diff --git "a/parse/train/79zWncwO2p/79zWncwO2p_middle.json" "b/parse/train/79zWncwO2p/79zWncwO2p_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/79zWncwO2p/79zWncwO2p_middle.json"
@@ -0,0 +1,34736 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 121,
+ 97,
+ 491,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 97,
+ 491,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 97,
+ 491,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Collaborating with Humans without Human Data",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 122,
+ 159,
+ 489,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 159,
+ 480,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 159,
+ 480,
+ 172
+ ],
+ "score": 1.0,
+ "content": "DJ Strouse⇤, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett⇤",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 282,
+ 170,
+ 329,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 282,
+ 170,
+ 329,
+ 182
+ ],
+ "score": 1.0,
+ "content": "DeepMind",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 120,
+ 180,
+ 491,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 180,
+ 491,
+ 195
+ ],
+ "score": 1.0,
+ "content": "{strouse, kevinrmckee, botvinick, edwardhughes, reverett}@deepmind.com",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 221,
+ 328,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 220,
+ 330,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 220,
+ 330,
+ 235
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 246,
+ 469,
+ 465
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 246,
+ 471,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 246,
+ 471,
+ 260
+ ],
+ "score": 1.0,
+ "content": "Collaborating with humans requires rapidly adapting to their individual strengths,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 257,
+ 471,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 257,
+ 471,
+ 270
+ ],
+ "score": 1.0,
+ "content": "weaknesses, and preferences. Unfortunately, most standard multi-agent reinforce-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 268,
+ 470,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 268,
+ 470,
+ 281
+ ],
+ "score": 1.0,
+ "content": "ment learning techniques, such as self-play (SP) or population play (PP), produce",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 280,
+ 471,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 280,
+ 471,
+ 292
+ ],
+ "score": 1.0,
+ "content": "agents that overfit to their training partners and do not generalize well to humans.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 290,
+ 470,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 290,
+ 470,
+ 302
+ ],
+ "score": 1.0,
+ "content": "Alternatively, researchers can collect human data, train a human model using behav-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 302,
+ 470,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 302,
+ 470,
+ 313
+ ],
+ "score": 1.0,
+ "content": "ioral cloning, and then use that model to train “human-aware” agents (“behavioral",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 325
+ ],
+ "score": 1.0,
+ "content": "cloning play”, or BCP). While such an approach can improve the generalization",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 322,
+ 470,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 322,
+ 470,
+ 336
+ ],
+ "score": 1.0,
+ "content": "of agents to new human co-players, it involves the onerous and expensive step of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 142,
+ 335,
+ 469,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 335,
+ 469,
+ 347
+ ],
+ "score": 1.0,
+ "content": "collecting large amounts of human data first. Here, we study the problem of how",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 345,
+ 471,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 345,
+ 471,
+ 357
+ ],
+ "score": 1.0,
+ "content": "to train agents that collaborate well with human partners without using human data.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 355,
+ 470,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 355,
+ 470,
+ 369
+ ],
+ "score": 1.0,
+ "content": "We argue that the crux of the problem is to produce a diverse set of training part-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 367,
+ 470,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 367,
+ 470,
+ 380
+ ],
+ "score": 1.0,
+ "content": "ners. Drawing inspiration from successful multi-agent approaches in competitive",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 378,
+ 470,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 378,
+ 470,
+ 390
+ ],
+ "score": 1.0,
+ "content": "domains, we find that a surprisingly simple approach is highly effective. We train",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 389,
+ 470,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 389,
+ 470,
+ 401
+ ],
+ "score": 1.0,
+ "content": "our agent partner as the best response to a population of self-play agents and their",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 399,
+ 470,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 399,
+ 470,
+ 412
+ ],
+ "score": 1.0,
+ "content": "past checkpoints taken throughout training, a method we call Fictitious Co-Play",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 410,
+ 470,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 410,
+ 470,
+ 423
+ ],
+ "score": 1.0,
+ "content": "(FCP). Our experiments focus on a two-player collaborative cooking simulator that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 422,
+ 471,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 422,
+ 471,
+ 434
+ ],
+ "score": 1.0,
+ "content": "has recently been proposed as a challenge problem for coordination with humans.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 432,
+ 469,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 432,
+ 469,
+ 444
+ ],
+ "score": 1.0,
+ "content": "We find that FCP agents score significantly higher than SP, PP, and BCP when",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 444,
+ 470,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 444,
+ 470,
+ 456
+ ],
+ "score": 1.0,
+ "content": "paired with novel agent and human partners. Furthermore, humans also report a",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 455,
+ 453,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 455,
+ 453,
+ 466
+ ],
+ "score": 1.0,
+ "content": "strong subjective preference to partnering with FCP agents over all baselines.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 487,
+ 190,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 192,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 192,
+ 502
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 512,
+ 505,
+ 600
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "Generating agents which collaborate with novel partners is a longstanding challenge for Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 434,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 434,
+ 523,
+ 451,
+ 534
+ ],
+ "score": 0.26,
+ "content": "\\bar { 6 6 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 522,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "is especially",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "important in situations where an AI must generalize to novel human partners [6, 61]. Many successful",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 108,
+ 556,
+ 506,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 556,
+ 506,
+ 567
+ ],
+ "score": 1.0,
+ "content": "[12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach human-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "level using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64].",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "This begs the question: Can model-free RL without human data generate agents that can collaborate",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 188,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 589,
+ 188,
+ 600
+ ],
+ "score": 1.0,
+ "content": "with novel humans?",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 605,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "We seek an answer to this question in the space of common-payoff games, where all agents work",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "repeated games played against copies of itself, does not produce agents that generalize well to novel",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "coordinate with themselves, and so make for brittle and stubborn collaborators with new partners",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "who act differently. Population play (PP) trains a population of agents, all of whom interact with each",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 130,
+ 684
+ ],
+ "score": 1.0,
+ "content": "other",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 670,
+ 147,
+ 682
+ ],
+ "score": 0.65,
+ "content": "\\pmb { \\| 3 9 \\| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 148,
+ 669,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": ". While PP can generate agents capable of cooperation with humans in competitive team",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 683,
+ 134,
+ 693
+ ],
+ "score": 1.0,
+ "content": "games",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 682,
+ 152,
+ 693
+ ],
+ "score": 0.43,
+ "content": "\\textcircled { 1 3 4 } \\textcircled { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 681,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": ", it still fails to produce robust partners for novel humans in pure common-payoff settings",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 37.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 731,
+ 385,
+ 742
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 730,
+ 386,
+ 743
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 730,
+ 386,
+ 743
+ ],
+ "score": 1.0,
+ "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 120,
+ 702,
+ 192,
+ 712
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 700,
+ 194,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 700,
+ 194,
+ 713
+ ],
+ "score": 1.0,
+ "content": "⇤Equal contribution.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 121,
+ 97,
+ 491,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 97,
+ 491,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 97,
+ 491,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Collaborating with Humans without Human Data",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 122,
+ 159,
+ 489,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 159,
+ 480,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 159,
+ 480,
+ 172
+ ],
+ "score": 1.0,
+ "content": "DJ Strouse⇤, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett⇤",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 282,
+ 170,
+ 329,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 282,
+ 170,
+ 329,
+ 182
+ ],
+ "score": 1.0,
+ "content": "DeepMind",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 120,
+ 180,
+ 491,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 180,
+ 491,
+ 195
+ ],
+ "score": 1.0,
+ "content": "{strouse, kevinrmckee, botvinick, edwardhughes, reverett}@deepmind.com",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 120,
+ 159,
+ 491,
+ 195
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 221,
+ 328,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 220,
+ 330,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 220,
+ 330,
+ 235
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 246,
+ 469,
+ 465
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 246,
+ 471,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 246,
+ 471,
+ 260
+ ],
+ "score": 1.0,
+ "content": "Collaborating with humans requires rapidly adapting to their individual strengths,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 257,
+ 471,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 257,
+ 471,
+ 270
+ ],
+ "score": 1.0,
+ "content": "weaknesses, and preferences. Unfortunately, most standard multi-agent reinforce-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 268,
+ 470,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 268,
+ 470,
+ 281
+ ],
+ "score": 1.0,
+ "content": "ment learning techniques, such as self-play (SP) or population play (PP), produce",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 280,
+ 471,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 280,
+ 471,
+ 292
+ ],
+ "score": 1.0,
+ "content": "agents that overfit to their training partners and do not generalize well to humans.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 290,
+ 470,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 290,
+ 470,
+ 302
+ ],
+ "score": 1.0,
+ "content": "Alternatively, researchers can collect human data, train a human model using behav-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 302,
+ 470,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 302,
+ 470,
+ 313
+ ],
+ "score": 1.0,
+ "content": "ioral cloning, and then use that model to train “human-aware” agents (“behavioral",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 325
+ ],
+ "score": 1.0,
+ "content": "cloning play”, or BCP). While such an approach can improve the generalization",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 322,
+ 470,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 322,
+ 470,
+ 336
+ ],
+ "score": 1.0,
+ "content": "of agents to new human co-players, it involves the onerous and expensive step of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 142,
+ 335,
+ 469,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 335,
+ 469,
+ 347
+ ],
+ "score": 1.0,
+ "content": "collecting large amounts of human data first. Here, we study the problem of how",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 345,
+ 471,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 345,
+ 471,
+ 357
+ ],
+ "score": 1.0,
+ "content": "to train agents that collaborate well with human partners without using human data.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 355,
+ 470,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 355,
+ 470,
+ 369
+ ],
+ "score": 1.0,
+ "content": "We argue that the crux of the problem is to produce a diverse set of training part-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 367,
+ 470,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 367,
+ 470,
+ 380
+ ],
+ "score": 1.0,
+ "content": "ners. Drawing inspiration from successful multi-agent approaches in competitive",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 378,
+ 470,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 378,
+ 470,
+ 390
+ ],
+ "score": 1.0,
+ "content": "domains, we find that a surprisingly simple approach is highly effective. We train",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 389,
+ 470,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 389,
+ 470,
+ 401
+ ],
+ "score": 1.0,
+ "content": "our agent partner as the best response to a population of self-play agents and their",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 399,
+ 470,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 399,
+ 470,
+ 412
+ ],
+ "score": 1.0,
+ "content": "past checkpoints taken throughout training, a method we call Fictitious Co-Play",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 410,
+ 470,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 410,
+ 470,
+ 423
+ ],
+ "score": 1.0,
+ "content": "(FCP). Our experiments focus on a two-player collaborative cooking simulator that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 422,
+ 471,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 422,
+ 471,
+ 434
+ ],
+ "score": 1.0,
+ "content": "has recently been proposed as a challenge problem for coordination with humans.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 432,
+ 469,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 432,
+ 469,
+ 444
+ ],
+ "score": 1.0,
+ "content": "We find that FCP agents score significantly higher than SP, PP, and BCP when",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 444,
+ 470,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 444,
+ 470,
+ 456
+ ],
+ "score": 1.0,
+ "content": "paired with novel agent and human partners. Furthermore, humans also report a",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 455,
+ 453,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 455,
+ 453,
+ 466
+ ],
+ "score": 1.0,
+ "content": "strong subjective preference to partnering with FCP agents over all baselines.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 141,
+ 246,
+ 471,
+ 466
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 487,
+ 190,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 192,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 192,
+ 502
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 512,
+ 505,
+ 600
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "Generating agents which collaborate with novel partners is a longstanding challenge for Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 434,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 434,
+ 523,
+ 451,
+ 534
+ ],
+ "score": 0.26,
+ "content": "\\bar { 6 6 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 522,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "is especially",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "important in situations where an AI must generalize to novel human partners [6, 61]. Many successful",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 108,
+ 556,
+ 506,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 556,
+ 506,
+ 567
+ ],
+ "score": 1.0,
+ "content": "[12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach human-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "level using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64].",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "This begs the question: Can model-free RL without human data generate agents that can collaborate",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 188,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 589,
+ 188,
+ 600
+ ],
+ "score": 1.0,
+ "content": "with novel humans?",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 512,
+ 506,
+ 600
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 605,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "We seek an answer to this question in the space of common-payoff games, where all agents work",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "repeated games played against copies of itself, does not produce agents that generalize well to novel",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "coordinate with themselves, and so make for brittle and stubborn collaborators with new partners",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "who act differently. Population play (PP) trains a population of agents, all of whom interact with each",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 130,
+ 684
+ ],
+ "score": 1.0,
+ "content": "other",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 670,
+ 147,
+ 682
+ ],
+ "score": 0.65,
+ "content": "\\pmb { \\| 3 9 \\| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 148,
+ 669,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": ". While PP can generate agents capable of cooperation with humans in competitive team",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 683,
+ 134,
+ 693
+ ],
+ "score": 1.0,
+ "content": "games",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 682,
+ 152,
+ 693
+ ],
+ "score": 0.43,
+ "content": "\\textcircled { 1 3 4 } \\textcircled { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 681,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": ", it still fails to produce robust partners for novel humans in pure common-payoff settings",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "[12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 249,
+ 410,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 389,
+ 262
+ ],
+ "score": 1.0,
+ "content": "strategic diversity and producing agents not so different from self-play",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 390,
+ 249,
+ 407,
+ 261
+ ],
+ "score": 0.63,
+ "content": "\\dot { \\mathbb { B } } \\dot { \\mathbb { 4 } } \\mathbb { I }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 408,
+ 250,
+ 410,
+ 262
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 605,
+ 506,
+ 695
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "70b2b8c1e9478aed36abef766805e05c03561ba29533d4a7de7ad061fa2d261f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 109.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 117,
+ 109.0,
+ 493,
+ 148.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 117,
+ 148.0,
+ 493,
+ 187.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 194,
+ 505,
+ 229
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "Figure 1: In this work, we evaluate a variety of agent training methods (Section 2) in zero-shot",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 205,
+ 505,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 205,
+ 505,
+ 218
+ ],
+ "score": 1.0,
+ "content": "coordination with agents (Section 4). We then run a human-agent collaborative study designed to",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 217,
+ 300,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 217,
+ 300,
+ 229
+ ],
+ "score": 1.0,
+ "content": "elicit human preferences over agents (Section 5)",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 239,
+ 503,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "[12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 249,
+ 410,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 389,
+ 262
+ ],
+ "score": 1.0,
+ "content": "strategic diversity and producing agents not so different from self-play",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 390,
+ 249,
+ 407,
+ 261
+ ],
+ "score": 0.63,
+ "content": "\\dot { \\mathbb { B } } \\dot { \\mathbb { 4 } } \\mathbb { I }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 250,
+ 410,
+ 262
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 266,
+ 505,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 279
+ ],
+ "score": 1.0,
+ "content": "Our approach starts with the intuition that the key to producing robust agent collaborators is exposure",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 277,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 277,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "to diverse training partners. We find that a surprisingly simple strategy is effective in generating",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 222,
+ 301
+ ],
+ "score": 1.0,
+ "content": "sufficient diversity. We train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 289,
+ 233,
+ 298
+ ],
+ "score": 0.78,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 288,
+ 506,
+ 301
+ ],
+ "score": 1.0,
+ "content": "self-play agents varying only their random seed for neural network",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "initialization. Periodically during training, we save agent “checkpoints” representing their strategy",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "at that point in time. Then, we train an agent partner as the best-response to both the fully-trained",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "agents and their past checkpoints. The different checkpoints simulate different skill levels, and the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "different random seeds simulate breaking symmetries in different ways. We refer to this agent training",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 487,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 202,
+ 355
+ ],
+ "score": 1.0,
+ "content": "procedure as Fictitious",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 343,
+ 216,
+ 353
+ ],
+ "score": 0.26,
+ "content": "\\mathbf { C o }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 342,
+ 487,
+ 355
+ ],
+ "score": 1.0,
+ "content": "-Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69].",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 359,
+ 505,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 506,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 506,
+ 371
+ ],
+ "score": 1.0,
+ "content": "We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 237,
+ 383
+ ],
+ "score": 1.0,
+ "content": "Based on the game Overcooked",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 369,
+ 254,
+ 381
+ ],
+ "score": 0.63,
+ "content": "\\boldsymbol { \\left[ \\left[ 2 5 \\right] \\right] }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 369,
+ 506,
+ 383
+ ],
+ "score": 1.0,
+ "content": ", it has recently been proposed as a coordination challenge for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 207,
+ 405
+ ],
+ "score": 1.0,
+ "content": "humans was achieved in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 391,
+ 225,
+ 403
+ ],
+ "score": 0.66,
+ "content": "[ \\mathbb { 1 2 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 391,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "via behavioral cloning (BC) of human data. More precisely, BC was",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "used to produce models that can stand in as human proxies during training in simulation, a method",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 411,
+ 505,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 411,
+ 505,
+ 427
+ ],
+ "score": 1.0,
+ "content": "we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "both novel agent and human partners, and that humans express a significant preference for partnering",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 403,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 403,
+ 458
+ ],
+ "score": 1.0,
+ "content": "data for training, while achieving better outcomes for humans at test time.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 462,
+ 358,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 360,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 360,
+ 475
+ ],
+ "score": 1.0,
+ "content": "We summarize the novel contributions of this paper as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 129,
+ 483,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 494,
+ 230,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 494,
+ 230,
+ 506
+ ],
+ "score": 1.0,
+ "content": "humans (Section 2.1).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 128,
+ 508,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 508,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": "2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 519,
+ 381,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 519,
+ 381,
+ 532
+ ],
+ "score": 1.0,
+ "content": "coordination with a variety of held-out agents (Section 4.2).",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 128,
+ 532,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 532,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "3. We propose a rigorous human-agent interaction study with behavioral analysis and partici-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 543,
+ 255,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 545,
+ 235,
+ 556
+ ],
+ "score": 1.0,
+ "content": "pant feedback (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 543,
+ 252,
+ 556
+ ],
+ "score": 0.32,
+ "content": "\\boxed { 5 . 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 545,
+ 255,
+ 556
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 128,
+ 557,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 557,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 141,
+ 568,
+ 352,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 569,
+ 332,
+ 581
+ ],
+ "score": 1.0,
+ "content": "score and in human partner preference (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 568,
+ 350,
+ 581
+ ],
+ "score": 0.57,
+ "content": "5 . 2 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 569,
+ 352,
+ 581
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 595,
+ 171,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 592,
+ 172,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 592,
+ 172,
+ 611
+ ],
+ "score": 1.0,
+ "content": "2 Methods",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 619,
+ 236,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 237,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 237,
+ 634
+ ],
+ "score": 1.0,
+ "content": "2.1 Fictitious Co-Play (FCP)",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 640,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Diverse training conditions have been shown to make agents more robust, from environmental",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 264,
+ 663
+ ],
+ "score": 1.0,
+ "content": "variations (i.e. domain randomization",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 650,
+ 313,
+ 662
+ ],
+ "score": 0.43,
+ "content": "\\textcircled { 1 5 4 } , \\textcircled { 5 6 } , \\textcircled { 6 7 } \\textcircled { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 314,
+ 650,
+ 466,
+ 663
+ ],
+ "score": 1.0,
+ "content": "to heterogeneity in training partners",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 650,
+ 484,
+ 662
+ ],
+ "score": 0.42,
+ "content": "\\left[ \\left[ 6 9 \\right] \\right]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 650,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": ". We",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 661,
+ 506,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 661,
+ 506,
+ 674
+ ],
+ "score": 1.0,
+ "content": "seek to train agents that are robust partners for humans in common-payoff games, and so extend this",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 217,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 217,
+ 685
+ ],
+ "score": 1.0,
+ "content": "line of work to that setting.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 466,
+ 702
+ ],
+ "score": 1.0,
+ "content": "One important challenge in collaborating with novel partners is dealing with symmetries",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 688,
+ 484,
+ 700
+ ],
+ "score": 0.33,
+ "content": "\\textcircled { \\left| 3 1 \\right| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": ". For",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "example, two agents A and B facing each other may move past each other by A going left and B going",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 753
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 8
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 187
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "70b2b8c1e9478aed36abef766805e05c03561ba29533d4a7de7ad061fa2d261f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 117,
+ 70,
+ 493,
+ 109.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 117,
+ 109.0,
+ 493,
+ 148.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 117,
+ 148.0,
+ 493,
+ 187.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 194,
+ 505,
+ 229
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "Figure 1: In this work, we evaluate a variety of agent training methods (Section 2) in zero-shot",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 205,
+ 505,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 205,
+ 505,
+ 218
+ ],
+ "score": 1.0,
+ "content": "coordination with agents (Section 4). We then run a human-agent collaborative study designed to",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 217,
+ 300,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 217,
+ 300,
+ 229
+ ],
+ "score": 1.0,
+ "content": "elicit human preferences over agents (Section 5)",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 239,
+ 503,
+ 262
+ ],
+ "lines": [],
+ "index": 6.5,
+ "bbox_fs": [
+ 106,
+ 237,
+ 505,
+ 262
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 266,
+ 505,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 279
+ ],
+ "score": 1.0,
+ "content": "Our approach starts with the intuition that the key to producing robust agent collaborators is exposure",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 277,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 277,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "to diverse training partners. We find that a surprisingly simple strategy is effective in generating",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 222,
+ 301
+ ],
+ "score": 1.0,
+ "content": "sufficient diversity. We train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 289,
+ 233,
+ 298
+ ],
+ "score": 0.78,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 288,
