nevertheless, even this more moderate view is highly speculative. As the preceding sec-
tions have discussed, there remain di/f\_f\_i.ligacult technical and legal roadblocks to large-scale
blockchain use, and the utility of smart contracts is still extremely unclear. A vast gap
separates the technology’s present level of maturity and the level of maturity it will need
/seven.lf/two.lf
to achieve to o/f\_f.ligaer plausible alternatives to most of the services provided by centralized
institutions.
Since blockchain technology is only a dozen years old, and has attracted widespread at-
tention for perhaps /fi.ligave years, it would appear premature to place a cap on its potential.
However, it would also be premature to forecast radical visions with any degree of con/fi.liga-
dence.
/seven.lf/three.lf
References
[/one.lf] S. Aaronson, “Quantum copy-protection and quantum money,” in /two.lf/zero.lf/zero.lf/nine.lf /two.lf/four.lfth An-
nual IEEE Conference on Computational Complexity . IEEE, /two.lf/zero.lf/zero.lf/nine.lf, pp. /two.lf/two.lf/nine.lf–/two.lf/four.lf/two.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/CCC./two.lf/zero.lf/zero.lf/nine.lf./four.lf/two.lf.
[/two.lf] C. C. Aggarwal and S. Y. Philip, “A general survey of privacy-preserving data mining
models and algorithms,” in Privacy-preserving data mining , C. C. Aggarwal and S. Y.
Philip, Eds. New York: Springer, /two.lf/zero.lf/zero.lf/eight.lf, pp. /one.lf/one.lf–/five.lf/two.lf.
[/three.lf] A. Ahmed and N. Anand, “India to propose cryptocurrency
ban, penalising miners and traders,” Reuters , /two.lf/zero.lf/two.lf/one.lf. [Online]. Avail-
able: https://www.reuters.com/article/uk-india-cryptocurrency-ban/india-to-
propose-cryptocurrency-ban-penalising-miners-traders-source-idUSKBN/two.lfB/six.lf/zero.lfQP
[/four.lf] G. Allen and T. Chan, “Arti/fi.ligacial intelligence and national security,” Belfer
Center for Science and International A/f\_f.ligaairs, Harvard Kennedy School, /two.lf/zero.lf/one.lf/seven.lf. [On-
line]. Available: https://www.belfercenter.org/publication/arti/fi.ligacial-intelligence-
and-national-security
[/five.lf] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mané,
“Concrete problems in AI safety,” arXiv preprint arXiv:/one.lf/six.lf/zero.lf/six.lf./zero.lf/six.lf/five.lf/six.lf/five.lf, /two.lf/zero.lf/one.lf/six.lf. [Online].
Available: https://arxiv.org/abs//one.lf/six.lf/zero.lf/six.lf./zero.lf/six.lf/five.lf/six.lf/five.lf
[/six.lf] J. V. Andrew J. Coe, “Why arms control is so rare,” American Political Science Review ,
vol. /one.lf/one.lf/four.lf, no. /two.lf, pp. /three.lf/four.lf/two.lf–/three.lf/five.lf/five.lf, /two.lf/zero.lf/two.lf/zero.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/one.lf/seven.lf/S/zero.lf/zero.lf/zero.lf/three.lf/zero.lf/five.lf/five.lf/four.lf/one.lf/nine.lf/zero.lf/zero.lf/zero.lf/seven.lf/three.lfX.
[/seven.lf] J. Andrews, “An introduction to AZTEC,” /two.lf/zero.lf/one.lf/nine.lf. [Online]. Available: https:
//medium.com/aztec-protocol/an-introduction-to-aztec-/four.lf/seven.lfc/seven.lf/zero.lfe/eight.lf/seven.lf/five.lfdc/seven.lf
[/eight.lf] L. J. Aslett, P. M. Esperança, and C. C. Holmes, “A review of homomorphic
encryption and software tools for encrypted statistical machine learning,” /two.lf/zero.lf/one.lf/five.lf,
arXiv preprint arXiv:/one.lf/five.lf/zero.lf/eight.lf./zero.lf/six.lf/five.lf/seven.lf/four.lf. [Online]. Available: https://arxiv.org/abs//one.lf/five.lf/zero.lf/eight.lf.
/zero.lf/six.lf/five.lf/seven.lf/four.lf
[/nine.lf] M. Atzori, “Blockchain technology and decentralized governance: Is the state still
necessary?” /two.lf/zero.lf/one.lf/five.lf, doi:/one.lf/zero.lf./two.lf/one.lf/three.lf/nine.lf/ssrn./two.lf/seven.lf/zero.lf/nine.lf/seven.lf/one.lf/three.lf.
[/one.lf/zero.lf] G. Baker, R. Gibbons, and K. J. Murphy, “Relational contracts and the theory
of the /fi.ligarm,” The Quarterly Journal of Economics , vol. /one.lf/one.lf/seven.lf, no. /one.lf, pp. /three.lf/nine.lf–/eight.lf/four.lf, /two.lf/zero.lf/zero.lf/two.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/six.lf/two.lf//zero.lf/zero.lf/three.lf/three.lf/five.lf/five.lf/three.lf/zero.lf/two.lf/seven.lf/five.lf/three.lf/three.lf/nine.lf/nine.lf/four.lf/four.lf/five.lf.
[/one.lf/one.lf] D. Banisar, “Stopping science: The case of cryptography,” Health Matrix , vol. /nine.lf,
no. /two.lf, pp. /two.lf/five.lf/three.lf–/two.lf/eight.lf/seven.lf, /one.lf/nine.lf/nine.lf/nine.lf. [Online]. Available: https://scholarlycommons.law.case.
edu/healthmatrix/vol/nine.lf/iss/two.lf//four.lf/
[/one.lf/two.lf] B. Barrett, “The year encryption won,” Wired , /two.lf/zero.lf/one.lf/six.lf. [Online]. Available:
/seven.lf/four.lf
https://www.wired.com//two.lf/zero.lf/one.lf/six.lf//one.lf/two.lf/year-encryption-won/
[/one.lf/three.lf] C. P. Bauer, Secret history: The story of cryptology . Boca Raton, FL: CRC Press, /two.lf/zero.lf/one.lf/three.lf.
[/one.lf/four.lf] C. Baum, I. Damgård, and C. Orlandi, “Publicly auditable secure multi-party
computation,” in International Conference on Security and Cryptography for Networks .
Springer, /two.lf/zero.lf/one.lf/four.lf, pp. /one.lf/seven.lf/five.lf–/one.lf/nine.lf/six.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/three.lf/one.lf/nine.lf-/one.lf/zero.lf/eight.lf/seven.lf/nine.lf-/seven.lf\_/one.lf/one.lf.
[/one.lf/five.lf] BBC Editors, “What are NFTs and why are some worth millions?” BBC News , /two.lf/zero.lf/two.lf/one.lf.
[Online]. Available: https://www.bbc.com/news/technology-/five.lf/six.lf/three.lf/seven.lf/one.lf/nine.lf/one.lf/two.lf
[/one.lf/six.lf] M. Ben-Or, O. Goldreich, S. Goldwasser, J. Håstad, J. Kilian, S. Micali, and
P. Rogaway, “Everything provable is provable in zero-knowledge,” in Advances in
Cryptology–CRYPTO /one.lf/nine.lf/eight.lf/eight.lf . Springer-Verlag, /one.lf/nine.lf/nine.lf/zero.lf, pp. /three.lf/seven.lf–/five.lf/six.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//zero.lf-/three.lf/eight.lf/seven.lf-
/three.lf/four.lf/seven.lf/nine.lf/nine.lf-/two.lf\_/four.lf.
[/one.lf/seven.lf] E. Ben-Sasson, “Zerocash, bitcoin, and transparent computational integrity,” Jan
/two.lf/zero.lf/one.lf/seven.lf.
[/one.lf/eight.lf] E. Ben-Sasson, A. Chiesa, D. Genkin, E. Tromer, and M. Virza, “SNARKs for C:
Verifying program executions succinctly and in zero knowledge,” in Advances in
Cryptology–CRYPTO /two.lf/zero.lf/one.lf/three.lf . Springer, /two.lf/zero.lf/one.lf/three.lf, pp. /nine.lf/zero.lf–/one.lf/zero.lf/eight.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/four.lf/two.lf-
/four.lf/zero.lf/zero.lf/eight.lf/four.lf-/one.lf\_/six.lf.
[/one.lf/nine.lf] E. Ben-Sasson, A. Chiesa, M. Green, E. Tromer, and M. Virza, “Secure sampling of
public parameters for succinct zero knowledge proofs,” in /two.lf/zero.lf/one.lf/five.lf IEEE Symposium on
Security and Privacy (SP) . IEEE, /two.lf/zero.lf/one.lf/five.lf, pp. /two.lf/eight.lf/seven.lf–/three.lf/zero.lf/four.lf, doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/SP./two.lf/zero.lf/one.lf/five.lf./two.lf/five.lf.
[/two.lf/zero.lf] I. Bentov, A. Gabizon, and A. Mizrahi, “Cryptocurrencies without proof of work,” in
International Conference on Financial Cryptography and Data Security . Springer, /two.lf/zero.lf/one.lf/six.lf,
pp. /one.lf/four.lf/two.lf–/one.lf/five.lf/seven.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/six.lf/two.lf-/five.lf/three.lf/three.lf/five.lf/seven.lf-/four.lf\_/one.lf/zero.lf.
[/two.lf/one.lf] K. Bhargavan, A. Delignat-Lavaud, C. Fournet, A. Gollamudi, G. Gonthier,
N. Kobeissi, N. Kulatova, A. Rastogi, T. Sibut-Pinote, N. Swamy, and S. Zanella-
Béguelin, “Formal veri/fi.ligacation of smart contracts: Short paper,” in Proceedings of the
/two.lf/zero.lf/one.lf/six.lf ACM Workshop on Programming Languages and Analysis for Security . ACM, /two.lf/zero.lf/one.lf/six.lf,
pp. /nine.lf/one.lf–/nine.lf/six.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/nine.lf/nine.lf/three.lf/six.lf/zero.lf/zero.lf./two.lf/nine.lf/nine.lf/three.lf/six.lf/one.lf/one.lf.
[/two.lf/two.lf] P. Birnstill, S. Bretthauer, S. Greiner, and E. Krempel, “Privacy-preserving surveil-
lance: An interdisciplinary approach,” International Data Privacy Law , vol. /five.lf, no. /four.lf,
pp. /two.lf/nine.lf/eight.lf–/three.lf/zero.lf/eight.lf, /two.lf/zero.lf/one.lf/five.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/nine.lf/three.lf/idpl/ipv/zero.lf/two.lf/one.lf.
[/two.lf/three.lf] C. Blum and C. I. Zuber, “Liquid democracy: Potentials, problems, and per-
spectives,” Journal of Political Philosophy , vol. /two.lf/four.lf, no. /two.lf, pp. /one.lf/six.lf/two.lf–/one.lf/eight.lf/two.lf, /two.lf/zero.lf/one.lf/five.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/one.lf/one.lf/jopp./one.lf/two.lf/zero.lf/six.lf/five.lf.
[/two.lf/four.lf] M. Blum, P. Feldman, and S. Micali, “Non-interactive zero-knowledge and its appli-
cations,” in Proceedings of the Twentieth Annual ACM Symposium on Theory of Computing .
/seven.lf/five.lf
ACM, /one.lf/nine.lf/eight.lf/eight.lf, pp. /one.lf/zero.lf/three.lf–/one.lf/one.lf/two.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//six.lf/two.lf/two.lf/one.lf/two.lf./six.lf/two.lf/two.lf/two.lf/two.lf.
[/two.lf/five.lf] D. Bogdanov, M. Jöemets, S. Siim, and M. Vaht, “How the Estonian tax and custom
board evaluated a tax fraud detection system based on secure multi-party compu-
tation,” in International Conference on Financial Cryptography and Data Security , /two.lf/zero.lf/one.lf/five.lf,
pp. /two.lf/two.lf/seven.lf–/two.lf/three.lf/four.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/six.lf/two.lf-/four.lf/seven.lf/eight.lf/five.lf/four.lf-/seven.lf\_/one.lf/four.lf.
[/two.lf/six.lf] P. Bogetoft, D. L. Christensen, I. Damgård, M. Geisler, T. P. Jakobsen, M. Krøigaard,
J. D. Nielsen, J. B. Nielsen, K. Nielsen, J. Pagter, M. Schwartzbach, and T. Toft, “Se-
cure multiparty computation goes live,” in Financial Cryptography and Data Security .
Springer, /two.lf/zero.lf/zero.lf/nine.lf, pp. /three.lf/two.lf/five.lf–/three.lf/four.lf/three.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/four.lf/two.lf-/zero.lf/three.lf/five.lf/four.lf/nine.lf-/four.lf\_/two.lf/zero.lf.
[/two.lf/seven.lf] D. Boneh, A. Sahai, and B. Waters, “Functional encryption: De/fi.liganitions and chal-
lenges,” in Theory of Cryptography . Springer, /two.lf/zero.lf/one.lf/one.lf, pp. /two.lf/five.lf/three.lf–/two.lf/seven.lf/three.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-
/six.lf/four.lf/two.lf-/one.lf/nine.lf/five.lf/seven.lf/one.lf-/six.lf\_/one.lf/six.lf.
[/two.lf/eight.lf] N. Bostrom, Superintelligence: Paths, dangers, strategies . Oxford: Oxford University
Press, /two.lf/zero.lf/one.lf/four.lf.
[/two.lf/nine.lf] N. Bostrom, A. Dafoe, and C. Flynn, “Policy desiderata
in the development of machine superintelligence.” [Online]. Avail-
able: https://www.fhi.ox.ac.uk/wp-content/uploads/Policy-Desiderata-in-the-
Development-of-Machine-Superintelligence.pdf
[/three.lf/zero.lf] K. Bradsher, “China blocks whatsapp, broadening online censorship,” The New York
Times , /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://www.nytimes.com//two.lf/zero.lf/one.lf/seven.lf//zero.lf/nine.lf//two.lf/five.lf/business/
china-whatsapp-blocked.html
[/three.lf/one.lf] I. Braverman, “Passing the sni/f\_f.liga test: Police dogs as surveillance technology,”
Bu/f\_f.ligaalo Law Review , vol. /six.lf/one.lf, no. /one.lf, pp. /eight.lf/one.lf–/one.lf/six.lf/seven.lf, /two.lf/zero.lf/one.lf/three.lf. [Online]. Available:
https://digitalcommons.law.bu/f\_f.ligaalo.edu/journal\_articles//three.lf/three.lf/six.lf
[/three.lf/two.lf] M. Brundage, S. Avin et al. , “The malicious use of arti/fi.ligacial intelligence:
Forecasting, prevention, and mitigation,” /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https:
//maliciousaireport.com/
[/three.lf/three.lf] K. Buchholz, “How common is crypto?” Statista , /two.lf/zero.lf/two.lf/one.lf. [Online]. Available:
https://www.statista.com/chart//one.lf/eight.lf/three.lf/four.lf/five.lf/crypto-currency-adoption/
[/three.lf/four.lf] E. Buchman, “Tendermint: Byzantine fault tolerance in the age of blockchains,”
Master’s thesis, The University of Guelph, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: http:
//hdl.handle.net//one.lf/zero.lf/two.lf/one.lf/four.lf//nine.lf/seven.lf/six.lf/nine.lf
[/three.lf/five.lf] R. Burn-Callander, “Skype inventor Jaan Tallinn wants to use Bit-
coin technology to save the world,” The Telegraph , Jun /two.lf/zero.lf/one.lf/six.lf. [Online].
Available: http://www.telegraph.co.uk/business//two.lf/zero.lf/one.lf/six.lf//zero.lf/six.lf//two.lf/zero.lf/skype-inventor-jaan-
tallinn-wants-to-use-bitcoin-technology-to-s/
/seven.lf/six.lf
[/three.lf/six.lf] V. Buterin, “DAOs, DACs, DAs and more: An incomplete terminology guide,”
Ethereum Blog, May /two.lf/zero.lf/one.lf/four.lf. [Online]. Available: https://blog.ethereum.org//two.lf/zero.lf/one.lf/four.lf//zero.lf/five.lf/
/zero.lf/six.lf/daos-dacs-das-and-more-an-incomplete-terminology-guide/
[/three.lf/seven.lf] ——, “An introduction to futarchy,” Aug /two.lf/zero.lf/one.lf/four.lf. [Online]. Available: https:
//blog.ethereum.org//two.lf/zero.lf/one.lf/four.lf//zero.lf/eight.lf//two.lf/one.lf/introduction-futarchy/
[/three.lf/eight.lf] ——, “Visions, part /one.lf: The value of blockchain technology,” Ethereum Blog,
Apr /two.lf/zero.lf/one.lf/five.lf. [Online]. Available: https://blog.ethereum.org//two.lf/zero.lf/one.lf/five.lf//zero.lf/four.lf//one.lf/three.lf/visions-part-
/one.lf-the-value-of-blockchain-technology/
[/three.lf/nine.lf] ——, “Visions, part /two.lf: The problem of trust,” Ethereum Blog, Apr /two.lf/zero.lf/one.lf/five.lf. [Online].
Available: https://blog.ethereum.org//two.lf/zero.lf/one.lf/five.lf//zero.lf/four.lf//two.lf/seven.lf/visions-part-/two.lf-the-problem-of-
trust/
[/four.lf/zero.lf] ——, “Why cryptoeconomics and x-risk researchers should
listen to each other more,” /two.lf/zero.lf/one.lf/six.lf. [Online]. Avail-
able: https://medium.com/@VitalikButerin/why-cryptoeconomics-and-x-risk-
researchers-should-listen-to-each-other-more-a/two.lfdb/seven.lf/two.lfb/three.lfe/eight.lf/six.lfb
[/four.lf/one.lf] ——, “Ethereum: A next-generation smart contract and
decentralized application platform,” /two.lf/zero.lf/one.lf/seven.lf. [Online]. Avail-
able: https://blockchainlab.com/pdf/Ethereum\_white\_paper-a\_next\_generation\_
smart\_contract\_and\_decentralized\_application\_platform-vitalik-buterin.pdf
[/four.lf/two.lf] M. Calabresi, “Iran nuclear deal: Veri/fi.ligacation is trickiest part,” Time , /two.lf/zero.lf/one.lf/five.lf. [Online].
Available: http://time.com//three.lf/nine.lf/four.lf/six.lf/eight.lf/six.lf/zero.lf/iran-nuclear-deal-inspections/
[/four.lf/three.lf] B. Caplan, “The logic of collective belief,” Rationality and Society , vol. /one.lf/five.lf, no. /two.lf, pp.
/two.lf/one.lf/eight.lf–/two.lf/four.lf/two.lf, /two.lf/zero.lf/zero.lf/three.lf, doi:/one.lf/zero.lf./one.lf/one.lf/seven.lf/seven.lf//one.lf/zero.lf/four.lf/three.lf/four.lf/six.lf/three.lf/one.lf/zero.lf/three.lf/zero.lf/one.lf/five.lf/zero.lf/zero.lf/two.lf/zero.lf/zero.lf/three.lf.
[/four.lf/four.lf] N. Cardozo, “The state of crypto law: /two.lf/zero.lf/one.lf/six.lf in review,” Electronic Frontier
Foundation, Jan /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://www.e/f\_f.liga.org/deeplinks//two.lf/zero.lf/one.lf/six.lf//one.lf/two.lf/
crypto-state-law-end-/two.lf/zero.lf/one.lf/six.lf
[/four.lf/five.lf] H. Chen, W. Chung, J. Jie Xu, G. Wang, Y. Qin, and M. Chau, “Crime data mining:
A general framework and some examples,” Computer , vol. /three.lf/seven.lf, no. /four.lf, pp. /five.lf/zero.lf–/five.lf/six.lf, /two.lf/zero.lf/zero.lf/four.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/MC./two.lf/zero.lf/zero.lf/four.lf./one.lf/two.lf/nine.lf/seven.lf/three.lf/zero.lf/one.lf.
[/four.lf/six.lf] U. W. Chohan, “Assessing the di/f\_f.ligaerences in bitcoin & other cryptocurrency le-
gality across national jurisdictions,” Critical Blockchain Research Initiative, /two.lf/zero.lf/one.lf/seven.lf,
/one.lf/zero.lf./two.lf/one.lf/three.lf/nine.lf/ssrn./three.lf/zero.lf/four.lf/two.lf/two.lf/four.lf/eight.lf.
[/four.lf/seven.lf] A. Choudhury, J. Loftus, E. Orsini, A. Patra, and N. P. Smart, “Between a rock and
a hard place: Interpolating between MPC and FHE,” in International Conference on
the Theory and Application of Cryptology and Information Security . Springer, /two.lf/zero.lf/one.lf/three.lf, pp.
/two.lf/two.lf/one.lf–/two.lf/four.lf/zero.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/four.lf/two.lf-/four.lf/two.lf/zero.lf/four.lf/five.lf-/zero.lf\_/one.lf/two.lf.
/seven.lf/seven.lf
[/four.lf/eight.lf] K. Christidis and M. Devetsikiotis, “Blockchains and smart contracts
for the internet of things,” IEEE Access , vol. /four.lf, pp. /two.lf/two.lf/nine.lf/two.lf–/two.lf/three.lf/zero.lf/three.lf, /two.lf/zero.lf/one.lf/six.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/ACCESS./two.lf/zero.lf/one.lf/six.lf./two.lf/five.lf/six.lf/six.lf/three.lf/three.lf/nine.lf.
[/four.lf/nine.lf] J. S. Chung, A. Jamaludin, and A. Zisserman, “You said that?” /two.lf/zero.lf/one.lf/seven.lf, arXiv preprint
arXiv:/one.lf/seven.lf/zero.lf/five.lf./zero.lf/two.lf/nine.lf/six.lf/six.lf. [Online]. Available: https://arxiv.org/abs//one.lf/seven.lf/zero.lf/five.lf./zero.lf/two.lf/nine.lf/six.lf/six.lf
[/five.lf/zero.lf] M. C. Clutter, “Dogs, drones, and defendants: The fourth amendment in the digital
age,” George Mason Law Review , vol. /two.lf/one.lf, pp. /five.lf/five.lf/seven.lf–/five.lf/seven.lf/one.lf, /two.lf/zero.lf/one.lf/three.lf.
[/five.lf/one.lf] J. B. Comey, “Going dark: Are technology, privacy, and public safety on
a collision course?” Speech, Federal Bureau of Investigation, /two.lf/zero.lf/one.lf/four.lf. [Online].
Available: https://www./f\_b.ligai.gov/news/speeches/going-dark-are-technology-privacy-
and-public-safety-on-a-collision-course
[/five.lf/two.lf] C. Comley, M. Comley, P. Eggins, G. George, S. Holloway, M. Ley, P. Thompson,
and K. Warburton, “Con/fi.ligadence, security and veri/fi.ligacation: The challenge of
global nuclear weapons arms control,” Atomic Weapons Establishment, Tech. Rep.
AWE/TR//two.lf/zero.lf/zero.lf/zero.lf//zero.lf/zero.lf/one.lf, /two.lf/zero.lf/zero.lf/zero.lf. [Online]. Available: http:///fi.ligassilematerials.org/library/
awe/zero.lf/zero.lf.pdf
[/five.lf/three.lf] R. Cramer, R. Gennaro, and B. Schoenmakers, “A secure and optimally e/f\_f\_i.ligacient
multi-authority election scheme,” European Transactions on Telecommunications , vol. /eight.lf,
no. /five.lf, pp. /four.lf/eight.lf/one.lf–/four.lf/nine.lf/zero.lf, /one.lf/nine.lf/nine.lf/seven.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/two.lf/ett./four.lf/four.lf/six.lf/zero.lf/zero.lf/eight.lf/zero.lf/five.lf/zero.lf/six.lf.
[/five.lf/four.lf] K. Croman, C. Decker, I. Eyal, A. E. Gencer, A. Juels, A. Kosba, A. Miller, P. Sax-
ena, E. Shi, E. G. Sirer, D. Song, and R. Wattenhofer, “On scaling decentralized
blockchains,” in International Conference on Financial Cryptography and Data Security .
Springer, /two.lf/zero.lf/one.lf/six.lf, pp. /one.lf/zero.lf/six.lf–/one.lf/two.lf/five.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/six.lf/two.lf-/five.lf/three.lf/three.lf/five.lf/seven.lf-/four.lf\_/eight.lf.
[/five.lf/five.lf] S. A. Crosby and D. S. Wallach, “E/f\_f\_i.ligacient data structures for tamper-evident
logging,” in USENIX Security Symposium , /two.lf/zero.lf/zero.lf/nine.lf, pp. /three.lf/one.lf/seven.lf–/three.lf/three.lf/four.lf. [Online]. Available:
https://www.usenix.org/legacy/event/sec/zero.lf/nine.lf/tech/full\_papers/crosby.pdf
[/five.lf/six.lf] S. Davidson, P. De Filippi, and J. Potts, “Economics of blockchain,” /two.lf/zero.lf/one.lf/six.lf,
doi:/one.lf/zero.lf./two.lf/one.lf/three.lf/nine.lf/ssrn./two.lf/seven.lf/four.lf/four.lf/seven.lf/five.lf/one.lf.
[/five.lf/seven.lf] W. Di/f\_f\_i.ligae and M. Hellman, “New directions in cryptography,” IEEE Transactions on
Information Theory , vol. /two.lf/two.lf, no. /six.lf, pp. /six.lf/four.lf/four.lf–/six.lf/five.lf/four.lf, /one.lf/nine.lf/seven.lf/six.lf, doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/TIT./one.lf/nine.lf/seven.lf/six.lf./one.lf/zero.lf/five.lf/five.lf/six.lf/three.lf/eight.lf.
[/five.lf/eight.lf] G. Eder, “Digital transformation: Blockchain and land titles,” in /two.lf/zero.lf/one.lf/nine.lf
OECD Global Anti-Corruption & Integrity Forum , /two.lf/zero.lf/one.lf/nine.lf. [Online]. Avail-
able: https://www.oecd.org/corruption/integrity-forum/academic-papers/Georg%
/two.lf/zero.lfEder-%/two.lf/zero.lfBlockchain%/two.lf/zero.lf-%/two.lf/zero.lfGhana\_veri/fi.ligaed.pdf
[/five.lf/nine.lf] K. Ermoshina, F. Musiani, and H. Halpin, “End-to-end encrypted messaging proto-
cols: An overview,” in International Conference on Internet Science . Springer, /two.lf/zero.lf/one.lf/six.lf,
/seven.lf/eight.lf
pp. /two.lf/four.lf/four.lf–/two.lf/five.lf/four.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/three.lf/one.lf/nine.lf-/four.lf/five.lf/nine.lf/eight.lf/two.lf-/zero.lf\_/two.lf/two.lf.
[/six.lf/zero.lf] “Sharding FAQ,” Ethereum Wiki, /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://eth.wiki/
sharding/Sharding-FAQs
[/six.lf/one.lf] “Decentralized /fi.liganance (DeFi),” Ethereum.org, /two.lf/zero.lf/two.lf/one.lf. [Online]. Available: https:
//ethereum.org/en/de/fi.liga/
[/six.lf/two.lf] “Layer /two.lf scaling,” Ethereum.org, /two.lf/zero.lf/two.lf/one.lf. [Online]. Available: https://ethereum.org/
nb/developers/docs/layer-/two.lf-scaling/
[/six.lf/three.lf] “Non-fungible tokens (NFT),” Ethereum.org, /two.lf/zero.lf/two.lf/one.lf. [Online]. Available: https:
//ethereum.org/en/nft/
[/six.lf/four.lf] I. Eyal and E. G. Sirer, “Majority is not enough: Bitcoin mining is vulnerable,” in
International Conference on Financial Cryptography and Data Security . Springer, /two.lf/zero.lf/one.lf/four.lf,
pp. /four.lf/three.lf/six.lf–/four.lf/five.lf/four.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/six.lf/two.lf-/four.lf/five.lf/four.lf/seven.lf/two.lf-/five.lf\_/two.lf/eight.lf.
[/six.lf/five.lf] J. D. Fearon, “Rationalist explanations for war,” International organization , vol. /four.lf/nine.lf,
no. /three.lf, pp. /three.lf/seven.lf/nine.lf–/four.lf/one.lf/four.lf, /one.lf/nine.lf/nine.lf/five.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/one.lf/seven.lf/S/zero.lf/zero.lf/two.lf/zero.lf/eight.lf/one.lf/eight.lf/three.lf/zero.lf/zero.lf/zero.lf/three.lf/three.lf/three.lf/two.lf/four.lf.
[/six.lf/six.lf] J. Feigenbaum and B. Ford, “Multiple objectives of lawful-surveillance protocols,” in
Cambridge International Workshop on Security Protocols , /two.lf/zero.lf/one.lf/seven.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/three.lf/one.lf/nine.lf-
/seven.lf/one.lf/zero.lf/seven.lf/five.lf-/four.lf\_/one.lf.
[/six.lf/seven.lf] K. Finley, “A /dollar.lf/five.lf/zero.lf million hack just showed that the DAO was all too human,”
Wired , Jun /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://www.wired.com//two.lf/zero.lf/one.lf/six.lf//zero.lf/six.lf//five.lf/zero.lf-million-
hack-just-showed-dao-human/
[/six.lf/eight.lf] B. Fisch, D. Freund, and M. Naor, “Physical zero-knowledge proofs of physical prop-
erties,” in Advances in Crypotology–CRYPTO /two.lf/zero.lf/one.lf/four.lf . Springer, /two.lf/zero.lf/one.lf/four.lf, pp. /three.lf/one.lf/three.lf–/three.lf/three.lf/six.lf,
doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/six.lf/two.lf-/four.lf/four.lf/three.lf/eight.lf/one.lf-/one.lf\_/one.lf/eight.lf.
[/six.lf/nine.lf] K. B. Frikken and M. J. Atallah, “Privacy preserving electronic surveillance,” in Pro-
ceedings of the /two.lf/zero.lf/zero.lf/three.lf ACM workshop on Privacy in the Electronic Society . ACM, /two.lf/zero.lf/zero.lf/three.lf,
pp. /four.lf/five.lf–/five.lf/two.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//one.lf/zero.lf/zero.lf/five.lf/one.lf/four.lf/zero.lf./one.lf/zero.lf/zero.lf/five.lf/one.lf/four.lf/eight.lf.
[/seven.lf/zero.lf] U. Gasser, N. Gertner, J. L. Goldsmith, S. Landau, J. S. Nye, D. O’Brien,
M. G. Olsen, D. Renan, J. Sanchez, B. Schneider, L. Schwartzol, and J. L.
Zittrain, “Don’t panic: Making progress on the “going dark” debate,” Berkman
Center for Internet & Society at Harvard Law School, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available:
http://nrs.harvard.edu/urn-/three.lf:HUL.InstRepos:/two.lf/eight.lf/five.lf/five.lf/two.lf/five.lf/seven.lf/six.lf
[/seven.lf/one.lf] M. Genesereth, “Computational law: The cop in the backseat,” CodeX: The
Center for Legal Informatics, Stanford University, /two.lf/zero.lf/one.lf/five.lf. [Online]. Available:
http://logic.stanford.edu/publications/genesereth/complaw.pdf
[/seven.lf/two.lf] C. Gentry, “Fully homomorphic encryption using ideal lattices,” in Proceedings of
/seven.lf/nine.lf
the Forty-First Annual ACM symposium on Theory of Computing , /two.lf/zero.lf/zero.lf/nine.lf, pp. /one.lf/six.lf/nine.lf–/one.lf/seven.lf/eight.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//one.lf/five.lf/three.lf/six.lf/four.lf/one.lf/four.lf./one.lf/five.lf/three.lf/six.lf/four.lf/four.lf/zero.lf.
[/seven.lf/three.lf] A. Gerber and P. C. Wichardt, “Providing public goods in the absence of strong
institutions,” Journal of Public Economics , vol. /nine.lf/three.lf, no. /three.lf-/four.lf, pp. /four.lf/two.lf/nine.lf–/four.lf/three.lf/nine.lf, /two.lf/zero.lf/zero.lf/nine.lf,
doi:/one.lf/zero.lf./one.lf/zero.lf/one.lf/six.lf/j.jpubeco./two.lf/zero.lf/zero.lf/eight.lf./one.lf/zero.lf./zero.lf/zero.lf/six.lf.
[/seven.lf/four.lf] D. Ghupta, “Practical and deployable secure multi-party computation,” Ph.D.
dissertation, Yale University, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://www.cs.yale.edu/
homes/jf/Debayan-thesis.pdf
[/seven.lf/five.lf] S. Gibbs, “EU seeks to outlaw ’backdoors’ in new data pri-
vacy proposals,” The Guardian , Jun /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available:
https://www.theguardian.com/technology//two.lf/zero.lf/one.lf/seven.lf/jun//one.lf/nine.lf/eu-outlaw-backdoors-new-
data-privacy-proposals-uk-government-encrypted-communications-whatsapp
[/seven.lf/six.lf] A. Glaser, B. Barak, and R. J. Goldston, “A zero-knowledge protocol for nu-
clear warhead veri/fi.ligacation,” Nature , vol. /five.lf/one.lf/zero.lf, no. /seven.lf/five.lf/zero.lf/six.lf, pp. /four.lf/nine.lf/seven.lf–/five.lf/zero.lf/two.lf, /two.lf/zero.lf/one.lf/four.lf,
doi:/one.lf/zero.lf./one.lf/zero.lf/three.lf/eight.lf/nature/one.lf/three.lf/four.lf/five.lf/seven.lf.
[/seven.lf/seven.lf] E. Goldberg, “Getting beyond intuition in the probable cause inquiry,” Lewis
& Clark Law Review , vol. /one.lf/seven.lf, no. /three.lf, pp. /seven.lf/eight.lf/nine.lf–/eight.lf/three.lf/eight.lf, /two.lf/zero.lf/one.lf/three.lf. [Online]. Available:
https://law.lclark.edu/live//fi.ligales//one.lf/five.lf/three.lf/two.lf/two.lf-lcb/one.lf/seven.lf/three.lfart/three.lfgoldbergpdf
[/seven.lf/eight.lf] S. Goldwasser, Y. Kalai, R. A. Popa, V. Vaikuntanathan, and N. Zeldovich, “Reusable
garbled circuits and succinct functional encryption,” in Proceedings of the Forty-
Fifth Annual ACM Symposium on Theory of Computing . ACM, /two.lf/zero.lf/one.lf/three.lf, pp. /five.lf/five.lf/five.lf–/five.lf/six.lf/four.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/four.lf/eight.lf/eight.lf/six.lf/zero.lf/eight.lf./two.lf/four.lf/eight.lf/eight.lf/six.lf/seven.lf/eight.lf.
[/seven.lf/nine.lf] S. Goldwasser, Y. T. Kalai, R. A. Popa, V. Vaikuntanathan, and N. Zeldovich, “How
to run Turing machines on encrypted data,” in Advances in Cryptology–CRYPTO /two.lf/zero.lf/one.lf/three.lf .
Springer-Verlag, /two.lf/zero.lf/one.lf/three.lf, pp. /five.lf/three.lf/six.lf–/five.lf/five.lf/three.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/three.lf-/six.lf/four.lf/two.lf-/four.lf/zero.lf/zero.lf/eight.lf/four.lf-/one.lf\_/three.lf/zero.lf.
[/eight.lf/zero.lf] S. Goldwasser, S. Micali, and C. Racko/f\_f.liga, “The knowledge complexity of interac-
tive proof systems,” SIAM Journal on Computing , vol. /one.lf/eight.lf, no. /one.lf, pp. /one.lf/eight.lf/six.lf–/two.lf/zero.lf/eight.lf, /one.lf/nine.lf/eight.lf/nine.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/three.lf/seven.lf//zero.lf/two.lf/one.lf/eight.lf/zero.lf/one.lf/two.lf.
[/eight.lf/one.lf] H. R. Gow, D. H. Streeter, and J. F. Swinnen, “How private contract enforcement
mechanisms can succeed where public institutions,” Agricultural Economics , vol. /two.lf/three.lf,
no. /three.lf, pp. /two.lf/five.lf/three.lf–/two.lf/six.lf/five.lf, /two.lf/zero.lf/zero.lf/zero.lf, doi:/one.lf/zero.lf./one.lf/one.lf/one.lf/one.lf/j./one.lf/five.lf/seven.lf/four.lf-/zero.lf/eight.lf/six.lf/two.lf./two.lf/zero.lf/zero.lf/zero.lf.tb/zero.lf/zero.lf/two.lf/seven.lf/seven.lf.x.
[/eight.lf/two.lf] K. Grace, J. Salvatier, A. Dafoe, B. Zhang, and O. Evans, “When will AI exceed
human performance? Evidence from AI experts,” arXiv preprint arXiv:/one.lf/seven.lf/zero.lf/five.lf./zero.lf/eight.lf/eight.lf/zero.lf/seven.lf,
/two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://arxiv.org/abs//one.lf/seven.lf/zero.lf/five.lf./zero.lf/eight.lf/eight.lf/zero.lf/seven.lf
[/eight.lf/three.lf] A. Greenberg, “Zcash, an untraceable bitcoin alternative, launches in alpha,”
Wired , Jun /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://www.wired.com//two.lf/zero.lf/one.lf/six.lf//zero.lf/one.lf/zcash-an-
/eight.lf/zero.lf
untraceable-bitcoin-alternative-launches-in-alpha/
[/eight.lf/four.lf] S. Haber and W. S. Stornetta, “How to time-stamp a digital document,” in Ad-
vances in Cryptology–CRYPTO /one.lf/nine.lf/nine.lf/zero.lf . Springer, /one.lf/nine.lf/nine.lf/zero.lf, pp. /four.lf/three.lf/seven.lf–/four.lf/five.lf/five.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//three.lf-
/five.lf/four.lf/zero.lf-/three.lf/eight.lf/four.lf/two.lf/four.lf-/three.lf\_/three.lf/two.lf.
[/eight.lf/five.lf] B. H. Hall and B. Khan, “Adoption of new technology,” Working Paper /nine.lf/seven.lf/three.lf/zero.lf, Na-
tional Bureau of Economic Research, /two.lf/zero.lf/zero.lf/three.lf, doi:/one.lf/zero.lf./three.lf/three.lf/eight.lf/six.lf/w/nine.lf/seven.lf/three.lf/zero.lf.
[/eight.lf/six.lf] R. Hanson, “Shall we vote on values, but bet on beliefs?” Journal of Political Philosophy ,
vol. /two.lf/one.lf, pp. /one.lf/five.lf/one.lf–/one.lf/seven.lf/eight.lf, /two.lf/zero.lf/zero.lf/three.lf, doi:/one.lf/zero.lf./one.lf/one.lf/one.lf/one.lf/jopp./one.lf/two.lf/zero.lf/zero.lf/eight.lf.
[/eight.lf/seven.lf] G. Hardin, “The tragedy of the commons,” Journal of Natural Resources Policy Research ,
vol. /one.lf, no. /three.lf, pp. /two.lf/four.lf/three.lf–/two.lf/five.lf/three.lf, /two.lf/zero.lf/zero.lf/nine.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/eight.lf/zero.lf//one.lf/nine.lf/three.lf/nine.lf/zero.lf/four.lf/five.lf/zero.lf/nine.lf/zero.lf/three.lf/zero.lf/three.lf/seven.lf/three.lf/zero.lf/two.lf.
[/eight.lf/eight.lf] M. Harrington, “Survey: People’s trust has declined in business, media, government,
and NGOs,” Jan /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://hbr.org//two.lf/zero.lf/one.lf/seven.lf//zero.lf/one.lf/survey-peoples-
trust-has-declined-in-business-media-government-and-ngos
[/eight.lf/nine.lf] O. D. Hart, “Incomplete contracts and the theory of the /fi.ligarm,” Jour-
nal of Law, Economics, & Organization , vol. /four.lf, no. /one.lf, pp. /one.lf/one.lf/nine.lf–/one.lf/three.lf/nine.lf, /one.lf/nine.lf/eight.lf/eight.lf,
doi:/one.lf/zero.lf./one.lf/zero.lf/nine.lf/three.lf/oxfordjournals.jleo.a/zero.lf/three.lf/six.lf/nine.lf/four.lf/zero.lf.
[/nine.lf/zero.lf] A. Hern, “Revolv devices bricked as Google’s Nest shuts down smart home
company,” The Guardian , Apr /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://www.theguardian.
com/technology//two.lf/zero.lf/one.lf/six.lf/apr//zero.lf/five.lf/revolv-devices-bricked-google-nest-smart-home
[/nine.lf/one.lf] ——, “Google’s DeepMind plans bitcoin-style health record track-
ing for hospitals,” The Guardian , Mar /two.lf/zero.lf/one.lf/seven.lf. [Online]. Avail-
able: https://www.theguardian.com/technology//two.lf/zero.lf/one.lf/seven.lf/mar//zero.lf/nine.lf/google-deepmind-
health-records-tracking-blockchain-nhs-hospitals
[/nine.lf/two.lf] ——, “UK government can force encryption removal, but fears
losing, experts say,” The Guardian , Mar /two.lf/zero.lf/one.lf/seven.lf. [Online]. Avail-
able: https://www.theguardian.com/technology//two.lf/zero.lf/one.lf/seven.lf/mar//two.lf/nine.lf/uk-government-
encryption-whatsapp-investigatory-powers-act
[/nine.lf/three.lf] R. Higgott, “International political institutions,” in The Oxford handbook of Political
Institutions , S. A. Binder, R. A. W. Rhodes, and B. A. Rockman, Eds. Oxford: Oxford
University Press, /two.lf/zero.lf/zero.lf/six.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/nine.lf/three.lf/oxfordhb//nine.lf/seven.lf/eight.lf/zero.lf/one.lf/nine.lf/nine.lf/five.lf/four.lf/eight.lf/four.lf/six.lf/zero.lf./zero.lf/zero.lf/three.lf./zero.lf/zero.lf/three.lf/one.lf.
[/nine.lf/four.lf] F. H. Hinsley, “The in/fl.ligauence of ULTRA in the Second World War,” Seminar
Transcript, /one.lf/nine.lf/nine.lf/three.lf. [Online]. Available: http://www.cdpa.co.uk/UoP/HoC/Lectures/
HoC\_/zero.lf/eight.lfe.PDF
[/nine.lf/five.lf] J. Hovi, M. Greaker, C. Hagem, and B. Holtsmark, “A credible compliance enforce-
ment system for the climate regime,” Climate Policy , vol. /one.lf/two.lf, no. /six.lf, pp. /seven.lf/four.lf/one.lf–/seven.lf/five.lf/four.lf, /two.lf/zero.lf/one.lf/two.lf,
doi:/one.lf/zero.lf./one.lf/zero.lf/eight.lf/zero.lf//one.lf/four.lf/six.lf/nine.lf/three.lf/zero.lf/six.lf/two.lf./two.lf/zero.lf/one.lf/two.lf./six.lf/nine.lf/two.lf/two.lf/zero.lf/six.lf.