+ 506,
+ 301
+ ],
+ "score": 1.0,
+ "content": "self-play agents varying only their random seed for neural network",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "initialization. Periodically during training, we save agent “checkpoints” representing their strategy",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "at that point in time. Then, we train an agent partner as the best-response to both the fully-trained",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "agents and their past checkpoints. The different checkpoints simulate different skill levels, and the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "different random seeds simulate breaking symmetries in different ways. We refer to this agent training",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 487,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 202,
+ 355
+ ],
+ "score": 1.0,
+ "content": "procedure as Fictitious",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 343,
+ 216,
+ 353
+ ],
+ "score": 0.26,
+ "content": "\\mathbf { C o }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 342,
+ 487,
+ 355
+ ],
+ "score": 1.0,
+ "content": "-Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69].",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 265,
+ 506,
+ 355
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 359,
+ 505,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 506,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 506,
+ 371
+ ],
+ "score": 1.0,
+ "content": "We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 237,
+ 383
+ ],
+ "score": 1.0,
+ "content": "Based on the game Overcooked",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 369,
+ 254,
+ 381
+ ],
+ "score": 0.63,
+ "content": "\\boldsymbol { \\left[ \\left[ 2 5 \\right] \\right] }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 369,
+ 506,
+ 383
+ ],
+ "score": 1.0,
+ "content": ", it has recently been proposed as a coordination challenge for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 207,
+ 405
+ ],
+ "score": 1.0,
+ "content": "humans was achieved in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 391,
+ 225,
+ 403
+ ],
+ "score": 0.66,
+ "content": "[ \\mathbb { 1 2 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 391,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "via behavioral cloning (BC) of human data. More precisely, BC was",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "used to produce models that can stand in as human proxies during training in simulation, a method",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 411,
+ 505,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 411,
+ 505,
+ 427
+ ],
+ "score": 1.0,
+ "content": "we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "both novel agent and human partners, and that humans express a significant preference for partnering",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 403,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 403,
+ 458
+ ],
+ "score": 1.0,
+ "content": "data for training, while achieving better outcomes for humans at test time.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 104,
+ 358,
+ 506,
+ 458
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 462,
+ 358,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 360,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 360,
+ 475
+ ],
+ "score": 1.0,
+ "content": "We summarize the novel contributions of this paper as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 106,
+ 461,
+ 360,
+ 475
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 129,
+ 483,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with",
+ "type": "text"
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 494,
+ 230,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 494,
+ 230,
+ 506
+ ],
+ "score": 1.0,
+ "content": "humans (Section 2.1).",
+ "type": "text"
+ }
+ ],
+ "index": 27,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 508,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 508,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": "2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 519,
+ 381,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 519,
+ 381,
+ 532
+ ],
+ "score": 1.0,
+ "content": "coordination with a variety of held-out agents (Section 4.2).",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 532,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 532,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "3. We propose a rigorous human-agent interaction study with behavioral analysis and partici-",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 543,
+ 255,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 545,
+ 235,
+ 556
+ ],
+ "score": 1.0,
+ "content": "pant feedback (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 543,
+ 252,
+ 556
+ ],
+ "score": 0.32,
+ "content": "\\boxed { 5 . 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 545,
+ 255,
+ 556
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 557,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 557,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 568,
+ 352,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 569,
+ 332,
+ 581
+ ],
+ "score": 1.0,
+ "content": "score and in human partner preference (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 568,
+ 350,
+ 581
+ ],
+ "score": 0.57,
+ "content": "5 . 2 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 569,
+ 352,
+ 581
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 128,
+ 483,
+ 506,
+ 581
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 595,
+ 171,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 592,
+ 172,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 592,
+ 172,
+ 611
+ ],
+ "score": 1.0,
+ "content": "2 Methods",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 619,
+ 236,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 237,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 237,
+ 634
+ ],
+ "score": 1.0,
+ "content": "2.1 Fictitious Co-Play (FCP)",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 640,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Diverse training conditions have been shown to make agents more robust, from environmental",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 264,
+ 663
+ ],
+ "score": 1.0,
+ "content": "variations (i.e. domain randomization",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 650,
+ 313,
+ 662
+ ],
+ "score": 0.43,
+ "content": "\\textcircled { 1 5 4 } , \\textcircled { 5 6 } , \\textcircled { 6 7 } \\textcircled { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 314,
+ 650,
+ 466,
+ 663
+ ],
+ "score": 1.0,
+ "content": "to heterogeneity in training partners",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 650,
+ 484,
+ 662
+ ],
+ "score": 0.42,
+ "content": "\\left[ \\left[ 6 9 \\right] \\right]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 650,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": ". We",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 661,
+ 506,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 661,
+ 506,
+ 674
+ ],
+ "score": 1.0,
+ "content": "seek to train agents that are robust partners for humans in common-payoff games, and so extend this",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 217,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 217,
+ 685
+ ],
+ "score": 1.0,
+ "content": "line of work to that setting.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 640,
+ 506,
+ 685
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 466,
+ 702
+ ],
+ "score": 1.0,
+ "content": "One important challenge in collaborating with novel partners is dealing with symmetries",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 688,
+ 484,
+ 700
+ ],
+ "score": 0.33,
+ "content": "\\textcircled { \\left| 3 1 \\right| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": ". For",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "example, two agents A and B facing each other may move past each other by A going left and B going",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "these conventions if a human clearly prefers one over the other. A second important challenge is",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 322,
+ 308,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 322,
+ 308,
+ 336
+ ],
+ "score": 1.0,
+ "content": "partners, as well as partners who are still learning.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 105,
+ 688,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "2372b76b2e16cbfb2db5439982c244e9b3e60b30aff99699aa5043d431598f6f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 119.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 119.66666666666666,
+ 495,
+ 169.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 118,
+ 169.33333333333331,
+ 495,
+ 218.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 225,
+ 505,
+ 291
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The four agent training methods we evaluate in this work. Self-play (SP) where an agent",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "learns with itself, population-play (PP) where a population of agents are co-trained together, and",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 260
+ ],
+ "score": 1.0,
+ "content": "behavioral cloning play (BCP) where data from human games is used to create a behaviorally",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 504,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 493,
+ 271
+ ],
+ "score": 1.0,
+ "content": "cloned agent with which an RL agent is then trained. In our method, Fictitious Co-Play (FCP),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 258,
+ 504,
+ 268
+ ],
+ "score": 0.68,
+ "content": "N",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "self-play agents are trained independently and checkpointed throughout training. An agent is then",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 438,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 438,
+ 292
+ ],
+ "score": 1.0,
+ "content": "trained to best respond to the entire population of SP agents and their checkpoints.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 300,
+ 503,
+ 334
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "these conventions if a human clearly prefers one over the other. A second important challenge is",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 322,
+ 308,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 322,
+ 308,
+ 336
+ ],
+ "score": 1.0,
+ "content": "partners, as well as partners who are still learning.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 338,
+ 505,
+ 426
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 506,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 506,
+ 352
+ ],
+ "score": 1.0,
+ "content": "Fictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 349,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 221,
+ 362
+ ],
+ "score": 1.0,
+ "content": "of these challenges (Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 349,
+ 231,
+ 362
+ ],
+ "score": 0.45,
+ "content": "\\bigstar",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 349,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": "right). In the first stage, we train a diverse pool of partners. To",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 432,
+ 373
+ ],
+ "score": 1.0,
+ "content": "allow the pool to represent different symmetry breaking conventions, we train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 361,
+ 443,
+ 370
+ ],
+ "score": 0.79,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 443,
+ 360,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "partner agents",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "in self-play. Since these partners are trained independently, they can arrive at different arbitrary",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "conventions for breaking symmetries. To allow the pool to represent different skill levels, we use",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 393,
+ 506,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 393,
+ 506,
+ 406
+ ],
+ "score": 1.0,
+ "content": "multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 506,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 506,
+ 417
+ ],
+ "score": 1.0,
+ "content": "fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 415,
+ 506,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 506,
+ 427
+ ],
+ "score": 1.0,
+ "content": "using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 431,
+ 505,
+ 497
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "In the second stage, we train an FCP agent as the best response to the pool of diverse partners created",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 452,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 452,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 506,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 506,
+ 477
+ ],
+ "score": 1.0,
+ "content": "lead of human partners, and learn a general policy across a range of strategies and skills. We call our",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 487
+ ],
+ "score": 1.0,
+ "content": "method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 456,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 456,
+ 498
+ ],
+ "score": 1.0,
+ "content": "trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69].",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 510,
+ 230,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 230,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 230,
+ 522
+ ],
+ "score": 1.0,
+ "content": "2.2 Baselines and ablations",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 529,
+ 505,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "We compare FCP agents to the three baseline training methods listed below, each varying only in",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 540,
+ 493,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 540,
+ 493,
+ 553
+ ],
+ "score": 1.0,
+ "content": "their set of training partners, with the RL algorithm and architecture consistent across all agents:",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 129,
+ 560,
+ 506,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 560,
+ 455,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 560,
+ 455,
+ 573
+ ],
+ "score": 1.0,
+ "content": "1. Self-play (SP), where agents learn solely through interaction with themselves.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 128,
+ 573,
+ 507,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 573,
+ 507,
+ 588
+ ],
+ "score": 1.0,
+ "content": "2. Population-play (PP), where a population of agents are co-trained through random pairings.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 128,
+ 587,
+ 506,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 587,
+ 506,
+ 602
+ ],
+ "score": 1.0,
+ "content": "3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12].",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 110,
+ 608,
+ 473,
+ 620
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 607,
+ 474,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 607,
+ 474,
+ 621
+ ],
+ "score": 1.0,
+ "content": "We also evaluate three variations on FCP to better understand the conditions for its success:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 627,
+ 505,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 627,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 627,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "1. To test the importance of including past checkpoints in training, we evaluate an ablation",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 141,
+ 639,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 639,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "of FCP in which agents are trained only with the converged checkpoints of their partners",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 142,
+ 650,
+ 273,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 650,
+ 145,
+ 661
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 650,
+ 177,
+ 662
+ ],
+ "score": 0.84,
+ "content": "\\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 650,
+ 273,
+ 661
+ ],
+ "score": 1.0,
+ "content": "for “FCP minus time”).",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 129,
+ 664,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 664,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "2. To test whether FCP would benefit from additional diversity in its partner population, we",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 142,
+ 676,
+ 505,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 676,
+ 505,
+ 687
+ ],
+ "score": 1.0,
+ "content": "evaluate an augmentation of FCP in which the population of SP partners varies not just in",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 141,
+ 686,
+ 485,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 686,
+ 294,
+ 697
+ ],
+ "score": 1.0,
+ "content": "random seed, but also in architecture",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 686,
+ 326,
+ 698
+ ],
+ "score": 0.88,
+ "content": "\\operatorname { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 686,
+ 485,
+ 697
+ ],
+ "score": 1.0,
+ "content": "for “FCP plus architectural variation”).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 128,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "3. To test whether architectural variation can serve as a full replacement for playing with past",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 141,
+ 710,
+ 455,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 710,
+ 399,
+ 724
+ ],
+ "score": 1.0,
+ "content": "checkpoints, we evaluate the combination of both modifications",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 711,
+ 450,
+ 723
+ ],
+ "score": 0.87,
+ "content": "( \\mathrm { F C P } _ { - T , + A } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 710,
+ 455,
+ 724
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 219
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "2372b76b2e16cbfb2db5439982c244e9b3e60b30aff99699aa5043d431598f6f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 70,
+ 495,
+ 119.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 119.66666666666666,
+ 495,
+ 169.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 118,
+ 169.33333333333331,
+ 495,
+ 218.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 225,
+ 505,
+ 291
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The four agent training methods we evaluate in this work. Self-play (SP) where an agent",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "learns with itself, population-play (PP) where a population of agents are co-trained together, and",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 260
+ ],
+ "score": 1.0,
+ "content": "behavioral cloning play (BCP) where data from human games is used to create a behaviorally",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 504,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 493,
+ 271
+ ],
+ "score": 1.0,
+ "content": "cloned agent with which an RL agent is then trained. In our method, Fictitious Co-Play (FCP),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 258,
+ 504,
+ 268
+ ],
+ "score": 0.68,
+ "content": "N",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "self-play agents are trained independently and checkpointed throughout training. An agent is then",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 438,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 438,
+ 292
+ ],
+ "score": 1.0,
+ "content": "trained to best respond to the entire population of SP agents and their checkpoints.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 300,
+ 503,
+ 334
+ ],
+ "lines": [],
+ "index": 10,
+ "bbox_fs": [
+ 106,
+ 299,
+ 505,
+ 336
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 338,
+ 505,
+ 426
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 506,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 506,
+ 352
+ ],
+ "score": 1.0,
+ "content": "Fictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 349,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 221,
+ 362
+ ],
+ "score": 1.0,
+ "content": "of these challenges (Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 349,
+ 231,
+ 362
+ ],
+ "score": 0.45,
+ "content": "\\bigstar",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 349,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": "right). In the first stage, we train a diverse pool of partners. To",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 432,
+ 373
+ ],
+ "score": 1.0,
+ "content": "allow the pool to represent different symmetry breaking conventions, we train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 361,
+ 443,
+ 370
+ ],
+ "score": 0.79,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 443,
+ 360,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "partner agents",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "in self-play. Since these partners are trained independently, they can arrive at different arbitrary",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "conventions for breaking symmetries. To allow the pool to represent different skill levels, we use",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 393,
+ 506,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 393,
+ 506,
+ 406
+ ],
+ "score": 1.0,
+ "content": "multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 506,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 506,
+ 417
+ ],
+ "score": 1.0,
+ "content": "fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 415,
+ 506,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 506,
+ 427
+ ],
+ "score": 1.0,
+ "content": "using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 338,
+ 506,
+ 427
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 431,
+ 505,
+ 497
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "In the second stage, we train an FCP agent as the best response to the pool of diverse partners created",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 452,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 452,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 506,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 506,
+ 477
+ ],
+ "score": 1.0,
+ "content": "lead of human partners, and learn a general policy across a range of strategies and skills. We call our",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 487
+ ],
+ "score": 1.0,
+ "content": "method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 456,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 456,
+ 498
+ ],
+ "score": 1.0,
+ "content": "trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69].",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 104,
+ 431,
+ 506,
+ 498
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 510,
+ 230,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 230,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 230,
+ 522
+ ],
+ "score": 1.0,
+ "content": "2.2 Baselines and ablations",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 529,
+ 505,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "We compare FCP agents to the three baseline training methods listed below, each varying only in",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 540,
+ 493,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 540,
+ 493,
+ 553
+ ],
+ "score": 1.0,
+ "content": "their set of training partners, with the RL algorithm and architecture consistent across all agents:",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 106,
+ 529,
+ 505,
+ 553
+ ]
+ },
+ {
+ "type": "index",
+ "bbox": [
+ 129,
+ 560,
+ 506,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 560,
+ 455,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 560,
+ 455,
+ 573
+ ],
+ "score": 1.0,
+ "content": "1. Self-play (SP), where agents learn solely through interaction with themselves.",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 573,
+ 507,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 573,
+ 507,
+ 588
+ ],
+ "score": 1.0,
+ "content": "2. Population-play (PP), where a population of agents are co-trained through random pairings.",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 587,
+ 506,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 587,
+ 506,
+ 602
+ ],
+ "score": 1.0,
+ "content": "3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12].",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 128,
+ 560,
+ 507,
+ 602
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 110,
+ 608,
+ 473,
+ 620
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 607,
+ 474,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 607,
+ 474,
+ 621
+ ],
+ "score": 1.0,
+ "content": "We also evaluate three variations on FCP to better understand the conditions for its success:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 107,
+ 607,
+ 474,
+ 621
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 627,
+ 505,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 627,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 627,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "1. To test the importance of including past checkpoints in training, we evaluate an ablation",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 141,
+ 639,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 639,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "of FCP in which agents are trained only with the converged checkpoints of their partners",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 142,
+ 650,
+ 273,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 650,
+ 145,
+ 661
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 650,
+ 177,
+ 662
+ ],
+ "score": 0.84,
+ "content": "\\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 650,
+ 273,
+ 661
+ ],
+ "score": 1.0,
+ "content": "for “FCP minus time”).",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 129,
+ 664,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 664,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "2. To test whether FCP would benefit from additional diversity in its partner population, we",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 142,
+ 676,
+ 505,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 676,
+ 505,
+ 687
+ ],
+ "score": 1.0,
+ "content": "evaluate an augmentation of FCP in which the population of SP partners varies not just in",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 141,
+ 686,
+ 485,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 686,
+ 294,
+ 697
+ ],
+ "score": 1.0,
+ "content": "random seed, but also in architecture",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 686,
+ 326,
+ 698
+ ],
+ "score": 0.88,
+ "content": "\\operatorname { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 686,
+ 485,
+ 697
+ ],
+ "score": 1.0,
+ "content": "for “FCP plus architectural variation”).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 128,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "3. To test whether architectural variation can serve as a full replacement for playing with past",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 141,
+ 710,
+ 455,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 710,
+ 399,
+ 724
+ ],
+ "score": 1.0,
+ "content": "checkpoints, we evaluate the combination of both modifications",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 711,
+ 450,
+ 723
+ ],
+ "score": 0.87,
+ "content": "( \\mathrm { F C P } _ { - T , + A } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 710,
+ 455,
+ 724
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 128,
+ 627,
+ 506,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 73,
+ 186,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 188,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 188,
+ 86
+ ],
+ "score": 1.0,
+ "content": "2.3 Environment",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 93,
+ 505,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": "Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 103,
+ 506,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 103,
+ 506,
+ 117
+ ],
+ "score": 1.0,
+ "content": "environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 322,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 322,
+ 127
+ ],
+ "score": 1.0,
+ "content": "in Carroll et al. [12]. For full details, see Appendix A.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "In this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 141,
+ 504,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 504,
+ 153
+ ],
+ "score": 1.0,
+ "content": "many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 153,
+ 505,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 153,
+ 505,
+ 165
+ ],
+ "score": 1.0,
+ "content": "high-level actions to which both players can contribute: collecting tomatoes, depositing them into",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 161,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 161,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 368,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 368,
+ 188
+ ],
+ "score": 1.0,
+ "content": "it. Upon a successful delivery, both players are rewarded equally.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 191,
+ 505,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 203
+ ],
+ "score": 1.0,
+ "content": "To effectively complete the task, players must learn to navigate the kitchen and interact with objects in",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 202,
+ 506,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 202,
+ 506,
+ 214
+ ],
+ "score": 1.0,
+ "content": "the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 495,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 495,
+ 225
+ ],
+ "score": 1.0,
+ "content": "This environment therefore presents the challenges of both movement and strategic coordination.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 229,
+ 504,
+ 263
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 242
+ ],
+ "score": 1.0,
+ "content": "Each player observes an egocentric RGB view of the world, and at every step can perform one",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 506,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 506,
+ 252
+ ],
+ "score": 1.0,
+ "content": "of six actions: stand still, move {up, down, left, right}, interact. The behavior ofLÈ¡áyÒįŘį\u000eºµòµÒ¡y®",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 250,
+ 476,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 476,
+ 263
+ ],
+ "score": 1.0,
+ "content": "interact varies based on the cell which the player is facing (e.g. place tomato on counter).",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "48e8af75423e0fa5e5c23e2088c49716c4c6044ee3579f26cdd9911dbb4b4925.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 321.6666666666667
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 118,
+ 321.6666666666667,
+ 492,
+ 370.33333333333337
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 118,
+ 370.33333333333337,
+ 492,
+ 419.00000000000006
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 426,
+ 503,
+ 448
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 504,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 504,
+ 438
+ ],
+ "score": 1.0,
+ "content": "Figure 3: The Overcooked environment: a two-player common-payoff game in which players must",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 436,
+ 254,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 436,
+ 254,
+ 450
+ ],
+ "score": 1.0,
+ "content": "coordinate to cook and deliver soup.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 17.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "score": 0.963,
+ "type": "image",
+ "image_path": "2b3253fcaac0677dbb15cc1ed5aa0663e561ff52605ba15bd61605f96c716872.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 479.6666666666667
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 120,
+ 479.6666666666667,
+ 492,
+ 500.33333333333337
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 120,
+ 500.33333333333337,
+ 492,
+ 521.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 104,
+ 527,
+ 504,
+ 550
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Layouts: the kitchens which agents and humans play in, each emphasizing different",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "coordination strategies. Highlighted in bold are the terms used to refer to each in the rest of this paper.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ }
+ ],