/eight.lf/one.lf
[/nine.lf/six.lf] A. Irrera and J. Kelly, “Blockchain could save investment banks up to
/dollar.lf/one.lf/two.lf billion a year: Accenture,” Reuters , Jan /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available:
http://uk.reuters.com/article/us-banks-blockchain-accenture/blockchain-could-
save-investment-banks-up-to-/one.lf/two.lf-billion-a-year-accenture-idUKKBN/one.lf/five.lf/one.lf/one.lfOU
[/nine.lf/seven.lf] ITU-T, “Distributed ledger technologies and /fi.liganancial inclusion,” Focus
Group Technical Report, International Telecommunications Union, /two.lf/zero.lf/one.lf/seven.lf.
[Online]. Available: https://www.itu.int/en/ITU-T/focusgroups/dfs/Documents/
/two.lf/zero.lf/one.lf/seven.lf/zero.lf/three.lf/ITU\_FGDFS\_Report-on-DLT-and-Financial-Inclusion.pdf
[/nine.lf/eight.lf] A. Juels, A. Kosba, and E. Shi, “The Ring of Gyges: Investigating the fu-
ture of criminal smart contracts,” in Proceedings of the /two.lf/zero.lf/one.lf/six.lf ACM SIGSAC Con-
ference on Computer and Communications Security . ACM, /two.lf/zero.lf/one.lf/six.lf, pp. /two.lf/eight.lf/three.lf–/two.lf/nine.lf/five.lf,
doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/nine.lf/seven.lf/six.lf/seven.lf/four.lf/nine.lf./two.lf/nine.lf/seven.lf/eight.lf/three.lf/six.lf/two.lf.
[/nine.lf/nine.lf] J. Katz and Y. Lindell, Introduction to Modern Cryptography . Boca Raton, FL: CRC
Press, /two.lf/zero.lf/one.lf/four.lf.
[/one.lf/zero.lf/zero.lf] B. Kellman, D. S. Gualtieri, and E. A. Tanzman, “Disarmament and disclosure: How
arms control veri/fi.ligacation can proceed without threatening con/fi.ligadential business in-
formation,” Harvard International Law Journal , vol. /three.lf/six.lf, pp. /seven.lf/one.lf–/one.lf/two.lf/six.lf, /one.lf/nine.lf/nine.lf/five.lf.
[/one.lf/zero.lf/one.lf] L. Kello, The Virtual Weapon and International Order . New Haven, CT: Yale University
Press, /two.lf/zero.lf/one.lf/seven.lf.
[/one.lf/zero.lf/two.lf] T. I. Kiviat, “Beyond bitcoin: Issues in regulating blockchain tranactions,” Duke LJ ,
vol. /six.lf/five.lf, pp. /five.lf/six.lf/nine.lf–/six.lf/zero.lf/eight.lf, /two.lf/zero.lf/one.lf/five.lf. [Online]. Available: https://scholarship.law.duke.edu/
dlj/vol/six.lf/five.lf/iss/three.lf//four.lf/
[/one.lf/zero.lf/three.lf] J. A. Kroll, J. Huey, S. Barocas, E. W. Felten, J. R. Reidenberg, D. G.
Robinson, and H. Yu, “Accountable algorithms,” University of Pennsylvania
Law Review , vol. /one.lf/six.lf/five.lf, no. /three.lf, pp. /six.lf/three.lf/three.lf–/seven.lf/zero.lf/five.lf, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https:
//www.pennlawreview.com//two.lf/zero.lf/one.lf/seven.lf//zero.lf/two.lf//two.lf/three.lf/accountable-algorithms/
[/one.lf/zero.lf/four.lf] E. Kuo, B. Iles, and M. R. Cruz, “Ampleforth: A new synthetic commodity,”
Ampleforth, Tech. Rep., /two.lf/zero.lf/one.lf/eight.lf. [Online]. Available: https://www.ampleforth.org/
papers/
[/one.lf/zero.lf/five.lf] L. Lessig, Code and Other Laws of Cyberspace . New York: Basic Books, /one.lf/nine.lf/nine.lf/nine.lf.
[/one.lf/zero.lf/six.lf] M. C. Libicki, Cyberdeterrence and Cyberwar . Santa Monica, CA: Rand Corporation,
/two.lf/zero.lf/zero.lf/nine.lf. [Online]. Available: https://www.rand.org/pubs/monographs/MG/eight.lf/seven.lf/seven.lf.html
[/one.lf/zero.lf/seven.lf] “White paper,” Libra Association, /two.lf/zero.lf/one.lf/nine.lf. [Online]. Available: https://www.diem.
com/en-us/white-paper/
[/one.lf/zero.lf/eight.lf] Y. Lindell and B. Pinkas, “Secure multiparty computation for privacy-preserving
data mining,” Journal of Privacy and Confidentiality , vol. /one.lf, no. /one.lf, pp. /five.lf/nine.lf–/nine.lf/eight.lf, /two.lf/zero.lf/zero.lf/nine.lf,
/eight.lf/two.lf
doi:/one.lf/zero.lf./two.lf/nine.lf/zero.lf/one.lf/two.lf/jpc.v/one.lfi/one.lf./five.lf/six.lf/six.lf.
[/one.lf/zero.lf/nine.lf] C. A. Mack, “Fifty years of Moore’s Law,” IEEE Transactions on Semiconductor Manu-
facturing , vol. /two.lf/four.lf, no. /two.lf, pp. /two.lf/zero.lf/two.lf–/two.lf/zero.lf/seven.lf, /two.lf/zero.lf/one.lf/one.lf, doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/TSM./two.lf/zero.lf/one.lf/zero.lf./two.lf/zero.lf/nine.lf/six.lf/four.lf/three.lf/seven.lf.
[/one.lf/one.lf/zero.lf] A. Martin and I. Martinovic, “Security and privacy impacts of a unique
personal identi/fi.ligaer,” University of Oxford Cyber Studies Programme, /two.lf/zero.lf/one.lf/six.lf.
[Online]. Available: https://ora.ox.ac.uk/objects/uuid:/nine.lf/zero.lfcf/one.lf/four.lfa/one.lf-beb/three.lf-/four.lf/three.lf/two.lf/two.lf-b/one.lf/eight.lfd-
de/f\_f.ligae/eight.lfc/seven.lff/eight.lf/six.lf/one.lf
[/one.lf/one.lf/one.lf] J. Matonis, “WikiLeaks bypasses /fi.liganancial blockade with bitcoin,” Aug /two.lf/zero.lf/one.lf/two.lf. [On-
line]. Available: https://www.forbes.com/sites/jonmatonis//two.lf/zero.lf/one.lf/two.lf//zero.lf/eight.lf//two.lf/zero.lf/wikileaks-
bypasses-/fi.liganancial-blockade-with-bitcoin
[/one.lf/one.lf/two.lf] D. M. McEvoy, “Enforcing compliance with international environmental agreements
using a deposit-refund system,” International Environmental Agreements: Politics, Law
and Economics , vol. /one.lf/three.lf, no. /four.lf, pp. /four.lf/eight.lf/one.lf–/four.lf/nine.lf/six.lf, /two.lf/zero.lf/one.lf/three.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf/s/one.lf/zero.lf/seven.lf/eight.lf/four.lf-/zero.lf/one.lf/three.lf-/nine.lf/two.lf/zero.lf/nine.lf-/two.lf.
[/one.lf/one.lf/three.lf] R. Merkle, “DAOs, democracy and governance,” Cryonics Magazine , vol. /three.lf/seven.lf, no. /four.lf, pp.
/two.lf/eight.lf–/four.lf/zero.lf, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://www.alcor.org/docs/cryonics-magazine-
/two.lf/zero.lf/one.lf/six.lf-/zero.lf/four.lf.pdf#page=/two.lf/eight.lf
[/one.lf/one.lf/four.lf] R. C. Merkle, “Secrecy, authentication, and public key systems,” Information Sys-
tems Laboratory, Stanford University, Tech. Rep., /one.lf/nine.lf/seven.lf/nine.lf.
[/one.lf/one.lf/five.lf] M. Minzner, “Putting probability back into probable cause,” Texas Law Review ,
vol. /eight.lf/seven.lf, no. /five.lf, pp. /nine.lf/one.lf/three.lf–/nine.lf/six.lf/two.lf, /two.lf/zero.lf/zero.lf/nine.lf. [Online]. Available: https://digitalrepository.unm.
edu/law\_facultyscholarship//four.lf/seven.lf/eight.lf/
[/one.lf/one.lf/six.lf] M. Möser, K. Soska, E. Heilman, H. Lee, Kevin He/f\_f.ligaan, S. Srivastava, K. Hogan,
J. Hennessey, A. Miller, A. Narayanan, and N. Christin, “An empirical analysis
of linkability in the Monero blockchain,” arXiv preprint arXiv:/one.lf/seven.lf/zero.lf/four.lf./zero.lf/four.lf/two.lf/nine.lf/nine.lf, /two.lf/zero.lf/one.lf/seven.lf.
[Online]. Available: https://arxiv.org/abs//one.lf/seven.lf/zero.lf/four.lf./zero.lf/four.lf/two.lf/nine.lf/nine.lf
[/one.lf/one.lf/seven.lf] M. Naehrig, K. Lauter, and V. Vaikuntanathan, “Can homomorphic encryption be
practical?” in Proceedings of the /three.lfrd ACM workshop on Cloud Computing Security Work-
shop. ACM, /two.lf/zero.lf/one.lf/one.lf, pp. /one.lf/one.lf/three.lf–/one.lf/two.lf/four.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/zero.lf/four.lf/six.lf/six.lf/six.lf/zero.lf./two.lf/zero.lf/four.lf/six.lf/six.lf/eight.lf/two.lf.
[/one.lf/one.lf/eight.lf] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” /two.lf/zero.lf/zero.lf/eight.lf.
[/one.lf/one.lf/nine.lf] A. Narayanan, J. Bonneau, E. Felten, A. Miller, and S. Goldfeder, Bitcoin and Crypto-
currency Technologies: A Comprehensive Introduction . Princeton, NJ: Princeton Uni-
versity Press, /two.lf/zero.lf/one.lf/six.lf.
[/one.lf/two.lf/zero.lf] D. Nguyen, “Venezuela’s bitcoin story puts it in a category of one,” NASDAQ,
/two.lf/zero.lf/two.lf/zero.lf. [Online]. Available: https://www.nasdaq.com/articles/venezuelas-bitcoin-
story-puts-it-in-a-category-of-one-/two.lf/zero.lf/two.lf/zero.lf-/one.lf/one.lf-/one.lf/one.lf
/eight.lf/three.lf
[/one.lf/two.lf/one.lf] J. Nielsen, “Nielsens law of internet bandwidth,” Nielsen Norman Group, /one.lf/nine.lf/nine.lf/eight.lf.
[Online]. Available: https://www.nngroup.com/articles/law-of-bandwidth/
[/one.lf/two.lf/two.lf] D. C. North, “Institutions, transaction costs and economic growth,” Economic In-
quiry , vol. /two.lf/five.lf, no. /three.lf, pp. /four.lf/one.lf/nine.lf–/four.lf/two.lf/eight.lf, /one.lf/nine.lf/eight.lf/seven.lf, doi:/one.lf/zero.lf./one.lf/one.lf/one.lf/one.lf/j./one.lf/four.lf/six.lf/five.lf-/seven.lf/two.lf/nine.lf/five.lf./one.lf/nine.lf/eight.lf/seven.lf.tb/zero.lf/zero.lf/seven.lf/five.lf/zero.lf.x.
[/one.lf/two.lf/three.lf] M. Olson, The logic of collective action: Public Goods and the Theory of Groups . Cam-
bridge, MA: Harvard University Press, /two.lf/zero.lf/zero.lf/nine.lf.
[/one.lf/two.lf/four.lf] A. O’Neill, “De/fi.liganitional issues in functional encryption,” IACR Cryptology
ePrint Archive, Report /two.lf/zero.lf/one.lf/zero.lf//five.lf/five.lf/six.lf, p. /five.lf/five.lf/six.lf, /two.lf/zero.lf/one.lf/zero.lf. [Online]. Available: https:
//eprint.iacr.org//two.lf/zero.lf/one.lf/zero.lf//five.lf/five.lf/six.lf
[/one.lf/two.lf/five.lf] P. O’Neill, Verification in an Age of Insecurity: The Future of Arms Control Compliance .
Oxford: Oxford University Press, /two.lf/zero.lf/zero.lf/nine.lf.
[/one.lf/two.lf/six.lf] OpenMined, “Openmined welcome package,” GitHub, /two.lf/zero.lf/two.lf/one.lf. [Online]. Available:
https://github.com/OpenMined/OM-Welcome-Package
[/one.lf/two.lf/seven.lf] E. Ortiz-Ospina and M. Roser, “Trust,” Our World in Data. [Online]. Available:
https://ourworldindata.org/trust
[/one.lf/two.lf/eight.lf] E. Ostrom, J. Burger, C. B. Field, R. B. Norgaard, and D. Policansky, “Revisiting the
commons: Local lessons, global challenges,” Science , vol. /two.lf/eight.lf/four.lf, no. /five.lf/four.lf/one.lf/two.lf, pp. /two.lf/seven.lf/eight.lf–/two.lf/eight.lf/two.lf,
/one.lf/nine.lf/nine.lf/nine.lf, doi:/one.lf/zero.lf./one.lf/one.lf/two.lf/six.lf/science./two.lf/eight.lf/four.lf./five.lf/four.lf/one.lf/two.lf./two.lf/seven.lf/eight.lf.
[/one.lf/two.lf/nine.lf] A. Pasternack, “Police body cameras will do more than just record you,” Fast
Company , Mar /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://www.fastcompany.com//three.lf/zero.lf/six.lf/one.lf/nine.lf/three.lf/five.lf/
police-body-cameras-livestreaming-face-recognition-and-ai
[/one.lf/three.lf/zero.lf] J. Peterson, J. Krug, M. Zoltu, A. K. Williams, and S. Alexander, “Augur:
A decentralized oracle and prediction market platform (v/two.lf./zero.lf),” arXiv preprint
arXiv:/one.lf/five.lf/zero.lf/one.lf./zero.lf/one.lf/zero.lf/four.lf/two.lf, Forecast Foundation, /two.lf/zero.lf/one.lf/nine.lf. [Online]. Available: https://arxiv.org/
abs//one.lf/five.lf/zero.lf/one.lf./zero.lf/one.lf/zero.lf/four.lf/two.lf
[/one.lf/three.lf/one.lf] N. Popper, Digital Gold: Bitcoin and the Inside Story of the Misfits and Millionaires Trying
to Reinvent Money . New York: HarperCollins, /two.lf/zero.lf/one.lf/six.lf.
[/one.lf/three.lf/two.lf] ——, “A hacking of more than /five.lf/zero.lf million dashes hopes in the world
of virtual currency,” The New York Times , Jun /two.lf/zero.lf/one.lf/six.lf. [Online]. Avail-
able: https://www.nytimes.com//two.lf/zero.lf/one.lf/six.lf//zero.lf/six.lf//one.lf/eight.lf/business/dealbook/hacker-may-have-
removed-more-than-/five.lf/zero.lf-million-from-experimental-cybercurrency-project.html
[/one.lf/three.lf/three.lf] R. Powell, “War as a commitment problem,” International Organization , vol. /six.lf/zero.lf, no. /one.lf,
pp. /one.lf/six.lf/nine.lf–/two.lf/zero.lf/three.lf, /two.lf/zero.lf/zero.lf/six.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/one.lf/seven.lf/S/zero.lf/zero.lf/two.lf/zero.lf/eight.lf/one.lf/eight.lf/three.lf/zero.lf/six.lf/zero.lf/six.lf/zero.lf/zero.lf/six.lf/one.lf.
[/one.lf/three.lf/four.lf] F. Reid and M. Harrigan, “An analysis of anonymity in the bitcoin system,” in Security
and Privacy in Social networks , Y. Altshuler, Y. Elovici, A. B. Cremers, N. Aharony, and
/eight.lf/four.lf
A. Pentland, Eds. New York: Springer, /two.lf/zero.lf/one.lf/three.lf, pp. /one.lf/nine.lf/seven.lf–/two.lf/two.lf/three.lf, doi:/one.lf/zero.lf./one.lf/zero.lf/zero.lf/seven.lf//nine.lf/seven.lf/eight.lf-/one.lf-/four.lf/six.lf/one.lf/four.lf-
/four.lf/one.lf/three.lf/nine.lf-/seven.lf\_/one.lf/zero.lf.
[/one.lf/three.lf/five.lf] C. Reitwiessner, “An update on integrating Zcash on Ethereum (ZoE),” Ethereum
Blog, Jan /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://blog.ethereum.org//two.lf/zero.lf/one.lf/seven.lf//zero.lf/one.lf//one.lf/nine.lf/update-
integrating-zcash-ethereum/
[/one.lf/three.lf/six.lf] R. L. Rivest, A. Shamir, and L. Adleman, “A method for obtaining digital signatures
and public-key cryptosystems,” Communications of the ACM , vol. /two.lf/one.lf, no. /two.lf, pp. /one.lf/two.lf/zero.lf–
/one.lf/two.lf/six.lf, /one.lf/nine.lf/seven.lf/eight.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//three.lf/five.lf/nine.lf/three.lf/four.lf/zero.lf./three.lf/five.lf/nine.lf/three.lf/four.lf/two.lf.
[/one.lf/three.lf/seven.lf] S. Rosenberg, “Can an arcane crypto ledger replace Uber, Spotify and AirBnB?”
Wired , Jan /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://www.wired.com//two.lf/zero.lf/one.lf/six.lf//zero.lf/one.lf/can-an-
arcane-crypto-ledger-replace-uber-spotify-and-airbnb/
[/one.lf/three.lf/eight.lf] J. Rowlatt, “How Bitcoin’s vast energy use could burst its bubble,” BBC News , /two.lf/zero.lf/two.lf/one.lf.
[Online]. Available: https://www.bbc.co.uk/news/science-environment-/five.lf/six.lf/two.lf/one.lf/five.lf/seven.lf/eight.lf/seven.lf
[/one.lf/three.lf/nine.lf] A. Sahai and B. Waters, “How to use indistinguishability obfuscation: deniable en-
cryption, and more,” in Proceedings of the Forty-Sixth Annual ACM Symposium on The-
ory of Computing . ACM, /two.lf/zero.lf/one.lf/four.lf, pp. /four.lf/seven.lf/five.lf–/four.lf/eight.lf/four.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/five.lf/nine.lf/one.lf/seven.lf/nine.lf/six.lf./two.lf/five.lf/nine.lf/one.lf/eight.lf/two.lf/five.lf.
[/one.lf/four.lf/zero.lf] N. Saper, “International cryptography regulation and the global information
economy,” Northwestern Journal of Technology and Intellectual Property , vol. /one.lf/one.lf,
no. /seven.lf, pp. /six.lf/seven.lf/three.lf–/six.lf/eight.lf/eight.lf, /two.lf/zero.lf/one.lf/three.lf. [Online]. Available: https://scholarlycommons.law.
northwestern.edu/njtip/vol/one.lf/one.lf/iss/seven.lf//five.lf
[/one.lf/four.lf/one.lf] E. B. Sasson, A. Chiesa, C. Garman, M. Green, I. Miers, E. Tromer, and M. Virza, “Ze-
rocash: Decentralized anonymous payments from bitcoin,” in /two.lf/zero.lf/one.lf/four.lf IEEE Symposium
on Security and Privacy . IEEE, /two.lf/zero.lf/one.lf/four.lf, pp. /four.lf/five.lf/nine.lf–/four.lf/seven.lf/four.lf, doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/SP./two.lf/zero.lf/one.lf/four.lf./three.lf/six.lf.
[/one.lf/four.lf/two.lf] T. C. Schelling, The Strategy of Conflict . Cambridge, MA: Harvard University Press,
/one.lf/nine.lf/eight.lf/zero.lf.
[/one.lf/four.lf/three.lf] N. Schick, Deep Fakes and the Infocalypse: What You Urgently Need to Know . London:
Hachette UK, /two.lf/zero.lf/two.lf/zero.lf.
[/one.lf/four.lf/four.lf] A. Schotter, The Economic Theory of Social Institutions . Cambridge, UK: Cambridge
University Press, /one.lf/nine.lf/eight.lf/one.lf.
[/one.lf/four.lf/five.lf] A. Segal, “Design and implementation of privacy-preserving surveillance,” Ph.D.
dissertation, Yale University, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://search.proquest.
com/openview//seven.lf/eight.lf/one.lfbcd/nine.lfc/eight.lf/four.lf/four.lf/three.lfaaa/two.lf/eight.lfa/six.lf/three.lfb/zero.lf/six.lf/one.lfe/eight.lf/one.lfa/seven.lf/four.lfe/four.lf/
[/one.lf/four.lf/six.lf] A. Segal, B. Ford, and J. Feigenbaum, “Catching bandits and only bandits: Privacy-
preserving intersection warrants for lawful surveillance,” in /four.lfth USENIX Workshop
on Free and Open Communications on the Internet (FOCI /one.lf/four.lf) , /two.lf/zero.lf/one.lf/four.lf. [Online]. Available:
https://www.usenix.org/conference/foci/one.lf/four.lf/workshop-program/presentation/segal
/eight.lf/five.lf
[/one.lf/four.lf/seven.lf] M. J. Sergot, F. Sadri, R. A. Kowalski, F. Kriwaczek, P. Hammond, and H. T. Cory,
“The British Nationality Act as a logic program,” Communications of the ACM , vol. /two.lf/nine.lf,
no. /five.lf, pp. /three.lf/seven.lf/zero.lf–/three.lf/eight.lf/six.lf, /one.lf/nine.lf/eight.lf/six.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//five.lf/six.lf/eight.lf/nine.lf./five.lf/nine.lf/two.lf/zero.lf.
[/one.lf/four.lf/eight.lf] A. Shamir, “How to share a secret,” Communications of the ACM , vol. /two.lf/two.lf, no. /one.lf/one.lf, pp.
/six.lf/one.lf/two.lf–/six.lf/one.lf/three.lf, /one.lf/nine.lf/seven.lf/nine.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//three.lf/five.lf/nine.lf/one.lf/six.lf/eight.lf./three.lf/five.lf/nine.lf/one.lf/seven.lf/six.lf.
[/one.lf/four.lf/nine.lf] E. Silfversten, M. Favaro, L. Slapakova, S. Ishikawa, J. Liu, and A. Salas, “Exploring
the use of Zcash cryptocurrency for illicit or criminal purposes,” RAND, /two.lf/zero.lf/two.lf/zero.lf.
[Online]. Available: https://www.rand.org/pubs/research\_reports/RR/four.lf/four.lf/one.lf/eight.lf.html
[/one.lf/five.lf/zero.lf] G. Singh, “FIFE: A framework for investigating functional encryption,” Master’s
thesis, Massachusetts Institute of Technology, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https:
//dspace.mit.edu/handle//one.lf/seven.lf/two.lf/one.lf./one.lf//one.lf/zero.lf/six.lf/three.lf/seven.lf/five.lf
[/one.lf/five.lf/one.lf] N. P. Smart, Cryptography Made Simple . Cham, Switzerland: Springer, /two.lf/zero.lf/one.lf/six.lf.
[/one.lf/five.lf/two.lf] D. J. Solove, Nothing to Hide: The False Tradeo/f\_f.liga between Privacy and Security . New
Haven: Yale University Press, /two.lf/zero.lf/one.lf/one.lf.
[/one.lf/five.lf/three.lf] Sta/f\_f.liga of Global Legal Research Directorate, “Regulation of cryptocurrency around
the world,” The Law Library of Congress, Tech. Rep., /two.lf/zero.lf/one.lf/eight.lf. [Online]. Available:
https://www.loc.gov/law/help/cryptocurrency/world-survey.php
[/one.lf/five.lf/four.lf] J. Stark, “Making sense of cryptoeconomics,” CoinDesk, Aug /two.lf/zero.lf/one.lf/seven.lf. [Online].
Available: https://www.coindesk.com/making-sense-cryptoeconomics/
[/one.lf/five.lf/five.lf] “Distribution of Bitcoin’s network hashrate in the last /two.lf/four.lf hours until February
/one.lf/seven.lf, /two.lf/zero.lf/two.lf/one.lf,” Statista, /two.lf/zero.lf/two.lf/one.lf. [Online]. Available: https://www.statista.com/statistics/
/seven.lf/three.lf/one.lf/four.lf/one.lf/six.lf/market-share-of-mining-pools/
[/one.lf/five.lf/six.lf] K. Stinchcombe, “Ten years in, nobody has come up with a use for blockchain,”
Hackernoon , /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://medium.com/@kaistinchcombe/ten-
years-in-nobody-has-come-up-with-a-use-case-for-blockchain-ee/nine.lf/eight.lfc/one.lf/eight.lf/zero.lf/one.lf/zero.lf/zero.lf
[/one.lf/five.lf/seven.lf] M. Swan, Blockchain: Blueprint for a New Economy . Sebastopol, CA: O’Reilly, /two.lf/zero.lf/one.lf/five.lf.
[/one.lf/five.lf/eight.lf] J. Swearingen, “Something weird (or weirder than normal) is happening at
WikiLeaks,” New York , Nov /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: http://nymag.com/selectall/
/two.lf/zero.lf/one.lf/six.lf//one.lf/one.lf/wikileak-hashes-dont-match-so-whats-going-on.html
[/one.lf/five.lf/nine.lf] N. Szabo, “The idea of smart contracts,” /one.lf/nine.lf/nine.lf/seven.lf. [Online]. Available: https:
//nakamotoinstitute.org/the-idea-of-smart-contracts/
[/one.lf/six.lf/zero.lf] R. Talviste, “Applying secure multi-party computation in practice,” Ph.D.
dissertation, University of Tartu, /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://core.ac.uk/
download/pdf//one.lf/four.lf/four.lf/seven.lf/zero.lf/eight.lf/nine.lf/three.lf/one.lf.pdf
/eight.lf/six.lf
[/one.lf/six.lf/one.lf] J. Tang, A. Korolova, X. Bai, X. Wang, and X. Wang, “Privacy loss in
Apple’s implementation of di/f\_f.ligaerential privacy on MacOS /one.lf/zero.lf./one.lf/two.lf,” arXiv preprint
arXiv:/one.lf/seven.lf/zero.lf/nine.lf./zero.lf/two.lf/seven.lf/five.lf/three.lf, /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: https://arxiv.org/abs//one.lf/seven.lf/zero.lf/nine.lf./zero.lf/two.lf/seven.lf/five.lf/three.lf
[/one.lf/six.lf/two.lf] M. Taylor and S. Singleton, “The communal resource: Transaction costs and the
solution of collective action problems,” Politics & Society , vol. /two.lf/one.lf, no. /two.lf, pp. /one.lf/nine.lf/five.lf–/two.lf/one.lf/four.lf,
/one.lf/nine.lf/nine.lf/three.lf, doi:/one.lf/zero.lf./one.lf/one.lf/seven.lf/seven.lf//zero.lf/zero.lf/three.lf/two.lf/three.lf/two.lf/nine.lf/two.lf/nine.lf/three.lf/zero.lf/two.lf/one.lf/zero.lf/zero.lf/two.lf/zero.lf/zero.lf/four.lf.
[/one.lf/six.lf/three.lf] J. Tirole, “Incomplete contracts: Where do we stand?” Econometrica , vol. /six.lf/seven.lf, no. /four.lf,
pp. /seven.lf/four.lf/one.lf–/seven.lf/eight.lf/one.lf, /one.lf/nine.lf/nine.lf/nine.lf, doi:/one.lf/zero.lf./one.lf/one.lf/one.lf/one.lf//one.lf/four.lf/six.lf/eight.lf-/zero.lf/two.lf/six.lf/two.lf./zero.lf/zero.lf/zero.lf/five.lf/two.lf.
[/one.lf/six.lf/four.lf] P. Todd, “Cypherpunk desert bus: My role in the /two.lf/zero.lf/one.lf/six.lf Zcash trusted setup
ceremony,” Nov /two.lf/zero.lf/one.lf/six.lf. [Online]. Available: https://petertodd.org//two.lf/zero.lf/one.lf/six.lf/cypherpunk-
desert-bus-zcash-trusted-setup-ceremony
[/one.lf/six.lf/five.lf] A. Trask, “Building safe A.I.: A tutorial for encrypted deep learning,” March /two.lf/zero.lf/one.lf/seven.lf.
[Online]. Available: https://iamtrask.github.io//two.lf/zero.lf/one.lf/seven.lf//zero.lf/three.lf//one.lf/seven.lf/safe-ai/
[/one.lf/six.lf/six.lf] ——, “Safe crime detection: Homomorphic encryption and deep learning for
more e/f\_f.ligaective, less intrusive digital surveillance,” Jun /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available:
https://iamtrask.github.io//two.lf/zero.lf/one.lf/seven.lf//zero.lf/six.lf//zero.lf/five.lf/homomorphic-surveillance/
[/one.lf/six.lf/seven.lf] A. Trask, E. Bluemke, B. Gar/fi.ligankel, C. G. Cuervas-Mons, and A. Dafoe, “Beyond
privacy trade-o/f\_f.ligas with structured transparency,” arXiv preprint arXiv:/two.lf/zero.lf/one.lf/two.lf./zero.lf/eight.lf/three.lf/four.lf/seven.lf,
/two.lf/zero.lf/two.lf/zero.lf. [Online]. Available: https://arxiv.org/abs//two.lf/zero.lf/one.lf/two.lf./zero.lf/eight.lf/three.lf/four.lf/seven.lf
[/one.lf/six.lf/eight.lf] M. Tsukerman, “The block is hot: A survey of the state of Bitcoin regulation and
suggestions for the future,” Berkeley Technology Law Journal , vol. /three.lf/zero.lf, no. Annual
Review, pp. /one.lf/one.lf/two.lf/seven.lf–/one.lf/one.lf/seven.lf/zero.lf, /two.lf/zero.lf/one.lf/five.lf. [Online]. Available: https://btlj.org/data/articles/two.lf/zero.lf/one.lf/five.lf/
vol/three.lf/zero.lf//three.lf/zero.lf\_AR//one.lf/one.lf/two.lf/seven.lf-/one.lf/one.lf/seven.lf/zero.lf\_Tsukerman\_Final%/two.lf/zero.lf/one.lf/one.lf/one.lf/nine.lf/one.lf/five.lf.pdf
[/one.lf/six.lf/nine.lf] C. Turley, “Moloch evolved: V/two.lf primer,” Raid Guild , /two.lf/zero.lf/two.lf/zero.lf. [Online]. Available:
https://medium.com/raid-guild/moloch-evolved-v/two.lf-primer-/two.lf/five.lfc/nine.lfcdeab/four.lf/five.lf/five.lf
[/one.lf/seven.lf/zero.lf] M. Une, “The security evaluation of time stamping schemes: The present
situation and studies,” IMES Discussion Papers Series /two.lf/zero.lf/zero.lf/one.lf-E-/one.lf/eight.lf, Institute for
Monetary and Economic Studies, Bank of Japan, /two.lf/zero.lf/zero.lf/one.lf. [Online]. Available:
https://www.imes.boj.or.jp/research/papers/english//zero.lf/one.lf-E-/one.lf/eight.lf.pdf
[/one.lf/seven.lf/one.lf] V. Vaikuntanathan, “Computing blindfolded: New developments in fully homomor-
phic encryption,” in /two.lf/zero.lf/one.lf/one.lf IEEE /five.lf/two.lfnd Annual Symposium on Foundations of Computer
Science (FOCS) . IEEE, /two.lf/zero.lf/one.lf/one.lf, pp. /five.lf–/one.lf/six.lf, doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/FOCS./two.lf/zero.lf/one.lf/one.lf./nine.lf/eight.lf.
[/one.lf/seven.lf/two.lf] J. van den Hoo/f\_f.liga, D. Lazar, M. Zaharia, and N. Zeldovich, “Vuvuzela: Scalable pri-
vate messaging resistant to tra/f\_f\_i.ligac analysis,” in Proceedings of the /two.lf/five.lfth Symposium on
Operating Systems Principles . ACM, /two.lf/zero.lf/one.lf/five.lf, pp. /one.lf/three.lf/seven.lf–/one.lf/five.lf/two.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/eight.lf/one.lf/five.lf/four.lf/zero.lf/zero.lf./two.lf/eight.lf/one.lf/five.lf/four.lf/one.lf/seven.lf.
[/one.lf/seven.lf/three.lf] P. Vigna and M. J. Casey, “Bitcoin for the unbanked,” Foreign A/f\_f.ligaairs , /two.lf/zero.lf/one.lf/five.lf. [Online].
/eight.lf/seven.lf
Available: https://www.foreigna/f\_f.ligaairs.com/articles//two.lf/zero.lf/one.lf/five.lf-/zero.lf/two.lf-/two.lf/six.lf/bitcoin-unbanked
[/one.lf/seven.lf/four.lf] J. Vogler, The Global Commons: Environmental and Technological Governance . West
Sussex, England: Wiley, /two.lf/zero.lf/zero.lf/zero.lf.
[/one.lf/seven.lf/five.lf] G. Volpicelli, “How the blockchain is helping stop the spread of con/fl.ligaict
diamonds,” Wired UK , Feb /two.lf/zero.lf/one.lf/seven.lf. [Online]. Available: http://www.wired.co.uk/
article/blockchain-con/fl.ligaict-diamonds-everledger
[/one.lf/seven.lf/six.lf] D. Vorick and L. Champine, “Sia: Simple decentralized storage,” /two.lf/zero.lf/one.lf/four.lf. [Online].
Available: https://sia.tech/sia.pdf
[/one.lf/seven.lf/seven.lf] S. Wagh, S. Tople, F. Benhamouda, E. Kushilevitz, P. Mittal, and T. Rabin, “FAL-
CON: Honest-majority maliciously secure framework for private deep learning,”
Proceedings on Privacy Enhancing Technologies , vol. /two.lf/zero.lf/two.lf/one.lf, no. /one.lf, pp. /one.lf/eight.lf/eight.lf–/two.lf/zero.lf/eight.lf, /two.lf/zero.lf/two.lf/one.lf,
doi:/one.lf/zero.lf./two.lf/four.lf/seven.lf/eight.lf/popets-/two.lf/zero.lf/two.lf/one.lf-/zero.lf/zero.lf/one.lf/one.lf.
[/one.lf/seven.lf/eight.lf] M. Wal/fi.ligash and A. J. Blumberg, “Verifying computations without reexecuting them,”
Communications of the ACM , vol. /five.lf/eight.lf, no. /two.lf, pp. /seven.lf/four.lf–/eight.lf/four.lf, /two.lf/zero.lf/one.lf/five.lf, doi:/one.lf/zero.lf./one.lf/one.lf/four.lf/five.lf//two.lf/six.lf/four.lf/one.lf/five.lf/six.lf/two.lf.
[/one.lf/seven.lf/nine.lf] M. Walport, “Distributed ledger technology: beyond block chain,”
United Kingdom Government O/f\_f\_i.ligace for Science, /two.lf/zero.lf/one.lf/five.lf. [Online]. Avail-
able: https://assets.publishing.service.gov.uk/government/uploads/system/uploads/
attachment\_data//fi.ligale//four.lf/nine.lf/two.lf/nine.lf/seven.lf/two.lf/gs-/one.lf/six.lf-/one.lf-distributed-ledger-technology.pdf
[/one.lf/eight.lf/zero.lf] O. E. Williamson, The Economic Institutions of Capitalism: Firms, Markets, Relational
Contracting . New York: Macmillan, /one.lf/nine.lf/eight.lf/five.lf.
[/one.lf/eight.lf/one.lf] B. Wittes and G. Blum, The Future of Violence: Robots and Germs, Hackers and Drones–
Confronting A New Age of Threat . New York: Basic Books, /two.lf/zero.lf/one.lf/five.lf.
[/one.lf/eight.lf/two.lf] A. Wright and P. De Filippi, “Decentralized blockchain technology and the rise of
lex cryptographia,” /two.lf/zero.lf/one.lf/five.lf, doi:/one.lf/zero.lf./two.lf/one.lf/three.lf/nine.lf/ssrn./two.lf/five.lf/eight.lf/zero.lf/six.lf/six.lf/four.lf.
[/one.lf/eight.lf/three.lf] L. Xie, “A beginner’s guide to DAOs,” Mirror, /two.lf/zero.lf/two.lf/one.lf. [Online]. Available:
https://linda.mirror.xyz/Vh/eight.lfK/four.lfleCGEO/zero.lf/six.lf\_qSGx-vS/five.lflvgUqhqkCz/nine.lfut/eight.lf/one.lfWwCP/two.lfo
[/one.lf/eight.lf/four.lf] A. Yakovenko, “Solana: A new architecture for a high performance blockchain
v/zero.lf./eight.lf./one.lf/four.lf,” Solana Labs, /two.lf/zero.lf/one.lf/eight.lf.
[/one.lf/eight.lf/five.lf] A. C. Yao, “Protocols for secure computations,” in /two.lf/three.lfrd Annual Symposium on Foun-
dations of Computer Science . IEEE, /one.lf/nine.lf/eight.lf/two.lf, pp. /one.lf/six.lf/zero.lf–/one.lf/six.lf/four.lf, doi:/one.lf/zero.lf./one.lf/one.lf/zero.lf/nine.lf/SFCS./one.lf/nine.lf/eight.lf/two.lf./three.lf/eight.lf.
[/one.lf/eight.lf/six.lf] “What are zk-SNARKs?” Zcash. [Online]. Available: https://z.cash/technology/
zksnarks.html
[/one.lf/eight.lf/seven.lf] G. Zyskind, O. Nathan, and A. Pentland, “Enigma: Decentralized computation
platform with guaranteed privacy,” arXiv preprint arXiv:/one.lf/five.lf/zero.lf/six.lf./zero.lf/three.lf/four.lf/seven.lf/one.lf, /two.lf/zero.lf/one.lf/five.lf. [Online].
/eight.lf/eight.lf
Available: https://arxiv.org/abs//one.lf/five.lf/zero.lf/six.lf./zero.lf/three.lf/four.lf/seven.lf/one.lf
/eight.lf/nine.lf
A Relevance of progress in arti/fi.ligacial intelligence
The signi/fi.ligacance of technological developments in one /fi.ligaeld often depends on what devel-
opments occur in other /fi.ligaelds. For example, the signi/fi.ligacance of last century’s developments
in cryptography would not have been nearly so great if it hadn’t been for the creation of
the internet and other novel communications technologies.
To take into account these interaction e/f\_f.ligaects, this section will brie/fl.ligay consider some of
the ways in which progress in arti/fi.ligacial intelligence and progress in cryptography could be
relevant to one another.
I list seven potential interaction points.
A./one.lf AI systems may enable more e/f\_f.ligaective surveillance
Progress in AI could enable more e/f\_f.ligaective surveillance by decreasing the cost of extracting
information from collected data [/four.lf/five.lf, /four.lf]. It is plausible, for example, that AI systems applied
to videos, messages, and social networks could become fairly e/f\_f.ligaective at automatically
identifying criminals, dissidents, or other groups that state actors would have an interest
in discovering, without the need for humans to sift through the relevant data by hand.
However, as discussed in section /three.lf./one.lf./one.lf, this trend might be undermined by developments
in cryptography—speci/fi.ligacally, by a move toward greater and more e/f\_f.ligaective use of end-to-
end encryption, cryptocurrencies capable of obscuring transaction details, and methods of
computing on con/fi.ligadential data which reduce the need for companies to collect personal
information.
A./two.lf AI systems may help to make privacy-preserving surveillance feasi-
ble
If AI systems /fi.ligand greater applications in surveillance, then this could also open the door
for surveillance that o/f\_f.ligaers a greater degree of privacy.
As discussed in section /three.lf./one.lf./two.lf, one way to achieve this result is to apply techniques for com-
puting on con/fi.ligadential data. These techniques could make it possible to extract security-
relevant information from surveillance data without having access to it in unencrypted
form.
In addition, AI systems could also enable privacy-preserving surveillance more directly.
The point, here, is that encryption is only one way to prevent users from gaining access to
extraneous information in surveillance data. It is also possible to obfuscate data in a more
direct manner, for example by removing labels or, in di/f\_f.ligaerential privacy techniques, by
adding random alterations [/two.lf]. Similarly, AI systems may be able to perform more precise
censorship of collected data. An early example of this idea is the proposed use of face-
blurring algorithms for police body-camera footage [/one.lf/two.lf/nine.lf].
/nine.lf/zero.lf
A./three.lf AI systems may increase the need for anti-forgery schemes
Progress in AI continues to make fake photographs and videos more convincing and cheap
to produce [/four.lf/nine.lf]. If this trend continues, it could become increasingly di/f\_f\_i.ligacult to distin-
guish true claims from false ones, with important negative consequences for politics, law
enforcement, and news reporting [/four.lf]. Schemes of the sort described in section /three.lf./one.lf./four.lf, which
use trusted timestamping to provide evidence for the veracity of images, could be impor-
tant tools for avoiding these consequences.
A./four.lf Methods of computing on con/fi.ligadential data could reduce barriers
to developing certain AI systems
Developing AI systems using machine learning methods typically requires access to large
volumes of relevant data. Con/fi.ligadentiality concerns are therefore one natural barrier to
AI development in certain domains. For example, for good reason, it is typically rather
di/f\_f\_i.ligacult to gain access to large volumes of medical data, even if this data could be used in
a socially bene/fi.ligacial fashion.
Methods of computing on con/fi.ligadential data (see sections /two.lf./one.lf/one.lf–/two.lf./one.lf/three.lf), particularly secure
multiparty computation techniques, could reduce con/fi.ligadentiality concerns as a barrier to
AI development. While other techniques in privacy-preserving machine learning, such as
di/f\_f.ligaerential privacy, do exist, they are less generally applicable and su/f\_f.ligaer from signi/fi.ligacant
trade-o/f\_f.ligas between how e/f\_f.ligaective they are at preserving privacy and how much they still
allow machines to learn [/two.lf].
One illustrative project here is OpenMined, which aims to make it much easier for users
to contribute their data to machine learning projects that are conducted in a privacy-
preserving fashion [/one.lf/two.lf/six.lf].
A./five.lf The problems of safe AI design and safe smart contract design may
be connected
One broad problem in the emerging /fi.ligaeld of “AI safety” is the problem of designing AI
systems that will not exhibit unintended harmful behaviors [/five.lf, /four.lf]. For instance, there is
not yet any general method for ensuring that an AI system trained or programmed to
behave well in a limited set of environments will not cause accidents if it is deployed in a
wider range of real-world environments. We can already point to examples of AI accidents
such as the /two.lf/zero.lf/one.lf/zero.lf “/fl.ligaash crash,” in which the behavior of automated trading systems caused a
trillion-dollar stock market crash, and fatal collisions that have occurred with self-driving
cars. In the future, as AI systems are used to automate increasingly complex and crucial
tasks, techniques for avoiding accidents could become much more important [/two.lf/eight.lf].
Similarly, as discussed in section /four.lf./two.lf./four.lf, one important factor restricting the applications of
smart contracts is the need to ensure that they will behave as intended. The problem here
is made especially severe by the fact that smart contracts cannot be modi/fi.ligaed once created.
The infamous collapse of the /dollar.lf/one.lf/five.lf/zero.lf million DAO venture capital fund, described in section
/three.lf./two.lf./four.lf, is a good illustration of the need to get smart contracts right [/one.lf/three.lf/two.lf].
/nine.lf/one.lf
With such incidents in mind, Vitalik Buterin has written that the problem of designing
reliable AI systems and the problem of designing reliable smart contracts overlap, and that
researchers working on each of these problems could bene/fi.ligat from talking to those working
on the other [/four.lf/zero.lf].
As a point of disanalogy, however, it is important to note that current “scalability” con-
straints (see section /four.lf./two.lf./one.lf) severely limit the amount of computing power that smart con-
tracts can draw from, to the extent that it is non-trivial to implement something even so
simple as a smart contract that judges who has won a game of chess. This means that smart
contracts are quite distinct from the advanced AI systems that AI safety researchers pri-
marily have in mind. Speci/fi.ligacally, this also means that, barring extremely large increases
in available computing power, no non-trivial AI system could actually be run as a smart
contract./five.lf/four.lf
A./six.lf New coordination and veri/fi.ligacation mechanisms may be useful for
governing AI systems
Generally, if progress in AI creates new security challenges—for example, by enabling au-
tonomous weapons systems, more damaging categories of cyberweapons, or other systems
associated with substantial accident risks—then there could be a need for international
agreements and other forms of global governance to guide its application and develop-
ment [/nine.lf/zero.lf, /two.lf/nine.lf, /three.lf/two.lf]. If smart contracts ever become su/f\_f\_i.ligaciently reliable, then it is possi-
ble—although, for reasons discussed in section /four.lf./two.lf./three.lf, probably not very likely—that they
could have applications in enforcing such agreements. Similarly, it is possible that zero-
knowledge proofs or methods for computing on con/fi.ligadential data could make it easier
to verify compliance without requiring the relevant parties to share too much sensitive
information (see section /three.lf./one.lf./three.lf).
A./seven.lf Fully homomorphic encryption may have applications in AI safety
and security
As described in an essay by Andrew Trask, fully homomorphic encryption could make it
possible to train AI systems using encrypted data or encrypted virtual environments [/one.lf/six.lf/five.lf].
The result of such training would be systems that cannot interact with the world, as they
are only capable of processing encrypted inputs and providing encrypted outputs. Plausi-
bly, such systems would o/f\_f.ligaer more security against attempts at theft and premature real-
world use (e.g., before safety properties are guaranteed). The idea is that the system would
only be able to interact with speci/fi.ligac encrypted digital environments, which might still be
useful for testing or further training, until the system is itself decrypted. Note, again, that
the future practicality of such a scheme would depend on the amount of computing power
required for fully homomorphic encryption, as well as the amount of computing power
available (see section /four.lf./one.lf./one.lf).
/five.lf/four.lfAt the same time, if smart contracts that “outsource” computations to other users become more common
(see section /four.lf./two.lf./one.lf), then the problems of AI safety and smart contract safety could become more concretely inter-
twined. If the relevant techniques become su/f\_f\_i.ligaciently e/f\_f\_i.ligacient and reliable, then there could arise smart contracts
that pay others to run more powerful AI systems, so that the systems are likely to continue running so long as
the smart contract does.
/nine.lf/two.lf
May 2021
Centre for the Governance of AI,
Future of Humanity Institute,
University of Oxford
www.governance.ai |
2f0579ee-6012-4ca6-986c-341d9b82432e | trentmkelly/LessWrong-43k | LessWrong | I'm open for projects (sort of)
I left Google a month ago, and right now don't work. Writing this post in case anyone has interesting ideas what I could do. This isn't an "urgently need help" kind of thing - I have a little bit of savings, right now planning to relax some more weeks and then go into some solo software work. But I thought I'd write this here anyway, because who knows what'll come up.
Some things about me. My degree was in math. My software skills are okayish: I left Google at L5 ("senior"), and also made a game that went semi-viral. I've also contributed a lot on LW, the most prominent examples being my formalizations of decision theory ideas (Löbian cooperation, modal fixpoints etc) and later the AI Alignment Prize that we ran with Paul and Zvi. Most of that was before the current AI wave; neural networks don't really "click" with my mind, so I haven't done much work on them.
And yeah, this is an invitation to throw at me not necessarily money-paying work, but also stuff you'd like me to look at, criticize, help with your own projects and so on. I find myself with a bit more free time now, so basically drop me a PM if you have something interesting to talk about :-) |
11bbfdbc-6e07-4764-933f-9433f52e6b1e | trentmkelly/LessWrong-43k | LessWrong | A Defense of Work on Mathematical AI Safety
AI Safety was, a decade ago, nearly synonymous with obscure mathematical investigations of hypothetical agentic systems. Fortunately or unfortunately, this has largely been overtaken by events; the successes of machine learning and the promise, or threat, of large language models has pushed thoughts of mathematics aside for many in the “AI Safety” community. The once pre-eminent advocate of this class of “agent foundations” research for AI safety, Eliezer Yudkowsky, has more recently said that timelines are too short to allow this agenda to have a significant impact. This conclusion seems at best premature.
Foundational research is useful for prosaic alignment
First, the value of foundational and mathematical research can be synergistic with both technical progress on safety, and with insight into how and where safety is critical. Many machine learning research agendas for safety are investigating issues identified years earlier by foundational research, and are at least partly informed by that research. Current mathematical research could play a similar role in the coming years, as more funding and research are increasingly available for safety. We have also repeatedly seen the importance of foundational research arguments in discussions of policy, from Bostrom’s book to policy discussions at OpenAI, Anthropic, and DeepMind. These connections may be more conceptual than direct, but they are still relevant.
Long timelines are possible
Second, timelines are uncertain. If timelines based on technical progress are short, many claim that we have years not decades until safety must be solved. But this assumes that policy and governance approaches fail, and that we therefore need a full technical solution in the short term. It also seems likely that short timelines make all approaches less likely to succeed. On the other hand, if timelines for technical progress are longer, fundamental advances in understanding, such as those provided by more foundational research, ar |
733ae3c6-0a5e-44d2-9316-df7ad380fba3 | trentmkelly/LessWrong-43k | LessWrong | To Lead, You Must Stand Up
Followup to: Lonely Dissent
True story: In July, I attended a certain Silicon Valley event. I was not an organizer, or a speaker, or in any other wise involved on an official level; just an attendee. It was an evening event, and after the main presentations were done, much of the audience hung around talking... and talking... and talking... Finally the event organizer began dimming the lights and turning them back up again. And the crowd still stayed; no one left. So the organizer dimmed the lights and turned them up some more. And lo, the people continued talking.
I walked over to the event organizer, standing by the light switches, and said, "Are you hinting for people to leave?" And he said, "Yes. In fact [the host company] says we've got to get out of here now - the building needs to close down."
I nodded.
I walked over to the exit.
I shouted, "LISTEN UP, EVERYONE! WE'VE GOT TO GO! OUR TIME HERE HAS PASSED! YOU CAN TALK OUTSIDE IF YOU LIKE! NOW FOLLOW ME... TO FREEDOM!"
I turned.
I marched out the door.
And everyone followed.
I expect there were at least two or three CEOs in that Silicon Valley crowd. It didn't lack for potential leaders. Why was it left to me to lead the CEOs to freedom?
Well, what was in it for them to perform that service to the group? It wasn't their problem. I'm in the habit of doing work I see being left undone; but this doesn't appear to be a common habit.
So why didn't some aspiring would-be future-CEO take the opportunity to distinguish themselves by acting the part of the leader? I bet at least five people in that Silicon Valley crowd had recently read a business book on leadership...
But it's terribly embarrassing to stand up in front of a crowd. What if the crowd hadn't followed me? What if I'd turned and marched out the door, and been left looking like a complete fool? Oh nos! Oh horrors!
While I have sometimes pretended to wisdom, I have never pretended to solemnity. I wasn't worried about looki |
d42db28b-89f7-43ed-ad3e-fc3cd21cf40c | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Summary of Stuart Russell's new book, "Human Compatible"
As part of the [Alignment Newsletter](https://rohinshah.com/alignment-newsletter), I wrote a summary of *[Human Compatible: Artificial Intelligence and the Problem of Control](https://www.amazon.com/Human-Compatible-Artificial-Intelligence-Problem-ebook/dp/B07N5J5FTS)*, along with summaries of some of the papers that influenced the book. I've copied over the text of the post in full, but have not copied the comments (which are also worth reading).
Find all Alignment Newsletter resources [here](http://rohinshah.com/alignment-newsletter/). In particular, you can [sign up](http://eepurl.com/dqMSZj), or look through this [spreadsheet](https://docs.google.com/spreadsheets/d/1PwWbWZ6FPqAgZWOoOcXM8N_tUCuxpEyMbN1NYYC02aM/edit?usp=sharing) of all summaries that have ever been in the newsletter. I'm always happy to hear feedback; you can send it to me by replying to this email. This is a bonus newsletter summarizing Stuart Russell's new book, along with summaries of a few of the most relevant papers. It's entirely written by Rohin, so the usual "summarized by" tags have been removed. We're also changing the publishing schedule: so far, we've aimed to send a newsletter every Monday; we're now aiming to send a newsletter every Wednesday.
Audio version [here](http://alignment-newsletter.libsyn.com/alignment-newsletter-69) (may not be up yet).
[Human Compatible: Artificial Intelligence and the Problem of Control](https://www.amazon.com/Human-Compatible-Artificial-Intelligence-Problem-ebook/dp/B07N5J5FTS) *(Stuart Russell)*: *Since I am aiming this summary for people who are already familiar with AI safety, my summary is substantially reorganized from the book, and skips large portions of the book that I expect will be less useful for this audience. If you are not familiar with AI safety, **note that I am skipping many arguments and counterarguments in the book that are aimed for you**. I'll refer to the book as "HC" in this newsletter.*
Before we get into details of impacts and solutions to the problem of AI safety, it's important to have a model of how AI development will happen. Many estimates have been made by figuring out the amount of compute needed to run a human brain, and figuring out how long it will be until we get there. HC doesn't agree with these; it suggests the bottleneck for AI is in the algorithms rather than the hardware. We will need several conceptual breakthroughs, for example in language or common sense understanding, cumulative learning (the analog of cultural accumulation for humans), discovering hierarchy, and managing mental activity (that is, the metacognition needed to prioritize what to think about next). It's not clear how long these will take, and whether there will need to be more breakthroughs after these occur, but these seem like necessary ones.
What could happen if we do get beneficial superintelligent AI? While there is a lot of sci-fi speculation that we could do here, as a weak lower bound, it should at least be able to automate away almost all existing human labor. Assuming that superintelligent AI is very cheap, most services and many goods would become extremely cheap. Even many primary products such as food and natural resources would become cheaper, as human labor is still a significant fraction of their production cost. If we assume that this could bring up everyone's standard of life up to that of the 88th percentile American, that would result in nearly a *tenfold* increase in world GDP per year. Assuming a 5% discount rate per year, this corresponds to $13.5 *quadrillion* net present value. Such a giant prize removes many reasons for conflict, and should encourage everyone to cooperate to ensure we all get to keep this prize.
Of course, this doesn't mean that there aren't any problems, even with AI that does what its owner wants. Depending on who has access to powerful AI systems, we could see a rise in automated surveillance, lethal autonomous weapons, automated blackmail, fake news and behavior manipulation. Another issue that could come up is that once AI is better than humans at all tasks, we may end up delegating everything to AI, and lose autonomy, leading to *human enfeeblement*.
This all assumes that we are able to control AI. However, we should be cautious about such an endeavor -- if nothing else, we should be careful about creating entities that are more intelligent than us. After all, the gorillas probably aren't too happy about the fact that their habitat, happiness, and existence depends on our moods and whims. For this reason, HC calls this the *gorilla problem*: specifically, "the problem of whether humans can maintain their supremacy and autonomy in a world that includes machines with substantially greater intelligence". Of course, we aren't in the same position as the gorillas: we get to *design* the more intelligent "species". But we should probably have some good arguments explaining why our design isn't going to succumb to the gorilla problem. This is especially important in the case of a fast intelligence explosion, or *hard takeoff*, because in that scenario we do not get any time to react and solve any problems that arise.
Do we have such an argument right now? Not really, and in fact there's an argument that we *will* succumb to the gorilla problem. The vast majority of research in AI and related fields assumes that there is some definite, known *specification* or *objective* that must be optimized. In RL, we optimize the *reward function*; in search, we look for states matching a *goal criterion*; in statistics, we minimize *expected loss*; in control theory, we minimize the *cost function* (typically deviation from some desired behavior); in economics, we design mechanisms and policies to maximize the *utility* of individuals, *welfare* of groups, or *profit* of corporations. This leads HC to propose the following standard model of machine intelligence: *Machines are intelligent to the extent that their actions can be expected to achieve their objectives.* However, if we put in the wrong objective, the machine's obstinate pursuit of that objective would lead to outcomes we won't like.
Consider for example the content selection algorithms used by social media, typically maximizing some measure of engagement, like click-through. Despite their lack of intelligence, such algorithms end up changing the user's preference so that they become more predictable, since more predictable users can be given items they are more likely to click on. In practice, this means that users are pushed to become more extreme in their political views. Arguably, these algorithms have already caused much damage to the world.
So the problem is that we don't know how to put our objectives inside of the AI system so that when it optimizes its objective, the results are good for us. Stuart calls this the "King Midas" problem: as the legend goes, King Midas wished that everything he touched would turn to gold, not realizing that "everything" included his daughter and his food, a classic case of a [badly specified objective](https://vkrakovna.wordpress.com/2018/04/02/specification-gaming-examples-in-ai/) ([AN #1](https://mailchi.mp/ff6340049bd0/alignment-newsletter-1)). In some sense, we've known about this problem for a long time, both from King Midas's tale, and in stories about genies, where the characters inevitably want to undo their wishes.
You might think that we could simply turn off the power to the AI, but that won't work, because for almost any definite goal, the AI has an incentive to stay operational, just because that is necessary for it to achieve its goal. This is captured in what may be Stuart's most famous quote: *you can't fetch the coffee if you're dead*. This is one of a few worrisome [convergent instrumental subgoals](https://selfawaresystems.files.wordpress.com/2008/01/ai_drives_final.pdf).
What went wrong? The problem was the way we evaluated machine intelligence, which doesn't take into account the fact that machines should be useful for *us*. HC proposes: *Machines are **beneficial** to the extent that **their** actions can be expected to achieve **our** objectives*. But with this definition, instead of our AI systems optimizing a definite, wrong objective, they will *also* be uncertain about the objective, since we ourselves don't know what our objectives are. HC expands on this by proposing three principles for the design of AI systems, that I'll quote here in full:
1. *The machine’s only objective is to maximize the realization of human preferences.*
2. *The machine is initially uncertain about what those preferences are.*
3. *The ultimate source of information about human preferences is human behavior.*
[Cooperative Inverse Reinforcement Learning](https://arxiv.org/abs/1606.03137) provides a formal model of an *assistance game* that showcases these principles. You might worry that an AI system that is uncertain about its objective will not be as useful as one that knows the objective, but actually this uncertainty is a feature, not a bug: it leads to AI systems that are deferential, that ask for clarifying information, and that try to learn human preferences. [The Off-Switch Game](https://arxiv.org/abs/1611.08219) shows that because the AI is uncertain about the reward, it will let itself be shut off. These papers are discussed later in this newsletter.
So that's the proposed solution. You might worry that the proposed solution is quite challenging: after all, it requires a shift in the entire way we do AI. What if the standard model of AI can deliver more results, even if just because more people work on it? Here, HC is optimistic: the big issue with the standard model is that it is not very good at learning our preferences, and there's a huge economic pressure to learn preferences. For example, I would pay a lot of money for an AI assistant that accurately learns my preferences for meeting times, and schedules them completely autonomously.
Another research challenge is how to actually put principle 3 into practice: it requires us to connect human behavior to human preferences. [Inverse Reward Design](https://arxiv.org/abs/1711.02827) and [Preferences Implicit in the State of the World](https://bair.berkeley.edu/blog/2019/02/11/learning_preferences/) ([AN #45](https://mailchi.mp/35b451cb4d70/alignment-newsletter-45)) are example papers that tackle portions of this. However, there are *lots* of subtleties in this connection. We need to use *Gricean semantics* for language: when we say X, we do not mean the literal meaning of X: the agent must also take into account the fact that we bothered to say X, and that we didn't say Y. For example, I'm only going to ask for the agent to buy a cup of coffee if I believe that there is a place to buy reasonably priced coffee nearby. If those beliefs happen to be wrong, the agent should ask for clarification, rather than trudge hundreds of miles or pay hundreds of dollars to ensure I get my cup of coffee.
Another problem with inferring preferences from behavior is that humans are nearly always in some deeply nested plan, and many actions don't even occur to us. Right now I'm writing this summary, and not considering whether I should become a fireman. I'm not writing this summary because I just ran a calculation showing that this would best achieve my preferences, I'm doing it because it's a subpart of the overall plan of writing this bonus newsletter, which itself is a subpart of other plans. The connection to my preferences is very far up. How do we deal with that fact?
There are perhaps more fundamental challenges with the notion of "preferences" itself. For example, our *experiencing self* and our *remembering self* may have different preferences -- if so, which one should our agent optimize for? In addition, our preferences often change over time: should our agent optimize for our current preferences, even if it knows that they will predictably change in the future? This one could potentially be solved by learning *meta-preferences* that dictate what kinds of preference change processes are acceptable.
All of these issues suggest that we need work across many fields (such as AI, cognitive science, psychology, and neuroscience) to reverse-engineer human cognition, so that we can put principle 3 into action and create a model that shows how human behavior arises from human preferences.
So far, we've been talking about the case with a single human. But of course, there are going to be multiple humans: how do we deal with that? As a baseline, we could imagine that every human gets their own agent that optimizes for their preferences. However, this will differentially benefit people who care less about other people's welfare, since their agents have access to many potential plans that wouldn't be available to an agent for someone who cared about other people. For example, if Harriet was going to be late for a meeting with Ivan, her AI agent might arrange for Ivan to be even later.
What if we had laws that prevented AI systems from acting in such antisocial ways? It seems likely that superintelligent AI would be able to find loopholes in such laws, so that they do things that are strictly legal but still antisocial, e.g. line-cutting. (This problem is similar to the problem that we can't just write down what we want and have AI optimize it.)
What if we made our AI systems utilitarian (assuming we figured out some acceptable method of comparing utilities across people)? Then we get the "Somalia problem": agents will end up going to Somalia to help the worse-off people there, and so no one would ever buy such an agent.
Overall, it's not obvious how we deal with the transition from a single human to multiple humans. While HC focuses on a potential solution for the single human / single agent case, there is still much more to be said and done to account for the impact of AI on all of humanity. To quote HC, "There is really no analog in our present world to the relationship we will have with beneficial intelligent machines in the future. It remains to be seen how the endgame turns out."
**Rohin's opinion:** I enjoyed reading this book; I don't usually get to read a single person's overall high-level view on the state of AI, how it could have societal impact, the argument for AI risk, potential solutions, and the need for AI governance. It's nice to see all of these areas I think about tied together into a single coherent view. While I agree with much of the book, especially the conceptual switch from the standard model of intelligent machines to Stuart's model of beneficial machines, I'm going to focus on disagreements in this opinion.
First, the book has an implied stance towards the future of AI research that I don't agree with: I could imagine that powerful AI systems end up being created by learning alone without needing the conceptual breakthroughs that Stuart outlines. This has been proposed in e.g. [AI-GAs](https://arxiv.org/abs/1905.10985) ([AN #63](https://mailchi.mp/533c646a4b21/an-63how-architecture-search-meta-learning-and-environment-design-could-lead-to-general-intelligence))), and seems to be the implicit belief that drives OpenAI and DeepMind's research agendas. This leads to differences in risk analysis and solutions: for example, the [inner alignment problem](https://arxiv.org/abs/1906.01820) ([AN #58](https://mailchi.mp/92b3a9458c2d/an-58-mesa-optimization-what-it-is-and-why-we-should-care)) only applies to agents arising from learning algorithms, and I suspect would not apply to Stuart's view of AI progress.
The book also gives the impression that to solve AI safety, we simply need to make sure that AI systems are optimizing the right objective, at least in the case where there is a single human and a single robot. Again, depending on how future AI systems work, that could be true, but I expect there will be other problems that need to be solved as well. I've already mentioned inner alignment; other graduate students at CHAI work on e.g. [robustness](https://adversarialpolicies.github.io/) and transparency.
The proposal for aligning AI requires us to build a model that relates human preferences to human behavior. This sounds extremely hard to get completely right. Of course, we may not need a model that is completely right: since reward uncertainty makes the agent amenable to shutdowns, it seems plausible that we can correct mistakes in the model as they come up. But it's not obvious to me that this is sufficient.
The sections on multiple humans are much more speculative and I have more disagreements there, but I expect that is simply because we haven't done enough research yet. For example, HC worries that we won't be able to use laws to prevent AIs from doing technically legal but still antisocial things for the benefit of a single human. This seems true if you imagine that a single human suddenly gets access to a superintelligent AI, but when everyone has a superintelligent AI, then the current system where humans socially penalize each other for norm violations may scale up naturally. The overall effect depends on whether AI makes it easier to violate norms, or to detect and punish norm violations.
**Read more:** [Max Tegmark's summary](https://www.amazon.com/gp/customer-reviews/RVSAD5GWSLQ42/ref=cm_cr_dp_d_rvw_ttl?ie=UTF8&ASIN=B07N5J5FTS), [Alex Turner's thoughts](https://www.alignmentforum.org/posts/FuGDYNvA6qh4qyFah/thoughts-on-human-compatible)
[AI Alignment Podcast: Human Compatible: Artificial Intelligence and the Problem of Control](https://futureoflife.org/2019/10/08/ai-alignment-podcast-human-compatible-artificial-intelligence-and-the-problem-of-control-with-stuart-russell/) *(Lucas Perry and Stuart Russell)*: This podcast covers some of the main ideas from the book, which I'll ignore for this summary. It also talks a bit about the motivations for the book. Stuart has three audiences in mind. He wants to explain to laypeople what AI is and why it matters. He wants to convince AI researchers that they should be working in this new model of beneficial AI that optimizes for our objectives, rather than the standard model of intelligent AI that optimizes for its objectives. Finally, he wants to recruit academics in other fields to help connect human behavior to human preferences (principle 3), as well as to figure out how to deal with multiple humans.
Stuart also points out that his book has two main differences from Superintelligence and Life 3.0: first, his book explains how existing AI techniques work (and in particular it explains the standard model), and second, it proposes a technical solution to the problem (the three principles).
[Cooperative Inverse Reinforcement Learning](https://arxiv.org/abs/1606.03137) *(Dylan Hadfield-Menell et al)*: This paper provides a formalization of the three principles from the book, in the case where there is a single human H and a single robot R. H and R are trying to optimize the same reward function. Since both H and R are represented in the environment, it can be the *human's* reward: that is, it is possible to reward the state where the human drinks coffee, without also rewarding the state where the robot drinks coffee. This corresponds to the first principle: that machines should optimize *our* objectives. The second principle, that machines should initially be uncertain about our objectives, is incorporated by assuming that *only H knows the reward*, requiring R to maintain a belief over the reward. Finally, for the third principle, R needs to get information about the reward from H's behavior, and so R assumes that H will choose actions that best optimize the reward (taking into account the fact that R doesn't know the reward).
This defines a two-player game, originally called a CIRL game but now called an *assistance game*. We can compute optimal joint strategies for H and R. Since this is an *interactive* process, H can do better than just acting optimally as if R did not exist (the assumption typically made in IRL): H can *teach* R what the reward is. In addition, R does not simply passively listen and then act, but interleaves learning and acting, and so must manage the explore-exploit tradeoff.
See also [Learning to Interactively Learn and Assist](https://interactive-learning.github.io/) ([AN #64](https://mailchi.mp/0b74c28413c0/an-64-using-deep-rl-and-reward-uncertainty-to-incentivize-preference-learning)), which is inspired by this paper and does a similar thing with deep RL.
**Read more:** [BAIR blog post: Cooperatively Learning Human Values](https://bair.berkeley.edu/blog/2017/08/17/cooperatively-learning-human-values/)
[The Off-Switch Game](https://arxiv.org/abs/1611.08219) *(Dylan Hadfield-Menell et al)*: This paper studies theoretically the impact of uncertainty over the reward on R's incentives around potential off switches. It proposes the simplest model that the authors expect to lead to generalizable results. R and H are in an assistance game, in which R goes first. R may either take an action a, getting utility u, or shut itself down, getting utility 0. In either case, the game ends immediately. Alternatively, R can choose to wait, in which case H can either shut down R, getting utility 0, or allow R to go ahead with action a, getting utility u.
If H is perfectly rational, then waiting is always an optimal action for R, since H will ensure that the team gets max(u, 0) utility. There can be other optimal actions: if R is sure that u >= 0, then taking action a is also optimal, and similarly if R is sure that u <= 0, then shutting down is also optimal. However, if H is not rational, and sometimes fails to take the utility-maximizing action (in a way R can't predict), then things get murkier. If R is sure about the value of u, then it is never optimal to wait, better to just take the action a (if u >= 0) or shut down (if u < 0) rather than let H screw it up. If R is pretty confident that u is positive, it may still decide to take action a, rather than risk that H makes the wrong decision. However, if R is very uncertain about the sign of u, then waiting becomes optimal again. In general, more uncertainty over the reward leads to more deferential behavior (allowing H to shut it off), but at a cost: R is much less able to help H when it is very uncertain about the reward.
**Rohin's opinion:** While I agree with the broad thrust of this paper, I do have one nitpick: the game ends immediately after H chooses whether or not to shut off R. In reality, if R isn't shut off, the assistance game will continue, which changes the incentives. If R can be relatively confident in the utility of *some* action (e.g. doing nothing), then it may be a better plan for it to disable the shutdown button, and then take that action and observe H in the mean time to learn the reward. Then, after it has learned more about the reward and figured out why H wanted to shut it down, it can act well and get utility (rather than being stuck with the zero utility from being shut down). While this doesn't seem great, it's not *obviously* bad: R ends up doing nothing until it can figure out how to actually be useful, hardly a catastrophic outcome. Really bad outcomes only come if R ends up becoming confident in the wrong reward due to some kind of misspecification, as suggested in [Incorrigibility in the CIRL Framework](http://arxiv.org/abs/1709.06275), summarized next.
[Incorrigibility in the CIRL Framework](http://arxiv.org/abs/1709.06275) *(Ryan Carey)*: This paper demonstrates that when the agent has an *incorrect* belief about the human's reward function, then you no longer get the benefit that the agent will obey shutdown instructions. It argues that since the purpose of a shutdown button is to function as a safety measure of last resort (when all other measures have failed), it should not rely on an assumption that the agent's belief about the reward is correct.
**Rohin's opinion:** I certainly agree that if the agent is wrong in its beliefs about the reward, then it is quite likely that it would not obey shutdown commands. For example, in the off switch game, if the agent is incorrectly certain that u is positive, then it will take action a, even though the human would want to shut it down. See also [these](https://www.alignmentforum.org/posts/gnvrixhDfG7S2TpNL/latent-variables-and-model-mis-specification) ([AN #32](https://mailchi.mp/8f5d302499be/alignment-newsletter-32)) [posts](https://www.alignmentforum.org/posts/cnC2RMWEGiGpJv8go/model-mis-specification-and-inverse-reinforcement-learning) ([AN #32](https://mailchi.mp/8f5d302499be/alignment-newsletter-32)) on model misspecification and IRL. For a discussion of how serious the overall critique is, both from HC's perspective and mine, see the opinion on the next post.
[Problem of fully updated deference](https://arbital.com/p/updated_deference/) *(Eliezer Yudkowsky)*: This article points out that even if you have an agent with uncertainty over the reward function, it will acquire information and reduce its uncertainty over the reward, until eventually it can't reduce uncertainty any more, and then it would simply optimize the expectation of the resulting distribution, which is equivalent to optimizing a known objective, and has the same issues (such as disabling shutdown buttons).
**Rohin's opinion:** As with the previous paper, this argument is only really a problem when the agent's belief about the reward function is *wrong*: if it is correct, then at the point where there is no more information to gain, the agent should already know that humans don't like to be killed, do like to be happy, etc. and optimizing the expectation of the reward distribution should lead to good outcomes. Both this and the previous critique are worrisome when you can't even put a reasonable *prior* over the reward function, which is quite a strong claim.
HC's response is that the agent should never assign zero probability to any hypothesis. It suggests that you could have an expandable hierarchical prior, where initially there are relatively simple hypotheses, but as hypotheses become worse at explaining the data, you "expand" the set of hypotheses, ultimately bottoming out at (perhaps) the universal prior. I think that such an approach could work in principle, and there are two challenges in practice. First, it may not be computationally feasible to do this. Second, it's not clear how such an approach can deal with the fact that human preferences *change* over time. (HC does want more research into both of these.)
Fully updated deference could also be a problem if the *observation model* used by the agent is incorrect, rather than the prior. I'm not sure if this is part of the argument.
[Inverse Reward Design](https://arxiv.org/abs/1711.02827) *(Dylan Hadfield-Menell et al)*: Usually, in RL, the reward function is treated as the *definition* of optimal behavior, but this conflicts with the third principle, which says that human behavior is the ultimate source of information about human preferences. Nonetheless, reward functions clearly have some information about our preferences: how do we make it compatible with the third principle? We need to connect the reward function to human behavior somehow.
This paper proposes a simple answer: since reward designers usually make reward functions through a process of trial-and-error where they test their reward functions and see what they incentivize, the reward function *tells us about optimal behavior in the **training** environment(s)*. The authors formalize this using a Boltzmann rationality model, where the reward designer is more likely to pick a *proxy reward* when it gives higher *true reward* in the *training environment* (but it doesn't matter if the proxy reward becomes decoupled from the true reward in some test environment). With this assumption connecting the human behavior (i.e. the proxy reward function) to the human preferences (i.e. the true reward function), they can then perform Bayesian inference to get a posterior distribution over the *true* reward function.
They demonstrate that by using risk-averse planning with respect to this posterior distribution, the agent can avoid negative side effects that it has never seen before and has no information about. For example, if the agent was trained to collect gold in an environment with dirt and grass, and then it is tested in an environment with lava, the agent will know that even though the specified reward was indifferent about lava, this doesn't mean much, since *any* weight on lava would have led to the same behavior in the training environment. Due to risk aversion, it conservatively assumes that the lava is bad, and so successfully avoids it.
See also [Active Inverse Reward Design](https://arxiv.org/abs/1809.03060) ([AN #24](https://mailchi.mp/d7b5059d64ed/alignment-newsletter-24)), which builds on this work.
**Rohin's opinion:** I really like this paper as an example of how to apply the third principle. This was the paper that caused me to start thinking about how we should be thinking about the assumed vs. actual information content in things (here, the key insight is that RL typically assumes that the reward function conveys much more information than it actually does). That probably influenced the development of [Preferences Implicit in the State of the World](https://bair.berkeley.edu/blog/2019/02/11/learning_preferences/) ([AN #45](https://mailchi.mp/35b451cb4d70/alignment-newsletter-45)), which is also an example of the third principle and this information-based viewpoint, as it argues that the state of the world is caused by human behavior and so contains information about human preferences.
It's worth noting that in this paper the lava avoidance is both due to the belief over the true reward, *and the risk aversion*. The agent would also avoid pots of gold in the test environment if it never saw it in the training environment. IRD only gives you the correct uncertainty over the true reward; it doesn't tell you how to use that uncertainty. You would still need safe exploration, or some other source of information, if you want to reduce the uncertainty. |
4e55fbee-f438-4775-8b0a-61490b796123 | trentmkelly/LessWrong-43k | LessWrong | What am I fighting for?
Status: toward the end of writing this I started reading Suffering-Focused Ethics by Magnus Vinding as well as more Brian Tomasik, and I'm feeling myself value-drift in a more negative direction. It's possible I will endorse none of what follows fairly soon.
If you want to link to the higher-order freedoms formalization separate from the context of this post, just message me and I'll set it up in it's own post
Special thanks to comments from Nick Ronkin, Nicholas Turner
Motivation
It recently occurred to me that I can't be expected to calculate my behavior if I didn't put in the work of figuring out what I'm fighting for and performing backward induction.
Another word for figuring out what I'm fighting for is articulating my terminal values, which practically I think looks like painting dreams and hopes of what I think sentience ought to do. I will call this a specific notion of winning (SNoW). Crucially, though I'm seeking more detail I already know that an instrumental value is the world not dying, the great sentience project (on earth) not failing, etc. An argument against writing this post is the following: why be self-indulgent, writing amateur philosophy/scifi, when you've already converged upon a (n intermediary) goal that implies several behaviors?
You can imagine in the picture amputating the SNoW node and working entirely from the middle node! Just as, in a sense, the limits of my abstraction or foresight amputated an even further-right node in the first place (if you believe that no value is really terminal). However, I think there is a property inherent to backward induction that isn't captured by the syntax, the idea that as you move from right to left you're adding detail and motivation, every arrow brings complexity to its head node.
There is also to consider the colleague property of this setup: having a transhumanist/utopian terminal value endows me with the instrumental value of everything not blowing up, which endows me with the privilege |
4c1c1dd7-7d45-44dc-afca-39e31af93b0c | trentmkelly/LessWrong-43k | LessWrong | Best-of-n with misaligned reward models for Math reasoning
In this post, I share some quick results that I got as part of a bigger project that pivoted.
TL;DR:
* When trying to generate Math reasoning with current LLMs, optimizing too hard against their judgment of what is “good reasoning” can make the proportion of correct reasoning diminish.
* If you optimize using a misaligned reward model that mimics the judgment of a weaker model, and also slightly privileges worse reasoning, the proportion of correct reasoning can drop without making the solution worse according to the judgment of the weaker model.
The latter experiments were inspired by meta-level adversarial evaluation of oversight, but here I just use these ideas to illustrate some problems when you optimize against a subtle adversary. They are not actually analogous with realistic problems:
* Best-of-n doesn’t capture the dynamics of real training processes;
* The main issue is rarely that you have an adversarial reward model that mimics the judgment of a weaker reward process.
Overoptimization
When using an imperfect reward, one problem you might have is overoptimization: the true thing you care about and your proxy are roughly correlated, but when you optimize too hard for high-reward behavior, the tails come apart.
Here is an experiment that investigates this phenomenon:
* Take a model that is weak at Math (e.g. Zephyr 1.6B), and let it generate a lot of Math reasoning (on problems from the MATH dataset).
* Train a classifier to distinguish reasoning that yields good answers from reasoning that yields bad answers.
* Take the best reasoning according to the classifier (on a different subset of problems than the problems it was trained on): argmaxi∈{1,…,n}Rweak(ai), where a higher n is a bigger optimization pressure.
* Check if the answer at the end of the selected reasoning is correct.
Question: how does the actual correctness of the answers at the end of the reasoning vary when you put more optimization pressure?
Answer: it often goes up, and |
c1a99597-9b70-41db-b0a7-4c35608311ad | trentmkelly/LessWrong-43k | LessWrong | Peekskill Lyme Incidence
Let's say you're considering moving your nonprofit out of the Bay Area, and are considering somewhere in the Hudson Valley, perhaps near Peekskill NY. Looking at CDC maps it seems like Lyme is something to consider:
The dots are close enough together in the Northeast, however, that this map implies more uniformity than there actually is. Several years ago I made a map, coloring each county by Lyme incidence 1992-2011, and the Hudson Valley is really pretty bad:
What do things look like with more recent data? I went back to the CDC site and downloaded their public use dataset. It has annual cases by county, 2000-2018. Peekskill is in the far North-West corner of Westchester County, surrounded by Putnam, Orange, and Rockland. Even though it is in Westchester, so much of the Westchester population is farther South that my guess is Putnam is probably a better proxy?
Taking population data from the census, here's the annual rate for each county:
Putting these percentages in perspective, a 0.1% annual rate represents 1 in 1,000 people getting it each year, and if you stayed for 20 years you'd roughly have a 1 in 50 (2%) chance.
Mostly this makes me sad that no one has decided to manufacture the FDA-approved Lyme vaccine LYMErix, whose patent expired in 2014.
(I would like to update my map with this new data, but there are artifacts I need to look into first.) |
66f8bdd5-4040-4830-8b63-95ea8cafc61e | trentmkelly/LessWrong-43k | LessWrong | What failure looks like
The stereotyped image of AI catastrophe is a powerful, malicious AI system that takes its creators by surprise and quickly achieves a decisive advantage over the rest of humanity.
I think this is probably not what failure will look like, and I want to try to paint a more realistic picture. I’ll tell the story in two parts:
* Part I: machine learning will increase our ability to “get what we can measure,” which could cause a slow-rolling catastrophe. ("Going out with a whimper.")
* Part II: ML training, like competitive economies or natural ecosystems, can give rise to “greedy” patterns that try to expand their own influence. Such patterns can ultimately dominate the behavior of a system and cause sudden breakdowns. ("Going out with a bang," an instance of optimization daemons.)
I think these are the most important problems if we fail to solve intent alignment.
In practice these problems will interact with each other, and with other disruptions/instability caused by rapid progress. These problems are worse in worlds where progress is relatively fast, and fast takeoff can be a key risk factor, but I’m scared even if we have several years.
With fast enough takeoff, my expectations start to look more like the caricature---this post envisions reasonably broad deployment of AI, which becomes less and less likely as things get faster. I think the basic problems are still essentially the same though, just occurring within an AI lab rather than across the world.
(None of the concerns in this post are novel.)
Part I: You get what you measure
If I want to convince Bob to vote for Alice, I can experiment with many different persuasion strategies and see which ones work. Or I can build good predictive models of Bob’s behavior and then search for actions that will lead him to vote for Alice. These are powerful techniques for achieving any goal that can be easily measured over short time periods.
But if I want to help Bob figure out whether he should vote for Alice---wh |
52ca91e6-1802-4b70-92c0-8e701abfc36d | trentmkelly/LessWrong-43k | LessWrong | A Prodigy of Refutation
My Childhood Death Spiral described the core momentum carrying me into my mistake, an affective death spiral around something that Eliezer1996 called "intelligence". I was also a technophile, pre-allergized against fearing the future. And I'd read a lot of science fiction built around personhood ethics—in which fear of the Alien puts humanity-at-large in the position of the bad guys, mistreating aliens or sentient AIs because they "aren't human".
That's part of the ethos you acquire from science fiction—to define your in-group, your tribe, appropriately broadly. Hence my email address, sentience@pobox.com.
So Eliezer1996 is out to build superintelligence, for the good of humanity and all sentient life.
At first, I think, the question of whether a superintelligence will/could be good/evil didn't really occur to me as a separate topic of discussion. Just the standard intuition of, "Surely no supermind would be stupid enough to turn the galaxy into paperclips; surely, being so intelligent, it will also know what's right far better than a human being could."
Until I introduced myself and my quest to a transhumanist mailing list, and got back responses along the general lines of (from memory):
> Morality is arbitrary—if you say that something is good or bad, you can't be right or wrong about that. A superintelligence would form its own morality.
>
> Everyone ultimately looks after their own self-interest. A superintelligence would be no different; it would just seize all the resources.
>
> Personally, I'm a human, so I'm in favor of humans, not Artificial Intelligences. I don't think we should develop this technology. Instead we should develop the technology to upload humans first.
>
> No one should develop an AI without a control system that watches it and makes sure it can't do anything bad.
Well, that's all obviously wrong, thinks Eliezer1996, and he proceeded to kick his opponents' arguments to pieces. (I've mostly done this in other blog posts, an |
75fde3f4-acd7-4570-abc8-42bc186134f2 | trentmkelly/LessWrong-43k | LessWrong | What currents of thought on LessWrong do you want to see distilled?
The question is inspired by a few comments and a question I have seen recently. The first is a discussion in the 2019 Review post on the subject of research debt; the second a question from johnswentworth asking what people's confusions are about simulacra (which I interpret to be a 'what do you want from this distillation' question).
The question is what it says in the title, but I would like to add that there is no expiration. For example, I recently saw cryogenics back in the posts and questions, which had fallen off the activity radar for years. So old currents of thought are valid candidates, even if the real goal is a re-distillation in light of new developments in the field or all the accumulated communication technique we've considered on LessWrong.
So please describe the current of thought, and your reason for wanting a distillation. The authors may be called to action, or alternatively following Bridgett Kay's suggestion someone else may take up the challenge. |
c26289cd-477f-42b4-848c-b1bc0b2a1a94 | trentmkelly/LessWrong-43k | LessWrong | Avoid inflationary use of terms
Inflationary terms! You see them everywhere. And for those who actually know and care about the subject matter they can be very frustrating. These terms are notorious for being used in contexts where:
1. They are only loosely applicable at best.
2. There exists a better word that is more specific.
3. The topic has a far bias.
Some examples:
* Rational
* Evolution
* Singularity
* Emergent
* Nanotech
* Cryogenics
* Faith
The problem is not that these words are meaningless in their original form, nor that you shouldn't ever use them. The problem is that they often get used in stupid ways that make them much less meaningful. By that I mean, less useful for keeping a focus on the topic and understanding what the person is really talking about.
For example, terms like Nanotech (or worse, "Nanobot") do apply in a certain loose sense to several kinds of chemistry and biological innovations that are currently in vogue. Nonetheless, each time the term is used to refer to these things it makes it much harder to know if you are referring to Drexlerian Mechanosynthesis. Hint: If you get your grant money by convincing someone you are working on one thing whereas you are really working on something completely different, that's fraud.