+ "index": 22.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 570,
+ 229,
+ 582
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 230,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 230,
+ 584
+ ],
+ "score": 1.0,
+ "content": "2.4 Implementation details",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 591,
+ 505,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 507,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 507,
+ 603
+ ],
+ "score": 1.0,
+ "content": "Here we highlight several key implementation details for our training methods. For full details,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 460,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 460,
+ 614
+ ],
+ "score": 1.0,
+ "content": "including the architectures, hyperparameters, and compute used, please see Appendix B.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 617,
+ 505,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 504,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 341,
+ 630
+ ],
+ "score": 1.0,
+ "content": "For our reinforcement learning agents, we use the V-MPO",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 617,
+ 359,
+ 629
+ ],
+ "score": 0.45,
+ "content": "\\pmb { \\Vert 6 5 \\Vert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 617,
+ 504,
+ 630
+ ],
+ "score": 1.0,
+ "content": "algorithm along with a ResNet [26]",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 155,
+ 641
+ ],
+ "score": 1.0,
+ "content": "plus LSTM",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 628,
+ 173,
+ 640
+ ],
+ "score": 0.7,
+ "content": "\\mathbb { \\left| \\bigstar \\bigstar \\right\\| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "architecture which we found led to optimal behavior across all layouts. Agents are",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 375,
+ 653
+ ],
+ "score": 1.0,
+ "content": "trained using a distributed set of environments running in parallel",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 639,
+ 393,
+ 651
+ ],
+ "score": 0.69,
+ "content": "\\textcircled { 1 1 7 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": ", each sampling two agents",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 650,
+ 346,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 650,
+ 346,
+ 663
+ ],
+ "score": 1.0,
+ "content": "from the training population to play together every episode.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 507,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 324,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Both PP and FCP are trained with a population size of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 667,
+ 357,
+ 677
+ ],
+ "score": 0.89,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 667,
+ 507,
+ 680
+ ],
+ "score": 1.0,
+ "content": "agents which are sampled uniformly.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "(1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 410,
+ 713
+ ],
+ "score": 1.0,
+ "content": "agent), and (3) at the middle of training, defined as when the agent reaches",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 700,
+ 430,
+ 711
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 431,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "of its final reward",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 455,
+ 723
+ ],
+ "score": 1.0,
+ "content": "(i.e. an average-skilled agent). When varying architecture for the training partners of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 711,
+ 487,
+ 722
+ ],
+ "score": 0.9,
+ "content": "\\operatorname { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 11,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 73,
+ 186,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 188,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 188,
+ 86
+ ],
+ "score": 1.0,
+ "content": "2.3 Environment",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 93,
+ 505,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": "Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 103,
+ 506,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 103,
+ 506,
+ 117
+ ],
+ "score": 1.0,
+ "content": "environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 322,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 322,
+ 127
+ ],
+ "score": 1.0,
+ "content": "in Carroll et al. [12]. For full details, see Appendix A.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 105,
+ 92,
+ 506,
+ 127
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "In this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 141,
+ 504,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 504,
+ 153
+ ],
+ "score": 1.0,
+ "content": "many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 153,
+ 505,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 153,
+ 505,
+ 165
+ ],
+ "score": 1.0,
+ "content": "high-level actions to which both players can contribute: collecting tomatoes, depositing them into",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 161,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 161,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 368,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 368,
+ 188
+ ],
+ "score": 1.0,
+ "content": "it. Upon a successful delivery, both players are rewarded equally.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 130,
+ 506,
+ 188
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 191,
+ 505,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 203
+ ],
+ "score": 1.0,
+ "content": "To effectively complete the task, players must learn to navigate the kitchen and interact with objects in",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 202,
+ 506,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 202,
+ 506,
+ 214
+ ],
+ "score": 1.0,
+ "content": "the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 495,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 495,
+ 225
+ ],
+ "score": 1.0,
+ "content": "This environment therefore presents the challenges of both movement and strategic coordination.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 191,
+ 506,
+ 225
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 229,
+ 504,
+ 263
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 242
+ ],
+ "score": 1.0,
+ "content": "Each player observes an egocentric RGB view of the world, and at every step can perform one",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 506,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 506,
+ 252
+ ],
+ "score": 1.0,
+ "content": "of six actions: stand still, move {up, down, left, right}, interact. The behavior ofLÈ¡áyÒįŘį\u000eºµòµÒ¡y®",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 250,
+ 476,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 476,
+ 263
+ ],
+ "score": 1.0,
+ "content": "interact varies based on the cell which the player is facing (e.g. place tomato on counter).",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 228,
+ 506,
+ 263
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 419
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "48e8af75423e0fa5e5c23e2088c49716c4c6044ee3579f26cdd9911dbb4b4925.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 273,
+ 492,
+ 321.6666666666667
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 118,
+ 321.6666666666667,
+ 492,
+ 370.33333333333337
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 118,
+ 370.33333333333337,
+ 492,
+ 419.00000000000006
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 426,
+ 503,
+ 448
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 504,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 504,
+ 438
+ ],
+ "score": 1.0,
+ "content": "Figure 3: The Overcooked environment: a two-player common-payoff game in which players must",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 436,
+ 254,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 436,
+ 254,
+ 450
+ ],
+ "score": 1.0,
+ "content": "coordinate to cook and deliver soup.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 17.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 521
+ ],
+ "score": 0.963,
+ "type": "image",
+ "image_path": "2b3253fcaac0677dbb15cc1ed5aa0663e561ff52605ba15bd61605f96c716872.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 120,
+ 459,
+ 492,
+ 479.6666666666667
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 120,
+ 479.6666666666667,
+ 492,
+ 500.33333333333337
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 120,
+ 500.33333333333337,
+ 492,
+ 521.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 104,
+ 527,
+ 504,
+ 550
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Layouts: the kitchens which agents and humans play in, each emphasizing different",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "coordination strategies. Highlighted in bold are the terms used to refer to each in the rest of this paper.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ }
+ ],
+ "index": 22.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 570,
+ 229,
+ 582
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 230,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 230,
+ 584
+ ],
+ "score": 1.0,
+ "content": "2.4 Implementation details",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 591,
+ 505,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 507,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 507,
+ 603
+ ],
+ "score": 1.0,
+ "content": "Here we highlight several key implementation details for our training methods. For full details,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 460,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 460,
+ 614
+ ],
+ "score": 1.0,
+ "content": "including the architectures, hyperparameters, and compute used, please see Appendix B.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 591,
+ 507,
+ 614
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 617,
+ 505,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 504,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 341,
+ 630
+ ],
+ "score": 1.0,
+ "content": "For our reinforcement learning agents, we use the V-MPO",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 617,
+ 359,
+ 629
+ ],
+ "score": 0.45,
+ "content": "\\pmb { \\Vert 6 5 \\Vert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 617,
+ 504,
+ 630
+ ],
+ "score": 1.0,
+ "content": "algorithm along with a ResNet [26]",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 155,
+ 641
+ ],
+ "score": 1.0,
+ "content": "plus LSTM",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 628,
+ 173,
+ 640
+ ],
+ "score": 0.7,
+ "content": "\\mathbb { \\left| \\bigstar \\bigstar \\right\\| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "architecture which we found led to optimal behavior across all layouts. Agents are",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 375,
+ 653
+ ],
+ "score": 1.0,
+ "content": "trained using a distributed set of environments running in parallel",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 639,
+ 393,
+ 651
+ ],
+ "score": 0.69,
+ "content": "\\textcircled { 1 1 7 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": ", each sampling two agents",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 650,
+ 346,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 650,
+ 346,
+ 663
+ ],
+ "score": 1.0,
+ "content": "from the training population to play together every episode.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 617,
+ 505,
+ 663
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 507,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 324,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Both PP and FCP are trained with a population size of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 667,
+ 357,
+ 677
+ ],
+ "score": 0.89,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 667,
+ 507,
+ 680
+ ],
+ "score": 1.0,
+ "content": "agents which are sampled uniformly.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "(1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 410,
+ 713
+ ],
+ "score": 1.0,
+ "content": "agent), and (3) at the middle of training, defined as when the agent reaches",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 700,
+ 430,
+ 711
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 431,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "of its final reward",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 455,
+ 723
+ ],
+ "score": 1.0,
+ "content": "(i.e. an average-skilled agent). When varying architecture for the training partners of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 711,
+ 487,
+ 722
+ ],
+ "score": 0.9,
+ "content": "\\operatorname { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 152,
+ 85
+ ],
+ "score": 0.85,
+ "content": "\\mathrm { F C P } _ { - T , + A }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 153,
+ 72,
+ 505,
+ 86
+ ],
+ "score": 1.0,
+ "content": "variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 95,
+ 499,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 95,
+ 306,
+ 107
+ ],
+ "score": 1.0,
+ "content": "combinations, leaving the total population size of",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 306,
+ 95,
+ 339,
+ 105
+ ],
+ "score": 0.89,
+ "content": "N = 3 2",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 340,
+ 95,
+ 499,
+ 107
+ ],
+ "score": 1.0,
+ "content": "unchanged, ensuring a fair comparison.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 667,
+ 507,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 73,
+ 505,
+ 106
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 152,
+ 85
+ ],
+ "score": 0.85,
+ "content": "\\mathrm { F C P } _ { - T , + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 72,
+ 505,
+ 86
+ ],
+ "score": 1.0,
+ "content": "variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 95,
+ 499,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 95,
+ 306,
+ 107
+ ],
+ "score": 1.0,
+ "content": "combinations, leaving the total population size of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 95,
+ 339,
+ 105
+ ],
+ "score": 0.89,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 95,
+ 499,
+ 107
+ ],
+ "score": 1.0,
+ "content": "unchanged, ensuring a fair comparison.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 505,
+ 188
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 257,
+ 124
+ ],
+ "score": 1.0,
+ "content": "To train agents via behavioral cloning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 110,
+ 274,
+ 122
+ ],
+ "score": 0.79,
+ "content": "\\left[ \\left[ 5 8 \\right] \\right]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 110,
+ 398,
+ 124
+ ],
+ "score": 1.0,
+ "content": ", we use the open-source Acme",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 110,
+ 416,
+ 122
+ ],
+ "score": 0.77,
+ "content": "\\pmb { \\mathbb { B } } 0 \\|",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 110,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "to learn a policy from",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 122,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 298,
+ 167
+ ],
+ "score": 1.0,
+ "content": "agent-agent evaluation. Following Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 154,
+ 316,
+ 166
+ ],
+ "score": 0.61,
+ "content": "\\mathbb { \\overline { { \\lVert \\lambda \\rVert } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 154,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": ", we use a set of feature-based observations for",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 452,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 452,
+ 189
+ ],
+ "score": 1.0,
+ "content": "(asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced).",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 206,
+ 194,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 196,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 196,
+ 221
+ ],
+ "score": 1.0,
+ "content": "3 Related work",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 233,
+ 456,
+ 245
+ ],
+ "score": 1.0,
+ "content": "Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 232,
+ 483,
+ 245
+ ],
+ "score": 0.73,
+ "content": " { \\mathbb { B } } , { \\mathbb { G } } 6 { \\mathbb { I } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 233,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": ", also",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 243,
+ 257
+ ],
+ "score": 1.0,
+ "content": "known as zero-shot coordination",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 244,
+ 260,
+ 256
+ ],
+ "score": 0.49,
+ "content": "[ \\overbrace { 3 \\mathrm { 1 } } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 244,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": ". Prior work based in game-theoretic settings has suggested",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 255,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 255,
+ 205,
+ 267
+ ],
+ "score": 1.0,
+ "content": "the benefits of planning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 255,
+ 223,
+ 267
+ ],
+ "score": 0.64,
+ "content": "\\mathbb { [ [ \\mathrm { { 7 1 } ] } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 255,
+ 291,
+ 267
+ ],
+ "score": 1.0,
+ "content": ", online learning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 255,
+ 309,
+ 267
+ ],
+ "score": 0.35,
+ "content": "\\mathbb { \\left[ \\left. 5 1 \\right] \\right. }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 255,
+ 428,
+ 267
+ ],
+ "score": 1.0,
+ "content": ", and novel solution concepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 255,
+ 441,
+ 266
+ ],
+ "score": 0.78,
+ "content": "\\bar { \\mathbb { D } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 442,
+ 255,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": ", to name a few",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "score": 1.0,
+ "content": "examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "to more complex gridworld or continuous control settings, leading to work on hierarchical social",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 144,
+ 302
+ ],
+ "score": 1.0,
+ "content": "planning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 288,
+ 162,
+ 299
+ ],
+ "score": 0.76,
+ "content": "\\widehat { \\left\\| 3 6 \\right\\| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 286,
+ 324,
+ 302
+ ],
+ "score": 1.0,
+ "content": ", adapting to existing social conventions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 288,
+ 356,
+ 299
+ ],
+ "score": 0.34,
+ "content": "\\checkmark",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 286,
+ 438,
+ 302
+ ],
+ "score": 1.0,
+ "content": ", trajectory diversity",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 439,
+ 288,
+ 456,
+ 299
+ ],
+ "score": 0.6,
+ "content": "\\lVert \\boldsymbol { \\mathsf { A } } \\boldsymbol { \\mathsf { S } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 286,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": ", and theory",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 345,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 345,
+ 333
+ ],
+ "score": 1.0,
+ "content": "to held-out agent partners as a proxy for human co-players.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 337,
+ 505,
+ 381
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "Collaborative play with novel humans has been evaluated more actively in the context of training",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 347,
+ 507,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 347,
+ 507,
+ 361
+ ],
+ "score": 1.0,
+ "content": "agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the state-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "of-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 426,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 426,
+ 381
+ ],
+ "score": 1.0,
+ "content": "gridworld environment, as measured by the challenge tasks in Carroll et al. [12].",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 386,
+ 505,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "Diversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "that train with behaviorally diverse populations of game partners tend to demonstrate stronger",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 421
+ ],
+ "score": 1.0,
+ "content": "performance than their self-play counterparts. For example, across a range of multi-agent games,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "generalization to held-out populations can be improved by training larger and more diverse populations",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "[13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 472,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 422,
+ 464
+ ],
+ "score": 1.0,
+ "content": "optimized through selective matchmaking between increasingly diverse agents",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 451,
+ 469,
+ 463
+ ],
+ "score": 0.6,
+ "content": "\\pm \\boxed { 1 2 4 } \\boxed { 3 9 } \\boxed { 6 9 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 451,
+ 472,
+ 464
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 468,
+ 505,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "Despite the increased focus on improving multi-agent performance, evaluation has typically been",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "constrained to agent-agent settings. High-performing agents have infrequently been evaluated with",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 308,
+ 504
+ ],
+ "score": 1.0,
+ "content": "humans, particularly in non-competitive domains",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 489,
+ 326,
+ 501
+ ],
+ "score": 0.66,
+ "content": "\\dot { \\left. \\overline { { \\dot { \\left. \\dot { \\theta } \\right\\| } } } \\right. }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 489,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": ". We add to this growing literature, showing",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 471,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 471,
+ 513
+ ],
+ "score": 1.0,
+ "content": "that training with diversity is a powerful approach for effective human-agent collaboration.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 531
+ ],
+ "score": 1.0,
+ "content": "Human-agent interaction In recent years, increased attention has been directed toward designing",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 465,
+ 541
+ ],
+ "score": 1.0,
+ "content": "machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 528,
+ 482,
+ 540
+ ],
+ "score": 0.76,
+ "content": "\\textcircled { 1 1 6 } \\textcircled { }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 527,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "for a",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 304,
+ 552
+ ],
+ "score": 1.0,
+ "content": "broader review on Cooperative AI). Tylkin et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 305,
+ 539,
+ 322,
+ 550
+ ],
+ "score": 0.64,
+ "content": "\\lVert \\overline { { 6 8 } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 537,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "is particularly notable in also demonstrating",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "that partially trained agents can be useful learning targets for human helpers, although in a different",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 331,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 331,
+ 595
+ ],
+ "score": 1.0,
+ "content": "our agents’ performance (Tables 1 and 2 and Figure 7b).",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 599,
+ 505,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 343,
+ 612
+ ],
+ "score": 1.0,
+ "content": "A key preceding entry in this research area is Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 599,
+ 361,
+ 610
+ ],
+ "score": 0.58,
+ "content": "[ \\mathbb { 1 2 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 361,
+ 599,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": ", who similarly investigated human-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 507,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 507,
+ 623
+ ],
+ "score": 1.0,
+ "content": "agent coordination in Overcooked. We use their method (BCP) as a baseline throughout our exper-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 172,
+ 633
+ ],
+ "score": 1.0,
+ "content": "iments (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 620,
+ 191,
+ 633
+ ],
+ "score": 0.44,
+ "content": "\\boxed { 2 . 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 191,
+ 621,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": ". Relative to BCP, our approach removes the need for the expensive step of",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 632,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 632,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "human data collection for agent training. Furthermore, through our novel human-agent experimental",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "design, we go beyond objective performance metrics to compare the subjective preferences that",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 441,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 441,
+ 666
+ ],
+ "score": 1.0,
+ "content": "agents generate. For a detailed comparison of methods and results, see Appendix E.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 44.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 684,
+ 305,
+ 697
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 306,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 306,
+ 700
+ ],
+ "score": 1.0,
+ "content": "4 Zero-shot coordination with agents",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 710,
+ 487,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 489,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 489,
+ 723
+ ],
+ "score": 1.0,
+ "content": "In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 73,
+ 505,
+ 106
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 72,
+ 505,
+ 107
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 505,
+ 188
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 257,
+ 124
+ ],
+ "score": 1.0,
+ "content": "To train agents via behavioral cloning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 110,
+ 274,
+ 122
+ ],
+ "score": 0.79,
+ "content": "\\left[ \\left[ 5 8 \\right] \\right]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 110,
+ 398,
+ 124
+ ],
+ "score": 1.0,
+ "content": ", we use the open-source Acme",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 110,
+ 416,
+ 122
+ ],
+ "score": 0.77,
+ "content": "\\pmb { \\mathbb { B } } 0 \\|",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 110,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "to learn a policy from",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 122,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 298,
+ 167
+ ],
+ "score": 1.0,
+ "content": "agent-agent evaluation. Following Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 154,
+ 316,
+ 166
+ ],
+ "score": 0.61,
+ "content": "\\mathbb { \\overline { { \\lVert \\lambda \\rVert } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 154,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": ", we use a set of feature-based observations for",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 452,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 452,
+ 189
+ ],
+ "score": 1.0,
+ "content": "(asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced).",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 110,
+ 506,
+ 189
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 206,
+ 194,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 196,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 196,
+ 221
+ ],
+ "score": 1.0,
+ "content": "3 Related work",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 233,
+ 456,
+ 245
+ ],
+ "score": 1.0,
+ "content": "Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 232,
+ 483,
+ 245
+ ],
+ "score": 0.73,
+ "content": " { \\mathbb { B } } , { \\mathbb { G } } 6 { \\mathbb { I } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 233,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": ", also",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 243,
+ 257
+ ],
+ "score": 1.0,
+ "content": "known as zero-shot coordination",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 244,
+ 260,
+ 256
+ ],
+ "score": 0.49,
+ "content": "[ \\overbrace { 3 \\mathrm { 1 } } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 244,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": ". Prior work based in game-theoretic settings has suggested",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 255,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 255,
+ 205,
+ 267
+ ],
+ "score": 1.0,
+ "content": "the benefits of planning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 255,
+ 223,
+ 267
+ ],
+ "score": 0.64,
+ "content": "\\mathbb { [ [ \\mathrm { { 7 1 } ] } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 255,
+ 291,
+ 267
+ ],
+ "score": 1.0,
+ "content": ", online learning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 255,
+ 309,
+ 267
+ ],
+ "score": 0.35,
+ "content": "\\mathbb { \\left[ \\left. 5 1 \\right] \\right. }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 255,
+ 428,
+ 267
+ ],
+ "score": 1.0,
+ "content": ", and novel solution concepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 255,
+ 441,
+ 266
+ ],
+ "score": 0.78,
+ "content": "\\bar { \\mathbb { D } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 442,
+ 255,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": ", to name a few",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "score": 1.0,
+ "content": "examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "to more complex gridworld or continuous control settings, leading to work on hierarchical social",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 144,
+ 302
+ ],
+ "score": 1.0,
+ "content": "planning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 288,
+ 162,
+ 299
+ ],
+ "score": 0.76,
+ "content": "\\widehat { \\left\\| 3 6 \\right\\| }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 286,
+ 324,
+ 302
+ ],
+ "score": 1.0,
+ "content": ", adapting to existing social conventions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 288,
+ 356,
+ 299
+ ],
+ "score": 0.34,
+ "content": "\\checkmark",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 286,
+ 438,
+ 302
+ ],
+ "score": 1.0,
+ "content": ", trajectory diversity",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 439,
+ 288,
+ 456,
+ 299
+ ],
+ "score": 0.6,
+ "content": "\\lVert \\boldsymbol { \\mathsf { A } } \\boldsymbol { \\mathsf { S } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 286,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": ", and theory",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 345,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 345,
+ 333
+ ],
+ "score": 1.0,
+ "content": "to held-out agent partners as a proxy for human co-players.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 232,
+ 506,
+ 333
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 337,
+ 505,
+ 381
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "Collaborative play with novel humans has been evaluated more actively in the context of training",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 347,
+ 507,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 347,
+ 507,
+ 361
+ ],
+ "score": 1.0,
+ "content": "agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the state-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "of-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 426,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 426,
+ 381
+ ],
+ "score": 1.0,
+ "content": "gridworld environment, as measured by the challenge tasks in Carroll et al. [12].",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 104,
+ 335,
+ 507,
+ 381
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 386,
+ 505,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "Diversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "that train with behaviorally diverse populations of game partners tend to demonstrate stronger",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 421
+ ],
+ "score": 1.0,
+ "content": "performance than their self-play counterparts. For example, across a range of multi-agent games,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "generalization to held-out populations can be improved by training larger and more diverse populations",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "[13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 472,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 422,
+ 464
+ ],
+ "score": 1.0,
+ "content": "optimized through selective matchmaking between increasingly diverse agents",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 451,
+ 469,
+ 463
+ ],
+ "score": 0.6,
+ "content": "\\pm \\boxed { 1 2 4 } \\boxed { 3 9 } \\boxed { 6 9 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 451,
+ 472,
+ 464
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 384,
+ 506,
+ 464
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 468,
+ 505,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "Despite the increased focus on improving multi-agent performance, evaluation has typically been",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "constrained to agent-agent settings. High-performing agents have infrequently been evaluated with",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 308,
+ 504
+ ],
+ "score": 1.0,
+ "content": "humans, particularly in non-competitive domains",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 489,
+ 326,
+ 501
+ ],
+ "score": 0.66,
+ "content": "\\dot { \\left. \\overline { { \\dot { \\left. \\dot { \\theta } \\right\\| } } } \\right. }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 489,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": ". We add to this growing literature, showing",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 471,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 471,
+ 513
+ ],
+ "score": 1.0,
+ "content": "that training with diversity is a powerful approach for effective human-agent collaboration.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 468,
+ 505,
+ 513
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 531
+ ],
+ "score": 1.0,
+ "content": "Human-agent interaction In recent years, increased attention has been directed toward designing",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 465,
+ 541
+ ],
+ "score": 1.0,
+ "content": "machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 528,
+ 482,
+ 540
+ ],
+ "score": 0.76,
+ "content": "\\textcircled { 1 1 6 } \\textcircled { }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 527,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "for a",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 304,
+ 552
+ ],
+ "score": 1.0,
+ "content": "broader review on Cooperative AI). Tylkin et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 305,
+ 539,
+ 322,
+ 550
+ ],
+ "score": 0.64,
+ "content": "\\lVert \\overline { { 6 8 } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 537,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "is particularly notable in also demonstrating",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "that partially trained agents can be useful learning targets for human helpers, although in a different",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 331,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 331,
+ 595
+ ],
+ "score": 1.0,
+ "content": "our agents’ performance (Tables 1 and 2 and Figure 7b).",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 515,
+ 506,
+ 595
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 599,
+ 505,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 343,
+ 612
+ ],
+ "score": 1.0,
+ "content": "A key preceding entry in this research area is Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 599,
+ 361,
+ 610
+ ],
+ "score": 0.58,
+ "content": "[ \\mathbb { 1 2 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 361,
+ 599,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": ", who similarly investigated human-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 507,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 507,
+ 623
+ ],
+ "score": 1.0,
+ "content": "agent coordination in Overcooked. We use their method (BCP) as a baseline throughout our exper-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 172,
+ 633
+ ],
+ "score": 1.0,
+ "content": "iments (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 620,
+ 191,
+ 633
+ ],
+ "score": 0.44,
+ "content": "\\boxed { 2 . 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 191,
+ 621,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": ". Relative to BCP, our approach removes the need for the expensive step of",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 632,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 632,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "human data collection for agent training. Furthermore, through our novel human-agent experimental",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "design, we go beyond objective performance metrics to compare the subjective preferences that",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 441,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 441,
+ 666
+ ],
+ "score": 1.0,
+ "content": "agents generate. For a detailed comparison of methods and results, see Appendix E.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 105,
+ 599,
+ 507,
+ 666
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 684,
+ 305,
+ 697
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 306,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 682,
+ 306,
+ 700
+ ],
+ "score": 1.0,
+ "content": "4 Zero-shot coordination with agents",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 710,
+ 487,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 489,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 489,
+ 723
+ ],
+ "score": 1.0,
+ "content": "In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 709,
+ 489,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 72,
+ 406,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 407,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 407,
+ 87
+ ],
+ "score": 1.0,
+ "content": "4.1 Evaluation method: collaborative evaluation with agent partners",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 92,
+ 505,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": "Our primary concern in this work is generalization to novel human partners (as investigated in",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 103,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 138,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 103,
+ 148,
+ 116
+ ],
+ "score": 0.35,
+ "content": "\\textcircled{5}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 148,
+ 103,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": ". However, just as collecting human-human data for behavioral cloning is expensive, so",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 125,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 125,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "partners as a cheap proxy of performance with humans. This is then used to guide our model selection",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "process, allowing us to be more targeted with the agents we select for our human-agent evaluations.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 153,
+ 286,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 151,
+ 286,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 151,
+ 286,
+ 166
+ ],
+ "score": 1.0,
+ "content": "We evaluate with three held-out populations:",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 129,
+ 172,
+ 506,
+ 246
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 171,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 171,
+ 283,
+ 185
+ ],
+ "score": 1.0,
+ "content": "1. A BC model trained on human data,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 284,
+ 173,
+ 312,
+ 185
+ ],
+ "score": 0.91,
+ "content": "H _ { \\mathrm { p r o x y } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 171,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": ", intended as a proxy of generalization to humans,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 182,
+ 261,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 182,
+ 240,
+ 196
+ ],
+ "score": 1.0,
+ "content": "as done by Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 183,
+ 258,
+ 195
+ ],
+ "score": 0.29,
+ "content": "[ \\overbrace { | 1 2 | }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 182,
+ 261,
+ 196
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 128,
+ 196,
+ 507,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 196,
+ 507,
+ 211
+ ],
+ "score": 1.0,
+ "content": "2. A set of self-play agents varying in seed, architecture, and training time (specifically, held-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 207,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 208,
+ 207,
+ 222
+ ],
+ "score": 1.0,
+ "content": "out seeds of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 209,
+ 241,
+ 219
+ ],
+ "score": 0.89,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 208,
+ 335,
+ 222
+ ],
+ "score": 1.0,
+ "content": "partners trained for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 209,
+ 367,
+ 220
+ ],
+ "score": 0.9,
+ "content": "\\mathrm { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 208,
+ 460,
+ 222
+ ],
+ "score": 1.0,
+ "content": "agent; see Section 2.4).",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 462,
+ 207,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "These are",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 142,
+ 219,
+ 423,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 219,
+ 423,
+ 232
+ ],
+ "score": 1.0,
+ "content": "intended to test generalization to a diverse yet still skillful population.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 129,
+ 234,
+ 464,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 234,
+ 464,
+ 246
+ ],
+ "score": 1.0,
+ "content": "3. Randomly initialized agents intended to test generalization to low-skill partners.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 506,
+ 298
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "For all results, we report the average number of deliveries made by both players within an episode,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 330,
+ 277
+ ],
+ "score": 1.0,
+ "content": "aggregated across the 5 different layouts from Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 331,
+ 264,
+ 341,
+ 277
+ ],
+ "score": 0.74,
+ "content": "\\boxed { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 264,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "(with the per-layout results reported in",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 146,
+ 288
+ ],
+ "score": 1.0,
+ "content": "Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 275,
+ 166,
+ 288
+ ],
+ "score": 0.51,
+ "content": "\\underline { { \\overline { { ( \\mathrm { C . 2 } ) } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 276,
+ 167,
+ 288
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 275,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "We estimate mean and standard deviation across 5 random seeds. For each seed, we",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 299
+ ],
+ "score": 1.0,
+ "content": "evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 310,
+ 162,
+ 322
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 163,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 163,
+ 323
+ ],
+ "score": 1.0,
+ "content": "4.2 Results",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 330,
+ 336,
+ 342
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 338,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 338,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Finding 1: FCP significantly outperforms all baselines",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 343,
+ 505,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "To begin, we compare our FCP agent and the baselines when partnered with the three held-out",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 354,
+ 323,
+ 367
+ ],
+ "score": 1.0,
+ "content": "populations introduced above. As can be seen in Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 353,
+ 334,
+ 367
+ ],
+ "score": 0.61,
+ "content": "\\boxed { 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 354,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "FCP significantly outperforms all baselines",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 103,
+ 361,
+ 502,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 361,
+ 473,
+ 381
+ ],
+ "score": 1.0,
+ "content": "when partnered with all three held-out populations. Notably, it performs better than BCP with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 365,
+ 502,
+ 377
+ ],
+ "score": 0.88,
+ "content": "H _ { \\mathrm { p r o x y } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 451,
+ 389
+ ],
+ "score": 1.0,
+ "content": "even though BCP trains with such a model and FCP does not. Similar to Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 451,
+ 375,
+ 469,
+ 387
+ ],
+ "score": 0.28,
+ "content": "\\mathbb { \\lVert 1 2 \\rVert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 469,
+ 375,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": ", we find",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 387,
+ 252,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 387,
+ 252,
+ 399
+ ],
+ "score": 1.0,
+ "content": "that BCP significantly outscores SP.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 506,
+ 415
+ ],
+ "score": 1.0,
+ "content": "When paired with a randomly initialized partner which behaves suboptimally, we see an even greater",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 506,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 506,
+ 426
+ ],
+ "score": 1.0,
+ "content": "difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "such agents, it may not be surprising that it does so well with partners that behave poorly. However,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "what is surprising is how brittle the other training methods are. This suggests that they may not",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 480,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 480,
+ 460
+ ],
+ "score": 1.0,
+ "content": "perform well with humans who are not highly skilled players, which we will see in Section 5.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "8083dcac415b6ef9624e1aa4f428a0a33c3e9d47ce8903df3798f3e80ef0427c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 499.6666666666667
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 101,
+ 499.6666666666667,
+ 495,
+ 529.3333333333334
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 101,
+ 529.3333333333334,
+ 495,
+ 559.0
+ ],
+ "spans": [],
+ "index": 31
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 565,
+ 506,
+ 622
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Agent-agent collaborative evaluation: Performance of each agent when partnered with",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 277,
+ 588
+ ],
+ "score": 1.0,
+ "content": "each of the held-out populations (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 575,
+ 296,
+ 588
+ ],
+ "score": 0.62,
+ "content": "4 . 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 576,
+ 387,
+ 588
+ ],
+ "score": 1.0,
+ "content": "in episodes of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 387,
+ 576,
+ 426,
+ 586
+ ],
+ "score": 0.9,
+ "content": "T = 5 4 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 576,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": ". Importantly, FCP",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "scores higher than all baselines with a variety of test partners. Error bars represent standard deviation",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "over five random training seeds. Plots aggregate data across kitchen layouts; results calculated by",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 607,
+ 302,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 302,
+ 623
+ ],
+ "score": 1.0,
+ "content": "individual layout can be found in Appendix C.2.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 637,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 491,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 491,
+ 652
+ ],
+ "score": 1.0,
+ "content": "Finding 2: Training with past checkpoints is the most beneficial variation for performance",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "Next, we investigate how the different training partner variations influence FCP’s performance. In",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 660,
+ 506,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 660,
+ 317,
+ 676
+ ],
+ "score": 1.0,
+ "content": "particular, we separately ablate the past checkpoints",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 662,
+ 332,
+ 673
+ ],
+ "score": 0.72,
+ "content": "( T )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 660,
+ 401,
+ 676
+ ],
+ "score": 1.0,
+ "content": "and architecture",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 662,
+ 416,
+ 673
+ ],
+ "score": 0.68,
+ "content": "( A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 660,
+ 506,
+ 676
+ ],
+ "score": 1.0,
+ "content": "variations, evaluating",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 498,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 498,
+ 687
+ ],
+ "score": 1.0,
+ "content": "them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 210,
+ 703
+ ],
+ "score": 1.0,
+ "content": "Comparing the FCP and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 689,
+ 242,
+ 700
+ ],
+ "score": 0.83,
+ "content": "\\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 687,
+ 506,
+ 703
+ ],
+ "score": 1.0,
+ "content": "columns, we see that removing past checkpoints from training",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 353,
+ 714
+ ],
+ "score": 1.0,
+ "content": "significantly reduces performance. Comparing the FCP and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 700,
+ 385,
+ 712
+ ],
+ "score": 0.9,
+ "content": "\\operatorname { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 699,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "columns, we see that adding",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "architectural variation to the training population offers no improvement over training with past",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 72,
+ 406,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 407,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 407,
+ 87
+ ],
+ "score": 1.0,
+ "content": "4.1 Evaluation method: collaborative evaluation with agent partners",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 92,
+ 505,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": "Our primary concern in this work is generalization to novel human partners (as investigated in",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 103,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 138,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 103,
+ 148,
+ 116
+ ],
+ "score": 0.35,
+ "content": "\\textcircled{5}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 148,
+ 103,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": ". However, just as collecting human-human data for behavioral cloning is expensive, so",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 125,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 125,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "partners as a cheap proxy of performance with humans. This is then used to guide our model selection",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "process, allowing us to be more targeted with the agents we select for our human-agent evaluations.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 93,
+ 506,
+ 149
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 153,
+ 286,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 151,
+ 286,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 151,
+ 286,
+ 166
+ ],
+ "score": 1.0,
+ "content": "We evaluate with three held-out populations:",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 106,
+ 151,
+ 286,
+ 166
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 129,
+ 172,
+ 506,
+ 246
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 171,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 171,
+ 283,
+ 185
+ ],
+ "score": 1.0,
+ "content": "1. A BC model trained on human data,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 284,
+ 173,
+ 312,
+ 185
+ ],
+ "score": 0.91,
+ "content": "H _ { \\mathrm { p r o x y } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 171,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": ", intended as a proxy of generalization to humans,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 182,
+ 261,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 182,
+ 240,
+ 196
+ ],
+ "score": 1.0,
+ "content": "as done by Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 183,
+ 258,
+ 195
+ ],
+ "score": 0.29,
+ "content": "[ \\overbrace { | 1 2 | }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 182,
+ 261,
+ 196
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 128,
+ 196,
+ 507,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 196,
+ 507,
+ 211
+ ],
+ "score": 1.0,
+ "content": "2. A set of self-play agents varying in seed, architecture, and training time (specifically, held-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 207,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 208,
+ 207,
+ 222
+ ],
+ "score": 1.0,
+ "content": "out seeds of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 209,
+ 241,
+ 219
+ ],
+ "score": 0.89,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 208,
+ 335,
+ 222
+ ],
+ "score": 1.0,
+ "content": "partners trained for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 209,
+ 367,
+ 220
+ ],
+ "score": 0.9,
+ "content": "\\mathrm { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 208,
+ 460,
+ 222
+ ],
+ "score": 1.0,
+ "content": "agent; see Section 2.4).",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 462,
+ 207,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "These are",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 142,
+ 219,
+ 423,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 219,
+ 423,
+ 232
+ ],
+ "score": 1.0,
+ "content": "intended to test generalization to a diverse yet still skillful population.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 129,
+ 234,
+ 464,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 234,
+ 464,
+ 246
+ ],
+ "score": 1.0,
+ "content": "3. Randomly initialized agents intended to test generalization to low-skill partners.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 128,
+ 171,
+ 507,
+ 246
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 506,
+ 298
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "For all results, we report the average number of deliveries made by both players within an episode,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 330,
+ 277
+ ],
+ "score": 1.0,
+ "content": "aggregated across the 5 different layouts from Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 331,
+ 264,
+ 341,
+ 277
+ ],
+ "score": 0.74,
+ "content": "\\boxed { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 264,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "(with the per-layout results reported in",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 146,
+ 288
+ ],
+ "score": 1.0,
+ "content": "Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 275,
+ 166,
+ 288
+ ],
+ "score": 0.51,
+ "content": "\\underline { { \\overline { { ( \\mathrm { C . 2 } ) } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 276,
+ 167,
+ 288
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 275,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "We estimate mean and standard deviation across 5 random seeds. For each seed, we",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 299
+ ],
+ "score": 1.0,
+ "content": "evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 253,
+ 507,
+ 299
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 310,
+ 162,
+ 322
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 163,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 163,
+ 323
+ ],
+ "score": 1.0,
+ "content": "4.2 Results",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 330,
+ 336,
+ 342
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 338,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 338,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Finding 1: FCP significantly outperforms all baselines",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 343,
+ 505,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "To begin, we compare our FCP agent and the baselines when partnered with the three held-out",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 354,
+ 323,
+ 367
+ ],
+ "score": 1.0,
+ "content": "populations introduced above. As can be seen in Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 353,
+ 334,
+ 367
+ ],
+ "score": 0.61,
+ "content": "\\boxed { 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 354,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "FCP significantly outperforms all baselines",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 103,
+ 361,
+ 502,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 361,
+ 473,
+ 381
+ ],
+ "score": 1.0,
+ "content": "when partnered with all three held-out populations. Notably, it performs better than BCP with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 365,
+ 502,
+ 377
+ ],
+ "score": 0.88,
+ "content": "H _ { \\mathrm { p r o x y } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 451,
+ 389
+ ],
+ "score": 1.0,
+ "content": "even though BCP trains with such a model and FCP does not. Similar to Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 451,
+ 375,
+ 469,
+ 387
+ ],
+ "score": 0.28,
+ "content": "\\mathbb { \\lVert 1 2 \\rVert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 469,
+ 375,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": ", we find",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 387,
+ 252,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 387,
+ 252,
+ 399
+ ],
+ "score": 1.0,
+ "content": "that BCP significantly outscores SP.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 103,
+ 343,
+ 506,
+ 399
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 506,
+ 415
+ ],
+ "score": 1.0,
+ "content": "When paired with a randomly initialized partner which behaves suboptimally, we see an even greater",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 506,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 506,
+ 426
+ ],
+ "score": 1.0,
+ "content": "difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "such agents, it may not be surprising that it does so well with partners that behave poorly. However,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "what is surprising is how brittle the other training methods are. This suggests that they may not",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 480,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 480,
+ 460
+ ],
+ "score": 1.0,
+ "content": "perform well with humans who are not highly skilled players, which we will see in Section 5.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 403,
+ 506,
+ 460
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 559
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "8083dcac415b6ef9624e1aa4f428a0a33c3e9d47ce8903df3798f3e80ef0427c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 101,
+ 470,
+ 495,
+ 499.6666666666667
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 101,
+ 499.6666666666667,
+ 495,
+ 529.3333333333334
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 101,
+ 529.3333333333334,
+ 495,
+ 559.0
+ ],
+ "spans": [],
+ "index": 31
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 565,
+ 506,
+ 622
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Agent-agent collaborative evaluation: Performance of each agent when partnered with",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 277,
+ 588
+ ],
+ "score": 1.0,
+ "content": "each of the held-out populations (Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 575,
+ 296,
+ 588
+ ],
+ "score": 0.62,
+ "content": "4 . 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 576,
+ 387,
+ 588
+ ],
+ "score": 1.0,
+ "content": "in episodes of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 387,
+ 576,
+ 426,
+ 586
+ ],
+ "score": 0.9,
+ "content": "T = 5 4 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 576,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": ". Importantly, FCP",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "scores higher than all baselines with a variety of test partners. Error bars represent standard deviation",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "over five random training seeds. Plots aggregate data across kitchen layouts; results calculated by",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 607,
+ 302,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 302,
+ 623
+ ],
+ "score": 1.0,
+ "content": "individual layout can be found in Appendix C.2.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 637,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 491,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 491,
+ 652
+ ],
+ "score": 1.0,
+ "content": "Finding 2: Training with past checkpoints is the most beneficial variation for performance",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "Next, we investigate how the different training partner variations influence FCP’s performance. In",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 660,
+ 506,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 660,
+ 317,
+ 676
+ ],
+ "score": 1.0,
+ "content": "particular, we separately ablate the past checkpoints",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 662,
+ 332,
+ 673
+ ],
+ "score": 0.72,
+ "content": "( T )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 660,
+ 401,
+ 676
+ ],
+ "score": 1.0,
+ "content": "and architecture",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 662,
+ 416,
+ 673
+ ],
+ "score": 0.68,
+ "content": "( A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 660,
+ 506,
+ 676
+ ],
+ "score": 1.0,
+ "content": "variations, evaluating",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 498,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 498,
+ 687
+ ],
+ "score": 1.0,
+ "content": "them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 210,
+ 703
+ ],
+ "score": 1.0,
+ "content": "Comparing the FCP and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 689,
+ 242,
+ 700
+ ],
+ "score": 0.83,
+ "content": "\\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 687,
+ 506,
+ 703
+ ],
+ "score": 1.0,
+ "content": "columns, we see that removing past checkpoints from training",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 353,
+ 714
+ ],
+ "score": 1.0,
+ "content": "significantly reduces performance. Comparing the FCP and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 700,
+ 385,
+ 712
+ ],
+ "score": 0.9,
+ "content": "\\operatorname { F C P } _ { + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 699,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "columns, we see that adding",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "architectural variation to the training population offers no improvement over training with past",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 104,
+ 636,
+ 506,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "score": 0.974,
+ "html": "| Partner | FCP | FCP-T | FCP+A | FCP-T,+A |
| Hproxy | 10.6± 0.5 | 4.7± 0.4 | 9.9±0.6 | 7.0±0.8 |
| Diverse SP | 11.2 ± 0.1 | 6.9 ± 0.1 | 11.1 ± 0.4 | 8.6 ± 0.4 |
| Random | 8.6± 0.2 | 1.0 ± 0.1 | 8.4±0.4 | 3.2 ± 0.5 |
",
+ "type": "table",
+ "image_path": "56b2734b6d3a9e5266ca9b0dd8d3ba9f6d27df29190a11ac23347bb9e9fb3c37.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 85.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 167,
+ 85.66666666666667,
+ 441,
+ 101.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 167,
+ 101.33333333333334,
+ 441,
+ 117.00000000000001
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 123,
+ 505,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "score": 1.0,
+ "content": "Table 1: Ablation results: Performance of each variation of FCP – training with past partner",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 135,
+ 505,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 135,
+ 159,
+ 146
+ ],
+ "score": 1.0,
+ "content": "checkpoints",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 135,