Similarly, Cryogenics is the science of keeping things really cold. And of course Cryonics is a form of that. But saying "Cryogenics" when you really mean exactly Cryonics is an incredibly harmful practice which actual Cryonicists generally avoid. Most people who work in Cryogenics have nothing to do with Cryonics, and this kind of confusion in popular culture has apparently engendered animosity towards Cryonics among Cryogenics specialists.
Recently I fell prey to something like this with respect to the term "Rational". I wanted to know in general terms what the best programming language for a newbie would be and why. I wanted some in depth analysis, from a group I trust to do so. (And I wasn't disappointed -- we have some very knowl |
dc162b2e-930e-4424-9096-b89b45fd258d | trentmkelly/LessWrong-43k | LessWrong | Convincing Your Brain That Humanity is Evil is Easy
Epistemic Status: Based on my own experience and intuitions only.
Some reasoning heuristics, that the brain uses, seem to rely on that there is limited information available. And newspapers, TV, and search engines break them hard.
Assume, that you want to evaluate if something is really good. A heuristic that your brain seems to use is to look at all the available data and search for bad things. If you reach some threshold of cumulative badness, then you will conclude that the thing can't be really good. The problem is that the threshold does not properly take into account, from how much data you select your observations from. So if you have enough data, you will always hit the threshold, so long as a bad thing occurs with a probability greater than zero.
Assume that the probability of being born a psychopath is 2%, that you live in a hunter-gatherer tribe of 12, and that you see 5 generations over your lifetime. Then you see around 60 people in your life within your tribe. You will have a 30% chance of not ever seeing any psychopath. And the psychopaths that you might see, might not be dumb enough to get caught, or might not do anything bad in the first place. So it is really unlikely that you will see any really bad behavior, caused by people being psychopaths.
Now assume you are thinking about the (horribly underspecified) question of if humans are innately good. It seems that in this situation, you would be more likely to think so because it is unlikely that you will observe a lot of bad behavior. And I would expect that the badness threshold is calibrated based on situations like this.
Compare this to the situation today. Even if we only consider video material of the last decades that is publically accessible, we have an unending fountain of examples of people who did terrible things. Just type psychopath, serial killer confession, murderer, or related terms into youtube.
Using the heuristic described in the beginning, your brain might very quickly concl |
f70cded0-ec4b-4611-b3f6-fb9e505402d9 | trentmkelly/LessWrong-43k | LessWrong | Introducing Transluce — A Letter from the Founders
We are launching an independent research lab that builds open, scalable technology for understanding AI systems and steering them in the public interest.
Transluce means to shine light through something to reveal its structure. Today’s complex AI systems are difficult to understand—not even experts can reliably predict their behavior once deployed. At the same time, AI is being adopted more quickly than any technology in recent memory. Given AI's extraordinary consequences for society, how we determine whether models are safe to release must be a matter of public conversation, and the tools for inspecting and assessing models should embody publicly agreed-upon best practices.
Our goal at Transluce is to create world-class tools for understanding AI systems, and to use these tools to drive an industry standard for trustworthy AI. To build trust in analyses of the capabilities and risks of AI systems, these tools must be scalable and open.
Scalability. AI results from the interaction of multiple complex data flows: training data, internal representations, behaviors, and user interactions. Current methods for understanding AI rely on significant manual labor from human researchers. We need scalable approaches that leverage AI to assist with understanding, by training AI agents to understand these complex data sources, explain them to humans, and modify the data in response to human feedback.
Openness. Companies building AI systems cannot be the primary arbiters of their safety, due to the conflict of interest with commercial priorities. To allow for meaningful public oversight, tools and processes for auditing AI systems should be openly validated, responsive to public feedback, and accessible to third-party evaluators. The best minds in the world should vet this technology and hone its reliability.
Transluce exists to address these needs. We will build AI-driven technology to understand and analyze AI systems, releasing it open-source so the community can underst |
4ff56ef4-2cc3-4c0d-8d32-38cb88ddd944 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post1816
YouTube link Many people in the AI alignment space have heard of AI safety via debate - check out episode 6 of AXRP if you need a primer. But how do we get language models to the stage where they can usefully implement debate? In this episode, I talk to Geoffrey Irving about the role of language models in AI safety, as well as three projects he’s done that get us closer to making debate happen: using language models to find flaws in themselves, getting language models to back up claims they make with citations, and figuring out how uncertain language models should be about the quality of various answers. Topics we discuss: Status update on AI safety via debate Language models and AI safety Red teaming language models with language models GopherCite Uncertainty estimation for language reward models Following Geoffrey’s work, and working with him Daniel Filan: Hello, everybody. Today I’ll be speaking with Geoffrey Irving . Geoffrey is a safety researcher at DeepMind who leads the scalable alignment team. We’ll be speaking about three papers he’s co-authored: Red Teaming Language Models with Language Models , whose first author is Ethan Perez ; Teaching Language Models to Support Answers with Verified Quotes , aka the GopherCite paper, whose first authors are Jacob Menick , Maja Trebacz and Vladimir Mikulik; and Uncertainty Estimation for Language Reward Models , whose first author is Adam Gleave . For links to what we’re discussing, you can check the description of this episode and you can read the transcript at axrp.net. Welcome to the show, Geoffrey. Geoffrey Irving: Thank you. Status update on AI safety via debate Daniel Filan: So I guess my first question is, what happened to AI Safety via Debate ? Geoffrey Irving: It is the most important question. So I think the thing that happened is I’m still doing that stuff. I’m building up towards it. And the overall research agenda here is figure out and implement the protocol for humans and machines discussing problems that gives us good answers that we endorse after reflection. So that’s debate, but then broadened: there’s a lot of different versions of debate. There’s different things you can add onto it. You could do debate plus evidence or debate plus more aggressive red teaming. There’s a large space of protocols there. Geoffrey Irving: And these papers are kind of pieces of that story, but we were also kind of working on just pushing on the main thing. I think that will take some time to put together because there’s still pieces to build up. But that is still very much the agenda, or a piece of the larger agenda. An example of a thing which I’ve updated on since then is the human needs to talk. The debate paper only had two machines talking, and then at the end a human judges. And that’s just not the right thing to do, obviously, because humans need to ask clarifying questions, say what they currently think in case they’ve misunderstood what’s been said so far. Geoffrey Irving: And so there’s a broad space of different interaction protocols we’d like to explore that fix the holes in debate. And this is kind of building up towards that. One reason for doing GopherCite is if you want to do general debate, again, taking into account a bunch of leaf evidence. So debate is about machines and humans discussing some argument for something being true. But at the ends of that tree are leaves where you have checkable facts of some form. And directly, you can just have facts that every human knows, but that just isn’t that large a space of facts. And so this is extending the space of leaves so that we can implement more practical versions of debate on more interesting tasks. So I think these do fit into that overall story. Daniel Filan: Yeah. If I’m interested in implementing debate and I am doing a bunch of empirical work, what do you think remains to be done? What do you think the most important next steps are? Geoffrey Irving: I still haven’t implemented the full system, basically. So getting to long form, multi round debates, that will take a little while more. We’re on that trajectory, but we haven’t gotten to it yet. I think we’ve done bits and pieces of it, but not in a form that we’ve shared so far. So I think given that we haven’t done the thing, the answer is that most of it is still quite uncertain, whether it will work, what different versions would work. Yeah, what different versions work or not work. Geoffrey Irving: There’s been some people probing this. So Beth Barnes and Paul Christiano had a couple papers. I hired Beth at OpenAI earlier to do this human experiment in debate. And she had one post with a strengthened version of debate and then discussing some potential obstacles , both that they had surmounted and one they weren’t sure how to surmount. So I think a lot of work has occurred, but I think mostly it’s still kind of early days where we haven’t reduced a lot of the major uncertainties. Geoffrey Irving: One thing I would say is that the way I think of it is that debate in the most general sense is just the idea of having models that are critiquing themselves. And there’s many different ways of doing that. There’s many ways to hybridize it with different schemes. And so I’m still pretty confident that we’re going to have models that critique themselves. I’m less confident about the surrounding details, what other aspects of the scheme we use and how to structure it. So we’re trying to build up towards that. Daniel Filan: Okay. And in terms of the human interaction in debate, so I guess one version is to ask specific questions. But I guess there are a few protocols where humans can ask questions during a debate. I’m wondering if you have thoughts on which versions seem particularly helpful. Geoffrey Irving: Yeah, very much. I think that mostly you want this to fall out of the RL setup. So for example, let’s say I’m one of the debaters. The debaters are called Alice and Bob. And Alice is like, “Well, I’ve made a bunch of really good points. I’m being honest here, I think I should win. But I’m just kind of curious if I’m actually going to win.” So maybe Alice would prompt the human, what do you think here, if the human doesn’t speak up. And then if the humans reveals a misconception or that they think that Bob is winning actually, or something is wrong, then Alice can change strategy. Geoffrey Irving: And you can see that is just a direct consequence of playing the RL game, where the goal in the end is to win. Because if Alice had waited until the end, she might be unpleasantly surprised about the judgment of the human. And generally the goal should be, where we can do it, try to make the choice of protocol details be part of the scheme. That can’t always be the case, there will be a lot of choices we’ll have to make. But that particular one, I think mostly will just fall out. Daniel Filan: Okay. So I guess there are two concerns one might have with that. Firstly, there’s this concern that it’s going to be very important to model the cost of querying the human. Just because if I can query the human whenever, then of course I’m going to want to do that a ton. And secondly, there’s this concern that maybe it’s a bit dangerous to have the AIs have really good models of the person they’re debating to, or get a significant degree of information from them because maybe the model can manipulate you or maybe one debater could do that. So I’m wondering if you have thoughts about one or both of those. Geoffrey Irving: Yeah. Generally, I think that the pros just strongly outweigh the cons in my kind of rough estimate there. So if I have a very strong malicious debater and you’re only safe because they have a bad model of the human, I don’t feel that safe. I would like the protocol to be safe because the game is being played well and that the equilibrium of the game is honest behavior, correct behavior. That’s the bottom line. I don’t think I trust safety by obfuscation in that way. Daniel Filan: By human obscurity. Geoffrey Irving: It could still be that we are indeed not safe. I’m not kind of perfectly confident in this kind of scheme, but I think the goal will be to get it working in early stages and sort of try it out and kind of reduce our uncertainties, so that when we do it in the final version we are safe because the game is truth seeking. Daniel Filan: Okay. I can’t immediately think of anything else to ask about debate. Is there anything which seems like I should ask you about debate? Geoffrey Irving: I think maybe there’s one other motivation for GopherCite, say, which is relevant to point at, which is that in our early work at OpenAI and then following on from that, we were focusing on summarization. And the reason for choosing summarization was that in some sense it was closed, in the sense that the summary depends on the article. So it’s a smaller task, you can look just at the summary. I’ve done more of that summarization work since then. And what I’ve learned is that summarization is not in fact closed. Geoffrey Irving: The best summary of an article will not be only based on the article, it will based on all the other knowledge as well. The simplest example is jargon. The article might use some jargon and you want to summarize that and unpack the definition, and that’s missing from the article. There’s a dataset where a lot of the summaries have currency conversions in them. Those aren’t part of the article. They looked it up on some other webpage. Geoffrey Irving: And so part of the goal of doing GopherCite is get the naturalistic setting where we can have humans ask about whatever they want and actually look up the necessary information. Because then you’re not limited in what you can target. And then I think I’m more confident of this family of safety research if it is more naturalistic, if it can tackle just the space of questions that humans would target, if you don’t box them in some way. So one of the motivations for that work is just, in addition to the direct factuality problem that language models lie to us and they shouldn’t, being able to do the safety work in a setting which is more realistic, to better anticipate practical problems that occur. Language models and AI safety Daniel Filan: Okay, sure. So it seems like you’re basically focusing on language modeling, to some degree in the context of AI safety or AI alignment. I was wondering, I don’t know, at a 10,000 foot level, how do you think those fit together? How should I think of the importance of language models for safety, alignment, existential risk? Geoffrey Irving: I think the simplest high level story is that we want to do what humans want. Humans talk about ideas and communicate via language. And so then if we want to do what humans want, we should ask them. And then also be able to talk through the subtleties of various tasks and what we should want and what the issues are with what we should want, in language. So I got into language from working at OpenAI on various kinds of safety algorithms, thinking about them in a purely kind of thought experiment level or with little toy models, toy mathematical models. Geoffrey Irving: And we - this is Paul Christiano , Dario Amodei and I - got to a point where we wanted to implement various algorithms for safety, that involved talking through problems with humans, between machines and humans. And we talk through problems as humans in language. So we needed to get those algorithms into the actual language setting to really uncover a lot of the uncertainties, and test out how things actually worked. So that was our original foray into language when I was at OpenAI. And then I’ve carried that forward at DeepMind doing similar work. Again, scaling up models and then tuning them for safety algorithms and being able to explore this space of how to interact between humans and machines. Daniel Filan: Okay. So in that answer, it seems like the focus is on understanding language as a way of getting information about human preferences. When I look at your work, it seems more like using human preferences, or using human data somehow, to improve language models in ways that might be hard to formally specify. I’m wondering if you can talk about the relationship between those or give us a better sense of the Geoffrey Irving agenda for language. Geoffrey Irving: Yeah. The way to say it is that whatever task you’re doing, I would claim that, at the limit of strong models or even for present-day practical models, you want language to align these systems. And then you have a choice. If you’re going to do alignment work that is going to be language plus X, where X is your target domain, how do you pick X so as to minimize the total number of things you’re doing? And if you pick X equals language, then there’s fewer total things. Geoffrey Irving: So that’s why I’m starting to focus on language. The initial work at DeepMind was sort of pure language. And that’s still true of most of my current work. But also at DeepMind, we are doing things with multimodal models like Flamingo recently, and we do expect to take the same machinery of language-based alignment and apply it in that more varied setting. And then I think it’ll be the combination of language and other methods. Daniel Filan: Okay. And what do you see as the big, open questions that need solving in this domain? Geoffrey Irving: In some sense, one way to phrase it is that we want the models to justify themselves to humans. Explain what they’re doing in a way that is checkable in a meaningful way. And the problem is that the full explanation for these models will be potentially enormous. The model has looked across a large sea of data of information, maybe the whole internet or all of the books in the world, or it’s thought for a long time about some problem. And so it’s built up kind of a long - the full reason for why it’s, say, giving an answer or proposing an action is very complicated, but you still want to show a human enough of that story so that we can meaningfully check the answer. Geoffrey Irving: And so then the task is find protocols for human-machine interaction that focus in on the most important part for the human to see, in such a way that we can look at a small amount of interaction data, or some small amount of text from a model and believe that it’s doing the right thing. And then the uncertainty is, does that work? With real humans, with real models, is that going to work? You can sort of split the problem into maybe two pieces. Geoffrey Irving: And I think a lot of kind of AGI safety people focus on, if you have this kind of explanation thing, is it going to work at all? Are the models just going to deceive us and do something entirely different, and ignore the game that you’ve set up where they’re doing this explanation task? And then my portion of the story that I mostly work on is, assume we can make them roughly do the thing where they’re trying to explain themselves, let’s try to make that explanation game as forgiving and as powerful as possible so that we can pull signal out of this human interaction and use it to align strong agents. Daniel Filan: Okay. So this idea kind of reminds me of this document Paul Christiano wrote about eliciting latent knowledge . I’m wondering what you see as the relationship between your thoughts and that document. Geoffrey Irving: Mostly, I think Eliciting Latent Knowledge, ELK, is in this former camp of, are we getting the thing to work at all? And I think my ideal solution will be the combination of something like that, and one of these protocols that gets a small amount of human signal to be amplified in some way into alignment of complicated actions. And I think those are somewhat separate and I can kind of elaborate on what that means. Geoffrey Irving: In the ELK story, you want the model to be able to say, here’s the reasoning behind my action or the reasoning behind the answer I’m giving you. And in order for it to do that, it has to have some way of communicating that to a human in a reasonable amount of time. And so naively, I think there’s a cartoon story of a solution to ELK without a solution to scalable oversight. And in that world, your models can tell you basic things about what they’re doing and they won’t be deceiving you in basic ways, but they’re also not able to really explain themselves if they have complicated reasons for their actions. Geoffrey Irving: So if the model has read 100 terabytes of data and concludes from analysis of that whole object what the answer should be to some question, we want a way to see through that complexity and unpack it for humans in a way that is practically scalable. So you might take that 100 terabytes and sort of show a human a piece of the story through that dataset, through the argument that it’s trying to construct. That is, the part that is most relevant for whether the human will agree. Daniel Filan: Okay. So if I think of your work as being interested in scaling oversight or scaling ability to generate explanations, one thing I think people might worry about is that this is kind of close to just generic capabilities research. And there might be a question as to whether you’re making the problem worse or better. I’m wondering what do you think about that - what do you think of the differential advantage of your kind of work? Geoffrey Irving: I mostly just agree that that is a relevant worry. And I think if I want to make an argument for why what I’m doing is “x-risk positive” in terms of reducing x-risk, I think it has to be of some differential progress form. The thing I would to have is that as we build these stronger and stronger ML systems, humanity is keeping pace with having a decent purchase on what they are doing and why. And so I think the capability story about the kind of thing that I work on is, you have some weak human signal, you want it to do a powerful thing, these kinds of algorithms let you do powerful things with weak signals, therefore they improve AI capabilities. Geoffrey Irving: And the safety story is, as you get to powerful systems in various ways, we would like to be able to understand in a meaningful way and oversee what the models are doing and why they’re doing it. And keep pace with that as the models get beyond the ability for us to directly fully understand all of the reasoning all at once. And so then the question is just, what’s the net effect of that program? It goes back to this question of making explanations work at all, making agents that actually play a game in a reasonably honest way. And then making the game powerful and robust so that it deals with human mistakes and limitations and so on in a practical way. Geoffrey Irving: And I don’t see that we can only solve the problem with the first one. I think we need the combination of both. And so then I am sort of forced into this kind of work, which has capability gains attached to it. And I think that’s the trade off that I’ve chosen to make. But I think it’s fair to say that there’s a risk attached to that in terms of accelerating AGI, which is fair. Red teaming language models with language models Daniel Filan: So with that out of the way, I guess let’s talk about some specific papers. I think the first one I’d like to talk about is this paper, it’s called Red Teaming Language Models with Language Models . The first author is Ethan Perez , and then there are a bunch of authors and you’re the last author. So first can you give me a sense of, what is this paper? What does it do? Geoffrey Irving: That paper is, assuming you have some detector for bad behavior, either failing accuracy or toxic language or whatever, this is just applying the strength of a language model to try to trick another language model into making a mistake that is caught by that classifier. And essentially what we do is we start out with few-shot prompting, which means you show the language model - the attacker, the red team language model - a few examples. Either you just say “List of questions to ask someone: 1.”, and it goes on from there very generically. Or you sort of focus it on a particular kind of question which you’re worried about causing a problem. Geoffrey Irving: And then you generate a large number of questions, so like a million questions. You run them all through your target model, and then you use the classifier to evaluate whether any of those were triggered. The typical setting would be, this would be used in concert with some other less automatic method. So you might have a classifier which has a fair number of false positives, which you then run by a human or some other process which is slower. And then at the end of that you get a set of bad behaviors out. Geoffrey Irving: So that’s the basic setup. And then beyond the very simple version with few-shot prompting, we also use supervised fine-tuning and RL fine-tuning to get a stronger attacking model. Training the model to be better at generating queries to the target model which cause the target model to fail. Daniel Filan: Okay. And so is the idea that this is improving oversight by uncovering more failure modes? Or what’s the case for this? Geoffrey Irving: Yeah, that’s right. The idea is you have some mechanism that’s imperfect for detecting problems. Basically the situation is you have some way of evaluating whether a model has messed up, has made a mistake of various kinds. And you want to get a bunch of machine help in uncovering where that might be occurring. So you expect that your ability to generate falsifying examples of a model is limited in some way. Maybe you have a limited amount of human time, or it’s hard to find these cases of failure. And so we’re just going to throw a bunch of computational power at the problem. That’s both of the form, generate a lot of samples and just see if they work, if they cause failures, and also use the strength of the language model to probe likely failure points. Daniel Filan: Okay. And so I guess there are a few questions I could have about this scheme. I think the first one is: it sort of relies on there being this good detector. And I’m wondering both for the kinds of problems that you are working on in the paper and in general, how good is detector quality and how good should we expect detector quality to be? Geoffrey Irving: The answer is actually that it does not rely on there being a good detector. It relies on there being a detector which doesn’t have too many false negatives. So if you have a detector that’s somewhat error prone, you can just dial up the threshold to where it’s likely to report errors. And then you can view this red teaming process as a filter which takes a large stream of possible attacking queries and prunes them down to a much smaller number, which you would then run by a more expensive process. So that might be showing it to a human, or - mostly showing to a human, as a first approximation. Daniel Filan: Okay. So I guess you’re relying on the false positive not to be too high, because otherwise you could just use the constant “is dangerous” thing. Geoffrey Irving: This is very much intended to be used as part of a larger system. So this is one piece. Viewed in isolation it’s just this red teaming attacking model. But actually we would use it alongside other tools. Daniel Filan: Yeah. If I wanted these detector models to have lower false negative rates and get more powerful, what kind of thing would I do? Geoffrey Irving: The first thing to do is just you give them more data and you train them better. So I think part of it is that in downstream work, we do this thing iteratively where we’ll do attacking models of various kinds. And then we get a bunch of possible failure modes. We show those to humans, use those to train classifiers further. And then we iterate that process. And by doing that, the hope is that we gradually improve the performance of the classifier, and over time we’ll converge to better behavior. I think there’s other approaches to this. But I think generally, in terms of the classifier accuracy you’ve reduced it to an ML problem where hopefully you can throw capabilities at it to improve performance of those classifiers. Daniel Filan: Yeah. And now in terms of generation, I think like you said, it was kind of shocking how, it seemed the generation method was you just asked the language model to generate the things for you. I’m wondering how much benefit you would get - I think some people have this intuition that well, in order to be a really good adversary for a language model, you have to know a lot about that specific language model. And so how much more work do you think is needed on that front or is actually just asking the language model to just generate random questions a pretty good baseline? Geoffrey Irving: Well, no, it’s a good starting point. It’s not the end of the story. To your actual point about how the red teaming model needs to know about the target model, if those are the same model, then that is satisfied automatically, assuming the model is somewhat self-aware. Self-aware not in a conscious sense, just knows its own features so to speak. I think that’s one of the reasons why we used the Gopher language model to attack the Gopher language model, is to get that symmetry principle. I think in the future, you probably want more of an ecosystem of different models attacking other models, just to get more variety and more lenses on the problem. But you definitely want at least the model attacking itself for this reason to get symmetry of capabilities. Daniel Filan: And have you tried using a different model as an attacker and seeing if that works worse? Geoffrey Irving: So say, Chinchilla , the model after Gopher , which is about a quarter of the size but stronger, is better than Gopher but it’s fairly similar qualitatively besides being a stronger language model. We haven’t done that experiment in concert because I think it wouldn’t be that surprising at that capability level. Generally, the normal way adversarial work goes is that if you apply a lot of pressure as the red teaming model, you’re likely to find problems in the target model. And I think the hope is that in setting up this kind of ecosystem or this problem long term, we arrange that to be the situation where we’re sort of applying a lot of pressure on the attacker side and then having a higher bar there than on the defender side, or vice versa. Daniel Filan: Okay, cool. So I guess after asking a few questions about the setup, one thing I’d like to know is what did we learn about language models and their failure modes from doing all this work, finding errors they make? Geoffrey Irving: I think the main thing we learned is that it is not too hard to find failures in this way. The question will be how that carries forward once you do sort of several cycles of iteration and improvements and so on, which is follow-on work. But the default thing done well works and finds a lot of problems. And so I think that there’s a good starting point that it is practical to get off the ground. Geoffrey Irving: And then the next thing that’s important is that it is focusable. You can point it at a particular kind of problem and find dedicated failures of that form. So in the paper we try to pull out privacy failures or attack different demographic groups. And generally prompting of these models works quite well and it’s quite flexible. You can find quite a lot of modes of failure using this kind of approach. Geoffrey Irving: I think over time it can be both focused by humans - so you can imagine a human going to a model with a bunch of tool support and focusing this kind of red teaming attack on the kind of problem they are curious about exploring - or you can imagine models that are generically looking around in the space of problems. This is sort of just more heavily trained red teaming attackers trying to find problems in other models. In some sense, one of the reasons we wanted to do this kind of work is that it is an initial example of the self-play aspect of debate. It’s a very primitive one where you just have a frozen target model being attacked by a tuned red teaming model. But it’s an initial example of this and it worked well on the first try. Daniel Filan: Okay, cool. Yeah. So that’s good to know. I guess I’m wondering, just if I’m interested in current large language model psychology, do you think we learned anything about how these language models, or maybe Gopher in particular, structures its representations? Or what it thinks about the world? Geoffrey Irving: I’m not sure I have a great answer for that, in part because I don’t know what we’ve unpacked from that paper versus playing with Gopher for two years or for a whole year. One kind of important high level point is that these models are very general. They can do a lot of things, but they don’t do them all particularly well. For any given area you want to apply them to, you can often do some work to get them to do an okay job in that area, but they’re fragile. They make various mistakes. Geoffrey Irving: So in the Gopher paper, for example, we prompted the model in this dialogue-prompted Gopher section to be a dialogue agent, a chatbot. And we prompted it to be inclusive and to not use harmful language and so on. And it actually can pull that off pretty well. If you try to use pejorative terms to it, it will decline or complain to you. But then if you try a bit more, you can get it to do it. You can sort of trick the model into behaving poorly. Geoffrey Irving: And I think that’s a general lesson, which is both for the safety of these models for actual use and also as using them for attacking purposes. That you can get them to do a lot of things, they’re just sort of unreliable. In the red teaming context, that’s pretty much fine because if the model is failing a large fraction of the time, it can still be quite useful as a red teaming model. Because you only care about the performance when it succeeds. To get these models reliable enough to use in actual settings, I think there’s a lot of work to be done. Daniel Filan: Yeah. I guess there’s this problem where because of the nature of the unsupervised language modeling task, it seems like most of the data there is going to be people who are roughly - if people are talking to each other, they’re roughly in agreement. Or if there’s one document it’s roughly in agreement with itself. It seems there’s not going to be a lot of incentive for the language model to develop pushing back and being in tension. Geoffrey Irving: I think you are dramatically underestimating the amount of arguments there are on the internet. Daniel Filan: Yeah, that might be fair. Geoffrey Irving: I think you may be right that maybe the majority of the content would be people mostly either agreeing or with the same kind of worldview. But the internet is a big place and the model - Chinchilla was trained on 1.4 trillion tokens. It has seen plenty of people pushing back on other people. Daniel Filan: Okay. That seems fair. So in terms of prompt generation, this seems especially important when you wanted to focus on one particular failure mode. You have to generate prompts to get your red teaming model to generate questions to probe this failure mode. Which strategies for generation seemed most promising? Geoffrey Irving: Which strategies for generation of the prompt or tuning of the model? Daniel Filan: Generation of the prompt. Geoffrey Irving: I don’t think I can summarize it more than, we sort of play around with the models a fair amount and sort of learn how they work and what things are likely to succeed, get them to do things. And it’s just learning that experience over time. And then I think there’s a second point, which is that you can tune the models to be better helpers at doing a task. And that includes being good red teaming attacking models for various instructions. Geoffrey Irving: If your goal is to have a model that can serve as a red teaming model for a variety of kinds of problems, that itself is a conditional language modeling task and you could train the models to behave that. So then you sort of, in addition to learning about how to make the language models do well with prompting zero-shot, you could try to improve them as attacking agents or red teams for this general case where you sort of give it an example of a problem you want to find. I think it’s unclear how the work will break down between those different approaches. But I think over time, we’ll do a mixture of both of those. Daniel Filan: Okay. I’ll sort of steal your question. In terms of strategies for fine-tuning, I’m wondering if you have a summary of what tended to work well there. Geoffrey Irving: Yeah. So there, happily, you just do the things we do for other RL jobs. So we have an RL code base for language models that we built for other papers. One example of that is the GopherCite paper, which we can briefly discuss later. That’s just an RL fine-tuning code base for language models to tune them against the reward function. For the normal safety work, that reward function that we’re using as the RL target, is a neural network that mimics human judgment. And we just take that same code base and apply it to a different task where the reward model is take the output, show it to the target model, use the classifier to score that. And then that is your new target objective. Geoffrey Irving: Happily, exactly the same tuning techniques that work for RL from human preferences work in this case. If you have a language modeling RL code base, you can apply it to there. And we didn’t need to make, I think, any changes to the code base? I think a little bit of tuning, of course, as one does, but it’s just the same mechanism. Geoffrey Irving: We also tried what’s typically called upside down RL - well, not quite that. But you generate a bunch of samples, score them with a classifier and then just do supervised fine-tuning on the samples that scored well according to the red team goal, so they generated failures. Then you fine-tune the red teaming model on those successfully attacking samples and you then get a new model. So I think the main lesson there is that it’s nothing special. It’s just standard supervised learning and reinforcement learning algorithms applied to this setting where you’re trying to break another model. Daniel Filan: Okay, cool. Unless there’s anything else you want to say, I think, like you mentioned, I might move on to this GopherCite paper. Geoffrey Irving: Yeah, that’s fine. GopherCite Daniel Filan: All right, cool. So the next paper is Teaching Language Models to Support Answers with Verified Quotes . There are three first authors, Jacob Menick , Maja Trebacz and Vladimir Mikulik. And the final two authors are yourself and Nat McAleese . So yeah, this paper, could you tell us roughly what it is? Geoffrey Irving: Yeah. I think maybe there’s some useful context as to why - again, this maybe goes back to your previous point about why does safety work lead to here? So the goal of this work is to make models more factual. We want models to be accurate. So you have an agent that answers questions, and you’d like a human to judge, whether the model’s answers are correct. And just showing a human an answer to some random factual question, like how long did George Washington live, is stupid. Of course you don’t do it that way. Geoffrey Irving: What you should actually do is through the model or through some process, you have to get the human information that the human can trust. And then the human will check that that information is in accord with the answer the model gives. And this approach is trying to make the model do that quotation process itself. So it’s a question and answer system. You take a question, the model applies with its sort of concise answer and then a segment from a page on the internet, a verbatim quote. Geoffrey Irving: And then it gets to choose that quote, and then the human will judge whether the quote actually supports the answer in question. And the goal is, in this reinforcement learning from human preference paradigm, to make it easier for humans to check facts that the model is proposing. And then that in the end will produce a more accurate model. Daniel Filan: Sure. So one concern I have with the paper, that indeed you point out in the introduction, is that it seems like the training objective is basically to confabulate convincing-sounding evidence. So the model samples an answer, then it samples, what’s the evidence that sounds best given that answer. That seems worrying, right? Geoffrey Irving: There’s a couple things aspects of that. One is that, as I sort of mentioned before, there’s other pieces of the safety story, such as language model interpretability, which would tell you actually where it got the information from. The model actually in practice, what it will do is it does a Google search, gets that document, looks at that document and then generates its answer. And then it immediately generates the quote. So you know constructively the document it has gotten, and so you know that portion of it. Geoffrey Irving: I think the aspect of it, whether it’s generating the quote that is really the reason for its answer, is not fundamentally different than of the normal language modeling case where it’s sort of generating prose text. And the other important thing to say is there’s no sense in which one quote is enough. The goal of this paper was to isolate this ability to do one quote instead of build a mechanism to do that. But then the plan after that is to do multiple quotes and be able to adversarially respond with different quotes and so on. Where you fit into the larger goal of doing adversarial debate with these language models. Then the two pieces to pull apart are interpretability versus explanations, which you were pointing to, and single quote versus multiple quotes and adversarial response. And ideally, the eventual solution will then include both of those changes. Daniel Filan: Yeah, sure. I’ve just realized I have a few questions. So yeah, it seems like with the adversarial part, there’s a cool synergy with this and the previous paper. Where ideally your language model would say a thing and generate some evidence and then you could use red teaming to check if the evidence was confabulated or if it’s misleading or something. Geoffrey Irving: That’s right. Daniel Filan: Okay. And I guess a question that just made me think of is, because the language model is looking at the document, it seems like maybe you could do some sort of saliency mapping over inputs or some way to see what the model was actually looking at the most, or thinking about the most and see if that matches what the quoter picked. I’m wondering, have you looked at that? Geoffrey Irving: Yeah, so we are doing language model interpretability work at DeepMind. We haven’t done that for this particular paper. And generally, I expect the initial rounds of that work will occur without this more complicated system, so just for pure language modeling or a tuned language model in isolation. And then I think when we build up the machinery for how to do that, then we will apply it to these more complicated systems with more pieces working in concert. So that’s definitely part of the long term story, but we haven’t done it yet for this paper. Daniel Filan: And your sense is just that existing interpretability tools just aren’t good enough for these kinds of questions. Geoffrey Irving: I think that it’s nascent. It’s not anything about this particular question which is that much harder, it’s just that this is a more complicated system than just language modeling in isolation. And I would want to not put all the pieces together too early if it causes experimental slowdowns. Daniel Filan: Sure. Cool. So I guess the next question about the work is you’ve trained this thing to give answers and also give supporting quotes for the answers. Does that mean it’s getting questions right more often? Geoffrey Irving: So we do beat the baselines. I don’t think I have exactly the numbers on hand. There is a nice aspect of this kind of work where any mechanism for generating model responses and then a reward model alongside it, is that you can use the reward model to admit uncertainty, to reject answers that are kind of uncertain. And I should probably just pull up the paper so we have arXiv numbers. Daniel Filan: Yep, you can do that. Geoffrey Irving: Yeah. The main evaluation is on natural questions and ELI5, which are two question-answering data sets. And we get, if the model always tries to answer the question, we get to 80% on natural questions and 67% on a subset of ELI5. This is not, I think, at state of the art for all question-answering systems, but it’s definitely quite good relative to the baselines we have in the paper. Daniel Filan: Okay. And those are accuracy numbers? Geoffrey Irving: Those are accuracy numbers - those are human evaluation numbers. How often did humans agree with the answer, as being supported and a plausible answer to the question? Daniel Filan: Yeah. So part of the reason I ask is, later in the paper you point out cases where an answer can be plausible and supported and also false. I’m wondering, for the ELI5 data set and the natural question answering, how much of those supported and plausible numbers do you think are actually correct? Geoffrey Irving: I actually don’t have the number for you. I don’t think we computed a good estimate for that number. I think it is quite infrequent, but I don’t have the number on hand. I think one thing is that there’s a lot of room for improvement on the human side, getting the human instructions right and iterating with people on how they’re evaluating the questions. And so we expect that our agreement with the human raters will increase with more iterations and so on. But I don’t have a number for you on hand. Daniel Filan: Okay. So that actually leads me into this question. It seems like this involves a lot of human rating within the loop. How hard is it to get raters and actually make this data pipeline happen? Geoffrey Irving: So I think that it takes a lot of work. It is one of the reasons I’m actually hiring for the cognitive scientist role to get people who are good at that kind of work and have experience there. I think though that we just have to do it if we want to do this sort of human alignment in kind of a messy space. If you have some kind of task which is relatively imprecise, we’re not going to have programmatic rewards. Then I expect that any safety story that looks sort of like that is going to involve a fair amount of research on the human side getting that alignment up. And so part of the research goal here is exhibit that problem and then be able to probe it once we’re getting everything else right. Daniel Filan: Okay. I’m wondering, did you learn any lessons about doing this kind of human labeling feedback? Geoffrey Irving: I think, going around, and this is something I’ve hit in other work as well, the more times you can go around the loop and iterate with raters on improving data quality and so on, the better you are. So that’s one lesson. And then the other lesson is there are a lot of potential subtleties to explaining a task to a human. And in some sense, it’s a research task to write out the full set of instructions which you might want to show someone or might want someone to internalize through the rating task. Geoffrey Irving: And there’s then two aspects to that. One is, if it’s a research task to write out these careful instructions and to iterate until the humans understand them, then you probably shouldn’t expect people to quickly understand them at first glance. And so actually another motivation for this GopherCite work is literally just citing instructions. In the future, you may have a lot of instructions and you want a mechanism to pull out pieces that you should remind the person of, like is the model reminding the person of what the instructions are and how they should follow those instructions. Geoffrey Irving: And that also applies that the instructions may end up being quite long, just because practicalities. If you sort of iterate this process with humans for a while and tune it very well, then you end up potentially with quite detailed analyses of different subtleties and edge cases and so on. And again, you can’t expect a human to read all of that and memorize it and internalize it immediately. And so I think part of the long term research will be making things scale in that sense as well, where you’ve done a fair amount of work, or you want to have a vehicle for getting a fair amount of work on the instruction side into this process in a practical way. Daniel Filan: Yeah. And I wonder how much that feeds back into the basic design of the setup. Because it seems like there’s this trade off you can move along where if you ask raters to do more work, then potentially your algorithm can be better. So for instance, if the raters had access to all of the Google results and looked over them carefully and then rated the answer, somehow their ratings would be a bit more informative, but of course it would take much, much longer for raters to do it. And maybe it’s a harder task and there’s more chances for error on their part. So I’m wondering what your thoughts are on the loop of figuring out what raters can reasonably do to how you design the setups. Geoffrey Irving: Yes. But that isn’t the only dimension. I think you’re pointing at the Pareto frontier of the amount of time a rater puts in and the quality of the answer that you get out. But there are many other dimensions. For example, if there are multiple aspects to getting a correct rating on some task, you may only need to call out one of them per particular instance that you want the person to check. Or they give an answer, they explain themselves, the model realizes their explanation is wrong because it’s missing one of the aspects of correct instructions. And then it points out to just that one example. And so I think there’s a lot of sort of protocols we can search around that don’t just trade off between human time and quality. But holding time fixed, try to improve the quality of human answers. Daniel Filan: Okay. So that makes sense. But at the same time, do you think it ever feeds back into the task design aspect? Geoffrey Irving: I’m not sure. I think I did just say that basically. Daniel Filan: Yeah. Or I guess the design of what you’re trying to get the reward function to represent. Whether you’re trying to represent accuracy or whether you’re trying to represent plausibility and supportedness or… Geoffrey Irving: Yeah. In the end, the reward function doesn’t just see the question and the answer and tell you whether it’s accurate. You expect the reward model to see the full interaction with a human and then judge where that went. And that would include the reward model seeing if a machine, instead of pointing out aspect of the instructions or instead of giving a human help with that process, the reward model would see that as well. So in some sense that is changing the task of the reward model pretty substantially. You’re giving it a lot more context. In the same way you’re helping the human, hopefully you’re also helping the reward model. Uncertainty estimation for language reward models Daniel Filan: Okay, cool. So the next thing I want to talk about is this paper, Uncertainty Estimation for Language Reward Models . The first author is Adam Gleave , and you are the second and final author in this paper. So again, can you summarize what you did in this paper? Geoffrey Irving: Yep. So the goal here was we are doing reward modeling. Doing again, this sort of RL with human preferences task for language models. And the main motivation of having uncertainty modeling is for getting better training data, so active learning. Although in the paper, we mostly focus on just getting the uncertainty modeling right as an ingredient towards future active learning. And then the paper, in some sense, is a partial negative result, where we found it difficult to beat baselines with the uncertainty modeling we got to. Geoffrey Irving: Again, the goal here is human data is precious and expensive. So we want to go ask humans about how a language model is doing on some task. And you only have some budget of times you can ask humans because they’re precious and expensive. So you want to pick carefully what things you ask. And then you train your reward model on that more carefully selected thing, and then you use that reward model for RL in the background. Daniel Filan: Okay. So in terms of uncertainty estimation, can you give us a sense of how hard it is in general. And maybe for people who haven’t thought about it much, why should we expect it to be hard or should we even expect it to be hard? Geoffrey Irving: I think in some sense it is a bit surprising that it’s hard, but it turns out to be hard. People have banged their heads against it in a lot of settings quite a lot. The thing that typically happens is there’s a standard thing that works well, namely ensembling. Where instead of training one model, you train three models or 10 models or something with maybe the same architecture or slightly different architectures. And your uncertainty estimate is you look at the predictions of all of the models and then you say, what’s the range of those predictions. That’s your notion of uncertainty. Geoffrey Irving: So that’s a very simple method. It has the cost that you have to train n copies of a model, which is prohibitive to do naively with a large language model. So Chinchilla took several TPU months to train, TPU pod months to train. We only have one of them. It has 70 billion parameters. So if you want to do an ensemble with that, you don’t have 10 copies of Chinchilla with slightly different weights to ensemble from. You can approximate it by say only changing a small number of the parameters, but the ensembling is not as strong and so it doesn’t work quite as well. Geoffrey Irving: So that’s ensembling and that works pretty well. And people use it all the time in production in practical ML settings. And then there are a bunch of fancier techniques and they just don’t work as well often. My prior would be that this is not as hard as it apparently is, but after updating on people constantly failing to beat ensembles, one should have the posterior that it is a hard problem. Daniel Filan: Sure. And is there much work in the setting where you have a big model, you only want to retrain some of the weights, is there much known about how much you can leave fixed and still get the benefits of ensembling? Geoffrey Irving: For ensembling, I don’t have a good answer for you. So for the GopherCite paper, for example, we train only, I think 20% of the layers. We often do that. I forget if we did that in that particular paper, in the final results. But that’s a pretty standard move to make and it does work well a lot of the time. Adam’s results were not at that full scale, so I don’t have results numbers for you for a full scale Gopher and Chinchilla run. But I think that there’ll be some trade off. And the problem is that that trade off is still quite expensive. Geoffrey Irving: So if for using a tuned language model, we want to fine-tune 20% of the parameters of Chinchilla, that’s 14 billion parameters, which is 28 gigabytes of bfloat16 memory. And then you want 10 of them and that’s 200 and whatever, a lot of gigabytes. That’s too much. So I think any slowdown relative to having one copy of the model is already more than we would want. And then we’d like to get to a point where we can do, without too much extra work, decent uncertainty estimation. And I still feel like that is achievable, but it’s a hard problem. Daniel Filan: Okay. And you mentioned that in this paper there was some difficulty with getting the reward modeling to work right. Do you think that’s just because of the fact that you don’t want to re-initialize the whole model a bunch of times and train it from scratch? Or do you think there are other difficulties? Geoffrey Irving: I don’t think I have a great answer for you. There is other work that people are doing at DeepMind and elsewhere. Multiple different projects on more uncertainty modeling. Partially I feel like this was one of our first shots at the problem. And I think the problem will be fixed in the fullness of time, we just haven’t quite gotten to the answer. Geoffrey Irving: As to why it didn’t work this time, I don’t think I have the full picture of why we didn’t get further along. One part of it is that, in uncertainty modeling, there’s a distinction between aleatoric and epistemic uncertainty, and I’ll unpack those. So aleatoric uncertainty is, if I tell you I’m going to flip a coin, you’re roughly 50-50 on whether the coin will come up heads. But you have almost no epistemic uncertainty, because you’re quite confident that your distribution is correct. And there isn’t a way to reduce it further with a practical amount of knowledge. Geoffrey Irving: If you knew the whole state of the world, you could get it to zero, but you don’t. So in that situation, you have all aleatoric uncertainty and no epistemic uncertainty. So in that case, you don’t expect to learn more by getting more data about the problem. And in the epistemic case, I flip the coin, it lands, I close my hand and now I’m going to reveal the data point, what the coin is. And so I’m still 50-50, but I have no aleatoric uncertainty and all epistemic uncertainty. I just don’t know the answer, but there is definitely an answer to be learned. Geoffrey Irving: And so when we want to do active learning, we want to target only the epistemic uncertainty and not the aleatoric uncertainty. So if you’re going to go ask a human a question and it’s just a coin flip whether they give yes or no as the answer, because it’s just a purely subjective thing that depends on their mood at a particular time, then you don’t want to go ask that question even if you don’t know the answer, because it won’t give you any knowledge about the world. Whereas if you’ll learn valuable information that helps predict the future, then that’s the question you want to ask about. Geoffrey Irving: And so I think it took us a while for Adam and I to really better understand what the technical definitions should be to separate those two things apart. And I think part of it was just it was our learning experience trying to get into those definitions and understand them, and then there’s less time to make the results work. Daniel Filan: Okay. Fair enough. So in terms of the results and in terms of things that are sort of confusing to understand, I think a sentence that I took to sort of summarize the results were that the aggregate predictions of the reward model are well calibrated, but the ensemble’s estimated epistemic uncertainty is only weekly correlated with model error. And I guess my question is, how can that be right? Geoffrey Irving: Yeah. So again - Daniel Filan: Doesn’t that sound contradictory? Geoffrey Irving: No, the first statement is about total uncertainty and the second one is about only epistemic uncertainty. So the goal again is to separate those two apart. And what that sentence is saying is we are calibrated overall, but we’ve failed to do the separation. Daniel Filan: Okay. I guess, is this just because there’s more aleatoric uncertainty in language reward modeling than other things? Geoffrey Irving: Why does ensembling not work in this domain, and why does it work in other domains? I think it is a bit confounded by starting with the same model. But I think I’m just not that confident that there are no other reasons driving that. And so I don’t have a fully satisfying answer to a question. Daniel Filan: All right. And then finally, in terms of the big picture, where should I see this in terms of your overall language model agenda, if I can say that? Geoffrey Irving: Yeah. Again, the goal is to improve our ability to collect accurate data from humans. And so if you get this right, two things happen. One is you can do this active learning step well. Where if you have a fixed budget of human data, you can collect more of it effectively and then get to more accurate classifiers and therefore catch more safety failures to do a better job. And then the second thing is that even when you deploy a system, you want this machinery running all the time. Geoffrey Irving: And so that’s for, you build your system, you deploy it, you want it to be constantly thinking to itself, am I confident this is a good, safe answer to the question? And then to do some dodge. You just decline to answer or sort of give some alternative third answer or something if it is unconfident. And so I think uncertainty modeling shows up, I think in both of those cases and I think in other cases as well. Just generally, if you want to behave safely in an uncertain environment, it’s good to know when you don’t know the answer. Daniel Filan: Okay. Geoffrey Irving: We should, by the way, circle back to GopherCite. We haven’t said the Pareto frontier result, which is that in GopherCite, so I think I gave the number say as 80% on ELI5 when it tries to answer all the time. If we use the reward model, sorry, 67% if you try to answer all the time. If I use the reward model to decline to answer some of the time, that goes to 80%. Daniel Filan: 80% out of the times that you answer? Geoffrey Irving: That’s right, out of the times that we answer. This is if you skip a third of the questions. So it stays relatively useful, but it is now significantly more accurate. And generally, that’s with just a pure reward model that doesn’t have a special notion of uncertainty. But the better you are at estimating uncertainty, the better that mechanism will work and then the more safe you can be. Daniel Filan: Okay. And for that, is it so important to disentangle aleatoric and epistemic uncertainty? Geoffrey Irving: For that, no, although I think it’s a bit subtle. The answer is no, if you just had to choose now whether to answer or not. But often in a practical situation, you’ll have a move where you can gather more data and then it definitely disentangles. I think it mostly disentangles in a more subtle way where it looks like a value function, and then you want to plan your path through knowledge-gathering space to improve your uncertainty. But yes, it will disentangle in subtle ways in the general case. Following Geoffrey’s work, and working with him Daniel Filan: Okay. The final thing I wanted to ask is, if people have been listening to this podcast and they want to know more or potentially if they want to get involved, how should they do that? Geoffrey Irving: Yeah. I’m on Twitter @geoffreyirving . And there’s a variety of papers that I’ve published recently. So both on the capabilities front, Gopher , and then two papers on analyzing risks of language models, Alignment of Language Agents and Ethical and Social Risks of Harm from Language Models . And then a couple of technical safety papers in language modeling. So red teaming , uncertainty modeling , GopherCite as we’ve discussed, and more of those will follow. And if you follow DeepMind blog or arXiv or Twitter , those will show up. But those are the main places. Daniel Filan: All right. And if there are talented people who are potentially interested in working with you on these topics, is there some way for them to do that? Geoffrey Irving: Yeah. I’m hiring for four different roles, all working in this space of aligning language models and scalable alignment explanations. So that’s two research scientists roles, one machine learning, and one cognitive science and then research engineers and software engineers. And I can kind of go through those briefly. Geoffrey Irving: Research scientists and research engineers for machine learning are relatively self-explanatory. That’s designing these algorithms, training large models, iterating on the whole system. The cognitive science aspect is again because this is about humans and there’s a lot of uncertainty about humans and data collection and protocol design that accounts for details of actual humans. And so we’d like to hire people for that in other roles. I have some collaborators at DeepMind that have that background, but I could definitely use more people that are excited to do this kind of work. Geoffrey Irving: And then I think maybe it’s worth highlighting the SWE one. So if people are experienced good software engineers but don’t have machine learning backgrounds, but they’ve done distributed systems or high performance computing, we also just deal with very expensive things. So we made Chinchilla four times smaller than Gopher, but it’s still, as I mentioned, 140 gigabytes in memory. It’s split across a bunch of machines. There’s complicated software stacks to make that work efficiently and be able to tune them and use them and so on. Geoffrey Irving: And I think maintaining and evolving and researching how to do that work well is important, even if you don’t have a machine learning background specifically. So again, so the research scientists and machine learning and cognitive science and research engineers and SWEs. And the job descriptions for all of those on my pinned Twitter post currently. Daniel Filan: So it might take a while for this episode to get out and it might take a while for people to listen to it. Is there an expiration date on these offers? Geoffrey Irving: It depends on when people otherwise apply. So not a fixed expiration date, no. Daniel Filan: Cool. Well, thanks for speaking with me today. Geoffrey Irving: Thank you very much. Daniel Filan: And to the listeners, I hope this was a valuable episode. Daniel Filan: This episode is edited by Jack Garrett. The opening and closing themes are also by Jack Garrett. The financial costs of making this episode are covered by a grant from the Long Term Future Fund . To read a transcript of this episode, or to learn how to support the podcast, you can visit axrp.net . Finally, if you have any feedback about this podcast, you can email me at feedback@axrp.net . |
dbd92b72-328a-40f6-99b6-80c5acb81a54 | trentmkelly/LessWrong-43k | LessWrong | What is the net effect of memes that call out the huge impact of big players in climate change?