+ 168,
+ 145
+ ],
+ "score": 0.63,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 135,
+ 379,
+ 146
+ ],
+ "score": 1.0,
+ "content": "for time) and adding partner variation in architecture",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 135,
+ 393,
+ 146
+ ],
+ "score": 0.49,
+ "content": "( A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 393,
+ 135,
+ 505,
+ 146
+ ],
+ "score": 1.0,
+ "content": ". Scores are mean deliveries",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 145,
+ 505,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 145,
+ 505,
+ 158
+ ],
+ "score": 1.0,
+ "content": "with standard deviation over 5 random seeds. Notably, we find that the inclusion of past checkpoints",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 156,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 244,
+ 169
+ ],
+ "score": 1.0,
+ "content": "is essential for strong performance",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 245,
+ 156,
+ 306,
+ 168
+ ],
+ "score": 0.89,
+ "content": "( \\mathrm { F C P } > \\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 306,
+ 156,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "), and additionally including architectural variation",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 167,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 167,
+ 234,
+ 180
+ ],
+ "score": 1.0,
+ "content": "does not improve performance",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 168,
+ 299,
+ 179
+ ],
+ "score": 0.9,
+ "content": "( \\mathrm { F C P } \\approx \\mathrm { F C P } _ { + A . }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 167,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "). However, architectural variation is better than no",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 501,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 405,
+ 191
+ ],
+ "score": 1.0,
+ "content": "variation, improving performance when past checkpoints are not available",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 178,
+ 495,
+ 190
+ ],
+ "score": 0.91,
+ "content": "( \\mathrm { F C P } _ { - T , + A } > \\mathrm { F C P } _ { - T , }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 178,
+ 501,
+ 191
+ ],
+ "score": 1.0,
+ "content": ").",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 208,
+ 505,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 258,
+ 222
+ ],
+ "score": 1.0,
+ "content": "checkpoints. However, comparing the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 209,
+ 290,
+ 220
+ ],
+ "score": 0.87,
+ "content": "\\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 207,
+ 308,
+ 222
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 209,
+ 353,
+ 221
+ ],
+ "score": 0.88,
+ "content": "\\mathrm { F C P } _ { - T , + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 353,
+ 207,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "columns, we see that without training",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 469,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 469,
+ 233
+ ],
+ "score": 1.0,
+ "content": "with past checkpoints, architectural variation in the population does improve performance.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 246,
+ 313,
+ 259
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 245,
+ 314,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 245,
+ 314,
+ 262
+ ],
+ "score": 1.0,
+ "content": "5 Zero-shot coordination with humans",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 213,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 213,
+ 305
+ ],
+ "score": 1.0,
+ "content": "play with human partners.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "score": 0.814,
+ "type": "image",
+ "image_path": "5856cbdd90eeef2a7a6093668472063862a759746c339aa0324e78efb5d67b22.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 343.6666666666667
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 91,
+ 343.6666666666667,
+ 521,
+ 372.33333333333337
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 91,
+ 372.33333333333337,
+ 521,
+ 401.00000000000006
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 454
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Human-agent collaborative study: For our human-agent collaboration study, we recruited",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 433
+ ],
+ "score": 1.0,
+ "content": "participants online to play games with FCP and baseline agents. Participants played a randomized",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 507,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 507,
+ 444
+ ],
+ "score": 1.0,
+ "content": "sequence of episodes with different agent partners and kitchen layouts. After every two episodes,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 498,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 498,
+ 455
+ ],
+ "score": 1.0,
+ "content": "participants reported the direction and strength of their preference between their last two partners.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5
+ }
+ ],
+ "index": 17.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 474,
+ 427,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 429,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 429,
+ 489
+ ],
+ "score": 1.0,
+ "content": "5.1 Evaluation method: collaborative evaluation with human participants",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 495,
+ 506,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "score": 1.0,
+ "content": "To test how effectively FCP’s performance generalizes to human partners, we recruited participants",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 159,
+ 518
+ ],
+ "score": 1.0,
+ "content": "from Prolific",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 505,
+ 192,
+ 517
+ ],
+ "score": 0.31,
+ "content": "\\mathbb { 1 1 8 } , \\lvert 5 5 \\rvert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 505,
+ 328,
+ 518
+ ],
+ "score": 1.0,
+ "content": "for an online collaboration study",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 506,
+ 367,
+ 516
+ ],
+ "score": 0.87,
+ "content": "N = 1 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 505,
+ 371,
+ 518
+ ],
+ "score": 1.0,
+ "content": ";",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 506,
+ 398,
+ 516
+ ],
+ "score": 0.83,
+ "content": "3 7 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 505,
+ 430,
+ 518
+ ],
+ "score": 1.0,
+ "content": "female,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 506,
+ 457,
+ 516
+ ],
+ "score": 0.86,
+ "content": "5 9 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 458,
+ 505,
+ 483,
+ 518
+ ],
+ "score": 1.0,
+ "content": "male,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 506,
+ 505,
+ 516
+ ],
+ "score": 0.85,
+ "content": "1 . 8 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 506,
+ 529
+ ],
+ "score": 1.0,
+ "content": "nonbinary; median age between 25–34 years). We used a within-participant design for the study:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "each participant played with a full cohort of agents (i.e. generated through every training method).",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 492,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 492,
+ 553
+ ],
+ "score": 1.0,
+ "content": "This design allowed us to evaluate both objective performance as well as subjective preferences.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 554,
+ 505,
+ 654
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "Participants first read game instructions and played a short tutorial episode guiding them through",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 565,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 297,
+ 578
+ ],
+ "score": 1.0,
+ "content": "the dish preparation sequence (see Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 565,
+ 325,
+ 579
+ ],
+ "score": 0.78,
+ "content": "\\underline { { \\overline { { \\mathbb { D . 1 . 1 } } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 565,
+ 506,
+ 578
+ ],
+ "score": 1.0,
+ "content": "for instruction text and study screenshots).",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Participants then played 20 episodes with a randomized sequence of agent partners and kitchen",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 209,
+ 600
+ ],
+ "score": 1.0,
+ "content": "layouts. Episodes lasted",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 588,
+ 249,
+ 598
+ ],
+ "score": 0.89,
+ "content": "T = 3 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "steps (1 minute) each. After every two episodes, participants",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": "reported their preference over the agent partners from those episodes on a five-point Likert-type scale.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "After playing all 20 episodes, participants completed a debrief questionnaire collecting standard",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "demographic information and open-ended feedback on the study. Our statistical analysis below",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 476,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 476,
+ 655
+ ],
+ "score": 1.0,
+ "content": "for additional details of our study design and analysis, including independent ethical review.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 667,
+ 162,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 163,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 163,
+ 680
+ ],
+ "score": 1.0,
+ "content": "5.2 Results",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 477,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 477,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Finding 1: FCP coordinates best with humans, achieving the highest score across maps",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 507,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 507,
+ 714
+ ],
+ "score": 1.0,
+ "content": "To begin, we compare the objective team performance supported by our FCP and baseline agents.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 507,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 507,
+ 724
+ ],
+ "score": 1.0,
+ "content": "The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration:",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 753
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 117
+ ],
+ "score": 0.974,
+ "html": "| Partner | FCP | FCP-T | FCP+A | FCP-T,+A |
| Hproxy | 10.6± 0.5 | 4.7± 0.4 | 9.9±0.6 | 7.0±0.8 |
| Diverse SP | 11.2 ± 0.1 | 6.9 ± 0.1 | 11.1 ± 0.4 | 8.6 ± 0.4 |
| Random | 8.6± 0.2 | 1.0 ± 0.1 | 8.4±0.4 | 3.2 ± 0.5 |
",
+ "type": "table",
+ "image_path": "56b2734b6d3a9e5266ca9b0dd8d3ba9f6d27df29190a11ac23347bb9e9fb3c37.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 70,
+ 441,
+ 85.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 167,
+ 85.66666666666667,
+ 441,
+ 101.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 167,
+ 101.33333333333334,
+ 441,
+ 117.00000000000001
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 123,
+ 505,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "score": 1.0,
+ "content": "Table 1: Ablation results: Performance of each variation of FCP – training with past partner",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 135,
+ 505,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 135,
+ 159,
+ 146
+ ],
+ "score": 1.0,
+ "content": "checkpoints",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 135,
+ 168,
+ 145
+ ],
+ "score": 0.63,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 135,
+ 379,
+ 146
+ ],
+ "score": 1.0,
+ "content": "for time) and adding partner variation in architecture",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 135,
+ 393,
+ 146
+ ],
+ "score": 0.49,
+ "content": "( A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 393,
+ 135,
+ 505,
+ 146
+ ],
+ "score": 1.0,
+ "content": ". Scores are mean deliveries",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 145,
+ 505,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 145,
+ 505,
+ 158
+ ],
+ "score": 1.0,
+ "content": "with standard deviation over 5 random seeds. Notably, we find that the inclusion of past checkpoints",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 156,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 244,
+ 169
+ ],
+ "score": 1.0,
+ "content": "is essential for strong performance",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 245,
+ 156,
+ 306,
+ 168
+ ],
+ "score": 0.89,
+ "content": "( \\mathrm { F C P } > \\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 306,
+ 156,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "), and additionally including architectural variation",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 167,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 167,
+ 234,
+ 180
+ ],
+ "score": 1.0,
+ "content": "does not improve performance",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 168,
+ 299,
+ 179
+ ],
+ "score": 0.9,
+ "content": "( \\mathrm { F C P } \\approx \\mathrm { F C P } _ { + A . }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 167,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "). However, architectural variation is better than no",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 501,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 405,
+ 191
+ ],
+ "score": 1.0,
+ "content": "variation, improving performance when past checkpoints are not available",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 178,
+ 495,
+ 190
+ ],
+ "score": 0.91,
+ "content": "( \\mathrm { F C P } _ { - T , + A } > \\mathrm { F C P } _ { - T , }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 178,
+ 501,
+ 191
+ ],
+ "score": 1.0,
+ "content": ").",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 122,
+ 506,
+ 191
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 208,
+ 505,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 258,
+ 222
+ ],
+ "score": 1.0,
+ "content": "checkpoints. However, comparing the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 209,
+ 290,
+ 220
+ ],
+ "score": 0.87,
+ "content": "\\mathrm { F C P } _ { - T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 207,
+ 308,
+ 222
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 209,
+ 353,
+ 221
+ ],
+ "score": 0.88,
+ "content": "\\mathrm { F C P } _ { - T , + A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 353,
+ 207,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "columns, we see that without training",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 469,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 469,
+ 233
+ ],
+ "score": 1.0,
+ "content": "with past checkpoints, architectural variation in the population does improve performance.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 207,
+ 506,
+ 233
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 246,
+ 313,
+ 259
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 245,
+ 314,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 245,
+ 314,
+ 262
+ ],
+ "score": 1.0,
+ "content": "5 Zero-shot coordination with humans",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 213,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 213,
+ 305
+ ],
+ "score": 1.0,
+ "content": "play with human partners.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 271,
+ 506,
+ 305
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 401
+ ],
+ "score": 0.814,
+ "type": "image",
+ "image_path": "5856cbdd90eeef2a7a6093668472063862a759746c339aa0324e78efb5d67b22.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 91,
+ 315,
+ 521,
+ 343.6666666666667
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 91,
+ 343.6666666666667,
+ 521,
+ 372.33333333333337
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 91,
+ 372.33333333333337,
+ 521,
+ 401.00000000000006
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 454
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Human-agent collaborative study: For our human-agent collaboration study, we recruited",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 433
+ ],
+ "score": 1.0,
+ "content": "participants online to play games with FCP and baseline agents. Participants played a randomized",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 507,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 507,
+ 444
+ ],
+ "score": 1.0,
+ "content": "sequence of episodes with different agent partners and kitchen layouts. After every two episodes,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 498,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 498,
+ 455
+ ],
+ "score": 1.0,
+ "content": "participants reported the direction and strength of their preference between their last two partners.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5
+ }
+ ],
+ "index": 17.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 474,
+ 427,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 429,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 429,
+ 489
+ ],
+ "score": 1.0,
+ "content": "5.1 Evaluation method: collaborative evaluation with human participants",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 495,
+ 506,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "score": 1.0,
+ "content": "To test how effectively FCP’s performance generalizes to human partners, we recruited participants",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 159,
+ 518
+ ],
+ "score": 1.0,
+ "content": "from Prolific",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 505,
+ 192,
+ 517
+ ],
+ "score": 0.31,
+ "content": "\\mathbb { 1 1 8 } , \\lvert 5 5 \\rvert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 505,
+ 328,
+ 518
+ ],
+ "score": 1.0,
+ "content": "for an online collaboration study",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 506,
+ 367,
+ 516
+ ],
+ "score": 0.87,
+ "content": "N = 1 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 505,
+ 371,
+ 518
+ ],
+ "score": 1.0,
+ "content": ";",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 506,
+ 398,
+ 516
+ ],
+ "score": 0.83,
+ "content": "3 7 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 505,
+ 430,
+ 518
+ ],
+ "score": 1.0,
+ "content": "female,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 506,
+ 457,
+ 516
+ ],
+ "score": 0.86,
+ "content": "5 9 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 458,
+ 505,
+ 483,
+ 518
+ ],
+ "score": 1.0,
+ "content": "male,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 506,
+ 505,
+ 516
+ ],
+ "score": 0.85,
+ "content": "1 . 8 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 506,
+ 529
+ ],
+ "score": 1.0,
+ "content": "nonbinary; median age between 25–34 years). We used a within-participant design for the study:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "each participant played with a full cohort of agents (i.e. generated through every training method).",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 492,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 492,
+ 553
+ ],
+ "score": 1.0,
+ "content": "This design allowed us to evaluate both objective performance as well as subjective preferences.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 493,
+ 506,
+ 553
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 554,
+ 505,
+ 654
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "Participants first read game instructions and played a short tutorial episode guiding them through",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 565,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 297,
+ 578
+ ],
+ "score": 1.0,
+ "content": "the dish preparation sequence (see Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 565,
+ 325,
+ 579
+ ],
+ "score": 0.78,
+ "content": "\\underline { { \\overline { { \\mathbb { D . 1 . 1 } } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 565,
+ 506,
+ 578
+ ],
+ "score": 1.0,
+ "content": "for instruction text and study screenshots).",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Participants then played 20 episodes with a randomized sequence of agent partners and kitchen",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 209,
+ 600
+ ],
+ "score": 1.0,
+ "content": "layouts. Episodes lasted",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 588,
+ 249,
+ 598
+ ],
+ "score": 0.89,
+ "content": "T = 3 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "steps (1 minute) each. After every two episodes, participants",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": "reported their preference over the agent partners from those episodes on a five-point Likert-type scale.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "After playing all 20 episodes, participants completed a debrief questionnaire collecting standard",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "demographic information and open-ended feedback on the study. Our statistical analysis below",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 476,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 476,
+ 655
+ ],
+ "score": 1.0,
+ "content": "for additional details of our study design and analysis, including independent ethical review.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 554,
+ 506,
+ 655
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 667,
+ 162,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 163,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 163,
+ 680
+ ],
+ "score": 1.0,
+ "content": "5.2 Results",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 477,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 477,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Finding 1: FCP coordinates best with humans, achieving the highest score across maps",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 507,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 507,
+ 714
+ ],
+ "score": 1.0,
+ "content": "To begin, we compare the objective team performance supported by our FCP and baseline agents.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 507,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 507,
+ 724
+ ],
+ "score": 1.0,
+ "content": "The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration:",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39,
+ "bbox_fs": [
+ 105,
+ 686,
+ 507,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 139
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 84
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 84
+ ],
+ "score": 1.0,
+ "content": "the FCP-human teams significantly outperform all other agent-human teams, achieving the highest",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 244,
+ 97
+ ],
+ "score": 1.0,
+ "content": "average scores across maps, every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 84,
+ 286,
+ 95
+ ],
+ "score": 0.9,
+ "content": "p < 0 . 0 0 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 83,
+ 318,
+ 97
+ ],
+ "score": 1.0,
+ "content": "(Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 83,
+ 333,
+ 95
+ ],
+ "score": 0.81,
+ "content": "\\lvert \\overline { { 7 \\mathrm { a } } } \\rvert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 83,
+ 506,
+ 97
+ ],
+ "score": 1.0,
+ "content": ", while performing as well as or better than",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 319,
+ 107
+ ],
+ "score": 1.0,
+ "content": "the other teams on each individual map (see Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 95,
+ 339,
+ 107
+ ],
+ "score": 0.6,
+ "content": "\\mathbf { D } . 3 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 94,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": ". Echoing the results from our agent-agent",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 218,
+ 119
+ ],
+ "score": 1.0,
+ "content": "ablation experiments (Table",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 105,
+ 228,
+ 117
+ ],
+ "score": 0.63,
+ "content": "^ { 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 105,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": ", the inclusion of past checkpoints in training proves critical to FCP’s",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 116,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 187,
+ 128
+ ],
+ "score": 1.0,
+ "content": "strong performance,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 117,
+ 230,
+ 128
+ ],
+ "score": 0.88,
+ "content": "p < 0 . 0 { \\overline { { 0 1 } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 116,
+ 261,
+ 128
+ ],
+ "score": 1.0,
+ "content": "(Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 261,
+ 116,
+ 276,
+ 129
+ ],
+ "score": 0.82,
+ "content": "\\textcircled { 7 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 116,
+ 374,
+ 128
+ ],
+ "score": 1.0,
+ "content": ". Similar to Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 116,
+ 392,
+ 128
+ ],
+ "score": 0.58,
+ "content": "\\mathbb { \\lVert \\rVert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 116,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": ", we find that BCP outscores",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 329,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 282,
+ 140
+ ],
+ "score": 1.0,
+ "content": "SP when collaborating with human players,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 128,
+ 325,
+ 139
+ ],
+ "score": 0.9,
+ "content": "p < 0 . 0 0 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 126,
+ 329,
+ 140
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 150,
+ 333,
+ 162
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 333,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 333,
+ 164
+ ],
+ "score": 1.0,
+ "content": "Finding 2: Participants prefer FCP over all baselines",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 163,
+ 506,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 507,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 507,
+ 177
+ ],
+ "score": 1.0,
+ "content": "FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 187
+ ],
+ "score": 1.0,
+ "content": "Participants expressed a significant preference for FCP partners over all other agents, including",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 183,
+ 506,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 183,
+ 176,
+ 198
+ ],
+ "score": 1.0,
+ "content": "BCP, with every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 185,
+ 215,
+ 196
+ ],
+ "score": 0.9,
+ "content": "p < 0 . 0 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 215,
+ 183,
+ 247,
+ 198
+ ],
+ "score": 1.0,
+ "content": "(Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 184,
+ 263,
+ 197
+ ],
+ "score": 0.83,
+ "content": "\\bar { 7 \\mathrm { c } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 183,
+ 506,
+ 198
+ ],
+ "score": 1.0,
+ "content": ". Notably, while human-BCP and human-PP teams did not",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 209
+ ],
+ "score": 1.0,
+ "content": "significantly differ in their completed deliveries, participants reported significantly preferring BCP",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 206,
+ 425,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 206,
+ 141,
+ 219
+ ],
+ "score": 1.0,
+ "content": "over PP,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 207,
+ 183,
+ 218
+ ],
+ "score": 0.9,
+ "content": "p = 0 . 0 0 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 184,
+ 206,
+ 425,
+ 219
+ ],
+ "score": 1.0,
+ "content": ", highlighting the informativeness of our subjective analysis.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "score": 0.927,
+ "type": "image",
+ "image_path": "14e048153d10325b3617f94b393393ed3ab6cf3e38198f8791acf0fc83b051df.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 270.6666666666667
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 107,
+ 270.6666666666667,
+ 504,
+ 311.33333333333337
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 311.33333333333337,
+ 504,
+ 352.00000000000006
+ ],
+ "spans": [],
+ "index": 14
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 357,
+ 506,
+ 403
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 507,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 507,
+ 370
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from human-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 505,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 233,
+ 379
+ ],
+ "score": 1.0,
+ "content": "agent play in episodes of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 368,
+ 270,
+ 379
+ ],
+ "score": 0.9,
+ "content": "T = 3 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 368,
+ 358,
+ 379
+ ],
+ "score": 1.0,
+ "content": ". Error bars represents",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 368,
+ 378,
+ 379
+ ],
+ "score": 0.87,
+ "content": "9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 368,
+ 505,
+ 379
+ ],
+ "score": 1.0,
+ "content": "confidence intervals, calculated",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 389,
+ 216,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 216,
+ 405
+ ],
+ "score": 1.0,
+ "content": "be found in Appendix D.3.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5
+ }
+ ],
+ "index": 14.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 421,
+ 266,
+ 433
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 420,
+ 267,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 420,
+ 267,
+ 435
+ ],
+ "score": 1.0,
+ "content": "5.3 Exploratory behavioral analysis",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "To better understand how the human-agent scores and preferences may have arisen, here we analyze",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 437,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 437,
+ 465
+ ],
+ "score": 1.0,
+ "content": "the resulting action trajectories of each human and agent player in our experiment.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "score": 0.519,
+ "type": "image",
+ "image_path": "65959c0ab23a18c0f8de72bc451d199e7d9d9cc2698e7ca2ac4f320e26dac829.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 517.6666666666666
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 138,
+ 517.6666666666666,
+ 474,
+ 557.3333333333333
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 138,
+ 557.3333333333333,
+ 474,
+ 596.9999999999999
+ ],
+ "spans": [],
+ "index": 24
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 601,
+ 506,
+ 657
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 507,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 412,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Behavioral analysis: (a) FCP is able to move most frequently",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 601,
+ 432,
+ 613
+ ],
+ "score": 0.85,
+ "content": "3 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 432,
+ 601,
+ 507,
+ 614
+ ],
+ "score": 1.0,
+ "content": "of the time), cor-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "responding to the best movement coordination with human partners. (b) FCP exhibits the most",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "equal preferences over cooking pots (0.11 difference), aligning with human preferences. Values are",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 649
+ ],
+ "score": 1.0,
+ "content": "calculated as the absolute difference in preferences between the two pots; 1 indicates that the player",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 645,
+ 487,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 487,
+ 659
+ ],
+ "score": 1.0,
+ "content": "only uses one of the two available pots, while 0 indicates that the player uses both pots equally.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 665,
+ 407,
+ 677
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 407,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 407,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Finding 1: FCP exhibits the best movement coordination with humans",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 678,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 405,
+ 691
+ ],
+ "score": 1.0,
+ "content": "First, we investigate how much each player moves in an episode (Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 677,
+ 421,
+ 691
+ ],
+ "score": 0.27,
+ "content": "\\textcircled { 8 \\mathrm { a } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": ", where moving in a",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 701
+ ],
+ "score": 1.0,
+ "content": "higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 740,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 740,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 139
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 84
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 84
+ ],
+ "score": 1.0,
+ "content": "the FCP-human teams significantly outperform all other agent-human teams, achieving the highest",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 83,
+ 506,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 244,
+ 97
+ ],
+ "score": 1.0,
+ "content": "average scores across maps, every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 84,
+ 286,
+ 95
+ ],
+ "score": 0.9,
+ "content": "p < 0 . 0 0 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 83,
+ 318,
+ 97
+ ],
+ "score": 1.0,
+ "content": "(Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 83,
+ 333,
+ 95
+ ],
+ "score": 0.81,
+ "content": "\\lvert \\overline { { 7 \\mathrm { a } } } \\rvert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 83,
+ 506,
+ 97
+ ],
+ "score": 1.0,
+ "content": ", while performing as well as or better than",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 319,
+ 107
+ ],
+ "score": 1.0,
+ "content": "the other teams on each individual map (see Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 95,
+ 339,
+ 107
+ ],
+ "score": 0.6,
+ "content": "\\mathbf { D } . 3 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 94,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": ". Echoing the results from our agent-agent",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 218,
+ 119
+ ],
+ "score": 1.0,
+ "content": "ablation experiments (Table",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 105,
+ 228,
+ 117
+ ],
+ "score": 0.63,
+ "content": "^ { 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 105,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": ", the inclusion of past checkpoints in training proves critical to FCP’s",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 116,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 187,
+ 128
+ ],
+ "score": 1.0,
+ "content": "strong performance,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 117,
+ 230,
+ 128
+ ],
+ "score": 0.88,
+ "content": "p < 0 . 0 { \\overline { { 0 1 } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 116,
+ 261,
+ 128
+ ],
+ "score": 1.0,
+ "content": "(Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 261,
+ 116,
+ 276,
+ 129
+ ],
+ "score": 0.82,
+ "content": "\\textcircled { 7 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 116,
+ 374,
+ 128
+ ],
+ "score": 1.0,
+ "content": ". Similar to Carroll et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 116,
+ 392,
+ 128
+ ],
+ "score": 0.58,
+ "content": "\\mathbb { \\lVert \\rVert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 116,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": ", we find that BCP outscores",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 329,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 282,
+ 140
+ ],
+ "score": 1.0,
+ "content": "SP when collaborating with human players,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 128,
+ 325,
+ 139
+ ],
+ "score": 0.9,
+ "content": "p < 0 . 0 0 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 126,
+ 329,
+ 140
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 73,
+ 506,
+ 140
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 150,
+ 333,
+ 162
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 333,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 333,
+ 164
+ ],
+ "score": 1.0,
+ "content": "Finding 2: Participants prefer FCP over all baselines",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 163,
+ 506,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 507,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 507,
+ 177
+ ],
+ "score": 1.0,
+ "content": "FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 187
+ ],
+ "score": 1.0,
+ "content": "Participants expressed a significant preference for FCP partners over all other agents, including",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 183,
+ 506,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 183,
+ 176,
+ 198
+ ],
+ "score": 1.0,
+ "content": "BCP, with every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 185,
+ 215,
+ 196
+ ],
+ "score": 0.9,
+ "content": "p < 0 . 0 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 215,
+ 183,
+ 247,
+ 198
+ ],
+ "score": 1.0,
+ "content": "(Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 184,
+ 263,
+ 197
+ ],
+ "score": 0.83,
+ "content": "\\bar { 7 \\mathrm { c } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 183,
+ 506,
+ 198
+ ],
+ "score": 1.0,
+ "content": ". Notably, while human-BCP and human-PP teams did not",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 209
+ ],
+ "score": 1.0,
+ "content": "significantly differ in their completed deliveries, participants reported significantly preferring BCP",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 206,
+ 425,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 206,
+ 141,
+ 219
+ ],
+ "score": 1.0,
+ "content": "over PP,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 207,
+ 183,
+ 218
+ ],
+ "score": 0.9,
+ "content": "p = 0 . 0 0 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 184,
+ 206,
+ 425,
+ 219
+ ],
+ "score": 1.0,
+ "content": ", highlighting the informativeness of our subjective analysis.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 105,
+ 162,
+ 507,
+ 219
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 352
+ ],
+ "score": 0.927,
+ "type": "image",
+ "image_path": "14e048153d10325b3617f94b393393ed3ab6cf3e38198f8791acf0fc83b051df.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 504,
+ 270.6666666666667
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 107,
+ 270.6666666666667,
+ 504,
+ 311.33333333333337
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 311.33333333333337,
+ 504,
+ 352.00000000000006
+ ],
+ "spans": [],
+ "index": 14
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 357,
+ 506,
+ 403
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 507,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 507,
+ 370
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from human-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 505,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 233,
+ 379
+ ],
+ "score": 1.0,
+ "content": "agent play in episodes of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 368,
+ 270,
+ 379
+ ],
+ "score": 0.9,
+ "content": "T = 3 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 368,
+ 358,
+ 379
+ ],
+ "score": 1.0,
+ "content": ". Error bars represents",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 368,
+ 378,
+ 379
+ ],
+ "score": 0.87,
+ "content": "9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 368,
+ 505,
+ 379
+ ],
+ "score": 1.0,
+ "content": "confidence intervals, calculated",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 389,
+ 216,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 216,
+ 405
+ ],
+ "score": 1.0,
+ "content": "be found in Appendix D.3.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5
+ }
+ ],
+ "index": 14.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 421,
+ 266,
+ 433
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 420,
+ 267,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 420,
+ 267,
+ 435
+ ],
+ "score": 1.0,
+ "content": "5.3 Exploratory behavioral analysis",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "To better understand how the human-agent scores and preferences may have arisen, here we analyze",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 437,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 437,
+ 465
+ ],
+ "score": 1.0,
+ "content": "the resulting action trajectories of each human and agent player in our experiment.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 440,
+ 505,
+ 465
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 597
+ ],
+ "score": 0.519,
+ "type": "image",
+ "image_path": "65959c0ab23a18c0f8de72bc451d199e7d9d9cc2698e7ca2ac4f320e26dac829.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 138,
+ 478,
+ 474,
+ 517.6666666666666
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 138,
+ 517.6666666666666,
+ 474,
+ 557.3333333333333
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 138,
+ 557.3333333333333,
+ 474,
+ 596.9999999999999
+ ],
+ "spans": [],
+ "index": 24
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 601,
+ 506,
+ 657
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 507,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 412,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Behavioral analysis: (a) FCP is able to move most frequently",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 601,
+ 432,
+ 613
+ ],
+ "score": 0.85,
+ "content": "3 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 432,
+ 601,
+ 507,
+ 614
+ ],
+ "score": 1.0,
+ "content": "of the time), cor-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "responding to the best movement coordination with human partners. (b) FCP exhibits the most",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "equal preferences over cooking pots (0.11 difference), aligning with human preferences. Values are",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 649
+ ],
+ "score": 1.0,
+ "content": "calculated as the absolute difference in preferences between the two pots; 1 indicates that the player",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 645,
+ 487,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 487,
+ 659
+ ],
+ "score": 1.0,
+ "content": "only uses one of the two available pots, while 0 indicates that the player uses both pots equally.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 601,
+ 507,
+ 659
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 665,
+ 407,
+ 677
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 407,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 407,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Finding 1: FCP exhibits the best movement coordination with humans",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 678,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 405,
+ 691
+ ],
+ "score": 1.0,
+ "content": "First, we investigate how much each player moves in an episode (Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 677,
+ 421,
+ 691
+ ],
+ "score": 0.27,
+ "content": "\\textcircled { 8 \\mathrm { a } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": ", where moving in a",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 701
+ ],
+ "score": 1.0,
+ "content": "higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 677,
+ 507,
+ 722
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 86
+ ],
+ "score": 1.0,
+ "content": "(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "movement strategy with its partner. This result was also reported by human participants, for example:",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 93,
+ 506,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 93,
+ 506,
+ 107
+ ],
+ "score": 1.0,
+ "content": "“I noticed that some of my partners seemed to know they needed to move around me, while others",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 106,
+ 473,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 106,
+ 473,
+ 119
+ ],
+ "score": 1.0,
+ "content": "seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples).",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 129,
+ 444,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 446,
+ 143
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 446,
+ 143
+ ],
+ "score": 1.0,
+ "content": "Finding 2: FCP’s preferences over cooking pots aligns best with that of humans",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 143,
+ 505,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 143,
+ 505,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 143,
+ 505,
+ 155
+ ],
+ "score": 1.0,
+ "content": "Next, we investigate whether there was a preference for a specific cooking pot in the layouts which",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 153,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 153,
+ 243,
+ 167
+ ],
+ "score": 1.0,
+ "content": "included two cooking pots (Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 154,
+ 258,
+ 167
+ ],
+ "score": 0.55,
+ "content": "\\textcircled { 8 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 153,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": ". To do this, we calculate the difference in the number of times",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "each pot was used by each player, where a high value indicates a strong preference for one pot and a",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 344,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 344,
+ 188
+ ],
+ "score": 1.0,
+ "content": "low value indicates more equal preference for the two pots.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 192,
+ 505,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 505,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 505,
+ 204
+ ],
+ "score": 1.0,
+ "content": "As can be seen in the FCP column, our agent typically has the most aligned preferences with that of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 213,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 230,
+ 227
+ ],
+ "score": 1.0,
+ "content": "one cooking pot over the other",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 214,
+ 258,
+ 225
+ ],
+ "score": 0.88,
+ "content": "5 5 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 213,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "of the time (i.e. a 0.11 point difference). In contrast, all other",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "score": 1.0,
+ "content": "agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 235,
+ 442,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 442,
+ 249
+ ],
+ "score": 1.0,
+ "content": "poorly to typical human behavior of using both pots, leading to worse performance.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 259,
+ 180,
+ 273
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 258,
+ 181,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 258,
+ 181,
+ 275
+ ],
+ "score": 1.0,
+ "content": "6 Discussion",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 279,
+ 505,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 292
+ ],
+ "score": 1.0,
+ "content": "Summary In this work, we investigated the challenging problem of zero-shot collaboration with",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "humans without using human data in the training pipeline. To accomplish this, we introduced",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "of training partners. We found that FCP agents scored significantly higher than all baselines when",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 336
+ ],
+ "score": 1.0,
+ "content": "partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "experimental design, we also found that humans reported a strong subjective preference to partnering",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 249,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 249,
+ 358
+ ],
+ "score": 1.0,
+ "content": "with FCP agents over all baselines.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Limitations and future work Our method currently relies on the manual process of initially",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "training and selecting a diverse set of partners. This is not only time consuming, but also prone to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "researcher biases that may negatively influence the behavior of the created agents. Additionally, while",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 393,
+ 507,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 393,
+ 302,
+ 408
+ ],
+ "score": 1.0,
+ "content": "we found FCP with a partner population size of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 394,
+ 336,
+ 405
+ ],
+ "score": 0.9,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 336,
+ 393,
+ 507,
+ 408
+ ],
+ "score": 1.0,
+ "content": "sufficient here, for more complex games,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 405,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "FCP may require an unrealistically large partner population size to represent sufficiently diverse",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "score": 1.0,
+ "content": "strategies. To address these concerns, methods for automatically generating partner diversity for",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "common-payoff games may be important. Possibilities include adaptive population matchmaking",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 295,
+ 451
+ ],
+ "score": 1.0,
+ "content": "as been used in competitive zero-sum games",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 437,
+ 313,
+ 449
+ ],
+ "score": 0.8,
+ "content": "\\mathbb { \\lVert 6 9 \\rVert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 437,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": ", as well as auxiliary objectives that explicitly",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 281,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 281,
+ 461
+ ],
+ "score": 1.0,
+ "content": "encourage behavioral diversity [19, 45, 46].",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 465,
+ 505,
+ 564
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 466,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 466,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "Our method requires a known and fixed reward function. We also focus on one domain in order to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 489
+ ],
+ "score": 1.0,
+ "content": "compare with prior work which has argued that human-in-the-loop training is necessary. Consequently,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 499
+ ],
+ "score": 1.0,
+ "content": "the resulting agents are only designed to adaptively collaborate on a single task, and not to infer",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 230,
+ 511
+ ],
+ "score": 1.0,
+ "content": "human preferences in general",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 497,
+ 273,
+ 509
+ ],
+ "score": 0.46,
+ "content": "\\textcircled { 1 1 } \\textcircled { 3 3 } \\textcircled { 5 9 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 497,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": ". Moreover, if a task’s reward function is poorly aligned",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": "with how humans approach the task, our method may well produce subpar partners, as would",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 505,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 505,
+ 531
+ ],
+ "score": 1.0,
+ "content": "any method without access to human data. Thus, additional domains and tasks should be studied",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 543
+ ],
+ "score": 1.0,
+ "content": "to better understand how our method generalizes. Targeted experiments to test specific forms of",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 328,
+ 554
+ ],
+ "score": 1.0,
+ "content": "generalization may be especially helpful in this regard",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 541,
+ 346,
+ 552
+ ],
+ "score": 0.76,
+ "content": "\\overline { { \\| 3 8 \\| } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 542,
+ 504,
+ 554
+ ],
+ "score": 1.0,
+ "content": ", as could approaches that procedurally",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 352,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 335,
+ 565
+ ],
+ "score": 1.0,
+ "content": "generate environment layouts requiring diverse solutions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 552,
+ 352,
+ 564
+ ],
+ "score": 0.39,
+ "content": "\\lVert 2 2 \\rVert",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 569,
+ 505,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "Finally, it may be possible to produce even stronger agent assistants by combining the strengths of",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 395,
+ 592
+ ],
+ "score": 1.0,
+ "content": "FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 396,
+ 579,
+ 413,
+ 591
+ ],
+ "score": 0.5,
+ "content": "\\pmb { \\Vert 3 8 \\Vert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 414,
+ 579,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "recently demonstrated",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 590,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "that modifying BCP to train with multiple BC partners produces more robust collaboration with",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 602,
+ 425,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 425,
+ 614
+ ],
+ "score": 1.0,
+ "content": "held-out agents, a finding that would be interesting to test with human partners.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 617,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 652
+ ],
+ "score": 1.0,
+ "content": "policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "expresses preferences or intentions to harm others). This could therefore produce negative societal",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 672,
+ 486,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 672,
+ 486,
+ 685
+ ],
+ "score": 1.0,
+ "content": "effects either if training leads to poor alignment or if agents are optimized for harmful metrics.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 46.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "One potential strategy for mitigating these risks is the use of human preference data [15]. Such data",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "could be used to fine-tune and filter trained agents before deployment, encouraging better alignment",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "with human values. A key question in this line of research is how human preference data should be",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 86
+ ],
+ "score": 1.0,
+ "content": "(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "movement strategy with its partner. This result was also reported by human participants, for example:",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 93,
+ 506,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 93,
+ 506,
+ 107
+ ],
+ "score": 1.0,
+ "content": "“I noticed that some of my partners seemed to know they needed to move around me, while others",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 106,
+ 473,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 106,
+ 473,
+ 119
+ ],
+ "score": 1.0,
+ "content": "seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples).",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5,
+ "bbox_fs": [
+ 104,
+ 73,
+ 506,
+ 119
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 129,
+ 444,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 446,
+ 143
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 446,
+ 143
+ ],
+ "score": 1.0,
+ "content": "Finding 2: FCP’s preferences over cooking pots aligns best with that of humans",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 143,
+ 505,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 143,
+ 505,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 143,
+ 505,
+ 155
+ ],
+ "score": 1.0,
+ "content": "Next, we investigate whether there was a preference for a specific cooking pot in the layouts which",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 153,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 153,
+ 243,
+ 167
+ ],
+ "score": 1.0,
+ "content": "included two cooking pots (Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 154,
+ 258,
+ 167
+ ],
+ "score": 0.55,
+ "content": "\\textcircled { 8 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 153,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": ". To do this, we calculate the difference in the number of times",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "each pot was used by each player, where a high value indicates a strong preference for one pot and a",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 344,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 344,
+ 188
+ ],
+ "score": 1.0,
+ "content": "low value indicates more equal preference for the two pots.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 143,
+ 506,
+ 188
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 192,
+ 505,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 505,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 505,
+ 204
+ ],
+ "score": 1.0,
+ "content": "As can be seen in the FCP column, our agent typically has the most aligned preferences with that of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 213,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 230,
+ 227
+ ],
+ "score": 1.0,
+ "content": "one cooking pot over the other",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 214,
+ 258,
+ 225
+ ],
+ "score": 0.88,
+ "content": "5 5 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 213,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "of the time (i.e. a 0.11 point difference). In contrast, all other",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 238
+ ],
+ "score": 1.0,
+ "content": "agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 235,
+ 442,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 442,
+ 249
+ ],
+ "score": 1.0,
+ "content": "poorly to typical human behavior of using both pots, leading to worse performance.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 191,
+ 506,
+ 249
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 259,
+ 180,
+ 273
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 258,
+ 181,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 258,
+ 181,
+ 275
+ ],
+ "score": 1.0,
+ "content": "6 Discussion",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 279,
+ 505,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 292
+ ],
+ "score": 1.0,
+ "content": "Summary In this work, we investigated the challenging problem of zero-shot collaboration with",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "humans without using human data in the training pipeline. To accomplish this, we introduced",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "of training partners. We found that FCP agents scored significantly higher than all baselines when",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 336
+ ],
+ "score": 1.0,
+ "content": "partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "experimental design, we also found that humans reported a strong subjective preference to partnering",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 249,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 249,
+ 358
+ ],
+ "score": 1.0,
+ "content": "with FCP agents over all baselines.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 280,
+ 506,
+ 358
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Limitations and future work Our method currently relies on the manual process of initially",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "training and selecting a diverse set of partners. This is not only time consuming, but also prone to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "researcher biases that may negatively influence the behavior of the created agents. Additionally, while",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 393,
+ 507,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 393,
+ 302,
+ 408
+ ],
+ "score": 1.0,
+ "content": "we found FCP with a partner population size of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 394,
+ 336,
+ 405
+ ],
+ "score": 0.9,
+ "content": "N = 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 336,
+ 393,
+ 507,
+ 408
+ ],
+ "score": 1.0,
+ "content": "sufficient here, for more complex games,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 405,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "FCP may require an unrealistically large partner population size to represent sufficiently diverse",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "score": 1.0,
+ "content": "strategies. To address these concerns, methods for automatically generating partner diversity for",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "common-payoff games may be important. Possibilities include adaptive population matchmaking",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 295,
+ 451
+ ],
+ "score": 1.0,
+ "content": "as been used in competitive zero-sum games",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 437,
+ 313,
+ 449
+ ],
+ "score": 0.8,
+ "content": "\\mathbb { \\lVert 6 9 \\rVert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 437,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": ", as well as auxiliary objectives that explicitly",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 281,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 281,
+ 461
+ ],
+ "score": 1.0,
+ "content": "encourage behavioral diversity [19, 45, 46].",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 104,
+ 361,
+ 507,
+ 461
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 465,
+ 505,
+ 564
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 466,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 466,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "Our method requires a known and fixed reward function. We also focus on one domain in order to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 489
+ ],
+ "score": 1.0,
+ "content": "compare with prior work which has argued that human-in-the-loop training is necessary. Consequently,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 499
+ ],
+ "score": 1.0,
+ "content": "the resulting agents are only designed to adaptively collaborate on a single task, and not to infer",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 230,
+ 511
+ ],
+ "score": 1.0,
+ "content": "human preferences in general",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 497,
+ 273,
+ 509
+ ],
+ "score": 0.46,
+ "content": "\\textcircled { 1 1 } \\textcircled { 3 3 } \\textcircled { 5 9 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 497,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": ". Moreover, if a task’s reward function is poorly aligned",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": "with how humans approach the task, our method may well produce subpar partners, as would",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 505,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 505,
+ 531
+ ],
+ "score": 1.0,
+ "content": "any method without access to human data. Thus, additional domains and tasks should be studied",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 543
+ ],
+ "score": 1.0,
+ "content": "to better understand how our method generalizes. Targeted experiments to test specific forms of",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 328,
+ 554
+ ],
+ "score": 1.0,
+ "content": "generalization may be especially helpful in this regard",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 541,
+ 346,
+ 552
+ ],
+ "score": 0.76,
+ "content": "\\overline { { \\| 3 8 \\| } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 542,
+ 504,
+ 554
+ ],
+ "score": 1.0,
+ "content": ", as could approaches that procedurally",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 352,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 335,
+ 565
+ ],
+ "score": 1.0,
+ "content": "generate environment layouts requiring diverse solutions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 552,
+ 352,
+ 564
+ ],
+ "score": 0.39,
+ "content": "\\lVert 2 2 \\rVert",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 466,
+ 506,
+ 565
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 569,
+ 505,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "Finally, it may be possible to produce even stronger agent assistants by combining the strengths of",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 395,
+ 592
+ ],
+ "score": 1.0,
+ "content": "FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 396,
+ 579,
+ 413,
+ 591
+ ],
+ "score": 0.5,
+ "content": "\\pmb { \\Vert 3 8 \\Vert }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 414,
+ 579,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "recently demonstrated",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 590,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "that modifying BCP to train with multiple BC partners produces more robust collaboration with",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 602,