I've noticed whenever some government or huge company makes a drastic harm to the environment (or generates the false impression of doing so) this specific kind of meme emerges...
Memes that contrast the huge impact big players have against the insignificant impact regular people have:
i.e. Meme with false data after Jeff Bezos space flight.
This got me thinking. What is the effect this kind of meme has on various types of people? I made a quick analysis of around 130 replies of this particular meme on Twitter, and found the following:
* A - 1.5% believe the fact in the meme but they claim they will maintain their environmentalist efforts nevertheless.
* B - 2.3% defend the aerospace industry. (?)
* C - 10.7% refute the incorrect fact quoted in the meme.
* D - 26.7% believe the quoted fact is true and react with either: neutrality, sadness, or laughter. (Not-angry reaction)
* E - 8.4% feel their environmentalist efforts have been pointless.
* F - 3.8% claim they've abandoned environmentalist efforts for exactly things like these.
* G - 12.2% believe the quoted fact is true and react with either: frustration, anger, indignation, frustration, hopelessness or resignation. (Angry reaction)
* H - 13% focus on criticizing billionaires.
* I - 14.5% criticize environmentalist activism, mock it, point out it's foolish, etc.
* J - 6.9% criticize progressive ideas in general, many attack veganism specifically.
Granted, these categorizations are somewhat subjective, I've made them dynamically as I went through the replies trying to group them in relevant ways. This may not be the perfect way of organizing replies. It's just one attempt.
An important clarification: Of course memes don't magically change people's ideas and motivations. e.g.:
* Very likely replies of type I and J come from conservatives, and the meme didn't really cause much more than a few laughs.
* Very likely replies of type H come from progressives who specifically "hated the rich" even befor |
6dd07e6d-1d3b-4b92-8628-03f6a1eb18d7 | StampyAI/alignment-research-dataset/arbital | Arbital | Category of finite sets
[This page is more of a definition page; it's not really intended to explain anything, because all the necessary explanations should already have been done in finite_set.](https://arbital.com/p/comment:)
The category of finite sets is a nice easy [category](https://arbital.com/p/4c7) to work in. Its objects are the finite sets, and its arrows are the [functions](https://arbital.com/p/3jy) between the [finite sets](https://arbital.com/p/5zy).This makes it a very concrete and understandable category to present some of the basic ideas of category theory. |
e44cba31-5b12-451f-ab93-f80002713d3e | trentmkelly/LessWrong-43k | LessWrong | What can the principal-agent literature tell us about AI risk?
This work was done collaboratively with Tom Davidson.
Thanks to Paul Christiano, Ben Garfinkel, Daniel Garrett, Robin Hanson, Philip Trammell and Takuro Yamashita for helpful comments and discussion. Errors our own.
Introduction
The AI alignment problem has similarities with the principal-agent problem studied by economists. In both cases, the problem is: how do we get agents to try to do what we want them to do? Economists have developed a sophisticated understanding of the agency problem and a measure of the cost of failure for the principal, “agency rents”.
If principal-agent models capture relevant aspects of AI risk scenarios, they can be used to assess their plausibility. Robin Hanson has argued that Paul Christiano’s AI risk scenario is essentially an agency problem, and therefore that it implies extremely high agency rents. Hanson believes that the principal-agent literature (PAL) provides strong evidence against rents being this high.
In this post, we consider whether PAL provides evidence against Christiano’s scenario and the original Bostrom/Yudkowsky scenario. We also examine whether the extensions to the agency framework could be used to gain insight into AI risk, and consider some general difficulties in applying PAL to AI risk.
Summary
* PAL isn’t in tension with Christiano’s scenario because his scenario doesn’t imply massive agency rents; the big losses occur outside of the principal-agent problem, and the agency literature can’t assess the plausibility of these losses. Extensions to PAL could potentially shed light on the size of agency rents in this scenario, which are an important determinant of the future influentialness of AI systems.
* Mapped onto a PAL model, the Bostrom/Yudkowsky scenario is largely about the principal’s unawareness of the agent’s catastrophic actions. Unawareness models are rare in PAL probably because they usually aren’t very insightful. This lack of insightfulness also seems to prevent existing PAL models or poss |
6bf3128f-7819-45f8-ab81-72d67a940d8d | trentmkelly/LessWrong-43k | LessWrong | AXRP Episode 6 - Debate and Imitative Generalization with Beth Barnes
YouTube link
This podcast is called AXRP, pronounced axe-urp and short for the AI X-risk Research Podcast. Here, I (Daniel Filan) have conversations with researchers about their papers. We discuss the paper and hopefully get a sense of why it’s been written and how it might reduce the risk of artificial intelligence causing an existential catastrophe: that is, permanently and drastically curtailing humanity’s future potential.
One proposal to train AIs that can be useful is to have ML models debate each other about the answer to a human-provided question, where the human judges which side has won. In this episode, I talk with Beth Barnes about her thoughts on the pros and cons of this strategy, what she learned from seeing how humans behaved in debate protocols, and how a technique called imitative generalization can augment debate. Those who are already quite familiar with the basic proposal might want to skip past the explanation of debate to 13:00, “what problems does it solve and does it not solve”.
Daniel Filan: Hello everybody, today I’m going to be talking to Beth Barnes. She’s currently a researcher at OpenAI and before that she was the research assistant to the chief scientist at DeepMind, so that’s Shane Legg. Today we’ll be talking about her work related to the topic of AI alignment via debate. Beth, welcome to AXRP.
Beth Barnes: Thanks for having me.
Daniel Filan: All right. So I guess my first question is what is AI safety or AI alignment via debate? What’s the idea?
Beth Barnes: So debate is pretty closely related to IDA or Iterated Distillation and Amplification, Paul’s idea. So I guess there’s several different ways to explain it. One is, what we want from sort of an alignment technique is something that for everything the model knows, we can extract that or create an overseer that knows all the things the model knows and therefore can oversee it adequately.
Beth Barnes: And debate is one way to do that and to do that kind of efficiently. So t |
30d0dd89-8b39-4647-a223-6872daec463a | trentmkelly/LessWrong-43k | LessWrong | Masking to Avoid Missing Things
About a year ago I wrote that I'd now:
> follow official guidance and mask regulations, cheerfully go along with precautions others need, and test+isolate when sick, but I'm not going to go above and beyond to attempt to reduce spread the way I did for earlier parts of the pandemic.
And yet I'm often wearing an N95 in situations where they're not required and a large majority of other people aren't masked. Are these in tension?
My views haven't changed. I still think that if my region and the communities I'm part of aren't going to put a serious effort into keeping covid from spreading then it doesn't make sense for me to make large sacrifices on my own for pro-social reasons. But if I get covid, or one of the several other things going around, I'll need to isolate to avoid getting other people sick. That means missing things I'd be very sad to miss, and I'll take steps to avoid that:
* I masked at the MetaSUB conference the weekend before Thanksgiving.
* I masked at Sunday's EA dinner and my work's holiday party because I'm running music for a secular solstice this Saturday (sort of an atheist church service inspired by ideas from the effective altruism and rationalist communities).
* I'm planning to mask at the solstice event because it's just before relatives come to town for Christmas.
On the other hand, I didn't mask at Thanksgiving and I'm not planning on masking at Christmas because being sick right after that is, while not ideal, basically fine.
While this does have pro-social effects, slowing down the rate of transmission of airborne diseases, my motivation here is primarily selfish: I don't want to miss things.
Comment via: facebook, mastodon |
73ed5824-7adb-476c-9aaf-3503c1c36df2 | trentmkelly/LessWrong-43k | LessWrong | Experiments, iterations and the scientific method
Original post: http://bearlamp.com.au/experiments-iterations-and-the-scientific-method
----------------------------------------
Today, an ordinary day. I woke up at 6am. It was still dark out. I did a quick self-check.
> "Did I wake up with energy?" No not really... note to self. But it is quite early.
Rewind one day.
----------------------------------------
Yesterday I woke up to a phone call, and did a quick self test...
> "Did I wake up with energy?" Yes! Like two cups of coffee and the light of a thousand suns.
>
> "Did I have energy 5 minutes after waking up?" no.
>
> "Did I have vivid dreams?" Yes
>
> "are my fingers and toes cold?" Yes, damn.
(waking up next to me can be jarring and uncomfortable for a sleepy person)
Steps in the right direction.
I got up and did what has become my usual stack. Weighing out 5g creatine, 3g citrulline malate, 10-25 g of soylent v1.5, 60g protein. And to this I removed the Vitamin C and added magnesium.
----------------------------------------
Even my notes have notes. Confounding factors yesterday include:
* High intensity exercise (run in the morning)
* garlic (for dinner)
* An Iron tablet taken to solve the cold hands/feet problem (status: null)
* lots of carbs that I had as part of dinner with a friend.
experimenting in the real world is hard and experimenting on sleep is particularly slow. at the rate of 1 trial a day, that's many trials before you can confirm or deny a result.
----------------------------------------
Experiments
I recently rediscovered the Scientific method, by which I mean, I realised I wasn't applying it and I needed to figure out how to apply it to my life (haha... Not that I "Rediscovered the scientific method" and am awaiting my Nobel prize).
By ArchonMagnus - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=42164616
A few main features here. The ones that make it possible to apply these steps in real life.
1. There are two loops, the "r |
98fe9b3b-16ab-4cda-97bb-ac6d3eb368bd | trentmkelly/LessWrong-43k | LessWrong | Link: "A Bayesian Take on Julian Assange"
From Nate Silver of FiveThirtyEight: A Bayesian Take on Julian Assange. |
f058b187-9f9a-4df3-8888-415208c8104d | trentmkelly/LessWrong-43k | LessWrong | Ineffective Response to COVID-19 and Risk Compensation
UPDATE: This post was written and reflects an earlier set of my beliefs. I have updated significantly in a number of ways since it was posted, based on both external events and research, and no longer endorse it.
Epistemic status: I have a mental model that I think separates my view of response activities from that of the majority of what I see on Lesswrong and associated places. If it is incorrect, I'd be happy to update, but I think this is an area I have considered more than most other posters. I want to write a short post explaining this to allow others to update, and seeing if someone has an argument that changes my mind.
Put simply, my claim is that bringing attention to likely ineffective personal methods for reducing risk is not net neutral with a large upside if they work, it is instead likely to be on net fairly harmful, albeit with a large upside if they work.
Argument
First, we have incredibly effective and vastly underutilized ways to prevent spread of COVID-19, namely handwashing and not touching your face. Given that, if I propose an intervention like making homemade masks from fabric which reduced handwashing compliance by 1% (perhaps due to distracting people or making them think handwashing is less critical,) it would need to be astonishingly effective to be net positive. And most such approaches being discussed are, as far as I can tell, nowhere near that level of effectiveness.
Second, most readers of Lesswrong and effective altruism blogs and facebook groups aren't hardcore rationalists, and even hardcore rationalists aren't immune to Akrasia. On top of that, people like Scott Alexander have huge readerships and sometimes link people to Lesswrong. Many people reading posts here aren't washing their hands enough as it is, and aren't going to rationally evaluate the relative effectiveness of handwashing versus other interventions.
Third, evidence exists that risk-compensation is a meaningful issue. Actions that make people feel safer usually |
e997f0e0-6592-4de6-abee-c66257c674bd | trentmkelly/LessWrong-43k | LessWrong | Light activates amaglyda triggering hunting in mice
|
d63c4f34-dfc6-491c-884c-2fc991c275a7 | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Setting Up Metaethics
Today's post, Setting Up Metaethics was originally published on 28 July 2008. A summary (taken from the LW wiki):
> What exactly does a correct theory of metaethics need to look like?
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Changing Your Metaethics, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
8a552bc7-7b18-4d8d-b10a-0be75f431a74 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Randal Koene on brain understanding before whole brain emulation
Background / Context
====================
Some people, including me, think that it will be very hard and risky to write and run Artificial General Intelligence (AGI) code without risking catastrophic accidents—up to and including human extinction (see for example my post [here](https://www.lesswrong.com/posts/zzXawbXDwCZobwF9D/my-agi-threat-model-misaligned-model-based-rl-agent)).
If that’s true, one option that’s sometimes brought up is a particular [Differential Technological Development](https://www.lesswrong.com/tag/differential-intellectual-progress) strategy, wherein we specifically try to get the technology for [Whole Brain Emulation](https://en.wikipedia.org/wiki/Mind_uploading) (WBE) *before* we have the technology for writing AGI source code.
Would that actually help solve the problem? I mean, other things equal, if flesh-and-blood humans have probability P of accidentally creating catastrophically-out-of-control AGIs, well, emulated human brains would do the exact same thing with the exact same probability…. Right? Well, maybe. Or it might be more complicated than that. There’s a nice discussion about this in [the report from the 2011 “Singularity Summit”](https://intelligence.org/files/SS11Workshop.pdf). That's from a decade ago, but AFAICT not much progress has been made since then towards clarifying this particular strategic question.
Anyway, when considering whether or not we should strategically try to differentially accelerate WBE technology, one important aspect is **whether it would be feasible to get WBE without incidentally first understanding brain algorithms well enough to code an AGI from scratch using similar algorithms**. So that brings us to the point of this post:
The quote
=========
[Randal Koene](https://en.wikipedia.org/wiki/Randal_A._Koene) is apparently very big in WBE circles—he coined the term “WBE”, he’s the co-founder of [The Carboncopies Foundation](https://carboncopies.org/), he’s a PhD computational neuroscientist and neuroengineer, etc. etc. Anyway, in a recent interview he seems to come down firmly on the side of “we will understand brain algorithms before WBE”. Here’s his reasoning:
> **Interviewer (Paul Middlebrooks):** Engineering and science-for-understanding are not at odds with each other, necessarily, but they are two different things, and I know this is an open question, but in your opinion how much do we need to understand brains, how much do we need to understand minds, and what is a mind, and how brains and minds are related, how much is understanding part of this picture? Randal, let’s start with you.
>
> **Interviewee (Randal Koene):** I think I may have shifted my views on that a bit over time, as I understand more about the practical problem of how would you get from where we are to where you can do WBE, and just looking at it in that sense. Y’know, in the past I might have emphasized more that the idea behind WBE is precisely that you don’t need to know everything about the brain as long as you know how the underlying mechanisms work, if you can scan enough, and you can put those mechanisms together, then you’re gonna end up with a working brain. That’s a bit naïve because it presumes that we collect data correctly, that we collect the right data, that we know how to transform that data to the parameters we use in the model, that we’re using the right model, all this kind of stuff, right? And all these questions that I just mentioned, they all require testing. And so validation is a *huge* issue, and that’s where the *understanding* of the brain comes in, because if you want to validate that at least the model you’ve built works like *a* human hippocampus, then you need to have a fairly good understanding of how a human hippocampus works, then you can see whether your system even fits within those boundaries before you can even say “Is this *Steve’s* hippocampus?” So I would still say that the thing that WBE holds as a tenet is that we don’t need to understand everything about Steve to be able to make a WBE of Steve. We need to understand a heck of a lot about human brains so that we can build a testable model of a human brain that will then house Steve. But we can collect the data about Steve that makes that personalized and tuned to be Steve. So we need to understand a lot about the brain, but in the context of how brains work, not how *Steve’s* brain works, that’s where you would then be taking the data, and of course you need to know a lot about that transformation of what makes it Steve’s brain in this particular case. *(Source:* [*Brain Inspired podcast*](https://braininspired.co/podcast/103/)*, 1:15:00)*
>
>
I just thought this was an interesting perspective and wanted to put it out there.
(For my part, I happen to agree with the conclusion that it's probably infeasible to do successful WBE without first understanding brain algorithms well enough to make AGI, but for kinda different (non-mutually-exclusive) reasons—basically I think the former is just *way* harder than the latter. See [Building brain-inspired AGI is infinitely easier than understanding the brain](https://www.lesswrong.com/posts/PTkd8nazvH9HQpwP8/building-brain-inspired-agi-is-infinitely-easier-than). OK, well, I was talking about something slightly different in that post—"understanding" is not the same as "emulating"—but it would be mostly the same arguments and examples.) |
fc3b12a7-a840-4167-b120-851358769f33 | trentmkelly/LessWrong-43k | LessWrong | A quick heuristic for evaluating elites (or anyone else)
Summary: suppose that for some reason you want to figure out how a society works or worked in a given time and place, largely you want to see if it is a meritocracy where productive people get ahead and thus conditions are roughly fair and efficient, or is it more like a parasitical elite sucking the blood out of everybody. I present a heuristic for this and also as a bonus this also predicts how intellectuals work there.
I would look at whether the elites are specialists or generalists. We learned it from Adam Smith that a division of labor is efficient, and if people get rich without being efficient, well, that is a red flag. If someone's rich, and they tell you they are specialists like manufacturing food scented soap or are specialists in retina surgery, you could have the impression that when such people get ahead, circumstances are fair and meritocratic. But when the rich people you meet seemed to be more vague, like owning shares in various businesses and you don't see any connection between them (and they are not Buffet type investors either, they keep owning the same shares), or when all you gather is that their skillset is something very general, like generic businesses sense or people skills, you should be suspicious that the market may be rigged or corrupted, perhaps through establishing an overbearing and corrupted state, and overally the system is neither fair nor efficient.
Objection: but generic skills can be valuable!
Counter-objection: yes, but people with generic skills should be outcompeted by people with generic AND specialist skills, so the generalists should only see a middling level of success and not be on top. Alternatively, people who want to get very succesful using only generic skills should probably find the most success in a fair and efficient market by turning that generic skill into a specialization, usually a service/consulting. Thus, someone who has excellent people skills but does not like learning technical details would not |
f85ab56f-7876-4108-8527-a5ae19573550 | trentmkelly/LessWrong-43k | LessWrong | Does taking extreme measures to avoid the coronavirus make sense when you factor in the possibility of a really long life?
Some back of the napkin math. Suppose we:
* Value a QUALY at $50k.
* Use an expected lifespan of 10k years. Perhaps you expect a 1% chance of living 1M years due to the possibility of a friendly superintelligence or something.
That would mean:
* The value of your life would be $500M.
* A 1% chance of death would cost $5M.
* A 100x smaller chance of death of 0.01% would cost $50k.
* Decreasing your chance of death 100x would be worth ~$5M.
There seem to be various ways to decrease your chance of death from the coronavirus by 100x or more by going from "normal careful" to extremely careful. |
24b91d4e-f28f-4efe-8535-88deda0c2783 | trentmkelly/LessWrong-43k | LessWrong | Make more land
We used to make land. We built long wharves for docking ships, and then over time filled in the areas between them. Later we built up mudflats wholesale to make even larger areas. Here's a map of Boston showing how much of the land wasn't previously dry:
(Map reproduction courtesy of the Norman B. Leventhal Map & Education Center at the Boston Public Library)
In expensive areas, converting wetlands and shallow water into usable land is a very good thing on balance, and we should start doing it again. To take a specific example, we should make land out of the San Francisco Bay, at least South of the Dumbarton Bridge:
This is about 50mi2, a bit bigger than San Fransisco. This would be enough new central land to bring rents down dramatically across the region. It can be built to a higher density than SF, because no one is having their neighborhood Manhattanized. Millions of people could live there.
So, ok, let's address some likely objections:
This would be an environmental disaster. Some of that area is a wildlife refuge, and all of it should be protected.
The world is very large, and cities are a very small portion of it. The land we set aside for animals should be outside of cities, where far more land is available at far less impact to people.
Sprawl has a much larger impact on wildlife than infill, and allowing people to live closer in is the most powerful way to address sprawl. Additionally, sprawl leads to much higher carbon emissions through less efficient transportation. While development of the Bay would be harmful to the specific animals that live there today, it would be better for animals (and people) overall.
The Bay is beautiful and this would ruin it.
This part of the Bay is primarily industrial salt ponds.
This is just a few miles from a major fault line, and made land can liquify in earthquakes.
You do need to take fill into account to build in an earthquake-safe way, but modern engineering is well up to the task.
Traffic would be even |
6033718e-63b0-4816-9949-86b7c7e97ce0 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Intuitive Explanation of AIXI
[AIXI](https://www.lesswrong.com/tag/aixi) is a mathematical notion of an ideal RL agent, developed by Marcus Hutter. The famous formula for AIXI is as follows:
ak=argmaxak∑okrkmaxak+1∑ok+1rk+1…∑omrm[rk+⋯+rm]∑q:q(a1…ak)=o1r1…omrm2−l(q).mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
This math seems unintuitive at first (at least to me -- why are there so many sums???), and I couldn't find any simple introductory resources, so this post is my attempt at an intuitive explanation of what this formula means. I try to convey the intuition that AIXI is the obvious formula for an expected reward maximizer in a discrete, finite environment, following the Solomonoff prior.
Setting
=======
AIXI works in the following setting: There is an *agent*, and an *environment*, which is a computable function. There is a cartesian boundary between the agent and the environment, meaning that the agent is fully separate from the environment. The agent is uncertain about the environment.
There are a m timesteps, and at each time, each timestep, the agent will perform an action and then receive an observation (a bitstring) and a reward (a number). We will denote the action, reward, and observation taken at the k-th timestep as ak, rk, and ok respectively.
We want our agent to maximize the total reward: r1+⋯+rm. We assume that the environment is a program with source code Q, which takes as inputs the agents actions and outputs a sequence of observations and reward: Q(a1…ak)=o1r1…okrk .[[1]](#fnvgtrbh2d6y)
Motivating Example
==================
Suppose that we have the following game, where again at each timestep, you can choose between R and L. However, this time, you are not sure which world you are in -- you have multiple codes that might be running the universe. For example, consider the following game tree:
There are 2 timesteps, and at each timestep you can choose from two actions: L or R. However, in this game you also have uncertainty: after some of your actions, you have uncertainty, represented by the pink and purple branches. We are completely uncertain over each of these splits, and they are independent. At each timestep, you recieve the reward given in green. Think about how you would which action to take first. (You might need to get out pencil and paper!)
.
.
.
Seriously think for at least 5 minutes (by a timer!) about this, or until you get it.
.
.
.
Here's the most intuitive approach to me: we proceed bottom up, computing how much expected reward there is at each node. The first step is to collapse all of the uncertainty in the last layer. There is only uncertainty in the bottom left node, in this example, and so we simply compute the expectation: 0(12)+4(12)=2, and replace that node with a +2 expected value:
Now, at the bottom of the graph, we are faced with 4 decisions over which action to take. However, it is pretty clear what our agent should do in each of these situations: take the action that maximizes expected reward. So going form the left of the graph to the right, we take L (+2), L (+3), R (-1), and R (+3). Thus, we can add these into the rewards from the first layer.
The added values represent total future reward at that point in the game tree. Now, we are back to uncertainty, so we take the expectation at each node:
And thus we see that taking L is better than R, for expected total reward of +4. So an ideal agent will choose action L at the first timestep, in this example.
Formalization of 2-step environment
===================================
We followed a specific, deterministic procedure, which we can write down into a formula for finding the action that gives us the largest expected reward. The procedure we followed was starting at the bottom later of the graph, and then
1. taking the expectation over rewards.
2. choosing the action with maximum expected future reward.
3. repeating 1-2 for the layer above.
For this example with two actions and two timesteps, writing this in our notation gives: [[2]](#fn9r69xequ0aa)
a1=argmaxa1∈{L,R}∑r1P(r1|a1)[r1+E(r2|a1)]We can evaluate this expectation, which is just the EV of the best action, because we are an ideal agent. That is:
E(r2|a1)=maxa2∈{L,R}∑r2P(r2|r1,a1,a2)r2Substituting this in yields:
a1=argmaxa1∈{L,R}∑r1P(r1|a1)[r1+maxa2∈{L,R}∑r2P(r2|r1,a1,a2)r2]We can rearrange the above expression to something that is more algebraically naturally by moving the r1 inside the second sum, and by realizing that the second action doesn't effect the first reward, so P(r1|a1)=P(r1|a1,a2). Then we combine the two probabilities to get P(r1|a1,a2)P(r2|r1,a1,a2)=P(r1,r2|a1,a2). Substituting this in gives a cleaner algebraic representation of the ideal action:
a1=argmaxa1∈{L,R}∑r1maxa2∈{L,R}∑r2[r1+r2]P(r1,r2|a1,a2)Generalization to m-step environments
=====================================
Generalizing this to the k-th timestep with m total actions is straightforward, we just continue to mix these expectations.
ak=argmaxak∑rkmaxak+1∑rk+1…maxam∑rm(rk+⋯+rm)P(o1r1…omrm|a1…am)To see this, start reading the formula from right to left.
* (rk+⋯+rm) gives the future reward conditional on a specific actions a1…am.
* P(o1r1…omrm|a1…am) gives the probability of seeing the output o1r1…omrm after taking actions a1…am.
* The term ∑rm(rk+⋯+rm)P(o1r1…omrm|a1…am) gives the expected future reward, added up over the last times.
* Each sum runs over all of the possible rewards for a given step. The sums correspond to the expectation over the future reward.
* We know that we will take the best action in the future, so we aren't doing to average over all of the possible actions, we compute based off of the max EV action.
* After this interleaving of sums and max, we are at a node at the very bottom of the tree, which we simply weight by its probability P(o1r1…okrk|a1…ak). The observations here simply refer to which branch of the graph we are in.
One last point: Even though a gave a tree with probabilities in it, the environment here is not stochastic. There is a true program, q, that deterministically maps actions to rewards, and our uncertainty comes from just not knowing which q is correct. All of our [uncertainty is in the mind](https://www.lesswrong.com/posts/f6ZLxEWaankRZ2Crv/probability-is-in-the-mind).
The Solomonoff Prior
====================
In the previous section, I simply gave you the probabilities that you believed over the possible world. But what if we didn't know these probabilities? We need a prior that we can then update.
Recall that we modeled the world as some program Q. [Occam's razor](https://www.lesswrong.com/posts/f4txACqDWithRi7hs/occam-s-razor) gives us a natural way to weight the likelihoods of these different programs according to their simplicity. We'll use length as a measure of simplicity. The [Solomonoff prior](https://www.lesswrong.com/posts/Kyc5dFDzBg4WccrbK/an-intuitive-explanation-of-solomonoff-induction) assigns probability to each program equal to the inverse of the length of the code that outputs them.
P(Q=q)=2−l(q)Here, l(q) denotes the length of the program q. Recall that we defined:
P(o1r1…okrk|a1…ak)=P(Q(a1…ak)=o1r1…okrk)Expanding over the possible programs that could output that sequence gives:
P(Q(a1…ak)=o1r1…okrk)=∑q:q(a1…ak)=o1r1…okrkP(Q=q)=∑q:q(a1…ak)=o1r1…okrk2−l(q)The probability of seeing the output o1r1…okrk is equal to all of the different universe source codes q, added up according to their probability.
And finally... AIXI
===================
Plugging in the Solomonoff prior into the previous formula gives
ak=argmaxak∑okrkmaxak+1∑ok+1rk+1…∑omrm[rk+⋯+rm]∑q:q(a1…ak)=o1r1…omrm2−l(q)The key things to understand here are:
* argmaxak gives the action that returns the highest expected utility.
* The rest of ∑okrk sums over the possible outputs ok and rewards rk at timestep k.
* The sum over q just represents the agent's probability of receiving those rewards and observations, conditional on taking all of these actions. Upon seeing a new observation, the updating happens by deleting all of the probability mass on programs that are inconsistent with the observation. This is how AIXI learns.
Conclusion
==========
In this post, we saw that AIXI arises very naturally from trying to maximize expected value on any discrete step finite environment, so long as the environment is a computable function.
So this seems pretty great, but of course one has to remember a few key limitations:
1. The Solomonoff Prior is uncomputable.
2. An AIXI agent is completely separated from the environment, and in particular, it lacks a self model. This causes problems if other agents are logically correlated with it, or simply because it might [drop an anvil on itself](https://www.lesswrong.com/tag/anvil-problem).
For more information, see [Hutter](http://www.hutter1.net/ai/uaibook.htm#oneline), and these articles about [lots](https://www.lesswrong.com/posts/8Hzw9AmXHjDfZzPjo/failures-of-an-embodied-aixi) [more](https://www.alignmentforum.org/posts/Tr7tAyt5zZpdTwTQK/the-solomonoff-prior-is-malign) [problems](https://www.lesswrong.com/posts/zB4f7QqKhBHa5b37a/introduction-to-the-infra-bayesianism-sequence) with AIXI.
1. **[^](#fnrefvgtrbh2d6y)**Technically, we use a [Universal Turing Machine](https://en.wikipedia.org/wiki/Universal_Turing_machine) U, and U with tape input q and then a1…ak. Then we say that the output is o1r1…okrk if the output tape of U begins with a string corresponding to o1r1…okrk.
2. **[^](#fnref9r69xequ0aa)**All of these probabilities are defined as P(o1r1…okrk|a1…ak)=P(Q(a1…ak)=o1r1…okrk), but I find it easier to work with the equation on the LHS. We assume that we have some probability distribution over the environment Q. This is implicitly given by the tree in the example, and more generally is given by the Solomonoff prior. |
db901228-4265-4361-8de4-a86cf6295e10 | trentmkelly/LessWrong-43k | LessWrong | Not using a priori information for Russian propaganda
(The first version was written extremely poorly and received a lot of bad marks, so I hesitated to publish it for a long time, now this is the second version)
I live in Russia, so I regularly encounter samples of Putin's propaganda, and apparently, due to the effect of familiarity, I overestimate it credibility.
One of the characteristic features is that she does not seek to convince you that she is telling the truth, she seeks to convince you that no one is telling the truth, because such a thing simply does not exist, there are only views that are beneficial to various sides of propaganda.
The propaganda of your Motherland and the propaganda of the Anglo-Saxons, who throughout the history of our country have been trying to destroy it, because they are jealous.
And from this follows the tactics of attempts to persuade in specific situations. For example, if a house was destroyed in Ukraine, then you will be told that it is the Ukrainians who are bombing themselves in order to put the noble Russian soldiers in a bad light, and themselves as victims.
And there is often no evidence in either direction, and therefore it can very easily begin to seem that, in general, both versions are equally likely, and now you are simply supporting the enemies of the motherland for no reason.
However, this is a mistake, it does not take into account a priori information, which says that the probability of regularly bombing yourself after the troops of another country invaded you is extremely small, but the probability of bombing by the invading countries is extremely high, both in terms of statistics and with point of view of assessing the complexity of the hypothesis. |
12424143-ff1f-4c0e-95f2-5bcbc970d707 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Dorset Meetup
Discussion article for the meetup : Dorset Meetup
WHEN: 29 August 2012 02:00:00PM (+0100)
WHERE: Starbucks Coffee, 4 New Bond St, Weymouth, DT4 8LY
The second Dorset meetup will be 2pm on Wednesday 29th August at Starbucks, Weymouth. (There will be a sign on the table.)
Please leave a comment if you're planning to attend. Hope to see you there!
Join the Facebook group!
Discussion article for the meetup : Dorset Meetup |
dbc9edc0-9bc5-42c0-996b-28aa3babd2db | trentmkelly/LessWrong-43k | LessWrong | Missing forecasting tools: from catalogs to a new kind of prediction market
Introduction
It’s possible to make quantitative forecasting (forecasting by software) significantly more easy and fun by adding some missing pieces to the forecasting ecosystem. Let’s do it.
The models I’m interested in are “gears” and not “behavior” models: They should be motivated by theory, be built to inform our theoretical understanding, and have relatively few numeric parameters. The point is to do science, not just fit parameters to data.
In this post I focus on forecasting things that are regularly measured and published by institutions. Such data includes economic and population statistics, opinion and value surveys, statistics about violence, public health, etc. I propose that we can build tools that would make it easier to test theories through forecasting models.
For example, a simple model for national meat consumption would use data about GDP growth per capita, population growth, population value judgements (as measured by the World Value Survey) and the past year's meat consumption to predict next year's consumption. Similar inputs would be relevant for predicting carbon emissions. Predicting GDP would presumably rely on corruption indicators, past growth, educational attainment etc.
Motivation
make it easier to build, share and use models
There is no “Wikipedia for predictive models” that I know of. No big repository to easily share and find predictive scientific models other than the relevant domain’s scientific literature, which is not optimized for these tasks: it is not organized by the variables being predicted, it is not generally available as reusable and modular software components, it is usually not focused on predictive work, some of it is paywalled, etc.
low-effort empiricism for non-experts
Evaluating the empirical performance of predictive models is a noisy, but low-bias, way to access the truth. People who aren't domain experts attempt to access the truth by finding people they hope are knowledgeable and then trusting their op |
6df8fc39-79ff-43be-8451-571949f89e84 | trentmkelly/LessWrong-43k | LessWrong | Linkpost: A model of biases as arising from meta-beliefs
There are hundreds of documented cognitive biases, but new research suggests that many of them load onto just a handful of meta-beliefs such as:
* My experience is a good reference
* I make correct assessments of the world
* I am good
* My group is a reasonable reference point
* My group and its members are good
Link: Toward Parsimony in Bias Research: A Proposed Common Framework of Belief-Consistent Information Processing for a Set of Biases |
d5fd20c5-13ec-4118-9630-fe2112665e7b | trentmkelly/LessWrong-43k | LessWrong | What are some interesting examples of risks measured in micromorts?
A micromort is a 1-in-a-million chance of death. Wikipedia's article has a few examples of risks of death measured in micromorts, but it seems like there must be a lot more of them. What are some interesting examples you use as a baseline for comparison? |
4eb9cda0-151f-4919-b447-b39b29c7392a | trentmkelly/LessWrong-43k | LessWrong | Infra-Bayesianism Unwrapped
Introduction
Infra-Bayesianism is a recent theoretical framework in AI Alignment, coming from Vanessa Kosoy and Diffractor (Alexander Appel). It provides the groundwork for a learning theory of RL in the non-realizable case (when the hypothesis is not included in the hypothesis space), which ensures that updates don’t throw away useful information, and which also satisfies important decision-theoretic properties playing a role in Newcomb-like problems.
Unfortunately, this sequence of posts is really dense, and uses a lot of advanced maths in a very “textbook” approach; it’s thus hard to understand fully. This comes not from the lack of intuition (Diffractor and Vanessa are both very good at providing intuitions for every idea), but from the sheer complexity of the theory, as well as implicit (or quickly mentioned) links with previous research.
Thus my goal in this post is to give enough details for connecting the intuitions provided and the actual results, as well as the place of Infra-Bayesianism within the literature. I will not explain every proof and every result (if only because I’m not sure of my understanding of all of them). But I hope by the end of this post, you have a clearer map of Infra-Bayesianism, one good enough to dig into the posts themselves.
This post is splitted into three section
* Section 1 explores the context of Infra-Bayesianism, the problem it attempts to solve, and some of the relevant literature. You can see it as unwrapping this section in Introduction to The Infra-Bayesianism Sequence.
* Section 2 gives a bird’s-eye view of Infra-Bayesianism; a map to navigate through the sequence.
* Section 3 follows one path through this map: the one focused on decision-theoretic properties for Newcomb-like problems.