+ 425,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 425,
+ 614
+ ],
+ "score": 1.0,
+ "content": "held-out agents, a finding that would be interesting to test with human partners.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 569,
+ 506,
+ 614
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 617,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 652
+ ],
+ "score": 1.0,
+ "content": "policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "expresses preferences or intentions to harm others). This could therefore produce negative societal",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 672,
+ 486,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 672,
+ 486,
+ 685
+ ],
+ "score": 1.0,
+ "content": "effects either if training leads to poor alignment or if agents are optimized for harmful metrics.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 617,
+ 506,
+ 685
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "One potential strategy for mitigating these risks is the use of human preference data [15]. Such data",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "could be used to fine-tune and filter trained agents before deployment, encouraging better alignment",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "with human values. A key question in this line of research is how human preference data should be",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "aggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 97
+ ],
+ "score": 1.0,
+ "content": "agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 225,
+ 107
+ ],
+ "score": 1.0,
+ "content": "beliefs and perceptions of AI",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 226,
+ 94,
+ 243,
+ 106
+ ],
+ "score": 0.47,
+ "content": "\\pm 8 \\jmath",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 244,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": ", and how they steer human-agent interaction, would help inform",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 104,
+ 506,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 104,
+ 506,
+ 120
+ ],
+ "score": 1.0,
+ "content": "agent design for positive societal impact. For instance, developers could incorporate specific priors",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 351,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 316,
+ 128
+ ],
+ "score": 1.0,
+ "content": "into agents to reinforce tendencies for fair outcomes",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 317,
+ 116,
+ 349,
+ 128
+ ],
+ "score": 0.28,
+ "content": "\\pm \\mathbb { Z } 0 . \\pm \\mathbb { B } 2 \\mathbb { I }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 349,
+ 115,
+ 351,
+ 128
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 51,
+ "bbox_fs": [
+ 106,
+ 689,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 73,
+ 505,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "aggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 97
+ ],
+ "score": 1.0,
+ "content": "agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 225,
+ 107
+ ],
+ "score": 1.0,
+ "content": "beliefs and perceptions of AI",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 94,
+ 243,
+ 106
+ ],
+ "score": 0.47,
+ "content": "\\pm 8 \\jmath",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": ", and how they steer human-agent interaction, would help inform",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 104,
+ 506,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 104,
+ 506,
+ 120
+ ],
+ "score": 1.0,
+ "content": "agent design for positive societal impact. For instance, developers could incorporate specific priors",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 351,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 316,
+ 128
+ ],
+ "score": 1.0,
+ "content": "into agents to reinforce tendencies for fair outcomes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 116,
+ 349,
+ 128
+ ],
+ "score": 0.28,
+ "content": "\\pm \\mathbb { Z } 0 . \\pm \\mathbb { B } 2 \\mathbb { I }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 115,
+ 351,
+ 128
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 132,
+ 505,
+ 177
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Conclusion We proposed a method which is both effective at collaborating with humans and simple",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "to implement. We also presented a rigorous and general methodology for evaluating with humans",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "and eliciting their preferences. Together, these establish a strong foundation for future research on",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 413,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 413,
+ 179
+ ],
+ "score": 1.0,
+ "content": "the important challenge of human-agent collaboration for benefiting society.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 192,
+ 207,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 208,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 208,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Acknowledgements",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 216,
+ 505,
+ 260
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 249,
+ 420,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 420,
+ 261
+ ],
+ "score": 1.0,
+ "content": "Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 274,
+ 205,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 207,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 207,
+ 291
+ ],
+ "score": 1.0,
+ "content": "Funding disclosure",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 298,
+ 459,
+ 310
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 462,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 462,
+ 313
+ ],
+ "score": 1.0,
+ "content": "This work was funded solely by DeepMind. The authors declare no competing interests.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 325,
+ 163,
+ 338
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 165,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 165,
+ 339
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 338,
+ 506,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 343,
+ 506,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 343,
+ 506,
+ 356
+ ],
+ "score": 1.0,
+ "content": "[1] J. Abramson, A. Ahuja, I. Barr, A. Brussee, F. Carnevale, M. Cassin, R. Chhaparia, S. Clark,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 126,
+ 354,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 354,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "B. Damoc, A. Dudzik, et al. Imitating interactive intelligence. arXiv preprint arXiv:2012.05672,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 126,
+ 365,
+ 153,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 365,
+ 153,
+ 378
+ ],
+ "score": 1.0,
+ "content": "2020.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 110,
+ 380,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 380,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "[2] S. V. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 127,
+ 390,
+ 505,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 390,
+ 505,
+ 404
+ ],
+ "score": 1.0,
+ "content": "for ad hoc coordination in multiagent systems. In International Conference on Autonomous",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 126,
+ 402,
+ 323,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 402,
+ 323,
+ 414
+ ],
+ "score": 1.0,
+ "content": "Agents and Multiagent Systems (AAMAS), 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 111,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "[3] B. Baker. Emergent reciprocity and team formation from randomized uncertain social prefer-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 127,
+ 426,
+ 398,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 426,
+ 398,
+ 441
+ ],
+ "score": 1.0,
+ "content": "ences. In Neural Information Processing Systems (NeurIPS), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 111,
+ 441,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 441,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "[4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 127,
+ 452,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 452,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "S. Moitra, E. Hughes, et al. The Hanabi challenge: A new frontier for AI research. Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 127,
+ 462,
+ 257,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 462,
+ 257,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Intelligence, 280:103216, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 110,
+ 476,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 476,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "[5] S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 126,
+ 488,
+ 369,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 488,
+ 369,
+ 500
+ ],
+ "score": 1.0,
+ "content": "new teammates. Artificial Intelligence, 242:132–171, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 111,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "[6] A. Bauer, D. Wollherr, and M. Buss. Human–robot collaboration: A survey. International",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 126,
+ 513,
+ 335,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 513,
+ 335,
+ 525
+ ],
+ "score": 1.0,
+ "content": "Journal of Humanoid Robotics, 5(01):47–66, 2008.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 110,
+ 527,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 527,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "[7] G. W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 127,
+ 539,
+ 262,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 539,
+ 262,
+ 550
+ ],
+ "score": 1.0,
+ "content": "allocation, 13(1):374–376, 1951.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 111,
+ 550,
+ 505,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 550,
+ 505,
+ 566
+ ],
+ "score": 1.0,
+ "content": "[8] N. Brown and T. Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 127,
+ 563,
+ 331,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 563,
+ 331,
+ 576
+ ],
+ "score": 1.0,
+ "content": "professionals. Science, 359(6374):418–424, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 110,
+ 576,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 576,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "[9] N. Brown and T. Sandholm. Superhuman AI for multiplayer poker. Science, 365(6456):885–890,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 127,
+ 589,
+ 153,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 589,
+ 153,
+ 600
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 602,
+ 506,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 506,
+ 616
+ ],
+ "score": 1.0,
+ "content": "[10] K. Bullard, F. Meier, D. Kiela, J. Pineau, and J. Foerster. Exploring zero-shot emergent",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 127,
+ 614,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 614,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "communication in embodied multi-agent populations. arXiv preprint arXiv:2010.15896, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 641
+ ],
+ "score": 1.0,
+ "content": "[11] K. Bullard, D. Kiela, J. Pineau, and J. Foerster. Quasi-equivalence discovery for zero-shot",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 126,
+ 639,
+ 397,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 639,
+ 397,
+ 651
+ ],
+ "score": 1.0,
+ "content": "emergent communication. arXiv preprint arXiv:2103.08067, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "[12] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 126,
+ 663,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 663,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "utility of learning about humans for human-AI coordination. In Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 127,
+ 676,
+ 233,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 676,
+ 233,
+ 687
+ ],
+ "score": 1.0,
+ "content": "Systems (NeurIPS), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "[13] R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Investigating partner diversification",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 127,
+ 700,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 700,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "methods in cooperative multi-agent deep reinforcement learning. In International Conference",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 127,
+ 711,
+ 294,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 711,
+ 294,
+ 723
+ ],
+ "score": 1.0,
+ "content": "on Neural Information Processing, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 73,
+ 505,
+ 127
+ ],
+ "lines": [],
+ "index": 2,
+ "bbox_fs": [
+ 104,
+ 73,
+ 506,
+ 128
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 132,
+ 505,
+ 177
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Conclusion We proposed a method which is both effective at collaborating with humans and simple",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "to implement. We also presented a rigorous and general methodology for evaluating with humans",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "and eliciting their preferences. Together, these establish a strong foundation for future research on",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 413,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 413,
+ 179
+ ],
+ "score": 1.0,
+ "content": "the important challenge of human-agent collaboration for benefiting society.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 132,
+ 505,
+ 179
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 192,
+ 207,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 208,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 208,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Acknowledgements",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 216,
+ 505,
+ 260
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 249,
+ 420,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 420,
+ 261
+ ],
+ "score": 1.0,
+ "content": "Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 215,
+ 505,
+ 261
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 274,
+ 205,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 207,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 207,
+ 291
+ ],
+ "score": 1.0,
+ "content": "Funding disclosure",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 298,
+ 459,
+ 310
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 462,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 462,
+ 313
+ ],
+ "score": 1.0,
+ "content": "This work was funded solely by DeepMind. The authors declare no competing interests.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 296,
+ 462,
+ 313
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 325,
+ 163,
+ 338
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 165,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 165,
+ 339
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 109,
+ 338,
+ 506,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 343,
+ 506,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 343,
+ 506,
+ 356
+ ],
+ "score": 1.0,
+ "content": "[1] J. Abramson, A. Ahuja, I. Barr, A. Brussee, F. Carnevale, M. Cassin, R. Chhaparia, S. Clark,",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 354,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 354,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "B. Damoc, A. Dudzik, et al. Imitating interactive intelligence. arXiv preprint arXiv:2012.05672,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 126,
+ 365,
+ 153,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 365,
+ 153,
+ 378
+ ],
+ "score": 1.0,
+ "content": "2020.",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 380,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 380,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "[2] S. V. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 390,
+ 505,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 390,
+ 505,
+ 404
+ ],
+ "score": 1.0,
+ "content": "for ad hoc coordination in multiagent systems. In International Conference on Autonomous",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 126,
+ 402,
+ 323,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 402,
+ 323,
+ 414
+ ],
+ "score": 1.0,
+ "content": "Agents and Multiagent Systems (AAMAS), 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "[3] B. Baker. Emergent reciprocity and team formation from randomized uncertain social prefer-",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 426,
+ 398,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 426,
+ 398,
+ 441
+ ],
+ "score": 1.0,
+ "content": "ences. In Neural Information Processing Systems (NeurIPS), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 441,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 441,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "[4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin,",
+ "type": "text"
+ }
+ ],
+ "index": 25,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 452,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 452,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "S. Moitra, E. Hughes, et al. The Hanabi challenge: A new frontier for AI research. Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 127,
+ 462,
+ 257,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 462,
+ 257,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Intelligence, 280:103216, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 27,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 476,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 476,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "[5] S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 488,
+ 369,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 488,
+ 369,
+ 500
+ ],
+ "score": 1.0,
+ "content": "new teammates. Artificial Intelligence, 242:132–171, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "[6] A. Bauer, D. Wollherr, and M. Buss. Human–robot collaboration: A survey. International",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 513,
+ 335,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 513,
+ 335,
+ 525
+ ],
+ "score": 1.0,
+ "content": "Journal of Humanoid Robotics, 5(01):47–66, 2008.",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 527,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 527,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "[7] G. W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 539,
+ 262,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 539,
+ 262,
+ 550
+ ],
+ "score": 1.0,
+ "content": "allocation, 13(1):374–376, 1951.",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 550,
+ 505,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 550,
+ 505,
+ 566
+ ],
+ "score": 1.0,
+ "content": "[8] N. Brown and T. Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 563,
+ 331,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 563,
+ 331,
+ 576
+ ],
+ "score": 1.0,
+ "content": "professionals. Science, 359(6374):418–424, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 576,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 576,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "[9] N. Brown and T. Sandholm. Superhuman AI for multiplayer poker. Science, 365(6456):885–890,",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 589,
+ 153,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 589,
+ 153,
+ 600
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 602,
+ 506,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 506,
+ 616
+ ],
+ "score": 1.0,
+ "content": "[10] K. Bullard, F. Meier, D. Kiela, J. Pineau, and J. Foerster. Exploring zero-shot emergent",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 614,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 614,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "communication in embodied multi-agent populations. arXiv preprint arXiv:2010.15896, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 641
+ ],
+ "score": 1.0,
+ "content": "[11] K. Bullard, D. Kiela, J. Pineau, and J. Foerster. Quasi-equivalence discovery for zero-shot",
+ "type": "text"
+ }
+ ],
+ "index": 40,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 639,
+ 397,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 639,
+ 397,
+ 651
+ ],
+ "score": 1.0,
+ "content": "emergent communication. arXiv preprint arXiv:2103.08067, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 41,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "[12] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 663,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 663,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "utility of learning about humans for human-AI coordination. In Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 127,
+ 676,
+ 233,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 676,
+ 233,
+ 687
+ ],
+ "score": 1.0,
+ "content": "Systems (NeurIPS), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "[13] R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Investigating partner diversification",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 700,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 700,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "methods in cooperative multi-agent deep reinforcement learning. In International Conference",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 127,
+ 711,
+ 294,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 711,
+ 294,
+ 723
+ ],
+ "score": 1.0,
+ "content": "on Neural Information Processing, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 47,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "[14] R. Choudhury, G. Swamy, D. Hadfield-Menell, and A. D. Dragan. On the utility of model",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 98
+ ],
+ "score": 1.0,
+ "content": "learning in HRI. In ACM/IEEE International Conference on Human-Robot Interaction (HRI),",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 126,
+ 92,
+ 155,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 92,
+ 155,
+ 109
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "[15] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 120,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 120,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": "learning from human preferences. In Proceedings of the 31st International Conference on",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 126,
+ 131,
+ 392,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 131,
+ 392,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Neural Information Processing Systems, pages 4302–4310, 2017.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 146,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 146,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "[16] A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 157,
+ 477,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 157,
+ 477,
+ 172
+ ],
+ "score": 1.0,
+ "content": "T. Graepel. Open problems in cooperative AI. arXiv preprint arXiv:2012.08630, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 507,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 507,
+ 187
+ ],
+ "score": 1.0,
+ "content": "[17] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 182,
+ 507,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 182,
+ 507,
+ 198
+ ],
+ "score": 1.0,
+ "content": "T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable distributed deep-RL",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 194,
+ 506,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 194,
+ 506,
+ 209
+ ],
+ "score": 1.0,
+ "content": "with importance weighted actor-learner architectures. In International Conference on Machine",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 127,
+ 206,
+ 226,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 206,
+ 226,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Learning (ICML), 2018.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "[18] P. Eyal, R. David, G. Andrew, E. Zak, and D. Ekaterina. Data quality of platforms and panels",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 231,
+ 447,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 231,
+ 447,
+ 245
+ ],
+ "score": 1.0,
+ "content": "for online behavioral research. Behavior Research Methods, pages 1–20, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 260
+ ],
+ "score": 1.0,
+ "content": "[19] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 258,
+ 507,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 258,
+ 507,
+ 272
+ ],
+ "score": 1.0,
+ "content": "without a reward function. In International Conference on Learning Representations (ICLR),",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 127,
+ 267,
+ 155,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 267,
+ 155,
+ 283
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 298
+ ],
+ "score": 1.0,
+ "content": "[20] E. Fehr and K. M. Schmidt. A theory of fairness, competition, and cooperation. The Quarterly",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 294,
+ 314,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 294,
+ 314,
+ 309
+ ],
+ "score": 1.0,
+ "content": "Journal of Economics, 114(3):817–868, 1999.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 507,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 507,
+ 324
+ ],
+ "score": 1.0,
+ "content": "[21] J. Foerster, F. Song, E. Hughes, N. Burch, I. Dunning, S. Whiteson, M. Botvinick, and M. Bowl-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 321,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 321,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "ing. Bayesian action decoder for deep multi-agent reinforcement learning. In International",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 126,
+ 331,
+ 325,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 331,
+ 325,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Conference on Machine Learning (ICML), 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 507,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 507,
+ 360
+ ],
+ "score": 1.0,
+ "content": "[22] M. Fontaine, Y.-C. Hsu, Y. Zhang, B. Tjanaka, and S. Nikolaidis. On the importance of",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 358,
+ 496,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 358,
+ 496,
+ 370
+ ],
+ "score": 1.0,
+ "content": "environments in human-robot coordination. In Robotics: Science and Systems (RSS), 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "[23] I. Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30(3):411–437,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 24,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 384,
+ 154,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 384,
+ 154,
+ 397
+ ],
+ "score": 1.0,
+ "content": "2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "[24] M. Garnelo, W. M. Czarnecki, S. Liu, D. Tirumala, J. Oh, G. Gidel, H. van Hasselt, and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 410,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 410,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "D. Balduzzi. Pick your battles: Interaction graphs as population-level objectives for strategic",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 126,
+ 419,
+ 507,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 419,
+ 507,
+ 435
+ ],
+ "score": 1.0,
+ "content": "diversity. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS),",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 126,
+ 431,
+ 155,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 431,
+ 155,
+ 446
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 29,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 291,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 291,
+ 460
+ ],
+ "score": 1.0,
+ "content": "[25] Ghost Town Games. Overcooked, 2016.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "[26] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 473,
+ 419,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 473,
+ 419,
+ 487
+ ],
+ "score": 1.0,
+ "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2016.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 32,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "[27] J. Heinrich and D. Silver. Deep reinforcement learning from self-play in imperfect-information",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 499,
+ 415,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 499,
+ 415,
+ 512
+ ],
+ "score": 1.0,
+ "content": "games. In NIPS Deep Reinforcement Learning (DRL) Workshop, 2016.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 34,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "[28] J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 35,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 525,
+ 380,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 525,
+ 380,
+ 538
+ ],
+ "score": 1.0,
+ "content": "International Conference on Machine Learning (ICML), 2015.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 36,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 541,
+ 506,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 506,
+ 554
+ ],
+ "score": 1.0,
+ "content": "[29] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 37,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 552,
+ 202,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 552,
+ 202,
+ 564
+ ],
+ "score": 1.0,
+ "content": "1735–1780, 1997.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 38,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 507,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 507,
+ 580
+ ],
+ "score": 1.0,
+ "content": "[30] M. Hoffman, B. Shahriari, J. Aslanides, G. Barth-Maron, F. Behbahani, T. Norman, A. Abdol-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 39,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 577,
+ 507,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 577,
+ 507,
+ 591
+ ],
+ "score": 1.0,
+ "content": "maleki, A. Cassirer, F. Yang, K. Baumli, S. Henderson, A. Novikov, S. G. Colmenarejo, S. Cabi,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 127,
+ 588,
+ 506,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 588,
+ 506,
+ 603
+ ],
+ "score": 1.0,
+ "content": "C. Gulcehre, T. L. Paine, A. Cowie, Z. Wang, B. Piot, and N. de Freitas. Acme: A research",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 126,
+ 599,
+ 492,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 599,
+ 492,
+ 613
+ ],
+ "score": 1.0,
+ "content": "framework for distributed reinforcement learning. arXiv preprint arXiv:2006.00979, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "[31] H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “Other-Play” for zero-shot coordination. In",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 625,
+ 381,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 625,
+ 381,
+ 639
+ ],
+ "score": 1.0,
+ "content": "International Conference on Machine Learning (ICML), 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 507,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 507,
+ 654
+ ],
+ "score": 1.0,
+ "content": "[32] E. Hughes, J. Z. Leibo, M. G. Phillips, K. Tuyls, E. A. Duéñez-Guzmán, A. G. Castañeda,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 651,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 651,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": "I. Dunning, T. Zhu, K. R. McKee, R. Koster, H. Roff, and T. Graepel. Inequity aversion",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 127,
+ 663,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 663,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "improves cooperation in intertemporal social dilemmas. In Neural Information Processing",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 127,
+ 674,
+ 232,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 674,
+ 232,
+ 686
+ ],
+ "score": 1.0,
+ "content": "Systems (NeurIPS), 2018.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 48,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 703
+ ],
+ "score": 1.0,
+ "content": "[33] B. Ibarz, J. Leike, T. Pohlen, G. Irving, S. Legg, and D. Amodei. Reward learning from human",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 49,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 714
+ ],
+ "score": 1.0,
+ "content": "preferences and demonstrations in Atari. In Neural Information Processing Systems (NeurIPS),",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 126,