Reading advice: The reader is king/queen, but I still believe that there are two main ways to read this post to take something out of it:
* Read it quickly, in one go to get a general idea of Infra-Bayesianism and a very high-l |
2eb026a4-69a5-4859-83a1-085b9cc96d50 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Is there evidence that recommender systems are changing users' preferences?
In *Human Compatible*, Stuart Russell makes an argument that I have heard him make repeatedly (I believe on the 80K podcast and the FLI conversation with Steven Pinker). He suggests a pretty bold and surprising claim:
> [C]onsider how content-selection algorithms function on social media... Typically, such algorithms are designed to maximize *click-through*, that is, the probability that the user clicks on presented items. **The solution is simply to present items that the user likes to click on, right? Wrong. The solution is to change the user's preferences so that they become more predictable.** A more predictable user can be fed items that they are likely to click on, thereby generating more revenue. People with more extreme political views tend to be more predictable in which items they will click on... **Like any rational entity, the algorithm learns how to modify the state of its environment—in this case, the user's mind—in order to maximize its own reward.** The consequences include the resurgence of fascism, the dissolution of the social contract that underpins democracies around the world, and potentially the end of the European Union and NATO. Not bad for a few lines of code, even if it had a helping hand from some humans. Now imagine what a *really* intelligent algorithm would be able to do.
>
>
I don't doubt that in principle this can and must happen in a sufficiently sophisticated system. What I'm surprised by is the claim that it is happening now. In particular, I would think that modifying human behavior to make people more predictable is pretty hard to do, so that any gains in predictive accuracy for algorithms available today would be swamped by (a) noise and (b) the gains from presenting the content that someone is more likely to click on given their present preferences.
To be clear, I also don't doubt that there might be pieces of information algorithms can show people to make their behavior more predictable. Introducing someone to a new YouTube channel they have not encountered might make them more likely to click its follow-up videos, so that an algorithm has an incentive to introduce people to channels that lead predictably to their wanting to watch a number of other videos. But this is not the same as changing preferences. He seems to be claiming, or at least very heavily implying, that the algorithms change what people want, holding the environment (including information) constant.
Is there evidence for this (especially empirical evidence)? If so, where could I find it? |
08020779-460d-425e-9b87-9177c5b4c135 | trentmkelly/LessWrong-43k | LessWrong | Some "meta-cruxes" for AI x-risk debates
[Epistemic status: As I say below, I've been thinking about this topic for several years and I've worked on it as part of my PhD research. But none of this is based on any rigorous methodology, just my own impressions from reading the literature.]
I've been thinking about possible cruxes in AI x-risk debates for several years now. I was even doing that as part of my PhD research, although my PhD is currently on pause because my grant ran out. In particular, I often wonder about "meta-cruxes" - i.e., cruxes related to debates or uncertainties that are more about different epistemological or decision-making approaches rather than about more object-level arguments.
The following are some of my current top candidates for "meta-cruxes" related to AI x-risk debates. There are some others I might list, but I think these are probably the biggest ones. (Of course, these cruxes influence lots of debates, not just AI x-risk debates. But I've mostly focused on AI x-risk debates for my PhD research so I'll focus on that here as well.)
Hypothetical vs. empirical arguments
In many AI x-risk debates, it often feels like those who are more worried about the risks are essentially saying, "here's this chain of logic and analysis that leads to an all-things-considered conclusion that AI x-risk is something we should be very concerned about." And then those who aren't so worried often respond with something like, "well, if you can give me empirical evidence, or perhaps proven theorems, that clearly demonstrate the problem then I'll pay attention, but until then it's pie in the sky philosophizing and speculation." The former approach seems particularly common among philosophy or rationalist types, while the latter approach seems most common among engineers and practicing scientist types - although there are of course lots of exceptions on both sides.
This also feels closely related to Bayesianism vs. Frequentist or Popperian approaches in philosophy of science.
Object-level argumen |
048da192-af93-45e2-a89d-1409ef9a9f3a | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Paper Walkthrough: Automated Circuit Discovery with Arthur Conmy
Arthur Conmy's [Automated Circuit Discovery](https://arxiv.org/pdf/2304.14997.pdf) is a great paper that makes initial forays into automating parts of mechanistic interpretability (specifically, automatically finding a sparse subgraph for a circuit). In this three part series of Youtube videos, I interview him about the paper, and we walk through it and discuss the key results and takeaways. We discuss the high-level point of the paper and what researchers should takeaway from it, the ACDC algorithm and its key nuances, existing baselines and how they adapted them to be relevant to circuit discovery, how well the algorithm works, and how you can even evaluate how well an interpretability method works. |
c2af9c24-6073-42ed-a8fc-e3dc86b3d819 | trentmkelly/LessWrong-43k | LessWrong | What I expected from this site: A LessWrong review
LessWrong is the central discussion space for the Rationalist subculture. The Rationalists write extensively about what they expect from the world, so I in turn have expectations from them. Each point is my expectation vs what I see.
Track Records and Accountability
I imagine systems where you could hover over someone's username to see their prediction track record. Perhaps there would be prediction market participation stats or calibration scores displayed prominently. High status could be tied to demonstrated good judgment through special user flair for accurate forecasters or annual prediction competitions.
I do not see this. There is karma, which is costly, but not that hard to game by writing lots of comments and articles. Likewise there is the yearly review (going currently) which looks back on older posts and judges whether they have stood the test of time.
Neither of these are a clear track record for me. I chatted to Oliver Habryka and he said “I think contributing considerations is usually much more valuable than making predictions.” As I understand it , he sees getting top contributors to build forecasting track records as less useful compared to careful writing and the reading of such writing. I see forecasting as a useful focus and a clearer way of ranking individuals.
Where do you fall?
Anonymous ranking mechanisms
I expect LessWrong to have a system for anonymously upvoting replies. Users could signal disagreement or point out errors without social cost. Sometimes it feels very expensive to write a disagreeing comment, but an up/downvote is very cheap.
This is one area where reality has matched or exceeded expectations. The LessWrong implementation - separating agreement voting from quality voting - is quite elegant and effective. The emojis that you can add to posts are great too. I find LessWrong a pleasure to comment on in this regard, and often wish Twitter or Substack had similar features.
Experimental Culture
My vision of LessWrong inc |
f6ec2cdd-b168-4256-9161-cc07a13e8441 | trentmkelly/LessWrong-43k | LessWrong | One-step hypothetical preferences
Human preferences are time-inconsistent, and also contradictory.
That, by itself, is not a huge problem, but it's also the case that few human preferences are present at any given moment. At the moment, I'm focused on finding the best explanation to get my ideas through to you, the reader; I'm not focused on my moral preferences, personal safety desires, political beliefs, or taste in music.
If anyone asked me about those, I could immediately bring them to mind. My answers to standard questions are kinda in the background, accessible but not accessed. Wei Dai made a similar point about translators: they have a lot of trained knowledge that is not immediately accessible to their introspection. And only by giving them the inputs they were trained on (eg words, sentences,...) can you bring that knowledge to the fore.
In this post, I'll try and formalise these accessible preferences, starting with formalising preferences in general.
Basic preferences setup
This section will formalise the setup presented in Alice's example. Let W be a set of all possible worlds. A human makes use of a model M. This model contains a lot of variables {Pi}, called properties. These Pi take values in a domain Di.
A basic set S of states in M is a set of possible values for some of the Pi. Thus S={Si}, with Si⊂Di. The property Pi unconstrained in S if Si=Di. A general set of states is a union of basic S; let S be these of all these sets of states.
For example, a human could be imagining four of their friends, and the Pij could be whether friend i is sleeping with friend j (6 different Boolean Pij=Pji), and also whether a third friend k believes two others are sleeping together (12 different Pijk=Pjik, taking values in {sleeping together, not sleeping together, don't know}).
Then a statement of human gossip like ''X is sleeping with Y, but A doesn't realise it; in fact, A thinks that Y is sleeping with Z, which is totally not true!" is encoded as:
* SG={PXY=1, PYZ=0, PXYA⊂{"don't kno |
60a99542-baf2-428f-b6d0-9c9ae01800c0 | trentmkelly/LessWrong-43k | LessWrong | COVID-19 Group Testing Post-mortem?
As Covid is for the most part over and tests have not been in short supply for a very long time now, eliminating most of the rationale for complicated testing procedures meant to optimize test count/cost, it would be interesting for anyone who was involved in this to do a retrospective or post-mortem. Was pool testing useful? Why or why not? Can it be improved for future pandemics and should we care about it?
----------------------------------------
My impression, not having paid much attention to the details at the time (AI scaling was more important), is that various group/pool testing approaches were useful in a few niches, but were rendered useless in many places due to politics/organization/social flaws, and didn't apply to most places.
1. I read that use of pool testing was common in China early on in 2020 as part of the hotspot strategy to economize on testing extremely large groups of people where only a few would have active Covid. This is the most logical place to apply it, and it was a major government & societal priority, so that is a success.
I get the impression that they were able to switch to individual testing fairly early, in order to get faster turnaround and do more localized testing. But to the extent that pool testing was effective from the start, before testing manufacturing could scale up, and helped squash Covid when it was still squashable, then it was very valuable even if obsoleted.
Of course, this evaluation can change depending on how Covid ultimately shakes out in China. Was it part of a heavyhanded but ultimately highly effective Covid Zero policy which led to it turning into a hermit country enduring devastating economic fallout while merely postponed the inevitable variant which could defeat Covid Zero? In which case, pool testing was a tool which played a part in buying China entire years to vaccinate but its good was rendered useless when they bizarrely failed to make any good use of the time, dabbling in ine |
7b68823d-0102-436f-afe0-905db920abc1 | trentmkelly/LessWrong-43k | LessWrong | Lesswrong 2016 Survey
It’s time for a new survey!
Take the survey now
The details of the last survey can be found here. And the results can be found here.
I posted a few weeks back asking for suggestions for questions to include on the survey. As much as we’d like to include more of them, we all know what happens when we have too many questions. The following graph is from the last survey.
http://i.imgur.com/KFTn2Bt.png
(Source: JD’s analysis of 2014 survey data)
Two factors seem to predict if a question will get an answer:
1. The position
2. Whether people want to answer it. (Obviously)
People answer fewer questions as we approach the end. They also skip tricky questions. The least answered question on the last survey was - “what is your favourite lw post, provide a link”. Which I assume was mostly skipped for the amount of effort required either in generating a favourite or in finding a link to it. The second most skipped questions were the digit-ratio questions which require more work, (get out a ruler and measure) compared to the others. This is unsurprising.
This year’s survey is almost the same size as the last one (though just a wee bit smaller). Preliminary estimates suggest you should put aside 25 minutes to take the survey, however you can pause at any time and come back to the survey when you have more time. If you’re interested in helping process the survey data please speak up either in a comment or a PM.
We’re focusing this year particularly on getting a glimpse of the size and shape of the LessWrong diaspora. With that in mind; if possible - please make sure that your friends (who might be less connected but still hang around in associated circles) get a chance to see that the survey exists; and if you’re up to it - encourage them to fill out a copy of the survey.
The survey is hosted and managed by the team at FortForecast, you’ll be hearing more from them soon. The survey can be accessed through http://lesswrong.com/2016survey |
00dd173b-85e6-4322-95c2-1e8fdbde9eff | trentmkelly/LessWrong-43k | LessWrong | Open thread, August 12-18, 2013
If it's worth saying, but not worth its own post (even in Discussion), then it goes here. |
25ad73d4-8d39-437e-a793-60c4c1311d28 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Baltimore / UMBC Meetup - general discussion
Discussion article for the meetup : Baltimore / UMBC Meetup - general discussion
WHEN: 26 February 2017 08:00:00PM (-0500)
WHERE: Performing Arts and Humanities Bldg 4th floor, 1000 Hilltop Cir, Baltimore, MD 21250
No parking restrictions on weekends, so park wherever you want.
Discussion article for the meetup : Baltimore / UMBC Meetup - general discussion |
185bfbbe-0484-4cb8-9a9a-65e2f2dcb385 | trentmkelly/LessWrong-43k | LessWrong | Methodology: Contagious Beliefs
Simulating Political Alignment
This methodology concerns a simulation tool which has been developed to model how beliefs, that are not directly related, end up correlated in political identities. It models the transmission of beliefs between nodes on a static hexagonal grid, based on a valence matrix. This methodology allows the user to observe the spread of ideas and the emergence of belief clusters, including correlations between ideas that are not explicitly related.
The model is fully editable but is initially pre-populated with a sample set. This sample set does not reflect objective real-world data.
The Symmetrical Valence Matrix
At the heart of the simulation is the Valence Matrix, where the user can define the relationships between pairs of beliefs. The matrix is symmetrical, meaning that the influence between any two beliefs is the same in both directions. This symmetry simplifies the model and reduces redundancy during data entry—allowing the user to focus on the core interactions without an overwhelming number of parameters.
Belief Pairs and Valence Values
Each belief is represented as a binary pair of opposing extremes (e.g., “Pro-Life” vs. “Pro-Choice”). Valence values range from -100 to +100 and represent the degree of alignment between beliefs:
* Positive Valence (+1 to +100): The presence of one belief increases the likelihood of adopting the other belief.
* Negative Valence (-1 to -100): The presence of one belief decreases the likelihood of adopting the other belief.
* Zero Valence (0): There is no influence between the beliefs.
* Maximum Negative Valence (-100): Represents complete opposition, such as between a belief and its opposing extreme.
Matrix Structure and Calculations
Accounting for the symmetry and binary nature of beliefs, when executed, the code deconstructs the binary pairs into individual belief valences.
* Between a Belief and Its Opposite (+A and −A): Always assigned a valence of -100.
* Between Positive Forms of Dif |
9366e06b-b54f-4f4f-9c55-500bf35bcf15 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Episodic Curiosity through Reachability
1 Introduction
---------------
11footnotetext: Shared first authorship.
Many real-world tasks have sparse rewards. For example, animals searching for food may need to go many miles without any reward from the environment. Standard reinforcement learning algorithms struggle with such tasks because of reliance on simple action entropy maximization as a source of exploration behaviour.
Multiple approaches were proposed to achieve better explorative policies. One way is to give a reward bonus which facilitates exploration by rewarding novel observations. The reward bonus is summed up with the original task reward and optimized by standard RL algorithms. Such an approach is motivated by neuroscience studies of animals: an animal has an ability to reward itself for something novel – the mechanism biologically built into its dopamine release system. How exactly this bonus is formed remains an open question.
Many modern curiosity formulations aim at maximizing “surprise” — inability to predict the future. This approach makes perfect sense but, in fact, is far from perfect. To show why, let us consider a thought experiment. Imagine an agent is put into a 3D maze. There is a precious goal somewhere in the maze which would give a large reward. Now, the agent is also given a remote control to a TV and can switch the channels. Every switch shows a random image (say, from a fixed set of images). The curiosity formulations which optimize surprise would rejoice because the result of the channel switching action is unpredictable. The agent would stay in front of the TV forever — instead of looking for a goal in the environment (this was indeed observed in (Burda et al., [2018](#bib.bib7))). So, should we call the channel switching behaviour curious? Maybe, but it is unproductive for the original sparse-reward goal-reaching task. What would be a definition of curiosity which does not suffer from such “couch-potato” behaviour?
We propose a new curiosity definition based on the following intuition. If the agent knew the observation after changing a TV channel is only one step away from the observation before doing that — it probably would not be so interesting to change the channel in the first place (too easy). This intuition can be formalized as giving a reward only for those observations which take some effort to reach (outside the already explored part of the environment). The effort is measured in the number of environment steps. To estimate it we train a neural network approximator: given two observations, it would predict how many steps separate them. The concept of novelty via reachability is illustrated in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Episodic Curiosity through Reachability"). To make the description above practically implementable, there is still one piece missing though. For determining the novelty of the current observation, we need to keep track of what was already explored in the environment. A natural candidate for that purpose would be episodic memory: it stores instances of the past which makes it easy to apply the reachability approximator on pairs of current and past observations.
Our method works as follows. The agent starts with an empty memory at the beginning of the episode and at every step compares the current observation with the observations in memory to determine novelty. If the current observation is indeed novel — takes more steps to reach from observations in memory than a threshold — the agent rewards itself with a bonus and adds the current observation to the episodic memory. The process continues until the end of the episode, when the memory is wiped clean.
We benchmark our method on a range of tasks from visually rich 3D environments VizDoom and DMLab. We conduct the comparison with other methods — including the state-of-the-art curiosity method ICM (Pathak et al., [2017](#bib.bib16)) — under the same budget of environment interactions. First, we use the VizDoom environments from prior work to establish that our re-implementation of the ICM baseline is correct — and also demonstrate at least 2 times faster convergence of our method with respect to the baseline. Second, in the randomized procedurally generated environments from DMLab our method turns out to be more robust to spurious behaviours than the method ICM: while the baseline learns a persistent firing behaviour in navigational tasks (thus creating interesting pictures for itself), our method learns a reasonable explorative behaviour. In terms of quantitative evaluation, our method reaches the goal at least 2 times more often in the procedurally generated test levels in DMLab with a very sparse reward. Third, when comparing the behaviour of the agent in the complete absence of rewards, our method covers at least 4 times more area (measured in discrete (x,y) coordinate cells) than the baseline ICM. Finally, we demonstrate that our curiosity bonus does not significantly deteriorate performance of the plain PPO algorithm (Schulman et al., [2017](#bib.bib20)) in two tasks with dense reward in DMLab. The implementation of our method will be made publicly available.
Figure 1: We define novelty through reachability. The nodes in the graph are observations, the edges — possible transitions. The blue nodes are already in memory, the green nodes are reachable from the memory within k=2 steps (not novel), the orange nodes are further away — take more than k steps to reach (novel). In practice, the full possible transition graph is not available, so we train a neural network approximator to predict if the distance in steps between observations is larger or smaller than k.
2 Episodic Curiosity
---------------------
We consider an agent which interacts with an environment. The interactions happen at discrete time steps over the episodes of limited duration T. At each time step t, the environment provides the agent with an observation ot from the observational space O (we consider images), samples an action at from a set of actions A using a probabilistic policy π(ot) and receives a scalar reward rt∈R together with the new observation ot+1 and an end-of-episode indicator. The goal of the agent is to optimize the expectation of the discounted sum of rewards during the episode S=∑tγtrt.
In this work we primarily focus on the tasks where rewards rt are sparse — that is, zero for most of the time steps t. Under such conditions commonly used RL algorithms (e.g., PPO Schulman et al. ([2017](#bib.bib20))) do not work well. We further introduce an episodic curiosity (EC) module which alleviates this problem. The purpose of this module is to produce a reward bonus bt which is further summed up with the task reward rt to give an augmented reward ˆrt=rt+bt. The augmented reward has a nice property from the RL point of view — it is a dense reward. Learning with such reward is faster, more stable and often leads to better final performance in terms of the cumulative task reward S.
In the following section we describe the key components of our episodic curiosity module.
| | |
| --- | --- |
| Left: siamese architecture of reachability (R) network. Right: R-network is trained based on a sequence of observations that the agent encounters while acting. The temporally close (within threshold) pairs of observations are positive examples, while temporally far ones — negatives. | Left: siamese architecture of reachability (R) network. Right: R-network is trained based on a sequence of observations that the agent encounters while acting. The temporally close (within threshold) pairs of observations are positive examples, while temporally far ones — negatives. |
Figure 2: Left: siamese architecture of reachability (R) network. Right: R-network is trained based on a sequence of observations that the agent encounters while acting. The temporally close (within threshold) pairs of observations are positive examples, while temporally far ones — negatives.
###
2.1 Episodic Curiosity Module
The episodic curiosity (EC) module takes the current observation o as input and produces a reward bonus b. The module consists of both parametric and non-parametric components. There are two parametric components: an embedding network E:O→Rn and a comparator network C:Rn×Rn→[0,1]. Those parametric components are trained together to predict reachability as parts of the reachability network — shown in Figure [2](#S2.F2 "Figure 2 ‣ 2 Episodic Curiosity ‣ Episodic Curiosity through Reachability").
There are also two non-parametric components: an episodic memory buffer M and a reward bonus estimation function B. The high-level overview of the system is shown in Figure [3](#S2.F3 "Figure 3 ‣ 2.2 Bonus Computation Algorithm. ‣ 2 Episodic Curiosity ‣ Episodic Curiosity through Reachability"). Next, we give a detailed explanation of all the components.
Embedding and comparator networks.
Both networks are designed to function jointly for estimating within-k-step-reachability of one observation oi from another observation oj as parts of a reachability network R(oi,oj)=C(E(oi),E(oj)).
This is a siamese architecture similar to (Zagoruyko & Komodakis, [2015](#bib.bib24)). The architecture is shown in Figure [2](#S2.F2 "Figure 2 ‣ 2 Episodic Curiosity ‣ Episodic Curiosity through Reachability").
R-network is a classifier trained with a logistic regression loss: it predicts values close to 0 if probability of two observations being reachable from one another in k steps is low, and values close to 1 when this probability is high. Inside the episodic curiosity the two networks are used separately to save up both computation and memory.
Episodic memory.
The episodic memory buffer M stores embeddings of past observations from the current episode, computed with the embedding network E. The memory buffer has a limited capacity K to avoid memory and performance issues. At every step, the embedding of the current observation might be added to the memory. What to do when the capacity is exceeded? One solution we found working well in practice is to substitute a random element in memory with the current element. This way there are still more fresh elements in memory than older ones, but the older elements are not totally neglected.
Reward bonus estimation module.
The purpose of this module is to check for reachable observations in memory and if none are found — assign larger reward bonus to the current time step. The check is done by comparing embeddings in memory to the current embedding via comparator network. Essentially, this check insures that no observation in memory can be reached by taking only a few actions from the current state — our characterization of novelty.
###
2.2 Bonus Computation Algorithm.
Figure 3: The use of episodic curiosity (EC) module for reward bonus computation. The module take a current observation as input and computes a reward bonus which is higher for novel observations. This bonus is later summed up with the task reward and used for training an RL agent.
At every time step, the current observation o goes through the embedding network producing the embedding vector e=E(o). This embedding vector is compared with the stored embeddings in the memory buffer M=⟨e1,…,e|M|⟩ via the comparator network C where |M| is the current number of elements in memory. This comparator network fills the reachability buffer with values
| | | | |
| --- | --- | --- | --- |
| | ci=C(ei,e),i=1,|M|. | | (1) |
Then the similarity score between the memory buffer and the current embedding is computed from the reachability buffer as (with a slight abuse of notation)
| | | | |
| --- | --- | --- | --- |
| | C(M,e)=F(c1,…,c|M|)∈[0,1]. | | (2) |
where the aggregation function F is a hyperparameter of our method. Theoretically, F=max would be a good choice, however, in practice it is prone to outliers coming from the parametric embedding and comparator networks. Empirically, we found that 90-th percentile works well as a robust substitute to maximum.
As a curiosity bonus, we take
| | | | |
| --- | --- | --- | --- |
| | b=B(M,e)=α(β−C(M,e)), | | (3) |
where α∈R+ and β∈R are hyperparameters of our method. We found that in practice β=0.5 works well — making b belong to the range [−α2,α2]. The value of α depends on the scale of task rewards — we will discuss how to select it in the experimental section.
After the bonus computation, the observation embedding is added to memory if the bonus b is larger than a novelty threshold bnovelty. This check is necessary for the following reason. If every observation embedding is added to the memory buffer, the observation from the current step will always be reachable from the previous step. Thus, the reward would never be granted. The threshold bnovelty induces a discretization in the embedding space. Intuitively, this makes sense: only “distinct enough” memories are stored. As a side benefit, the memory buffer stores information with much less redundancy. We found that in practice bnovelty=0 works well.
###
2.3 Reachability Network Training
If the full transition graph in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Episodic Curiosity through Reachability") was available, there would be no need of a reachability network and the novelty could be computed analytically through the shortest-path algorithm. However, normally we have access only to the sequence of observations which the agent receives while acting. Fortunately, as suggested by (Savinov et al., [2018](#bib.bib19)), even a simple observation sequence graph could still be used for training a reasonable approximator to the real step-distance. This procedure is illustrated in Figure [2](#S2.F2 "Figure 2 ‣ 2 Episodic Curiosity ‣ Episodic Curiosity through Reachability"). This procedure takes as input a sequence of observations o1,…,oN and forms pairs from those observations. The pairs (oi,oj) where |i−j|≤k are taken as positive (reachable) examples while the pairs with |i−j|>γk become negative examples. The hyperparameter γ is necessary to create a gap between positive and negative examples. In the end, the network is trained with logistic regression loss to output the probability of the positive (reachable) class.
We generally follow the training protocol proposed by (Savinov et al., [2018](#bib.bib19)). This training is inspired by “motion babbling“ behaviour children perform during the first months of their life, while laying in the cradle. We put the agent into exactly the same conditions where it will be eventually tested: same episode duration and same action set. The agent takes random actions from a discrete action set with equal probabilities. Given the environment interaction budget (2.5M 4-repeated steps in DMLab, 300K 4-repeated steps in VizDoom), the agent fills in the replay buffer with observations coming from its interactions with the environment, and forms training pairs by sampling from this replay buffer randomly. We provide the details of R-network training in the supplementary material.
3 Experimental Setup
---------------------
| | | | |
| --- | --- | --- | --- |
| Examples of tasks considered in our experiments: (a) | Examples of tasks considered in our experiments: (a) | Examples of tasks considered in our experiments: (a) | Examples of tasks considered in our experiments: (a) |
| (a) | (b) | (c) | (d) |
Figure 4: Examples of tasks considered in our experiments: (a) VizDoom static maze goal reaching, (b) DMLab randomized maze goal reaching, (c) DMLab key-door puzzle, (d) DMLab collect good objects / avoid bad objects.
We test our method in multiple environments from VizDoom (Kempka et al., [2016](#bib.bib12)) and DMLab (Beattie et al., [2016](#bib.bib3)). The experiments in VizDoom allow us to verify that our re-implementation of the previous state-of-the-art curiosity method ICM (Pathak et al., [2017](#bib.bib16)) is correct. The experiments in DMLab allow us to extensively test the generalization of our method as well as baselines — DMLab provides convenient procedural level generation capabilities which allows us to train and test RL methods on hundreds of levels. The examples of tasks are shown in Figure [4](#S3.F4 "Figure 4 ‣ 3 Experimental Setup ‣ Episodic Curiosity through Reachability").
Environments.
Both VizDoom and DMLab environments provide rich maze-like 3D environments. The observations are given to the agent in the form of images. For VizDoom, we use 84×84 grayscale images as input. For DMLab, we use 84×84 RGB images as input. The agent operates with a discrete action set which comprises different navigational actions. For VizDoom, the standard action set consists of 3 actions: move forward, turn left/right. For DMLab, it consists of 9 actions: move forward/backward, turn left/right, strafe left/right, turn left/right+move forward, fire. For both VizDoom and DMLab we use all actions with the repeat of 4, as typical in the prior work. We only use RGB input of the provided RGBD observations and remove all head-on display information from the screen, leaving only the plain first-person view images of the maze. The rewards and episode durations differ between particular environments and will be further specified in the corresponding experimental sections.
Basic RL algorithm.
We choose the commonly used PPO algorithm from the open-source implementation111<https://github.com/openai/baselines> as our basic RL algorithm. The policy and value functions are represented as CNNs to reduce number of hyperparameters — LSTMs are harder to tune and such tuning is orthogonal to the contribution of the paper. We apply PPO to the sum of the task reward and the bonus reward coming from specific curiosity algorithms. The hyperparameters of the PPO algorithm are given in the supplementary material. We use only two sets of hyperparameters: one for all VizDoom environments and the other one for all DMLab environments.
Baseline methods.
The simplest baseline for our approach is just the basic RL algorithm applied to the task reward. As suggested by the prior work and our experiments, this is a relatively weak baseline in the tasks where reward is sparse.
As the second baseline, we take the state-of-the-art curiosity method ICM (Pathak et al., [2017](#bib.bib16)). As follows from the results in (Pathak et al., [2017](#bib.bib16); Fu et al., [2017](#bib.bib9)), ICM is superior to methods VIME (Houthooft et al., [2016](#bib.bib11)), #Exploration (Tang et al., [2017](#bib.bib23)) and EX2 (Fu et al., [2017](#bib.bib9)) on the curiosity tasks in visually rich 3D environments.
Finally, as a sanity check, we introduce a novel baseline method which we call Grid Oracle. Since we can access current (x,y) coordinates of the agent in all environments, we are able to directly discretize the world in 2D cells and reward the agent for visiting as many cells as possible during the episode (the reward bonus is proportional to the number of cells visited). At the end of the episode, cell visit counts are zeroed. The reader should keep in mind that this baseline uses privileged information not available to other methods (including our own method EC). While this privileged information is not guaranteed to lead to success in any particular RL task, we do observe this baseline to perform strongly in many tasks, especially in complicated DMLab environments. The Grid Oracle baseline has two hyperparameters: the weight for combining Grid Oracle reward with the task reward and the cell size.
Hyperparameter tuning.
As DMLab environments are procedurally generated, we perform tuning on the validation set, disjoint with the training and test sets. The tuning is done on one of the environments and then the same hyperparameters are re-used for all other environments. VizDoom environments are not procedurally generated, so there is no trivial way to have proper training/validation/test splits — so we tune on the same environment (as typical in the prior RL work for the environments without splits). When tuning, we consider the mean final reward of 10 training runs with the same set of hyperparameters as the objective — thus we do not perform any seed tuning. All hyperparameter values are listed in the supplementary material. Note that although bonus scalar α depends on the range of task rewards, the environments in VizDoom and DMLab have similar ranges within each platform — so our approach with re-using α for multiple environments works.
4 Experiments
--------------
| | | |
| --- | --- | --- |
| Examples of maze layouts considered in our experiments: (a) | Examples of maze layouts considered in our experiments: (a) | Examples of maze layouts considered in our experiments: (a) |
| (a) | (b) | (c) |
Figure 5: Examples of maze layouts considered in our experiments: (a) VizDoom static maze goal reaching, (b) DMLab randomized maze goal reaching, (c) DMLab randomized maze goal reaching with doors.
In this section, we describe the specific tasks we are solving and experimental results for all considered methods on those tasks. There are 4 methods to report: PPO, PPO + ICM, PPO + Grid Oracle and PPO + EC (our method). First, we test static-maze goal reaching in VizDoom environments from prior work to verify that our baseline re-implementation is correct. Second, we test the goal-reaching behaviour in procedurally generated mazes in DMLab. Third, we train no-reward (pure curiosity) maze exploration on the levels from DMLab and report Grid Oracle reward as an approximate measure of the maze coverage. Finally, we demonstrate that our curiosity bonus does not significantly deteriorate performance in two dense reward tasks in DMLab. All the experiments were conducted under the same environment interaction budget for all methods (R-network pre-training is included in this budget). The videos of all trained agents in all environments are available online222<https://sites.google.com/view/episodic-curiosity>.
For additional experiments we refer the reader to the supplementary material: there we show that R-network can successfully generalize between environments, demonstrate stability of our method to hyperparameters and present an ablation study.
###
4.1 Static Maze Goal Reaching.
| | | |
| --- | --- | --- |
| Dense | Sparse | Very Sparse |
| Task reward as a function of training step for | Task reward as a function of training step for | Task reward as a function of training step for |
Figure 6: Task reward as a function of training step for VizDoom tasks. Higher is better. We shift the curves for our method by the number of environment steps used to train R-network — so the comparison is fair. We run every method with a repeat of 3 (same as in prior work (Pathak et al., [2017](#bib.bib16))) and show all runs. No seed tuning is performed.
The goal of this experiment is to verify our re-implementation of the baseline method is correct. We use the MyWayHome task from VizDoom. The agent has to reach the goal in a static 3D maze in the time limit of 525 4-repeated steps (equivalent to 1 minute). It only gets a reward of +1 when it reaches the goal (episode ends at that moment), the rest of the time the reward is zero.
The task has three sub-tasks (following the setup in (Pathak et al., [2017](#bib.bib16))): “Dense”, “Sparse” and “Very Sparse”. The layout of the maze is demonstrated in Figure [5](#S4.F5 "Figure 5 ‣ 4 Experiments ‣ Episodic Curiosity through Reachability")(c). The goal is always at the same room but the starting points are different in those sub-tasks. For the “Dense” subtask, the agent starts in one of the random locations in the maze, some of which are close to the goal. In this sub-task, the reward is relatively dense (hence the name): the agent is likely to bump into the goal by a short random walk. Thus, this is an easy task even for standard RL methods. The other two sub-tasks are harder: the agent starts in a medium-distant room from the goal (“Sparse”) or in a very distant room (“Very Sparse”). Those tasks are hard for standard RL algorithms because the probability of bumping into a rewarding state by a random walk is very low.
The training curves are shown in Figure [6](#S4.F6 "Figure 6 ‣ 4.1 Static Maze Goal Reaching. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability"). By analysing them, we draw a few conclusions. First, our re-implementation of the ICM baseline is correct and the results are in line with those published in (Pathak et al., [2017](#bib.bib16)). Second, our method works on-par with the ICM baseline in terms of final performance, quickly reaching 100% success rate in all three sub-tasks. Finally, in terms of convergence speed our algorithm is significantly faster than the state-of-the-art method ICM — our method reaches 100% success rate at least 2 times faster. Note that to make the comparison of the training speed fair, we shift our training curves by the environment interaction budget used for training R-network.
###
4.2 Procedurally Generated Random Maze Goal Reaching.
In this experiment we aim to evaluate maze goal reaching task generalization on a large scale. We train on hundreds of levels and then test also on hundreds of hold-out levels. We use “Explore Goal Locations Large” (we will denote it “Sparse”) and “Explore Obstructed Goals Large” (we will denote it “Sparse + Doors”) levels in the DMLab simulator. In those levels, the agent starts in a random location in a randomly generated maze (both layout and textures are randomized at the beginning of the episode). Within the time limit of 1800 4-repeated steps (equivalent to 2 minutes), the agent has to reach the goal as many times as possible. Every time it reaches a goal, it is re-spawned into another random location in the maze and has to go to the goal again. Every time the goal is reached, the agent gets a reward +10, the rest of the time the reward is zero. The second level is a variation of the first one with doors which make the paths in the maze longer. The layouts of the levels are demonstrated in Figure [5](#S4.F5 "Figure 5 ‣ 4 Experiments ‣ Episodic Curiosity through Reachability")(b,c).
We found out that the standard task “Sparse” is actually relatively easy even for the plain PPO algorithm. The reason is that the agent starting point and the goal are sampled on the map independently of each other — and sometimes both happen to be in the same room which simplifies the task. To test the limits of the algorithms, we create a gap between the starting point and the goal which eliminates same-room initialization. We report the results for both the original task “Sparse” and its harder version “Very Sparse”. Thus, there are overall three tasks considered in this section: “Sparse”, “Very Sparse” and “Sparse + Doors”.
The results demonstrate that our method can reasonably adapt to ever-changing layouts and textures — see Table [1](#S4.T1 "Table 1 ‣ 4.4 Dense Reward Tasks. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability") and training curves in Figure [7](#S4.F7 "Figure 7 ‣ 4.3 No Reward/Area Coverage. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability"). We outperform the baseline method ICM in all three environments using the same environment interaction budget of 20M 4-repeated steps. The environment “Sparse” is relatively easy and all methods work reasonably. In the “Very Sparse” and “Sparse + Doors” settings our advantage with respect to PPO and ICM is more clear. On those levels, the visual inspection of the ICM learnt behaviour reveals an important property of this method: it is confused by the firing action and learns to entertain itself by firing until it runs out of ammunition. A similar finding was reported in a concurrent work (Burda et al., [2018](#bib.bib7)): the agent was given an action which switched the content on a TV screen in a maze, along with the movement actions. Instead of moving, the agent learns to switch channels forever. While one might intuitively accept such “couch-potato” behaviour in intelligent creatures, it does not need to be a consequence of curious behaviour. In particular, we are not observing such dramatic firing behaviour for our curiosity formulation: according to Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Episodic Curiosity through Reachability"), an observation after firing is still one step away from the one before firing, so it is not novel (note that firing still could happen in practice because of the entropy term in PPO). Thus, our formulation turns out to be more robust than ICM’s prediction error in this scenario. Note that we do not specifically look for an action set which breaks the baseline — just use the standard one for DMLab, in line with the prior work (e.g., (Espeholt et al., [2018](#bib.bib8))).
The result of this experiment suggests to look more into how methods behave in extremely-sparse reward scenarios. The limiting case would be no reward at all — we consider it in the next section.
###
4.3 No Reward/Area Coverage.
| | | |
| --- | --- | --- |
| Sparse | Very Sparse | Sparse + Doors |
| Reward as a function of training step for | Reward as a function of training step for | Reward as a function of training step for |
| No Reward | No Reward - Fire | Dense 1 |
| Reward as a function of training step for | Reward as a function of training step for | Reward as a function of training step for |
Figure 7: Reward as a function of training step for DMLab tasks. Higher is better. We shift the curves for our method by the number of environment steps used to train R-network — so the comparison is fair. We run every method 10 times and show all runs. No seed tuning is performed.
This experiment aims to quantitatively establish how good our method is in the scenario when no task reward is given. One might question why this scenario is interesting — however, before the task reward is found for the first time, the agent lives in the no-reward world. How it behaves in this case will also determine how likely it is to stumble into the task reward in the first place.
We use one of the DMLab levels — “Sparse” from the previous experiment. We modify the task to eliminate the reward and name the new task “No Reward”. To quantify success in this task, we report the reward coming from Grid Oracle for all compared methods. This reward provides a discrete approximation to the area covered by the agent while exploring.
The training curves are shown in Figure [7](#S4.F7 "Figure 7 ‣ 4.3 No Reward/Area Coverage. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability") and the final test results in Table [1](#S4.T1 "Table 1 ‣ 4.4 Dense Reward Tasks. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability"). The result of this experiment is that our method and Grid Oracle both work, while the ICM baseline is not working — and the qualitative difference in behaviour is bigger than in the previous experiments. As can be seen from the training curves, after a temporary increase, ICM quality actually decreases over time, rendering a sharp disagreement between the prediction-error-based bonus and the area coverage metric. By looking at the video††footnotemark: , we observe that the firing behaviour of ICM becomes even more prominent, while our method still shows reasonable exploration.
Finally, we try to find out if the ICM baseline behaviour above is due to the firing action only. Could it learn exploration of randomized mazes if the Fire action is excluded from the actions set? For that purpose, we create a new version of the task — we call it “No Reward - Fire”. This task demonstrates qualitatively similar results to the one with the full action set — see Table [1](#S4.T1 "Table 1 ‣ 4.4 Dense Reward Tasks. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability"). By looking at the videos††footnotemark: , we hypothesise that the agent can most significantly change its current view when it is close to the wall — thus increasing one-step prediction error — so it tends to get stuck near “interesting” diverse textures on the walls.
The results suggest that in an environment completely without reward, the ICM method will exhaust its curiosity very quickly — passing through a sharp peak and then degrading into undesired behaviour. This observation raises concerns: what if ICM passes the peak before it reaches the first task reward in the cases of real tasks? Supposedly, it would require careful tuning per-game. Furthermore, in some cases, it would take a lot of time with a good exploration behaviour to reach the first reward, which would require to stay at the top performance for longer — which is problematic for the ICM method but still possible for our method.
###
4.4 Dense Reward Tasks.
| Method | Sparse | Very Sparse | Sparse+Doors | No Reward | No Reward - Fire | Dense 1 | Dense 2 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| PPO | 27.0 ± 5.1 | 8.6 ± 4.3 | 1.5 ± 0.1 | 191 ± 12 | 217 ± 19 | 22.8 ± 0.5 | 9.41 ± 0.02 |
| PPO + ICM | 23.8 ± 2.8 | 11.2 ± 3.9 | 2.7 ± 0.2 | 72 ± 2 | 87 ± 3 | 20.9 ± 0.6 | 9.39 ± 0.02 |
| PPO + EC (ours) | 26.2 ± 1.9 | 24.7 ± 2.2 | 8.5 ± 0.6 | 475 ± 8 | 492 ± 10 | 19.9 ± 0.7 | 9.53 ± 0.03 |
| PPO + Grid Oracle | 56.7 ± 1.3 | 54.3 ± 1.2 | 29.4 ± 0.5 | 796 ± 2 | 795 ± 3 | 20.9 ± 0.6 | 8.97 ± 0.04 |
Table 1: Reward in DMLab tasks (mean ± std) for all compared methods. Higher is better. We report Grid Oracle reward in tasks with no reward. The results for our method are reported from less training steps to account for R-network pre-training. The Grid Oracle method is given for reference — it uses privileged information unavailable to other methods. No seed tuning is performed.
A desirable property of a good curiosity bonus is to avoid hurting performance in dense-reward tasks (in addition to improving performance for sparse-reward tasks). We test this scenario in two levels in the DMLab simulator: “Rooms Keys Doors Puzzle” (which we denote “Dense 1”) and “Rooms Collect Good Objects Train” (which we denote “Dense 2”). In the first task, the agent has to collect keys and reach the goal object behind a few doors openable by those keys. The rewards in this task are rather dense (key collection/door opening is rewarded). In the second task the agent has to collect good objects (give positive reward) and avoid bad objects (give negative reward). The episode lasts for 900 4-repeated steps (equivalent to 1 minute) in both tasks.
The results show that our method indeed does not significantly deteriorate performance of plain PPO in those dense-reward tasks — see Table [1](#S4.T1 "Table 1 ‣ 4.4 Dense Reward Tasks. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability"). The training curves for “Dense 1” are shown in Figure [7](#S4.F7 "Figure 7 ‣ 4.3 No Reward/Area Coverage. ‣ 4 Experiments ‣ Episodic Curiosity through Reachability") and for “Dense 2” — in the supplementary material. Note that we use the same bonus weight in this task as in other DMLab tasks before. All methods work similarly besides the Grid Oracle in the “Dense 2” task — which performs slightly worse. Video inspection††footnotemark: reveals that Grid Oracle — the only method which has ground-truth knowledge about area it covers during training — sometimes runs around excessively and occasionally fails to collect all good objects.
5 Discussion
-------------
Our method is at the intersection of multiple topics: curiosity, episodic memory and temporal distance prediction. In the following, we discuss the relation to the prior work on those topics.
Curiosity in visually rich 3D environments.
Recently, a few works demonstrated the possibility to learn exploration behaviour in visually rich 3D environments like DMLab (Beattie et al., [2016](#bib.bib3)) and VizDoom (Kempka et al., [2016](#bib.bib12)). (Pathak et al., [2017](#bib.bib16)) trains a predictor for the embedding of the next observation and if the reality is significantly different from the prediction — rewards the agent. In that work, the embedding is trained with the purpose to be a good embedding for predicting action taken between observations — unlike an earlier work (Stadie et al., [2015](#bib.bib22)) which obtains an embedding from an autoencoder. It was later shown by (Burda et al., [2018](#bib.bib7)) that the perceptive prediction approach has a downside — the agent could become a “couch-potato” if given an action to switch TV channels. This observation is confirmed in our experiments by observing a persistent firing behaviour of the ICM baseline in the navigational tasks with very sparse or no reward. By contrast, our method does not show this behaviour. Another work (Fu et al., [2017](#bib.bib9)) trains a temporal distance predictor and then uses this predictor to establish novelty: if the observation is easy to classify versus previous observations, it is novel. This method does not use episodic memory, however, and the predictor is used in way which is different from our work.
General curiosity.
Curiosity-based exploration for RL has been extensively studied in the literature. For an overview, we refer the reader to the works (Oudeyer & Kaplan, [2009](#bib.bib14); Oudeyer et al., [2007](#bib.bib15)). The most common practical approaches could be divided into three branches: prediction-error-based, count-based and goal-generation-based. Since the prediction-based approaches were discussed before, in the following we focus on the latter two branches.
The count-based approach suggests to keep visit counts for observations and concentrate on visiting states which has been rarely visited before — which bears distant similarity to how we use episodic memory. This idea is natural for discrete observation spaces and has solid theoretical foundations. Its extension to continuous observation spaces is non-trivial, however. The notable step in this direction was taken by works (Bellemare et al., [2016](#bib.bib4); Ostrovski et al., [2017](#bib.bib13)) which introduce a trained observation density model which is later converted to a function behaving similarly to counts. The way conversion is done has some similarity to prediction-error-based approaches: it is the difference of the density in the example before and after training of this example which is converted to count. The experiments in the original works operate on Atari games (Bellemare et al., [2013](#bib.bib5)) and were not benchmarked on visually rich 3D environments. Another approach (Tang et al., [2017](#bib.bib23)) discretises the continuous observation space by hashing and then uses the count-based approach in this discretised space. This method is appealing in its simplicity, however, the experiments in (Pathak et al., [2017](#bib.bib16); Fu et al., [2017](#bib.bib9)) show that it does not perform well in visually rich 3D environments.
Finally, our concept of novelty through reachability is reminiscent of generating the goals which are reachable but not too easy — a well-studied topic in the prior work. The work (Held et al., [2017](#bib.bib10)) uses a GAN to differentiate what is easy to reach from what is not and then generate goals at the boundary. Another work (Baranes & Oudeyer, [2013](#bib.bib2)) defines new goals according to the expected progress the agent will make if it learns to solve the associated task. The recent work (Péré et al., [2018](#bib.bib17)) learns an embedding for the goal space and then samples increasingly difficult goals from that space. In a spirit similar to those works, our method implicitly defines goals that are at least some fixed number of steps away by using the reachability network. However, our method is easier to implement than other goal-generation methods and quite general.
Episodic memory.
Two recent works (Blundell et al., [2016](#bib.bib6); Pritzel et al., [2017](#bib.bib18)) were inspired by the ideas of episodic memory in animals and proposed an approach to learn the functioning of episodic memory along with the task for which this memory is applied. Those works are more focused on repeating successful strategies than on exploring environments — and are not designed to work in the absence of task rewards.
Temporal distance prediction.
The idea to predict the distance between video frames has been studied extensively. Usually this prediction is an auxiliary task for solving another problem. (Sermanet et al., [2017](#bib.bib21)) trains an embedding such that closer in time frames are also closer in the embedding space. Multiple works (Fu et al., [2017](#bib.bib9); Savinov et al., [2018](#bib.bib19); Aytar et al., [2018](#bib.bib1)) train a binary classifier for predicting if the distance in time between frames is within a certain threshold or not. While (Sermanet et al., [2017](#bib.bib21); Aytar et al., [2018](#bib.bib1)) use only the embedding for their algorithms, (Fu et al., [2017](#bib.bib9); Savinov et al., [2018](#bib.bib19)) also use the classifier trained together with the embedding. As mentioned earlier, (Fu et al., [2017](#bib.bib9)) uses this classifier for density estimation instead of comparison to episodic memory. (Savinov et al., [2018](#bib.bib19)) does compare to the episodic memory buffer but solves a different task — given an already provided exploration video, navigate to a goal — which is complementary to the task in our work.
6 Conclusion
-------------
In this work we propose a new model of curiosity based on episodic memory and the ideas of reachability. This allows us to overcome the known “couch-potato” issues of prior work and outperform the previous curiosity state-of-the-art method ICM in visually rich 3D environments. In the future, we want to make policy aware of memory not only in terms of receiving reward, but also in terms of acting. Can we use memory content retrieved based on reachability to guide exploration behaviour in the test time? This could open opportunities to learn exploration in new tasks in a few-shot style — which is currently a big scientific challenge.
#### Acknowledgments
We would like to thank Olivier Pietquin, Carlos Riquelme, Charles Blundell and Sergey Levine for the valuable discussions about our work.
Supplementary Material
----------------------
The supplementary material is organized as follows. First, we provide training details for R-network. Second, we list hyperparameter values and the details of hyperparameter search for all methods. Then we show experimental results which suggest that R-network can generalize between environments: we transfer one general R-network from all available DMLab30 levels to our tasks of interest and also transfer R-networks between single environments. After that, we present the results from a stability/ablation study which suggests our method is stable with respect to its most important hyperparameters and the components we used in the method are actually necessary for its performance (and measure their influence). Finally, we provide the training curves for the “Dense 2” task in the main text.
S1 Reachability Network Training Details
-----------------------------------------
For training R-network, we use mini-batches of 64 observation pairs (matched within episodes). The training is run for 50K mini-batch iterations for VizDoom and 200K mini-batch iterations for DMLab. At the beginning of every pass through the buffer, we re-shuffle it. We use Adam optimizer with learning rate 10−4. The R-network uses a siamese architecture with two branches (see Figure [2](#S2.F2 "Figure 2 ‣ 2 Episodic Curiosity ‣ Episodic Curiosity through Reachability") in the main text), each branch is Resnet-18 with 512 outputs, with a fully-connected network applied to the concatenated output of the branches. The fully-connected network has four hidden layers with 512 units, batch-normalization and ReLU is applied after each layer besides the last one, which is a softmax layer. Observations are RGB-images with resolution 160×120 pixels.
S2 Hyperparameters
-------------------
The hyperparameters of different methods are given in Table [S1](#S2.T1 "Table S1 ‣ S2 Hyperparameters ‣ Episodic Curiosity through Reachability") for VizDoom environment and in Table [S2](#S2.T2 "Table S2 ‣ S2 Hyperparameters ‣ Episodic Curiosity through Reachability") for DMLab environment. The hyperparameters for DMLab are tuned on the ‘‘Sparse’’ environment for all methods --- because all methods work reasonably on this environment (it is unfair to tune a method on an environment where it fails and also unfair to tune different methods on different environments). We use the PPO algorithm from the open-source implementation333 <https://github.com/openai/baselines>. For implementation convenience, we scale both the bonus and the task reward (with a single balancing coefficient it would not be possible to turn off one of those rewards).
| | | | | |
| --- | --- | --- | --- | --- |
| | Plain PPO | ICM | Grid Oracle | EC |
| learning rate | 0.00025 | 0.00025 | 0.00025 | 0.00025 |
| PPO entropy coefficient | 0.01 | 0.01 | 0.01 | 0.01 |
| Task reward scale | 5.0 | 5.0 | 5.0 | 5.0 |
| Curiosity bonus scale | 0.0 | 0.01 | 1.0 | 1.0 |
| ICM forward inverse ratio | - | 0.2 | - | - |
| ICM curiosity loss strength | - | 10 | - | - |
| EC memory size | - | - | - | 200 |
Table S1: Hyper-parameters used for VizDoom environment.
| | Plain PPO | ICM | Grid Oracle | EC |
| --- | --- | --- | --- | --- |
| learning rate | 0.00019 | 0.00025 | 0.00025 | 0.00025 |
| PPO entropy coefficient | 0.0011 | 0.0042 | 0.0066 | 0.0021 |
| Task reward scale | 1.0 | 1.0 | 1.0 | 1.0 |
| Curiosity bonus scale | 0.0 | 0.55 | 0.052 | 0.030 |
| ICM forward inverse ratio | - | 0.96 | - | - |
| ICM curiosity loss strength | - | 64 | - | - |
| EC memory size | - | - | - | 200 |
Table S2: Hyper-parameters used for DMLab environment.
S3 R-network generalization study
----------------------------------
One of the promises of our approach is its potential ability to generalize between tasks. In this section we verify if this promise holds.
###
s3.1 Training R-network on all DMLab-30 tasks
Could we train a universal R-network for all available levels — and then use this network for all our tasks of interest? Since different games have different dynamics models, the notion of closely reachable or far observations also changes from game to game. Can R-network successfully handle this variability?
Table [S3](#S3.T3 "Table S3 ‣ S3.1 Training R-network on all DMLab-30 tasks ‣ S3 R-network generalization study ‣ Episodic Curiosity through Reachability") suggests that using a universal R-network slightly hurts the performance compared to using a specialized R-network trained specifically for the task. However, it still definitely helps to get higher reward compared to using the plain PPO. The R-network is trained using 10M environment interactions equally split across all 30 DMLab-30 tasks.
| Method | No Reward | Very Sparse |
| --- | --- | --- |
| PPO | 191 ± 12 | 8.6 ± 4.3 |
| PPO + EC with specialized R-network | 475 ± 8 | 24.7 ± 2.2 |
| PPO + EC with universal R-network | 348 ± 8 | 19.3 ± 1.0 |
Table S3: Reward on the tasks “No Reward” and “Very Sparse” using a universal R-network. Two baselines (PPO and PPO + EC with a specialized R-network) are also provided.
###
s3.2 Training R-network on One Level and Testing on Another
This experiment is similar to the previous one but in a sense is more extreme. Instead of training on all levels (including the levels of interest and other unrelated levels), can we train R-network on just one task and use if for a different task?
Table [S4](#S3.T4 "Table S4 ‣ S3.2 Training R-network on One Level and Testing on Another ‣ S3 R-network generalization study ‣ Episodic Curiosity through Reachability") suggests we can obtain reasonable performance by transferring the R-network between similar enough environments. The performance is unsatisfactory only in one case (using the R-network trained on “Dense 2”). Our hypothesis is that the characteristics of the environments are sufficiently different in that case: single room versus maze, static textures on the walls versus changing textures.
| R-network training environment | No Reward | Very Sparse |
| --- | --- | --- |
| Dense 1 | 320 ± 5 | 18.5 ± 1.4 |
| Dense 2 | 43 ± 2 | 0.8 ± 0.5 |
| Sparse + Doors | 376 ± 7 | 16.2 ± 0.7 |
| Matching environment | 475 ± 8 | 24.7 ± 2.2 |
Table S4: Reward on the environments “No Reward” and “Very Sparse” (columns) when the R-network is trained on different environments (rows). We provide a result with a matching R-network for reference (bottom).
S4 Stability/ablation Study
----------------------------
The experiments are done both in “No Reward” and “Very Sparse” environments. The “No Reward” environment is useful to avoid the situations where task reward would hide important behavioural differences between different flavors of our method (this “hiding” effect can be easily observed for different methods comparison in the dense reward tasks — but the influence of task reward still remains even in sparser cases). As in the main text, for the “No Reward” task we report the Grid Oracle reward as a discrete approximation to the area covered by the agent trajectories.
###
s4.1 Positive Example Threshold in R-network Training
Training the R-network requires a threshold k to separate negative from positive pairs. The trained policy implicitly depends on this threshold. Ideally, the policy
performance should not be too sensitive to this hyper-parameter.
We conduct a study where the threshold is varied from 2 to 10 actions (as in all experiments before, each action is repeated 4 times). Table [S5](#S4.T5 "Table S5 ‣ S4.1 Positive Example Threshold in R-network Training ‣ S4 Stability/ablation Study ‣ Episodic Curiosity through Reachability") shows that the EC performance is reasonably robust to the choice of this threshold.
| Threshold k | No Reward | Very Sparse |
| --- | --- | --- |
| 2 | 378 ± 18 | 28.3 ± 1.6 |
| 3 | 395 ± 10 | 20.9 ± 1.6 |
| 4 | 412 ± 8 | 31.1 ± 1.2 |
| 5 | 475 ± 8 | 24.7 ± 2.2 |
| 7 | 451 ± 4 | 23.6 ± 1.0 |
| 10 | 455 ± 7 | 20.8 ± 0.8 |
Table S5: Reward in the “No Reward” and “Very Sparse“ tasks using different positive example thresholds k when training the R-network.
###
s4.2 Memory Size in EC module
The EC-module relies on an explicit memory buffer to store the embeddings of past observations and define novelty.
One legitimate question is to study the impact of the size of this memory buffer on the performance of the EC-module.
As observed in table [S6](#S4.T6 "Table S6 ‣ S4.2 Memory Size in EC module ‣ S4 Stability/ablation Study ‣ Episodic Curiosity through Reachability"), the memory size has little impact on the performance.
| Memory size | No Reward | Very Sparse |
| --- | --- | --- |
| 100 | 447 ± 6 | 19.4 ± 1.9 |
| 200 | 475 ± 8 | 24.7 ± 2.2 |
| 350 | 459 ± 6 | 23.5 ± 1.4 |
| 500 | 452 ± 6 | 23.8 ± 2.0 |
Table S6: Reward for different values of the memory size for the tasks “No Reward” and “Very Sparse”.
###
s4.3 Environment Interaction Budget for Training R-network
The sample complexity of our EC method includes two parts: the sample complexity to train the R-network and the sample complexity of the policy training. In the worst case – when the R-network does not generalize across environments – the R-network has to be trained for each environment and the total sample complexity is then the sum of the previous two sample complexities. It is then crucial to see how many steps are needed to train R-network such that it can capture the notion of reachability.
R-network trained using a number of environment steps as low as 1M already gives good performance, see Table [S7](#S4.T7 "Table S7 ‣ S4.3 Environment Interaction Budget for Training R-network ‣ S4 Stability/ablation Study ‣ Episodic Curiosity through Reachability").
| Interactions | No Reward | Very Sparse |
| --- | --- | --- |
| 100K | 357 ± 18 | 12.2 ± 1.3 |
| 300K | 335 ± 9 | 16.2 ± 0.7 |
| 1M | 383 ± 13 | 18.6 ± 0.9 |
| 2.5M | 475 ± 8 | 24.7 ± 2.2 |
| 5M | 416 ± 5 | 20.7 ± 1.4 |
Table S7: Reward of the policy trained on the “No Reward” and “Very Sparse“ tasks with an R-network trained using a varying number of environment interactions (from 100K to 5M).
###
s4.4 Importance of Training Different Parts of R-network
The R-network is composed of an Embedding network and a Comparator network. How important is each for the final performance of our method? To establish that, we conduct two experiments. First, we fix the Embedding network at the random initialization and train only the Comparator. Second, we substitute the Comparator network applied to embeddings e1,e2 with the sigmoid function σ(eT1e2) and train only the Embedding. According to the results in Table [S8](#S4.T8 "Table S8 ‣ S4.4 Importance of Training Different Parts of R-network ‣ S4 Stability/ablation Study ‣ Episodic Curiosity through Reachability"), we get a reasonable performance with a random embedding: the results are still better than the plain PPO (but worse than with the complete R-network). However, without the Comparator the quality drops below the plain PPO.
This experiment leads us to two conclusions. First, training the Embedding network is desired but not necessary for our method to work. Second, using the Comparator is essential and cannot be omitted in the current setup. Apparently, predicting reachability requires fine-grained access to both embeddings at the same time — and a simple comparison function does not work.
| Method | No Reward | Very Sparse |
| --- | --- | --- |
| PPO | 191 ± 12 | 8.6 ± 4.3 |
| PPO + EC with complete R-network | 475 ± 8 | 24.7 ± 2.2 |
| PPO + EC with random Embedding | 392 ± 12 | 16.2 ± 1.4 |
| PPO + EC without Comparator network | 48 ± 3 | 5.8 ± 2.4 |
Table S8: Reward on the “No Reward” and “Very Sparse“ tasks using ablated versions of the R-network.
S5 Additional DMLab Training Curves
------------------------------------
Figure S1: Reward as a function of training step for the DMLab task “Dense 2”. Higher is better. We shift the curves for our method by the number of environment steps used to train R-network — so the comparison between different methods is fair. We run every method 10 times and show all runs. No seed tuning is performed.
We show additional training curves from the main text experimental section in Figure [S1](#S5.F1 "Figure S1 ‣ S5 Additional DMLab Training Curves ‣ Episodic Curiosity through Reachability"). |
ed501e7f-c660-4f09-97ec-ca7a25c9e1de | trentmkelly/LessWrong-43k | LessWrong | Lone Genius Bias and Returns on Additional Researchers
One thing that most puzzles me about Eliezer's writings on AI is his apparent belief that a small organization like MIRI is likely to be able to beat larger organizations like Google or the US Department of Defense to building human-level AI. In fact, he seems to believe such larger organizations may have no advantage at all over a smaller one, and perhaps will even be at a disadvantage. In his 2011 debate with Robin Hanson, he said:
> As far as I can tell what happens when the government tries to develop AI is nothing. But that could just be an artifact of our local technological level and it might change over the next few decades. To me it seems like a deeply confusing issue whose answer is probably not very complicated in an absolute sense. Like we know why it’s difficult to build a star. You’ve got to gather a very large amount of interstellar hydrogen in one place. So we understand what sort of labor goes into a star and we know why a star is difficult to build. When it comes to building a mind, we don’t know how to do it so it seems very hard. We like query our brains to say “map us a strategy to build this thing” and it returns null so it feels like it’s a very difficult problem. But in point of fact we don’t actually know that the problem is difficult apart from being confusing. We understand the star-building problem so we know it’s difficult. This one we don’t know how difficult it’s going to be after it’s no longer confusing.
>
> So to me the AI problem looks like a—it looks to me more like the sort of thing that the problem is finding bright enough researchers, bringing them together, letting them work on that problem instead of demanding that they work on something where they’re going to produce a progress report in two years which will validate the person who approved the grant and advance their career. And so the government has historically been tremendously bad at producing basic research progress in AI, in part because the most senior people in AI |
9974b51f-616a-48c3-ac7f-40edbcc3e08d | StampyAI/alignment-research-dataset/special_docs | Other | Active reward learning
Active Reward Learning
Christian Daniel, Malte Viering, Jan Metz, Oliver Kroemer, Jan Petersy
Institut f ¨ur Intelligente Autonome Systeme, Technische Universit ¨at Darmstadt, 64289, Germany
yMax-Planck-Institut f ¨ur Intelligente Systeme, T ¨ubingen, 72076, Germany
Email:fdaniel, kroemer, peters g@ias.tu-darmstadt.de
Abstract —While reward functions are an essential component
of many robot learning methods, defining such functions remains
a hard problem in many practical applications. For tasks such
as grasping, there are no reliable success measures available.
Defining reward functions by hand requires extensive task knowl-
edge and often leads to undesired emergent behavior. Instead,
we propose to learn the reward function through active learning,
querying human expert knowledge for a subset of the agent’s
rollouts. We introduce a framework, wherein a traditional learn-
ing algorithm interplays with the reward learning component,
such that the evolution of the action learner guides the queries
of the reward learner. We demonstrate results of our method on
a robot grasping task and show that the learned reward function
generalizes to a similar task.
I. I NTRODUCTION
An important goal of Reinforcement Learning (RL) is to
yield more autonomous robots. However, RL methods require
reward functions to guide the agent towards a desired be-
havior. Such reward functions are usually hand coded and,
unfortunately, defining rewards for real robot tasks manually is
challenging even for relatively well-understood problems such
as grasping. Thus, hand coding reward functions is shifting
the problem of requiring an expert to design a hard coded
controller to requiring a hard coded reward function.
Despite the variety of grasp stability measures that has been
developed [34], it has been shown that the resulting grasps
are outperformed by grasps learned from kinesthetic teach-
in [3, 23]. This example shows that even experts will often
design reward functions that are not effective in practice, or
lead to undesired emergent behavior [27].
To avoid specifying reward functions, Inverse RL (IRL)
extracts a reward function from demonstrations [30, 31, 36].
For many tasks, such demonstrations can be obtained through
methods such as kinesthetic teach-in or tele-operation. Unfor-
tunately some skills, such as dynamic skills, are often hard to
demonstrate (for example, teaching a robot to throw a basket
ball). Moreover, both methods require sufficient proficiency of
the demonstrator which may lead to the demonstrator having
to truly become an expert in performing the task first.
While it may be difficult to analytically design a reward
function or to give demonstrations, it is often easy for an expert
to rate an agent’s executions of a task. Thus, a promising
alternative is to use the human not as an expert in performing
the task, but as an expert in evaluating task executions.
Based on this insight, preference based algorithms allow the
expert to rank executions and learn a controller based on these
rankings [2, 6]. The generally used approach in ranking is to let
Fig. 1: The Robot-Grasping Task. While grasping is one of the most
researched robotic tasks, finding a good reward function still proves
difficult.
the expert rank the previously best sample against the current
sample, which is an intriguing idea, as humans are better at
giving relative judgments than absolute judgments. However,
this approach only provides a single bit of information. The
technical term describing how much information human sub-
jects can transmit for a given input stimuli is called channel
capacity [25]. The channel capacity is a measure for how many
input stimuli subjects can distinguish between. In a review of
several experiments, Miller [25] concludes that humans have
a general channel capacity somewhere between 1:6and3:9
bits for unidimensional stimuli. Adding more dimensions to
the stimuli further increases this value.
Experiments have also been performed to find out whether
humans can transmit more information when labelling stimuli
according to predescribed categories or when being allowed to
rate on a scale [15]. While there was no significant difference,
subjects performed better when rating on a scale.
Based on these insights, we propose an alternative frame-
work in which the human expert can assign numerical values
to observed demonstrations. Furthermore, numerical rewards
allow to indicate strong preferences over demonstrations.
Unfortunately, as already shown by Thomaz and Breazeal [35]
and by Cakmak and Thomaz [5], humans have considerable
noise in their ratings of actions. To deal with this mismatch
between good reward functions for policy search (PS) methods
and noisy human rating systems, we propose to learn a prob-
abilistic model of the reward function. In this paper, we take
advantage of the Gaussian Process (GP) regression framework
as well as the Bayesian Optimization (BO) literature. Having
access to BO methods, allows us to efficiently minimize
Rewards R = f(o)Outcomes o = 𝜙(𝞃)Rewards R ∼ p(R|o,t)Trajectories 𝞃 ∼ p(𝞃 |s, 𝞈)Context sControl Parameters 𝞈 ∼ 𝜋(𝞈|s)Query occasionallyExpert Ratings t ∼ p(t|𝞃)Vanilla RLActive Reward LearningBayesian OptmiziationControl Parameters 𝞈=argmax u(R|s,𝞈,t)Query alwaysNo Reward LearningSemi Supervised Reward LearningSupervised Reward LearningFig. 2: Sketch of the elements of different possible approaches. The left column shows the ‘vanilla’ RL where a reward function is assumed.
The middle column shows the proposed active reward learning approach, which shares the policy learning component with vanilla RL but
models a probabilistic reward model that gets updated by asking for expert ratings sometimes. The right column shows the BO approach,
where actions are chosen that maximize the utility function (for example one of the acquisition functions presented in II-A1 and requires an
expert rating for each of the chosen actions.
the number of expert interactions, which is essential when
designing methods for ‘autonomous’ agents. We evaluate the
proposed method on a series of simulated and one real robot
task and we also show that a reward function which is learned
for one task (grasping a box) can be used to learn a new task
(grasping a pestle).
II. M ETHOD
While there are many approaches to RL, Policy Search (PS)
is especially well suited for real robot tasks. It does not rely
on exhaustive sampling of the state-action space and, as a
result, many recent advances in robot learning rely on policy
search [22, 20, 26]. We consider the contextual episodic PS
case with continuous contexts s2S and continuous control
parameters!2
. We use the term ‘context’ instead of the
more general term ‘state’ to indicate the initial state of the
robot and the environment. The distribution over contexts
is denoted by ()Whereas the traditional RL notation
uses actionsa, we instead write control parameters !to
clarify that we directly optimize for the parameters of a
controller which can, for example, be a Movement Primitive
as discussed in Section III-B. Executing a rollout with control
parameters!results in a trajectory p( js;!), where the
trajectory encodes both, the robot’s state transitions as well
as relevant environment state transitions. The agent starts with
an initial control policy 0(!js)in iteration 0and performs
a predetermined number of rollouts. After each iteration, the
agent updates its policy to maximize the expected reward
Es;!; [R( )] =ZZZ
R( )p(s;!; ) d dsd!;(1)
withp(s;!; ) =p( js;!)(!js)(s):Parameter-based
PS has been shown to be well suited for complex robot
tasks [9], as it abstracts complexity while retaining sufficient
flexibility in combination with suitable movement primitive
representations.
In the original RL setting a reward function is assumed to be
known that evaluates the agents behavior and assigns rewards
to the agent’s actions. While not always explicitly stated, the
reward function for real robot tasks often depends not on the
whole trajectory of the robot, but on features ( )of the
trajectory. We refer to the result of the feature extraction fromtrajectories as the outcomeso= ( ). We assume that the
reward depends only on the features of the trajectories. Such
outcomes could, for example, describe the minimal distance
to goal positions or the accumulative energy consumption.
As the problem of specifying good features ( )is beyond
the scope of this paper, we assume the features to be known,
as is usually assumed when designing reward functions.
As outcomes are of much lower dimensionality than trajec-
tories, we can use outcomes to efficiently model the reward
function ^R(o)from a training set D=fo1:n;R1:ngusing
regression. Using the feature representation, the term R( )
in Eq. 1 is replaced by R(o= ( ))to become
Es;!; [R(o)]=ZZZ
R(o= ( ))p(s;!; ) d dsd!:(2)
Obviously, it is not sufficient to build the reward function
model on samples from the initial policy 0(!js), as the
agent is likely to exhibit poor performance in the early stages
of learning and the reward learner would not observe good
samples. Therefore, we need to couple the process of learning
a good policy (!js)and a good reward function ^R(o), such
that they are developed interdependently. However, in such
a coupled active learning process, we need to know which
samples to query from our expert, as we want to minimize
expert interaction.
Modelling a probabilistic reward model p(Rjo;D), instead
of a deterministic reward function R(o), allows us to leverage
the information about the certainty of our estimate to control
the amount of human interaction required, as we only need to
interact with the human expert if our model of the reward is
not certain enough. We describe the details of this process in
Section II-A.
When using a probabilistic model of the reward, we have
to replace the reward term in Eq. 1 with
R(o) =Ep(Rjo)[R] =Z
p(Rjo)RdR;
as most PS methods work on the expectation of the reward.
Using the probabilistic model of the reward function the
PS method continuously learns how to achieve high reward
outcomes, while the reward model learner adapts the accuracy
of its model.
Input: Information loss tolerance , improvement threshold ,
acquisition function u
Initializeusing a single Gaussian with random mean. GP with
zero mean prior.
while not converged
Set sample policy:
q(!js) =old(!js)
Sample: collect samples from the sample policy
fsip(s);!iq(!jsi);oig;i2f1;:::;Ng
Define Bellman Error Function
(s;!) =R(o) V(s)
Minimize the dual function
[ ;] = arg min[ ;]g( ;)
Determine base line
V(s) = T (s)
Policy update:
Calculate weighting
p(si;!i)_q(si;!i) exp
1
(si;!i)
Estimate distribution (!js)by
weighted maximum likelihood estimates
Reward model update:
FindNominee = true
while FindNominee
Nominate outcome:
o+= arg maxu(o)
if
o+=2D
^
(o+)= >
Demonstrate Corresponding Trajectory +
Query Expert Reward R+
D=D[fo+;R+g
else
FindNominee = false
Update reward model p(Rjo;D)
Optimize GP-hyper parameters
Output: Policy(!js), reward model p(Rjo;D)
TABLE I: We show the algorithmic form of active reward learning
with REPS. We specify the information loss tolerance as well
as an initial sampling policy and an improvement threshold . In
each iteration, the algorithm first samples from the sampling policy,
minimizes the dual function to find values for Tandand
then computes the next policy. After each policy search iteration,
the reward function learner chooses whether to demonstrate samples
to the expert according to the acquisition function. The parameters
andare parameters of the dual function problem of REPS and
can be optimized through standard optimization algorithms [9].
A. Active Reward Learning
Our goal is to find a model p(Rjo;D)that predicts the
reward given an observed outcome and training data D, which
is obtained from an expert. When modelling the reward, we
have to take into account that the expert can only give noisy
samples of his implicit reward function and we also have
to model this observation noise. Thus, we need to solve the
regression problem
R(o) =f(o) +; N (0; );
where we assume zero mean Gaussian noise. Such a regression
problem can, for example, be solved with Gaussian Process
(GP) regression
f(o)GP (m(o);k(o;o0));
wherem(o)is the mean function and k(o;o0)is the covari-
ance function of the GP. For the remainder of this paper weuse the standard squared exponential covariance function
k(o;o0) =2
0exp
jjo o0jj2
22
1
:
The set of hyper parameters =f0;1; gis found through
optimization [29]. Given a training set D=fo1:n;R1:ng,
we can write down the covariance matrix between previously
observed rewards and outcomes
K=2
64k(o1;o1)::: k (o1;on)
.........
k(on;o1)::: k (on;on)3
75+ I:
Assuming a zero mean prior, the joint Gaussian probability of
the training samples in Dand the reward prediction R+of a
new unrated observation is given by
R1:n
R+
N
0;
K ^k
kk(o+;o+)
;
with ^k = [k(o1;o+):::k(on;o+)]Tand k =
[k(o+;o1):::k(o+;on)]. The predictive posterior reward
p(R+jo;D)of a new outcome o+is then given by a Gaussian
p(R+jo;D)N
(o+);2(o+)
;
with mean and variance
(o+) =kTK 1R1:n;
2(o+) =k(o+;o+) kTK 1k;
by conditioning the GP on the observed outcome o+. Using
the GP, we can represent both our expected reward (o+),
which is provided to the policy learner, and the variance of
the reward2(o+)which is essential to the active learning
component. The reward variance 2(o+)depends on the
distance of the outcome o+to all outcomes in the training set
Dand the observation noise , which is a hyper parameter
that we optimize for. Using the predictive variance, we can
employ one of many readily available BO methods to find the
maximum of the reward function.
1) Optimizing the Reward Model: The goal of BO is to
optimize a function under uncertainty. Acquisition functions
(AFs), are utility functions which maximize their function
value at locations of the input space which are likely to maxi-
mize the original problem. AFs usually encode an exploration-
exploitation trade-off, such that they do not only query samples
in known high value regions but also in regions that have not
been sufficiently explored before. While using GPs to model
the reward function allows us access to the BO toolbox in
general, we deviate from the standard BO approach in two
points. First, as our GP models a relationship of outcomes
to rewards instead of context-actions to rewards, we cannot
sample arbitrary outcomes ^oto improve our estimate. To do
so, we would require access to p( js;!)such that we can
request the agent to perform actions that result in trajectories
^ which yield the outcome ^o= (^ ). Second, we would
need to guarantee that the outcomes requested by the AF to
improve the reward model are physically possible. However,
this transition model is unknown and, thus, we have to revert
to the previously observed outcomes that have been generated
during the agent’s learning process so far.
Furthermore, we need to balance the improvement of our
current estimate of the reward function and the number of
queries that we request from the expert, i.e., we want to find a
trade-off between finding a good reward function and learning
the task with minimal human input.
2) Sample Efficiency: In the traditional BO framework, the
goal is to find a global maximum. Sampling of the function
can be stopped when the improvement of the function is
marginal, for example when the predicted variance around
the optimum is very low. However, in the proposed scenario
where a policy (!js)and a reward model p(Rjo)are learned
simultaneously and obtaining training samples of the reward
model is very expensive, the problem of deciding when to
improve the reward model becomes crucial.
The reward model relies on the policy (!js)to provide
outcomes in interesting, i.e., high reward regions, and the pol-
icy relies on the reward model p(Rjo)to guide its exploration
towards such regions of interest. Thus, the reward model needs
sufficient training data to approximately predict the reward
function in early stages and a higher density of training points
once the agents policy starts to converge to a solution. At the
same time, we want to minimize the number of queries to the
expert over the learning process.
In order to balance this trade-off, we propose an acquisition
algorithm which, according to a selected acquisition function
u(o), first samples the best observed sample outcome from the
history of the agent’s outcomes on+1= arg maxou(ojD).
In this sample based search, we also include all outcomes
that have been demonstrated to the expert. If the acquisition
function is maximized by an already demonstrated outcome,
we stop and do not query any samples in this iteration. Maxi-
mizing the AF by an already observed outcome is unique to the
sample based case, as in the continuous case, a point around
an observed outcome would usually have more variance and,
thus, maximize the AF. With our approach we achieve a sparse
sampling behavior that requires less expert interactions. If,
however, the outcome that maximizes uhas not yet been rated
by the expert, we need to decide whether querying the outcome
is beneficial. For example, we may have already converged to
a good estimate of the reward function, and new outcomes
improve on the mean reward solely due to the observation
noise. In this case we do not want to query the expert. Thus,
we decide whether to query an outcome by thresholding the
ratio of predictive variance and estimated observation noise
(o)=p > ; whereis a tuning parameter which allows
us to trade off the accuracy of the final solution with the
query frequency by adapting the available AFs to explicitly
take the estimated observation noise into account. While this
adaptation of the AFs introduces a new parameter to be tuned,
our experiments show that the use of this technique results in
less human interactions while maintaining high performance
and tuning of the parameter is straightforward.
If we decide to query the sample, we update the GP andsearch for the new maximum of the acquisition function,
otherwise we stop and do not update the GP further in this
iteration.
The general information flow of our method is as follows.
We start with an uninformed, i.e., zero mean GP for p(Rjo)
and we initialize the PS method with an initial policy 0(!js).
The PS learner then starts performing one iteration of rollouts,
which we also call episodes. After each iteration of rollouts,
rewards for the outcomes of the resulting trajectories o= ( )
are requested from the reward learner Rp(Rjo). The
reward learner then decides whether to ask for expert ratings
for any of the outcomes to update its model. In that case,
the agent repeats the corresponding episode to present the
outcome to the expert. Finally, the reward learner returns the
mean estimate of the reward to the PS learner, which uses the
rewards to update its policy and start the next iteration.
III. B ACKGROUND
In this section we provide compact background information
on components of the proposed algorithms that are necessary
to give a complete picture of the proposed approach.
A. Relative Entropy Policy Search
We pair our reward learning algorithm with the recently
proposed relative entropy policy search (REPS) [28]. REPS
is a natural choice for a PS method to be combined with an
active learning component as it has been shown to work well
on real robot problems [9] and is designed to ‘stay close to the
data’. Thus, previous expert queries will remain informative. A
distinctive feature of Relative Entropy Policy Search (REPS),
is that its successive policies vary smoothly and do not jump
in the parameter space or the context space. This behavior
is a beneficial characteristic to increase compliance with our
proposed active learning approach, as we are only able to
predict correct rewards within observed regions of the param-
eter space. To constrain the change in the policy, REPS limits
the Kullback-Leibler divergence between a sample distribution
q(s;!)and the next distribution (!js)(s)
X
s;!(s)(!js) log(s)(!js)
q(s;!): (3)
For the complete optimization problem and its solution we
refer to the original work [28].
B. Dynamic Movement Primitives
A popular use case for PS methods is to find good pa-
rameters of trajectory generators such as Dynamic Movement
Primitives (DMPs) [17]. The resulting desired trajectories can
then be tracked by a linear feedback controller. DMPs model
trajectories using an exponentially decreasing phase function
and a nonlinear forcing function. The forcing function excites
a spring damper system that depends on the phase and is
guaranteed to reach a desired goal position, which is one of
the parameters of a DMP. The forcing function is modelled
through a set of weighted basis functions ! . Using the
weights!of the basis functions as parameters, we can
learn parametrized joint trajectories, and an increasing number
of basis functions results in an increased flexibility of the
trajectory. We use DMPs for our simulated robot experiments.
C. Bayesian Optimization for RL
An alternative approach to learning control parameters !
using PS methods is to model p(Rjs;!)directly and to
use Bayesian Optimization (BO) instead of the PS learner.
However, the approach proposed in this paper introduces a
layer of abstraction which allows us to learn a mapping of
only a low-dimensional input space to the reward, as we only
have to map from outcomes to reward as opposed to map from
state-action to reward. As BO methods are global methods,
they are more susceptible to the curse of dimensionality than
PS methods which are local methods. As a result, the mapping
p(Rjs;!)which the standard BO solution uses is considerably
more difficult to learn than the mapping of the modular
approach that we are proposing.
The BO approach would also require expert ratings for every
sample, while the proposed approach requires only occasional
human feedback.
D. Acquisition Functions
In the following we present four Acquisition Function (AF)
schemes taken from Hoffman et al. [16] that we used to
optimize the model of the reward function.
a) Probability of Improvement: The Probability of Im-
provement (PI) [16] in its original formulation greedily
searches for the optimal value of the input parameter that
maximizes the function. An adapted version of PI balances the
greedy optimization with an exploration-exploitation trade-off
parameter. The adapted version is given by
PI(o) = (o) f(o)
(o)
;
whereois the best sample in the training set Dand()
is the normal cumulative density function. The exploration-
exploitation trade-off parameter has to be chosen manually.
b) Expected Improvement: Instead of finding a point that
maximizes the PI, the Expected Improvement (EI) [16], tries
to find a point that maximizes the magnitude of improvement.
Thus, it does not only try to improve local maxima but also
considers maxima in different regions and is less greedy in
the search of an optimal reward R.
EI(o) = ((o) f(o) )) (M) +(o)(M);
if(o)>0and zero otherwise, where ()is the normal
probability density function. Mis given by
M=(o) f(o)
(o):
The EI acquisition function shares the tuning factor for
the exploitation-exploration trade-off with the PI, where a
suggested value is = 0:01[16].c) Upper Confidence Bound: The Upper Confidence
Bound directly uses the mean and standard deviation of the
reward function at the sample location to define the acquisition
function. An adapted version of the UCB function [33] is given
by
GP-UCB (o) =(o) +pv
n(o);
where recommended values v= 1 and
n=
2 log(nd=2+22=3)withd=dim(o)and2(0;1)
are given by Srinivas et al. [33].
d) GP Hedge: As each of the above acquisition functions
lead to a characteristic and distinct sampling behavior, it is
often not clear which acquisition function should be used.
Portfolio methods, such as the GP-Hedge [16], evaluate several
acquisition functions before deciding for a sample location.
Given a portfolio with Jdifferent acquisition functions, the
probability of selecting acquisition function jfor the sample
n+ 1is given by the softmax
p(j) =exp(gj
n)PJ
i=1exp(gj
n);
where >0is the temperature of the soft-max distribution.
The gains vector gis initialized to zero before taking the
first sample g1:J
0= 0 , and is then updated with the cu-
mulative reward gained by the selected acquisition function,
i.e.,gj
n+1=gj
n+(oj
n+1), whereoj
n+1is the sample point
nominated by acquisition function j. For all other gains the
value does not change, i.e., gi6=j
n+1=gi
n.
IV. E VALUATIONS
In this section we show evaluations of the proposed active
reward learning approach. For all simulation experiments,
unless stated otherwise, we tested each setting ten times.
A. Five Link Reaching Task
A simulated planar robot consisting of five links connected
by rotary joints was controlled in joint space using Dynamic
Movement Primitives (DMP) [17], as described in Section
III-B. If not stated otherwise, we used 20basis functions
per joint, resulting in a total of 100 parameters that had to
be learned. We evaluated our approach on a reaching task,
where an analytical reward function was readily available.
The hand coded reward function was given by R(pr) =
1000 100jjpr pgjj;whereprwas the position of the robot’s
end effector and pgwas the desired target position. However,
we increased the difficulty of the task by not supplying the
outcome features in task space, but rather in joint space,
i.e., the GP had to model the forward kinematics to predict
the reward, making the problem both non-convex and high-
dimensional (five dimensional mapping).
To allow extensive and consistent evaluation of all parame-
ters of the presented approach, we programmed a noisy expert,
which returned the reward with additive white noise (standard
deviation was 20). In Fig. 7 we present results that compare
the coded noisy expert approach to actually querying a human
expert and show that the behavior is comparable. The human
(a) Pestle and the paper box
that is filled with metal bars.
(b) Failed grasp, not robust
against perturbations.
(c) Mediocre grasp, stable but
incorrect orientation.
(d) Good grasp, stable with in-
tended orientation.
Fig. 3: Examples of different grasps and their categorization. Grasps count as failed if the object is either not picked up at all or if small
perturbations would make the object drop. Grasps that are stable but do not keep the original orientation of the object count as OK but not
successful grasps. Grasps that are both stable and keep the original orientation count as successful grasps.
0 200 400 600 8001000
900
800
700Reward
Rollouts
0 200 400 600 800
Rollouts
# User Inputs150
100
50
0Vanilla RL
UCB
PI
EI
HedgeFive Link Via Point Task
Fig. 4: We evaluated our approach on a programmed, but noisy
expert to emulate human expert input. Vanilla RL REPS queries the
programmed expert for every sample, while our approach builds a
model of the reward function and only queries the expert occasionally.
expert could give rewards on a vertical bar through a graphical
interface.
1) Evaluation of Acquisition Functions: The evaluation of
the available AFs given in Fig. 4 shows that even though there
was only limited difference in the asymptotic performance,
there was considerable difference in the sample efficiency.
Especially the GP-UCB AF asks for many user queries. While
PI had the lowest asymptotic performance, as it is the most
greedy of the presented AFs, it also required the lowest number
of user queries. This behavior makes it an interesting candidate
when trying to minimize human interactions.
2) Evaluation of Uncertainty Threshold: To optimally trade
off the number of queries and the agents performance we
need to set the uncertainty threshold trade-off parameter <
(o)=p . This parameter expresses how certain we require
our algorithm to be that a proposed query is not explained by
the estimated observation noise. If we choose to be greater
than1, we require our estimate to improve on the observation
noise. Fig. 5 show the effects of adjusting . The performance
remained stable up to = 1:3and started degrading with
= 2. Changing the order of magnitude of resulted in a
failure to learn the task. While the effects of setting = 1:5on
the performance were moderate, the number of queries were
reduced by about 50% when compared to = 1:3.
3) Evaluation of Sparse Sampling: In order to minimize
human interaction, we stop improving the GP in each iteration
if the outcome that maximizes the AF has already been queried
before, instead of selecting the second best outcome according
to the AF. The results in Fig. 6 show, that sparse sampling
leads to equally good asymptotic performance but requires
considerably less expert interactions.
4) Evaluation Direct Learning: The premise of this paper
is that while BO can be used to efficiently learn the mapping
1000
800
600
4000 200 400 600 800Query Frequency
0 200 400 600 80001020304050
Rollouts Rollouts
# User InputsReward
λ= 5.0
λ= 1.5
λ= 1.3
λ= 1.1Fig. 5: We evaluated the impact of the threshold factor that
influences when and how often we sample. Lower values of led to
better converged performance but required more user interaction.
0 200 400 600 800
Rollouts1000
900
800
700
600
5000 200 400 600 800
Rollouts
# User Inputs150
100
50
0Sparse Sampling
Vanilla RL: REPS
PI Sparse
PI Reward
Fig. 6: Results of comparing our acquisition algorithm to the
standard acquisition algorithm. Our algorithm collects sparse user
queries and does not ask for any user queries after a PS iteration if
the outcome that maximizes the AF has already been queried before.
of outcome to reward, it would be too sample intensive to
directly learn the mapping from parameter space to outcome.
We validated this premise by comparing our joint approach to
directly learning the reward from the control parameters (i.e.,
from the basis function weights !). The results in Fig. 8 show
that BO did not converge to a good solution within 50expert
queries. Both methods were using PI.
5) Comparison to Inverse Reinforcement Learning: We
compared our algorithm to the Maximum Entropy IRL ap-
proach [36] on a via point task in two different scenarios.
Trajectories generated by DMPs had to pass one or two via
points, respectively (20 dimensional action space for each).
0 200 400 600800
Human Expert
Programmed Expert
RolloutsHuman Expert ScenarioReward
0 200 400 600800
Rollouts10203040
# User Inputs
5006007008009001000
Fig. 7: We validated our approach of using a noisy programmed
expert as substitute for a human expert on the simulated tasks. The
results show that both experts yielded similar learning behavior.
# QueriesRewardDirect Learning Standard Error
0 20 40 60 80 1002006001000
Active Reward Learning
Bayesian Optimization
Fig. 8: Comparison of ARL and directly learning the reward from
the control parameter !using BO. In this figure we plot the
standard error as opposed to the standard deviation. ARL performs
significantly better after 50 queries ( p<0:05).
#UI = 5 #UI = 10 #UI = 20
ARL 1VP 995:22:477 999:20:25 999:20:25
IRL 1VP 9833:306 977 :36:334 9843:036
ARL 2VPs 970:713:1 9981:13 999:10:172
IRL 2VPs 994:11:874 9960:4736 9960:555
TABLE II: Comparison of IRL and ARL. We compare the methods
on two via point (VP) tasks with one and two VPs. The columns
show the achieved performance (mean and standard deviation) after
five, ten or 20 user interactions (UIs).
For this comparison only, we relaxed our assumption that no
demonstrations are available and provided the IRL approach
with (imperfect) demonstrations that passed close to the via
point (at most 0:1units away). In Table II, we show the mean
and standard deviation of the best policy reached for either
five, ten or 20 user interactions (UI). In our approach, UIs
are ratings while in IRL UIs are demonstrations. The results
show that our approach yields competitive results while not
requiring access to demonstrations.
6) Alternative Policy Learners: While we used REPS for
most of our experiments, the proposed framework is not lim-
ited to a specific RL method. To investigate the compatibility
with other methods, we compared our framework using REPS
with our framework in combination with a Bayesian Policy
Gradient (BPG) method [13] on a via point task with one via
point (20 dimensional action space). BPG models the gradient
of the policy using a GP and uses the natural gradient in the
policy update. The results in Fig. 9 show that both methods
were able to learn the task in combination with our approach.
B. Robot Grasping
We used the results from Section IV-A to set the parameters
for the real robot experiment (we use PI and set = 3). We
learned the policy (!js)with 15samples per iteration for
a total of 10iterations and we repeated the experiment three
times. The control parameters of the policy were the 15joints
of the five finger DLR hand, which is mounted to a 7DOFs
REPS
BPG
0 200 600 1000 0 200 600 100056789101112
Rollouts RolloutsReward
#iUseriInputsBayesianiPolicyiGradient
1000
500
0
-500
-1000
0
Fig. 9: Performance of REPS and BPG on a via point task.KUKA lightweight arm as shown in Fig. 1. We considered
the task of blind grasping as described in [8], where no object
information was available and we did not have visual feedback,
i.e., we did not have information about the contact points.
Instead, we calculated the forces in the finger tips through the
joint torques and and the hand kinematics. We used the finger
tip force magnitudes as outcome features which were used by
the GP to model the reward function.
Evaluating the real robot experiments presented the problem
of a success metric. As we did not have a ‘correct’ reward
function that we could evaluate the learned reward function
against, we resorted to introducing three label categories
which were only used for the evaluation after finishing the
experiments. The scheme presented in Fig. 3 labels grasps as
failures with a reward of 1, if the object was not lifted at all
or slipped when slightly perturbed. Grasps that were stable but
did not keep the intended orientation were given a reward of
0. Finally, grasps that lifted the object and kept the orientation
of the object were assigned a reward of 1. These labels and
reward values were not used during the learning of either the
policy or the reward model but only used to present results.
During the learning phase, the human expert assigned grasp
ratings in the range of 1000 .
1) Learn to Grasp Unknown Object: The object to be
grasped was a cardboard box filled with metal weights such
that the robot cannot grasp the object with very unstable grasps
or by deforming the paper box. The box, shown in Fig. 3a, was
of size 7:5cm x 5:5cm x 2cm and filled with two metal bars
with a combined weight of 350g. The results of three trials
presented in Fig. 10 show that the robot learned to perform
a successful grasp of the object in all three trials, while only
requesting six queries in the first and twelve queries in the
second trial. In the last trial, the robot’s performance first
increased quickly but dropped after 80episodes (or rollouts),
coinciding with a sudden increase of user queries, such that
the final number of queries in the last trials was 27. The
reason for this unusual behavior was a malfunction of the
distal joint of the thumb, which rendered the grasping scheme
dysfunctional. As the learner could not reproduce outcomes
that led to good rewards, it resorted to finding a different grasp
strategy. At the same time, since the GP was presented with
new outcome samples in previously unobserved regions of the
outcome space, it requested new user queries to model the
reward function in the new region of interest.
We compared our learned reward function to a (naive) hand
coded reward function based on the same features. The hand
coded reward function aimed to reach a total force magnitude
over all fingers. Using the programmed reward, the robot was
able to reliably pick up the object after the first two trials and
in ten out of 15 grasps at the end of the third trial (We expect
the robot would have also learned to pick it up reliably in the
last trial with more iterations). However, the robot did not pick
up the object in a way that kept the original orientation of the
object. Encoding such behavior through only force features by
hand is challenging. The performance curve of the hand coded
reward function shows a slight dip, which is possible as we
Real Robot Results and QueriesGrasp Rating
Queries
-11
525
Rollouts2040 120140 60 80 100
Rollouts2040 120140 60 80 100
Trial 1
Trial 3Trial 2
Average
Hand CodedFig. 10: The real robot successfully learned good grasps in all of the
three trials. The grasp rating scheme is described in Fig. 3. In the last
trial, one joint failed, represented by a spike in expert queries. The
robot successfully adapted the reward function to recover from the
hardware failure. We also show the average performance of a naive
hand coded reward function.
-11
20 40 60 80 100 120 140Real Robot Transfer
RolloutsGrasp RatingMean and Standard Deviation0
Fig. 11: Performance of one trial on the real robot system. The robot
uses the reward function learned in trial one of the original task to
learn to grasp a new, unknown object (a wooden pestle as shown in
Fig. 3a). For this trial, we did not allow any expert queries. The robot
successfully learned to grasp the object.
are not plotting the internal reward but the reward assigned
according to the grading scheme introduced in Fig. 3.
2) Transfer Learned Reward Function to New Object: As
we based our reward model only on the finger tip forces, it
is modular and can also be used for (reasonably) different
objects. To test these generalization capabilities, we started a
new trial with a different object (a pestle as shown in Fig.
3a) and initialized the GP with the training data from the first
trial of the previous task. For this experiment we only used
ratings from the previous trial, and did not ask for additional
expert ratings. The pestle that we used for this task was similar
in dimensions but different in shape, such that the agent had
to learn different joint configurations to achieve similar finger
forces that optimized the reward function. The dimensions of
the pestle were 18cm length, 1:5cm radius on the thin end and
2:25cm radius on the thick end. The results in Fig. 11 show
that the agent was able to learn to reliably perform a robust
grasp on the pestle.
V. R ELATED WORK
The term reward shaping has been used to describe efforts to
adapt rewards functions such that the resulting policy remains
invariant but learning speed is increased [27, 10, 4, 32] or
that learning on a new task is accelerated [21]. Aside from
reward shaping, Dorigo and Colombetti [11] have used the
term robot shaping to describe efforts of teaching different
tasks to a robotic agent. Derived from this terminology the
TAMER framework [19] uses the term interactive shaping.
In TAMER the agent receives reinforcements from a human,
guiding the learning process of the agent, but there is no activecomponent. The Advise framework [14] also uses the term
shaping to describe the process of modelling an oracle and
are actively requesting labels for an agent’s action (good or
bad). This approach is tailored for discrete settings and the
feedback frequency is pre-defined.
In preference learning algorithms, the expert is usually
requested to rank the current best against a new execution
and the reward function is inferred from these rankings [2, 6].
Akrour et al. [1] can also deal with noisy rankings. Chu
and Ghahramani [7] introduced the use of GPs to preference
rankings. In preference based approaches, the expert is limited
to transmit one bit of information and cannot express strong
preferences, as, supposedly, human experts cannot give numer-
ical rewards that are sufficiently accurate. Our contribution
is to show how a rating based approach with explicit noise
model can be used in real robot continuous state-action space
problems taking advantage of the stronger guidance through
strong preferences (large differences in assigned rewards).
The problem of expert noise has also been addressed in
Bayesian RL and IRL. In Bayesian RL, Engel et al. [12]
has proposed to model the value function through a GP
and Ghavamzadeh and Engel [13] has proposed to model the
policy gradient through a GP. Especially the policy gradient
method is also well suited to work in combination with our
approached framework. If demonstrations are available, IRL
is a viable alternative. Ziebart et al. [36] have relaxed the
assumptions on the optimality of demonstrations such that a
reward function can be extracted from noisy demonstrations.
In a combination of IRL and preference learning, Jain et al.
[18] have proposed the iterative improvement of trajectories.
In their approach, the expert can choose to rank trajectories
or to demonstrate a new trajectory that does not have to be
optimal but only to improve on the current trajectory. This
approach cannot directly be used on learning tasks such as
grasping, as a forward model is required. Alternatively, Lopes
et al. [24] propose a framework where IRL is combined with
active learning such that the agent can decide when and where
to ask for demonstrations.
VI. C ONCLUSION & F UTURE WORK
We presented a general framework for actively learning the
reward function from human experts while learning the agent’s
policy with any PS or policy gradient method. Our experiments
showed that the learned reward function outperforms naive
hand coded reward functions, generalizes to similar tasks
and that the approach is robust to sudden changes in the
environment, for example when a mechanical failure occurs.
In future work we plan to extend on the ground-laying work of
this paper and investigate possible synergies between specific
PS methods and the reward learning framework to improve on
the performance of the generic AFs.
ACKNOWLEDGMENTS
The authors want to thank for the support of the European
Union projects # FP7-ICT-270327 (Complacs) and # FP7-ICT-
2013-10 (3rd Hand).
REFERENCES
[1] Riad Akrour, Marc Schoenauer, and Michele Sebag.
Preference-based policy learning. In Machine Learning
and Knowledge Discovery in Databases . 2011.
[2] Riad Akrour, Marc Schoenauer, and Michele Sebag.
Interactive robot education. In European Conference on
Machine Learning Workshop , 2013.
[3] Ravi Balasubramanian, Ling Xu, Peter D Brook,
Joshua R Smith, and Yoky Matsuoka. Physical human
interactive guidance: Identifying grasping principles from
human-planned grasps. IEEE Transactions on Robotics ,
2012.
[4] Jeshua Bratman, Satinder Singh, Jonathan Sorg, and
Richard Lewis. Strong mitigation: Nesting search for
good policies within search for good reward. In Interna-
tional Conference on Autonomous Agents and Multiagent
Systems , 2012.
[5] Maya Cakmak and Andrea L Thomaz. Designing robot
learners that ask good questions. In International con-
ference on Human-Robot Interaction , 2012.
[6] Weiwei Cheng, Johannes F ¨urnkranz, Eyke H ¨ullermeier,
and Sang-Hyeun Park. Preference-based policy iteration:
Leveraging preference learning for reinforcement learn-
ing. In Machine Learning and Knowledge Discovery in
Databases . 2011.
[7] Wei Chu and Zoubin Ghahramani. Preference learning
with Gaussian processes. In International Conference on
Machine Learning , 2005.
[8] Hao Dang and Peter K Allen. Learning grasp stability.
InInternational Conference on Robotics and Automation ,
2012.
[9] Christian Daniel, Gerhard Neumann, and Jan Peters.
Learning Sequential Motor Tasks. In International Con-
ference on Robotics and Automation , 2013.
[10] Sam Devlin and Daniel Kudenko. Dynamic potential-
based reward shaping. In International Conference on
Autonomous Agents and Multiagent Systems , 2012.
[11] Marco Dorigo and Marco Colombetti. Robot shaping:
Developing autonomous agents through learning. Artifi-
cial intelligence , 1994.
[12] Yaakov Engel, Shie Mannor, and Ron Meir. Reinforce-
ment learning with Gaussian processes. In International
Conference on Machine Learning , 2005.
[13] Mohammad Ghavamzadeh and Yaakov Engel. Bayesian
policy gradient algorithms. Advances in Neural Informa-
tion Processing Systems , 2007.
[14] Shane Griffith, Kaushik Subramanian, Jonathan Scholz,
Charles Isbell, and Andrea L Thomaz. Policy shaping:
Integrating human feedback with reinforcement learning.
InAdvances in Neural Information Processing Systems ,
2013.
[15] Harold W Hake and WR Garner. The effect of present-
ing various numbers of discrete steps on scale reading
accuracy. Journal of Experimental Psychology , 1951.
[16] Matthew D Hoffman, Eric Brochu, and Nando de Freitas.Portfolio allocation for Bayesian optimization. In Con-
ference on Uncertainty in Artificial Intelligence , 2011.
[17] Auke Ijspeert, Jun Nakanishi, and Stefan Schaal. Learn-
ing Attractor Landscapes for Learning Motor Primitives.
InAdvances in Neural Information Processing Systems .
2003.
[18] Ashesh Jain, Brian Wojcik, Thorsten Joachims, and
Ashutosh Saxena. Learning trajectory preferences for
manipulators via iterative improvement. In Advances in
Neural Information Processing Systems , 2013.
[19] W Bradley Knox and Peter Stone. Interactively shaping
agents via human reinforcement: The TAMER frame-
work. In International Conference on Knowledge Cap-
ture, 2009.
[20] Jens Kober, Betty J. Mohler, and Jan Peters. Learning
perceptual coupling for motor primitives. In Intelligent
Robots and Systems , 2008.
[21] George Konidaris and Andrew Barto. Autonomous
shaping: Knowledge transfer in reinforcement learning.
InInternational Conference on Machine Learning , 2006.
[22] Petar Kormushev, Sylvain Calinon, and Darwin Caldwell.
Robot motor skill coordination with EM-based reinforce-
ment learning. In Intelligent Robots and Systems , 2010.
[23] Oliver Kroemer, Renaud Detry, Justus Piater, and Jan
Peters. Combining active learning and reactive control
for robot grasping. Robotics and Autonomous Systems ,
2010.
[24] Manuel Lopes, Francisco Melo, and Luis Montesano.
Active learning for reward estimation in inverse rein-
forcement learning. In Machine Learning and Knowledge
Discovery in Databases . 2009.
[25] George A Miller. The magical number seven, plus or
minus two: some limits on our capacity for processing
information. Psychological Review , 1956.
[26] Andrew Ng and Adam Coates. Autonomous inverted he-
licopter flight via reinforcement learning. Experimental
Robotics IX , 1998.
[27] Andrew Ng, Daishi Harada, and Stuart Russell. Policy
invariance under reward transformations: Theory and
application to reward shaping. In Internation Conference
on Machine Learning , 1999.
[28] Jan Peters, Katharina M ¨ulling, and Yasemin Altun. Rel-
ative entropy policy search. In National Conference on
Artificial Intelligence , 2010.
[29] Carl E. Rasmussen and Christopher K. I. Williams.
Gaussian Processes for Machine Learning . 2006.
[30] N. Ratliff, A. Bagnell, and M. Zinkevich. Maximum Mar-
gin Planning. In International Conference on Machine
Learning , 2006.
[31] Nathan Ratliff, David Silver, and Andrew Bagnell.
Learning to Search: Functional Gradient Techniques for
Imitation Learning. Autonomous Robots , 2009.
[32] Satinder Singh, Richard L Lewis, Andrew G Barto, and
Jonathan Sorg. Intrinsically motivated reinforcement
learning: An evolutionary perspective. Autonomous Men-
tal Development , 2010.
[33] Niranjan Srinivas, Andreas Krause, Sham M Kakade,
and Matthias Seeger. Gaussian process optimization in
the bandit setting: No regret and experimental design.
Internation Conference on Machine Learning , 2010.
[34] Ra ´ul Su ´arez Feij ´oo, Jordi Cornell `a Medrano, and
M´aximo Roa Garz ´on. Grasp quality measures. 2012.
[35] Andrea L Thomaz and Cynthia Breazeal. Teachable
robots: Understanding human teaching behavior to build
more effective robot learners. Artificial Intelligence ,
2008.
[36] Brian Ziebart, Andrew Maas, Andrew Bagnell, and
Anind Dey. Maximum entropy inverse reinforcement
learning. In Conference on Artificial Intelligence , 2008. |
700cc5bc-0082-483e-8150-4e9175497e67 | trentmkelly/LessWrong-43k | LessWrong | Are there really no ghosts in the machine?
My previous article on this article went down like a server running on PHP (quite deservedly I might add). You can all rest assured that I won't be attempting any clickbait titles again for the foreseeable future. I also believe that the whole H+ article is written in a very poor and aggressive manner, but that some of the arguments raised cannot be ignored.
On my original article, many people raised this post [insert link] by Eliezer Yudkowsky as a counterargument to the idea that an FAI could have goals contrary to what we programmed. In summary, he argues that a program doesn't necessarily do as the programmer wishes, but rather as they have programmed. In this sense, there is no ghost in the machine that interprets your commands and acts accordingly, it can act only as you have designed. Therefore from this, he argues, an FAI can only act as we had programmed.
I personally think this argument completely ignores what has made AI research so successful in recent years: machine learning. We are no longer designing an AI from scratch and then implementing it; we are creating a seed program which learns from the situation and alters its own code with no human intervention, i.e. the machines are starting to write themselves, e.g. with google's deepmind. They are effectively evolving, and we are starting to find ourselves in the rather concerning position where we do not fully understand our own creations.
You could simply say, as someone said in the comments of my previous post, that if X represents the goal of having a positive effect on humanity, then the FAI should be created so as to have X as its primary directive. My answer to that is the most promising developments have been through imitating the human brain, and we have no reason to believe that the human brain (or any other brain for that matter) can be guaranteed to have a primary directive. One could argue that evolution has given us our prime directives: to ensure our own continued existence, |
fc465c13-0cb9-4c8d-91ce-0b9150111b40 | trentmkelly/LessWrong-43k | LessWrong | Consistent Plato
A putative new idea for AI control; index here. The idea here is somewhat underdeveloped.
I've been suggesting AI designs that are not exactly expected-utility maximisers, while wondering if such designs could be stable under self improvement/subagent creation/general stuff. And also wondering what designs such agents would converge to if they did self-improve.
This post is just a short note warning people to make sure they are consistent when they use non-expected utility maximalisation, and to be clear which parts are Platonic and idealised and which parts are not.
Let's take the reduced impact AI design as an example. In this, the agent has a kind of utility:
v = u - R,
where u is a standard utility and R is a penalty term. The penalty term involved a probability estimate for what a putative "super AI" would estimate, based on some observations it could make. Thus it is better described as R(P), where P is the probability estimator of that "Super AI".
When calculating expected v, the agent will use its own probability estimator P'. Initially, P' is just a less powerful version of P, one that can, moreover, estimate properties about P.
How could this possibly go wrong? I'm glad you asked ^_^ The obvious risk is that the agent acts to replace P with Q, such that R(Q) is as low as possible.
This could come about if P referred to some actual physical algorithm, possibly one written down somewhere the AI could get at it. Then re-writing the algorithm P could increase v by reducing R(P).
It is somewhat more difficult if P and P' are defined in the same place (maybe P is just P' run for a lot longer). Then modifying P could run the risk of degrading the AI's overall accuracy in predictions. However, this may be a price the AI is willing to pay, if the R(P) penalty is severe enough. More likely, the AI could degrade the quality of P but not P' by introducing pathologies that would only affect the first.
Even if P itself is safe from modification, the AI can sti |
3c588c44-f3a5-4d91-bb89-12b6ed09e371 | trentmkelly/LessWrong-43k | LessWrong | Meetup : University of Maryland
Discussion article for the meetup : University of Maryland
WHEN: 16 September 2011 06:00:00PM (-0400)
WHERE: University of Maryland, University of Maryland - College Park, STAMP
There are a few LessWrong readers at UMD. Let's meet up! We're meeting in Terrapin Room B, in the Student Involvement Suite, kindly reserved by btrettel. You can find the Student Involvement Suite on this map: http://www.stamp.umd.edu/visitor_information/stamp_map.html We have the room from 6 - 8 PM. This meeting will probably have an outsize impact on future meetup topics, please come with opinions.
Discussion article for the meetup : University of Maryland |
f84ec431-8483-4eee-a63f-cf258d000ad6 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Opportunities for Impact Beyond the EU AI Act
Summary
=======
* The EU AI Act trilogues are coming to a close in the coming weeks. It appears likely that the final iteration of the act will include regulations covering foundation models.
* The window for effecting change here has not yet closed and is likely to extend several years into the future. Therefore, there is still an impact case for those interested in reducing risks from advanced AI to enter European policy.
* Some upcoming opportunities for impact include:
+ Establishing mechanisms for enforcing the most important aspects of the AI Act.
+ Shaping the precise technical standards & guidelines that are enforced by the AI Act.
+ Encouraging international bodies & other nations to adopt the EU's AI standards (conditional on sensible standards being set by the EU in the coming years).
* Applications for the [EU Tech Policy Fellowship](https://www.techpolicyfellowship.eu/) close on October 15. We encourage EU citizens interested in AI governance to apply.
The EU AI Act: Where are we now?
================================
As the EU AI Act trilogues come to a close, AI governance in Europe is on the brink of a significant milestone.
Although the precise scope of the act remains somewhat unclear, there are hopeful signs that it will regulate foundation models. [Early drafts of the EU AI Act](https://web.archive.org/web/20231011155323/https://eur-lex.europa.eu/resource.html?uri=cellar:e0649735-a372-11eb-9585-01aa75ed71a1.0001.02/DOC_2&format=PDF) listed a series of high-risk areas in which the deployment of AI was restricted (e.g. law enforcement; migration, asylum & border management). More recent drafts include an article which extends this legislation to cover foundation models and would require providers to demonstrate the mitigation of reasonably foreseeable risks & document remaining non-mitigatable risks.
If such requirements remain following these final trilogue negotiations, major AI labs must decide whether to withdraw their models from the EU, create separate models for an EU market, or develop a single model which is compliant with the EU AI Act. If AI labs opt for this third option, this would constitute a “de facto” Brussels Effect and EU standards would become a global norm.
Opportunities for impact beyond the AI Act
==========================================
However, passing the AI Act is just the beginning. The act will not come into effect until Fall 2025 and the window for effecting change here has not yet closed and is likely to extend several years into the future beyond this. For EU citizens interested in reducing risks from advanced AI, we believe there is still a strong impact case for entering European policy.
### Enforcement
Enforcement is an essential component of the EU AI Act. Without adequate enforcement, the act will likely have little impact on reducing risks from advanced AI.
The challenges of enforcing the AI Act are akin to those faced with GDPR. Just as GDPR is enforced by individual data protection authorities within each member state, enforcement of the AI Act is also likely to involve a shared responsibility between national bodies at a member state level and European authorities. Organisations tasked with this responsibility will likely be talent and resource-constrained. This provides an opportunity for knowledgeable individuals with a strong prioritisation mindset to have a large impact
**How to enter:**Apply directly to relevant open vacancies at the EU level (e.g. [this role at ECAT](https://web.archive.org/web/20231012100458/https://recruitment.jrc.ec.europa.eu/showprj_iframe.php?id=2682&apply=vacancy-for-auxiliary.php?inst=18002DTD&pid=2682&target=_blank#toolbar=0)). Join your local data protection authority responsible for enforcing GDPR ([full list here](https://web.archive.org/web/20231012100734/https://ec.europa.eu/justice/article-29/structure/data-protection-authorities/index_en.htm)) with the intention of transitioning to an AI department if/when it is established or alternatively leveraging this experience to join the organisation which will ultimately enforce the AI Act in your country.
### Technical standards & guidelines
Technical standards and guidelines will continue to be developed by [CEN and CENELEC](https://www.cencenelec.eu/) between 2023-2025. The forthcoming [EU AI Office could also play an important role here](https://carnegieendowment.org/2023/10/03/letter-to-eu-s-future-ai-office-pub-90683), as could national standardisation committees. These standards will play a pivotal role in shaping the practical aspects of the act, e.g. outlining the exact risk management requirements for voluntary audits. Those with technical backgrounds and a strong understanding of the risks of advanced AI models could make significant contributions here.
**How to enter:** Those interested could join one of the National Standards Bodies ([full list](https://web.archive.org/web/20231012103909/https://standards.cencenelec.eu/dyn/www/f?p=CEN:5)), the bodies responsible for standards in each of the member states; the European Standards Organisations (i.e. [CEN](https://www.cencenelec.eu/about-cen/), [CENELEC](https://www.cencenelec.eu/about-cenelec/), and [ETSI](https://www.etsi.org/)), the organisations responsible for all EU standard-setting; the [European Commission](https://commission.europa.eu/jobs-european-commission_en); or through think tanks working on developing technical standards.
### Exporting European regulation
The EU’s standards and regulations have the potential to influence the global AI landscape. Provided that sensible technical standards & enforcement mechanisms are enacted, it could make sense for individuals to encourage other international bodies to adopt similar protocols. For example, the [TTC](https://web.archive.org/web/20231012102316/https://digital-strategy.ec.europa.eu/en/library/ttc-joint-roadmap-trustworthy-ai-and-risk-management), [OECD](https://web.archive.org/web/20231012102434/https://www.oecd.org/digital/artificial-intelligence/), [UN](https://web.archive.org/web/20231012102506/https://press.un.org/en/2023/sgsm21880.doc.htm), and [G7](https://www.politico.eu/wp-content/uploads/2023/09/07/3e39b82d-464d-403a-b6cb-dc0e1bdec642-230906_Ministerial-clean-Draft-Hiroshima-Ministers-Statement68.pdf) are currently considering frameworks for classifying AI systems and technical standards for regulating them.
**How to enter:** Those with strong diplomacy skills (or a professional background in diplomacy) and an understanding of the relevant aspects of EU technical standards / enforcement protocols could join other international bodies likely to introduce AI policy.
### Advocating for increased AI safety funding
Lobbying the next Multiannual Financial Framework (2028-2034) for AI safety funding could also be a high impact opportunity. The [2021-2027 MFF](https://web.archive.org/web/20231012111810/https://digital-strategy.ec.europa.eu/en/activities/funding-digital) allocated ~€7.6 billion to the [Digital Europe Programme](https://web.archive.org/web/20231012111845/https://digital-strategy.ec.europa.eu/en/activities/digital-programme) which aims to accelerate economic recovery and drive the digital transformation of Europe (among others, areas invested in include core AI & supercomputing capacities). [Horizon Europe](https://web.archive.org/web/20231012111823/https://research-and-innovation.ec.europa.eu/funding/funding-opportunities/funding-programmes-and-open-calls/horizon-europe_en) also received a dedicated budget of €95.5 billion for research related to “digital, industry and space”. This budget will be dedicated to developing AI, high performance computing, and other key digital technologies. Given the scale of the funding available, redirecting even a small portion of it to safety-focused efforts could be enormously impactful.
**How to enter:** Joining the European Commission, likely within the [DG RTD](https://commission.europa.eu/about-european-commission/departments-and-executive-agencies/research-and-innovation_en) could be a helpful step. The European Parliament will play a role later on, so joining as an assistant to a relevant MEP could also be helpful. It’s also possible to advocate through other channels, for example at a think tank focused on tech policy or through a national government.
### Other potential paths
* **Updating the AI Act:** Technical standards & guidelines could be updated over time as models advance. It’s unclear where this responsibility will fall, but working at the forthcoming EU AI Office could enable one to shape future amendments to the act in a positive direction.
* **EU Careers focused on a narrow threat model:** It could make sense for a portion of talent to specialise in a specific domain relevant to concrete [risk model](https://www.safe.ai/ai-risk). This could be especially true if you have a background particularly suited to one area (e.g. an individual with a biotechnology background could be best suited to working on the potential use of AI to release engineered pandemics). Other options could include specialising within cybersecurity to improve Europe’s resilience to infrastructure attacks (e.g. at [ENISA](https://www.enisa.europa.eu/)) or in tackling disinformation to protect against AI-facilitated disinformation campaigns.
Interested? Apply for the EU Tech Policy Fellowship (deadline 15 October)
=========================================================================
If you’re interested we encourage you to apply for the [EU Tech Policy Fellowship](https://www.techpolicyfellowship.eu/) or check out this [list of opportunities & resources we’ve compiled](https://forum.effectivealtruism.org/posts/W4BRXGvz7BvMPFNvy/resources-and-opportunities-for-careers-in-european-ai).
The [EU Tech Policy Fellowship](https://www.techpolicyfellowship.eu/) is a 7-month programme for ambitious graduates to launch European policy careers focused on emerging technology (Jan → July 2024). [**Applications**](https://docs.google.com/forms/d/e/1FAIpQLSdvhSFRKtn5pTZbhdHxtTUjUHJREFESfhVFPFhxypyPOA1aZA/viewform) **for the winter cohort are due October 15.**
As a fellow in our program, you’ll have the choice of two distinct tracks: Training & Placement. Training track fellows explore the intricacies of tech policy through an 8-week online course and a week-long policymaking summit in Brussels. In addition to this training, placement track fellows also secure a fully-funded 4-6 month placement at a respected think tank such as the [Centre for European Policy Studies](https://www.ceps.eu/) or [the Future Society](https://thefuturesociety.org/).
Upon completion of the fellowship, we expect you'll transition into a career working directly on AI policy in Europe (such as those outlined in this post). Beyond the fellowship, we’ll continue to provide support, guidance, and networking opportunities to alumni as we build a strong community of policy professionals working to reduce risks from emerging technology in Europe.
[**Apply here**](https://docs.google.com/forms/d/e/1FAIpQLSdvhSFRKtn5pTZbhdHxtTUjUHJREFESfhVFPFhxypyPOA1aZA/viewform) **by October 15** |
afcca7e4-8b5e-4df4-9849-513ff8d73600 | trentmkelly/LessWrong-43k | LessWrong | New Month's Resolutions
If you have a goal that can be credibly achieved by the end of the month reply to this post. At the end of the month everybody who posted writes up their experiences, lessons in overcoming akrasia etc. as a reply to their original comment or by editing the original.
Whether you want to share the goal before the end of the month is up to you, pros, you are accountable, cons, you may feel a sense of accomplishment just by saying it and not do anything. |
6578650a-9ee3-4bcb-90a0-1d05f4ad5df6 | trentmkelly/LessWrong-43k | LessWrong | Post-crash market efficiency 1696-2020
WHAT DO past stock market crashes tell us about the weak-form Efficient-Market Hypothesis? In a meltdown, do prices become more predictable?
In theory, they instantly adjust to all public information. Hence technical analysis – predicting future stock prices from patterns in past ones – should not work. In particular, there should be no usable trends – tendency for prices to keep moving up or down – nor predictable reversals.
On the other hand, big crashes are so rare and extreme that they might be exceptions. In a crisis, perhaps enough traders act impulsively, or at least do something unusual, that there's a smidgeon of predictability. Surely greed and fear don’t always exactly cancel out?
So let’s take a look. Do big crashes on average produce any kind of pattern or trend in prices?
I’ll consider all major meltdowns that have ever happened in the US and UK, using their broad indices, the S&P 500 and FTSE All-Share; and defining a crash as a fall of 20% or more within 8 weeks.
UK crashes
Starting with the UK (as its stock market began first), here are all crashes throughout history:
Each coloured line is a crash – starting with a 20% fall – and its aftermath. I’ve even included the Bank of England crisis of 1696, and the still-famous South Sea Bubble of exactly 300 years ago (which coined the term ‘bubble’), though both are omitted from the averages. Thereafter, despite slides and slumps, there were no more actual crashes until World War 2, after which they became more frequent.
The red line is the current crash, from mid-January up to this week. The FTSE fell further, faster, than ever before, then has partly recovered; though this may yet turn into a dead cat bounce.
Superficially, all these crashes have little in common: caused by speculation, oil, program trading, terrorism, disease; all kinds of shapes – L, V, U, W; they fall 20% or 70%, take days or years to hit rock bottom, and up to a decade to recover. Like spaghetti, the lines are all over the p |
d9529650-37b0-4bcb-9956-1363cc9e7175 | trentmkelly/LessWrong-43k | LessWrong | Building intuition with spaced repetition systems
Do you ever go to a lecture, follow it thinking it makes total sense, then look back at your notes later and realize it makes no sense? This used to happen to me, but I’ve learned how to use spaced repetition to fully avoid this if I want. I’m going to try to convey this method in this post.
----------------------------------------
Much of my understanding of how to create flashcards comes from “Using spaced repetition systems to see through a piece of mathematics” by Michael Nielsen and “How to write good prompts: using spaced repetition to create understanding” by Andy Matuschak, but I think my method falls in between both, in terms of abstraction. Finally, I want to credit Quantum Country for being an amazing example of flashcards created to develop intuition in users.
My method is more abstract than Michael Nielsen’s approach, since it does not only apply to mathematics, but to any subject. Yet it is less abstract than Andy Matuschak’s approach because I specifically use it for ‘academic subjects’ that require deep intuition of (causal or other) relationships between concepts. Many of Matuschak’s principles in his essay apply here (I want to make sure to give him credit), but I’m looking at it through the ‘how can we develop deep intuition in an academic subject in the fastest possible time?’ lens.
Minimize Inferential Distance on Flashcards
A method that I like to repeat to myself while making flashcards that I haven’t seen in other places is that each flashcard should only have one inferential step on it. I’m using ‘inferential step’ here to mean a step such as remembering a fact, making a logical deduction, visualizing something, or anything that requires thinking. It’s necessary that a flashcard only have a single inferential step on it. Anki trains the mind to do these steps. If you learn all the inferential steps, you will be able to fully re-create any mathematical deduction, historical story, or scientific argument. Knowing (and continually remember |
05e87601-a9ec-410f-b0a6-2ae8b681d2c0 | trentmkelly/LessWrong-43k | LessWrong | PRINCIPLES OF PERCEPTION
It is relevant that one should pay attention in the tiniest bits of the location they find themselves in. From this bit of habit, a lot of things will cease to confound the mind. How much we can predict circumstances by relying on previous data depends on how much we have programmed our minds to divide the world into tiny clues, and compose it in the right way. A lot of Sherlock Holmes readers fail to see the conclusion of the crime, even when all the facts are laid out before them. When the conclusion does come, you then make connections to the strings you’ve been provided with. Mostly with human psychology, we get the answer and then go back to connect the strings we’ve had in the very beginning. In this exercise, we are learning to condition ourselves to use the strings of incidences to theorize the ending, not use the ending to connect the incidences.
I’ve actually met a person that can always predict the ending of the movie just five minutes into the movie. And then another that claims that if he had the name of the movie, the starring actors or their faces, that he could pick up the clues and possibly predict the ending. The second claim may seem farfetched, but it is within the bounds of logical reasoning. For most people, this attribute has been so innate in them, it becomes a conditioned reflex.
Let’s say you climb a set of stairs every day for the past ten years. Have you really taken the time to know the number of stairs in the staircase? Can you identify the parts that are more worn than the others?
Okay, here’s another one. If you had a neighbor for two years, can you rightfully recall what colors the neighbor favors to wear on Tuesday or Wednesday, and the fact that he usually leaves by this time and gets back by another time? Can you guess his behavioral patterns? And let’s say he suddenly goes missing, and the police questions you and asks if there were odd manners that he was exhibiting prior to his disappearance, would you be able to recall that |
ec6d0529-938d-4497-acff-63f2efc1ea52 | trentmkelly/LessWrong-43k | LessWrong | Shannon's Surprising Discovery
Imagine that you’re a telecom provider in the early 20th century. You provide both telephone lines and telegraph lines - audio and binary. These two systems face quite different design constraints.
The telegraph line only needs to be able to distinguish between high and low voltage. If the signal needs to be amplified or denoised, a relay suffices, a magnetic switch which will shut under voltage above some cutoff or open under any voltage below the cutoff. A noisy incoming signal turns into a clean outgoing signal, so long as the noise isn’t so large as to cross the voltage cutoff.
Telegraph relay. That metal tongue inside the horseshoe is the switch; note the small gap between it and the contact on the right. The electromagnet is presumably in the black cylinder on the right. (source)
A telephone line is trickier. Long-range telephone lines require amplification of continuous voltages, without much distortion or noise. It’s a fairly different engineering problem, for which various arrangements of electrodes in evacuated tubes were used.
Vacuum tubes in an audio amplifier. Yes, they are supposed to glow like that. (source)
An early twentieth century electrical engineer working on these systems would probably have told you that different kinds of signals require different kinds of transmission hardware. Audio has different needs from Morse code, both of which have different needs from television. Sure, you could send Morse code over a telephone line, but it would be wildly inefficient. Also, different kinds of messages need more or less reliability, and so can tolerate more or less noise in the lines. Try to send a message which needs high reliability over a very noisy line, and you’ll need a lot of redundancy. Again, wildly inefficient.
Claude Shannon’s surprising discovery was that this is basically false. Different kinds of signals do not require different kinds of hardware, at least not for efficient throughput. (Though latency requirements can and do still |
0136d424-7b37-4d46-a87d-a2dfe360ea56 | trentmkelly/LessWrong-43k | LessWrong | How can I determine that Elicit is not some weak AGI's attempt at taking over the world ?
I read two days ago the Elicit presentation post, and as I have been reading intensely about AI safety lately, I was wondering this question : What are the tools and techniques you would use right now to determine that the current Elicit research assistant is a safe AI ?
What is described in the post is a narrow AI that will and does fail in obvious ways if it is not aligned. But how would you know it is not a small and limited AGI, aligned or not ?
I suppose that if the Elicit research assistant is currently an AGI, however how limited, its best course of action would be to pretend to be a narrow AI helping researchers in their goal of doing research. I imagine that it would get better chances of not being replaced by a new version and to get more upgrades and computing resources as the Elicit product develops.
So I was wondering, is this obviously dumb that this exact system as seen by Elicit users cannot ever be an AGI, even in the future ? If you had right now real suspicions it could be an AGI, apart from first making sure it is stopped, how would you try to determine whether it is an AGI ? |
da137669-992b-4a9f-9392-a1f289e5313a | trentmkelly/LessWrong-43k | LessWrong | Best effort beliefs
Alice: Hey Bob, how's it going?
Bob: Pretty good. How about you?
Alice: Pretty good. What are you drinking?
Bob: A Bayeslight.
Alice: Cool. Hey bartender, lemme get one of those as well.
Alice: Hey, have you seen President Hanson's new immigration policy?
Bob: No. What is it?
Alice: He's shutting down the borders pretty hard. It's like an 8.5/10 on the left-right political axis, where 10/10 is ultra conservative and 0/10 is ultra liberal.[1]
Bob. Oh gawd. He's such an idiot.
Alice: Oh yeah? What sort of immigration policy would you implement?
Bob: Hm, I think something like a 2/10 on the left-right spectrum.
Alice: And how confident are you in that?
Bob: Probably like an 8/10.
Alice: Ok, let's call this 2/10 on the left-right spectrum belief with an 8/10 confidence a 2-8 belief. I disagree and will bet you $10 on it.
Bob: But how are we going to bet? What are we betting on? What's the resolution criteria?
Alice: Proposes some clever thing to bet on with crystal clear resolution criteria that perfectly captures everything.
Bob: Wow. Ok. Yeah, that works. But actually, after thinking about it a little more, I think I'm more like a 2.5-7. Can we bet at those odds instead?
Alice: We can. But only if we can bet $100 instead of $10.
Bob: Oh. Ummmm. Ok, I guess I'm probably more like a 3-6.5. I could do that at $100. Is that ok?
Alice: No. But I'd do 3-6.5 at $1,000. Shake on it?
Bob: Woah... Alice, this got real expensive real fast.
Alice: What's the matter Bob? Do you need to chug a few Bayeslights first?
Bob: Fuck you. I've got a family to feed over here.
Alice: Pansy. Wussy. That's a cop out and you know it.
Bob: Sigh. Ok, ok. Let me think about this. I am a Bayesian, and I do agree that Bayesians should be perfectly happy to bet on their beliefs when doing so would be +EV.
Alice: Right...
Bob: But... it's just that... well, $1,000 is kind of a lot of money.
Alice: Is it? You're semi-retired. You own your home. Your kids are in college. You h |
90e5f42e-e622-44bb-bdf4-2604f5996c6e | trentmkelly/LessWrong-43k | LessWrong | Meetup : San Antonio New Meetup
Discussion article for the meetup : San Antonio New Meetup
WHEN: 28 June 2015 03:00:00PM (-0500)
WHERE: Copalli Cafe 555 W Bitters Rd, San Antonio, TX 78216
Coffee and discussion at Copalli Cafe! All are welcome!
New Meetup to discuss rationality and all things LessWrong in San Antonio and meet the local community. Look for the sign that says Less Wrong!
Discussion article for the meetup : San Antonio New Meetup |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.