+ 709,
+ 155,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 709,
+ 155,
+ 725
+ ],
+ "score": 1.0,
+ "content": "2018.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 51,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "[34] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 97
+ ],
+ "score": 1.0,
+ "content": "N. C. Rabinowitz, A. S. Morcos, A. Ruderman, et al. Human-level performance in 3D mul-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 95,
+ 507,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 95,
+ 507,
+ 109
+ ],
+ "score": 1.0,
+ "content": "tiplayer games with population-based reinforcement learning. Science, 364(6443):859–865,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 104,
+ 154,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 104,
+ 154,
+ 119
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 120,
+ 507,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 120,
+ 507,
+ 134
+ ],
+ "score": 1.0,
+ "content": "[35] S. Javdani, S. S. Srinivasa, and J. A. Bagnell. Shared autonomy via hindsight optimization.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 131,
+ 277,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 131,
+ 277,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Robotics Science and Systems, 2015.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 147,
+ 507,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 507,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[36] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, and J. B. Tenenbaum. Co-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 157,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 157,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "ordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 127,
+ 168,
+ 371,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 168,
+ 371,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Conference of the Cognitive Science Society (CogSci), 2016.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 183,
+ 506,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 183,
+ 506,
+ 197
+ ],
+ "score": 1.0,
+ "content": "[37] G. Klien, D. D. Woods, J. M. Bradshaw, R. R. Hoffman, and P. J. Feltovich. Ten challenges for",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 195,
+ 505,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 195,
+ 505,
+ 209
+ ],
+ "score": 1.0,
+ "content": "making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 127,
+ 207,
+ 197,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 207,
+ 197,
+ 218
+ ],
+ "score": 1.0,
+ "content": "(6):91–95, 2004.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 236
+ ],
+ "score": 1.0,
+ "content": "[38] P. Knott, M. Carroll, S. Devlin, K. Ciosek, K. Hofmann, A. Dragan, and R. Shah. Evaluating",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 231,
+ 506,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 231,
+ 506,
+ 245
+ ],
+ "score": 1.0,
+ "content": "the robustness of collaborative agents. In International Conference on Autonomous Agents and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 127,
+ 244,
+ 276,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 244,
+ 276,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Multiagent Systems (AAMAS), 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "[39] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 269,
+ 506,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 269,
+ 506,
+ 282
+ ],
+ "score": 1.0,
+ "content": "T. Graepel. A unified game-theoretic approach to multiagent reinforcement learning. In Neural",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 127,
+ 280,
+ 317,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 280,
+ 317,
+ 294
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems (NIPS), 2017.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 296,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 296,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "[40] A. Lerer and A. Peysakhovich. Learning existing social conventions via observationally",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 306,
+ 505,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 306,
+ 505,
+ 320
+ ],
+ "score": 1.0,
+ "content": "augmented self-play. In AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 126,
+ 316,
+ 185,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 316,
+ 185,
+ 330
+ ],
+ "score": 1.0,
+ "content": "(AIES), 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 20,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 330,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 330,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "[41] E. Lockhart, N. Burch, N. Bard, S. Borgeaud, T. Eccles, L. Smaira, and R. Smith. Human-agent",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 342,
+ 416,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 342,
+ 416,
+ 357
+ ],
+ "score": 1.0,
+ "content": "cooperation in bridge bidding. arXiv preprint arXiv:2011.14124, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 22,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 357,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 357,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "[42] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch. Multi-agent actor-critic",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 23,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 368,
+ 505,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 368,
+ 505,
+ 383
+ ],
+ "score": 1.0,
+ "content": "for mixed cooperative-competitive environments. In Neural Information Processing Systems",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 126,
+ 380,
+ 186,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 380,
+ 186,
+ 394
+ ],
+ "score": 1.0,
+ "content": "(NIPS), 2017.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 408
+ ],
+ "score": 1.0,
+ "content": "[43] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. Learning to learn to communicate. In",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 406,
+ 352,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 406,
+ 352,
+ 419
+ ],
+ "score": 1.0,
+ "content": "ICML Adaptive & Multitask Learning Workshop, 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 27,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 507,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 507,
+ 435
+ ],
+ "score": 1.0,
+ "content": "[44] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. On the interaction between super-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 28,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 431,
+ 505,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 431,
+ 505,
+ 446
+ ],
+ "score": 1.0,
+ "content": "vision and self-play in emergent communication. In International Conference on Learning",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 443,
+ 254,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 443,
+ 254,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Representations (ICLR), 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 30,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "[45] A. Lupu, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 468,
+ 434,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 468,
+ 434,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 32,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "[46] A. Mahajan, T. Rashid, M. Samvelyan, and S. Whiteson. Maven: Multi-agent variational",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 495,
+ 421,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 495,
+ 421,
+ 508
+ ],
+ "score": 1.0,
+ "content": "exploration. In Neural Information Processing Systems (NeurIPS), 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 34,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 509,
+ 507,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 509,
+ 507,
+ 525
+ ],
+ "score": 1.0,
+ "content": "[47] K. R. McKee, I. Gemp, B. McWilliams, E. A. Duèñez-Guzmán, E. Hughes, and J. Z. Leibo.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 35,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 522,
+ 507,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 522,
+ 507,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Social diversity and social preferences in mixed-motive reinforcement learning. In International",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 532,
+ 435,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 532,
+ 435,
+ 545
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 507,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 507,
+ 562
+ ],
+ "score": 1.0,
+ "content": "[48] K. R. McKee, X. Bai, and S. Fiske. Understanding human impressions of artificial intelligence.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 557,
+ 196,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 557,
+ 196,
+ 572
+ ],
+ "score": 1.0,
+ "content": "PsyArXiv, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 39,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "[49] K. R. McKee, E. Hughes, T. O. Zhu, M. J. Chadwick, R. Koster, A. G. Castaneda, C. Beattie,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 40,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 584,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 584,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "T. Graepel, M. Botvinick, and J. Z. Leibo. Deep reinforcement learning models the emergent",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 128,
+ 596,
+ 425,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 596,
+ 425,
+ 609
+ ],
+ "score": 1.0,
+ "content": "dynamics of human cooperation. arXiv preprint arXiv:2103.04982, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "[50] K. R. McKee, J. Z. Leibo, C. Beattie, and R. Everett. Quantifying environment and population",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 621,
+ 482,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 621,
+ 482,
+ 635
+ ],
+ "score": 1.0,
+ "content": "diversity in multi-agent reinforcement learning. arXiv preprint arXiv:2102.08370, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "[51] F. Melo and A. Sardinha. Ad hoc teamwork by learning teammates’ task. International",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 648,
+ 441,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 648,
+ 441,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2015.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 46,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 507,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 507,
+ 677
+ ],
+ "score": 1.0,
+ "content": "[52] B. Mutlu, A. Terrell, and C.-M. Huang. Coordination mechanisms in human-robot collaboration.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 47,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 673,
+ 343,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 673,
+ 343,
+ 687
+ ],
+ "score": 1.0,
+ "content": "In HRI Collaborative Manipulation Workshop, 2013.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 48,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "[53] S. Nikolaidis and J. Shah. Human-robot cross-training: Computational formulation, modeling",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 49,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 715
+ ],
+ "score": 1.0,
+ "content": "and evaluation of a human team training strategy. In ACM/IEEE International Conference on",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 127,
+ 711,
+ 288,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 711,
+ 288,
+ 724
+ ],
+ "score": 1.0,
+ "content": "Human-Robot Interaction (HRI), 2013.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 51,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "[54] OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 82,
+ 507,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 82,
+ 507,
+ 99
+ ],
+ "score": 1.0,
+ "content": "M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 94,
+ 506,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 94,
+ 506,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Q. Yuan, W. Zaremba, and L. Zhang. Solving Rubik’s Cube with a robot hand. arXiv preprint",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 105,
+ 232,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 105,
+ 232,
+ 118
+ ],
+ "score": 1.0,
+ "content": "arXiv:1910.07113, 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 119,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 119,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "[55] E. Peer, L. Brandimarte, S. Samat, and A. Acquisti. Beyond the Turk: Alternative platforms for",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 132,
+ 507,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 132,
+ 507,
+ 146
+ ],
+ "score": 1.0,
+ "content": "crowdsourcing behavioral research. Journal of Experimental Social Psychology, 70:153–163,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 127,
+ 142,
+ 155,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 142,
+ 155,
+ 157
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 158,
+ 505,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 158,
+ 505,
+ 170
+ ],
+ "score": 1.0,
+ "content": "[56] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 168,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 168,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "control with dynamics randomization. In IEEE International Conference on Robotics and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 125,
+ 180,
+ 236,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 180,
+ 236,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Automation (ICRA), 2018.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 195,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 195,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "[57] P. M. Pilarski, A. Butcher, M. Johanson, M. M. Botvinick, A. Bolt, and A. S. Parker. Learned",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 207,
+ 507,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 207,
+ 507,
+ 221
+ ],
+ "score": 1.0,
+ "content": "human-agent decision-making, communication and joint action in a virtual reality environment.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 126,
+ 216,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 216,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "In Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM),",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 126,
+ 227,
+ 155,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 227,
+ 155,
+ 243
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": "[58] D. A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 256,
+ 289,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 256,
+ 289,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Neural Computation, 3(1):88–97, 1991.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 285
+ ],
+ "score": 1.0,
+ "content": "[59] S. Russell. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 279,
+ 155,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 279,
+ 155,
+ 295
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": "[60] D. Sadigh, N. Landolfi, S. S. Sastry, S. A. Seshia, and A. D. Dragan. Planning for cars that",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "coordinate with people: leveraging effects on human actions for planning and active information",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 126,
+ 319,
+ 459,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 319,
+ 459,
+ 332
+ ],
+ "score": 1.0,
+ "content": "gathering over human internal state. Autonomous Robots, 42(7):1405–1426, 2018.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 20,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "[61] N. Schurr, J. Marecki, M. Tambe, and P. Scerri. Towards flexible coordination of human-agent",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 345,
+ 348,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 345,
+ 348,
+ 358
+ ],
+ "score": 1.0,
+ "content": "teams. Multiagent and Grid Systems, 1(1):3–16, 2005.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 22,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "[62] A. Shih, A. Sawhney, J. Kondic, S. Ermon, and D. Sadigh. On the critical role of conventions in",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 23,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 372,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 372,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "adaptive human-AI collaboration. In International Conference on Learning Representations",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 126,
+ 382,
+ 187,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 382,
+ 187,
+ 397
+ ],
+ "score": 1.0,
+ "content": "(ICLR), 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": "[63] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 409,
+ 505,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 409,
+ 505,
+ 423
+ ],
+ "score": 1.0,
+ "content": "M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 125,
+ 418,
+ 507,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 418,
+ 507,
+ 435
+ ],
+ "score": 1.0,
+ "content": "D. Hassabis. Mastering the game of go without human knowledge. Nature, 550(7676):354–359,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 126,
+ 428,
+ 156,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 428,
+ 156,
+ 446
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 29,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "[64] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 457,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 457,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 125,
+ 466,
+ 507,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 466,
+ 507,
+ 484
+ ],
+ "score": 1.0,
+ "content": "learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 127,
+ 480,
+ 202,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 480,
+ 202,
+ 492
+ ],
+ "score": 1.0,
+ "content": "1140–1144, 2018.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 507,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 507,
+ 509
+ ],
+ "score": 1.0,
+ "content": "[65] H. F. Song, A. Abdolmaleki, J. T. Springenberg, A. Clark, H. Soyer, J. W. Rae, S. Noury,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 505,
+ 507,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 505,
+ 507,
+ 519
+ ],
+ "score": 1.0,
+ "content": "A. Ahuja, S. Liu, D. Tirumala, N. Heess, D. Belov, M. Riedmiller, and M. M. Botvinick. V-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 125,
+ 515,
+ 507,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 515,
+ 507,
+ 532
+ ],
+ "score": 1.0,
+ "content": "MPO: On-policy maximum a posteriori policy optimization for discrete and continuous control.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 527,
+ 418,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 527,
+ 418,
+ 542
+ ],
+ "score": 1.0,
+ "content": "In International Conference on Learning Representations (ICLR), 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 543,
+ 507,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 507,
+ 557
+ ],
+ "score": 1.0,
+ "content": "[66] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Collabo-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 554,
+ 470,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 554,
+ 470,
+ 568
+ ],
+ "score": 1.0,
+ "content": "ration without pre-coordination. In AAAI Conference on Artificial Intelligence, 2010.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 39,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "[67] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 40,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 582,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 582,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "transferring deep neural networks from simulation to the real world. In IEEE/RSJ International",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 128,
+ 592,
+ 370,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 592,
+ 370,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Conference on Intelligent Robots and Systems (IROS), 2017.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 621
+ ],
+ "score": 1.0,
+ "content": "[68] P. Tylkin, G. Radanovic, and D. C. Parkes. Learning robust helpful behaviors in two-player",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 618,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 618,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "cooperative Atari environments. In International Conference on Autonomous Agents and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 126,
+ 629,
+ 277,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 629,
+ 277,
+ 643
+ ],
+ "score": 1.0,
+ "content": "Multiagent Systems (AAMAS), 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 45,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 507,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 507,
+ 658
+ ],
+ "score": 1.0,
+ "content": "[69] O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 46,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 655,
+ 507,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 655,
+ 507,
+ 669
+ ],
+ "score": 1.0,
+ "content": "R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 127,
+ 667,
+ 507,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 667,
+ 507,
+ 681
+ ],
+ "score": 1.0,
+ "content": "T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V. Dalibard,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 125,
+ 676,
+ 507,
+ 692
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 676,
+ 507,
+ 692
+ ],
+ "score": 1.0,
+ "content": "D. Budden, Y. Sulsky, J. Molloy, T. L. Paine, C. Gulcehre, Z. Wang, T. Pfaff, Y. Wu, R. Ring,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 127,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "D. Yogatama, D. Wünsch, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, K. Kavukcuoglu,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "D. Hassabis, C. Apps, and D. Silver. Grandmaster level in StarCraft II using multi-agent",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 127,
+ 711,
+ 366,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 711,
+ 366,
+ 723
+ ],
+ "score": 1.0,
+ "content": "reinforcement learning. Nature, 575(7782):350–354, 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 52,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "[70] R. E. Wang, S. A. Wu, J. A. Evans, J. B. Tenenbaum, D. C. Parkes, and M. Kleiman-Weiner. Too",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 83,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 83,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "many cooks: Coordinating multi-agent collaboration through inverse planning. In International",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 93,
+ 437,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 93,
+ 437,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 109,
+ 505,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 109,
+ 505,
+ 122
+ ],
+ "score": 1.0,
+ "content": "[71] F. Wu, S. Zilberstein, and X. Chen. Online planning for ad hoc autonomous agent teams. In",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 120,
+ 414,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 120,
+ 414,
+ 132
+ ],
+ "score": 1.0,
+ "content": "International Joint Conference on Artificial Intelligence (IJCAI), 2011.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 135,
+ 505,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 135,
+ 505,
+ 148
+ ],
+ "score": 1.0,
+ "content": "[72] S. Zheng, A. Trott, S. Srinivasa, N. Naik, M. Gruesbeck, D. C. Parkes, and R. Socher. The AI",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 146,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 146,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "economist: Improving equality and productivity with AI-driven tax policies. arXiv preprint",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 158,
+ 231,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 158,
+ 231,
+ 169
+ ],
+ "score": 1.0,
+ "content": "arXiv:2004.13332, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 343,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 40,
+ 507,
+ 727
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "[14] R. Choudhury, G. Swamy, D. Hadfield-Menell, and A. D. Dragan. On the utility of model",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 98
+ ],
+ "score": 1.0,
+ "content": "learning in HRI. In ACM/IEEE International Conference on Human-Robot Interaction (HRI),",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 126,
+ 92,
+ 155,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 92,
+ 155,
+ 109
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "[15] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 126,
+ 120,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 120,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": "learning from human preferences. In Proceedings of the 31st International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 126,
+ 131,
+ 392,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 131,
+ 392,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Neural Information Processing Systems, pages 4302–4310, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 146,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 146,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "[16] A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 126,
+ 157,
+ 477,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 157,
+ 477,
+ 172
+ ],
+ "score": 1.0,
+ "content": "T. Graepel. Open problems in cooperative AI. arXiv preprint arXiv:2012.08630, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 507,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 507,
+ 187
+ ],
+ "score": 1.0,
+ "content": "[17] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 125,
+ 182,
+ 507,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 182,
+ 507,
+ 198
+ ],
+ "score": 1.0,
+ "content": "T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable distributed deep-RL",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 194,
+ 506,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 194,
+ 506,
+ 209
+ ],
+ "score": 1.0,
+ "content": "with importance weighted actor-learner architectures. In International Conference on Machine",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 127,
+ 206,
+ 226,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 206,
+ 226,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Learning (ICML), 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "[18] P. Eyal, R. David, G. Andrew, E. Zak, and D. Ekaterina. Data quality of platforms and panels",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 126,
+ 231,
+ 447,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 231,
+ 447,
+ 245
+ ],
+ "score": 1.0,
+ "content": "for online behavioral research. Behavior Research Methods, pages 1–20, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 260
+ ],
+ "score": 1.0,
+ "content": "[19] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 126,
+ 258,
+ 507,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 258,
+ 507,
+ 272
+ ],
+ "score": 1.0,
+ "content": "without a reward function. In International Conference on Learning Representations (ICLR),",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 127,
+ 267,
+ 155,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 267,
+ 155,
+ 283
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 298
+ ],
+ "score": 1.0,
+ "content": "[20] E. Fehr and K. M. Schmidt. A theory of fairness, competition, and cooperation. The Quarterly",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 126,
+ 294,
+ 314,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 294,
+ 314,
+ 309
+ ],
+ "score": 1.0,
+ "content": "Journal of Economics, 114(3):817–868, 1999.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 507,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 507,
+ 324
+ ],
+ "score": 1.0,
+ "content": "[21] J. Foerster, F. Song, E. Hughes, N. Burch, I. Dunning, S. Whiteson, M. Botvinick, and M. Bowl-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 126,
+ 321,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 321,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "ing. Bayesian action decoder for deep multi-agent reinforcement learning. In International",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 126,
+ 331,
+ 325,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 331,
+ 325,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Conference on Machine Learning (ICML), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 507,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 507,
+ 360
+ ],
+ "score": 1.0,
+ "content": "[22] M. Fontaine, Y.-C. Hsu, Y. Zhang, B. Tjanaka, and S. Nikolaidis. On the importance of",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 128,
+ 358,
+ 496,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 358,
+ 496,
+ 370
+ ],
+ "score": 1.0,
+ "content": "environments in human-robot coordination. In Robotics: Science and Systems (RSS), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "[23] I. Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30(3):411–437,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 127,
+ 384,
+ 154,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 384,
+ 154,
+ 397
+ ],
+ "score": 1.0,
+ "content": "2020.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "[24] M. Garnelo, W. M. Czarnecki, S. Liu, D. Tirumala, J. Oh, G. Gidel, H. van Hasselt, and",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 126,
+ 410,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 410,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "D. Balduzzi. Pick your battles: Interaction graphs as population-level objectives for strategic",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 126,
+ 419,
+ 507,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 419,
+ 507,
+ 435
+ ],
+ "score": 1.0,
+ "content": "diversity. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS),",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 126,
+ 431,
+ 155,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 431,
+ 155,
+ 446
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 291,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 291,
+ 460
+ ],
+ "score": 1.0,
+ "content": "[25] Ghost Town Games. Overcooked, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "[26] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 126,
+ 473,
+ 419,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 473,
+ 419,
+ 487
+ ],
+ "score": 1.0,
+ "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "[27] J. Heinrich and D. Silver. Deep reinforcement learning from self-play in imperfect-information",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 126,
+ 499,
+ 415,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 499,
+ 415,
+ 512
+ ],
+ "score": 1.0,
+ "content": "games. In NIPS Deep Reinforcement Learning (DRL) Workshop, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "[28] J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 126,
+ 525,
+ 380,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 525,
+ 380,
+ 538
+ ],
+ "score": 1.0,
+ "content": "International Conference on Machine Learning (ICML), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 541,
+ 506,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 506,
+ 554
+ ],
+ "score": 1.0,
+ "content": "[29] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 127,
+ 552,
+ 202,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 552,
+ 202,
+ 564
+ ],
+ "score": 1.0,
+ "content": "1735–1780, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 507,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 507,
+ 580
+ ],
+ "score": 1.0,
+ "content": "[30] M. Hoffman, B. Shahriari, J. Aslanides, G. Barth-Maron, F. Behbahani, T. Norman, A. Abdol-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 127,
+ 577,
+ 507,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 577,
+ 507,
+ 591
+ ],
+ "score": 1.0,
+ "content": "maleki, A. Cassirer, F. Yang, K. Baumli, S. Henderson, A. Novikov, S. G. Colmenarejo, S. Cabi,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 127,
+ 588,
+ 506,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 588,
+ 506,
+ 603
+ ],
+ "score": 1.0,
+ "content": "C. Gulcehre, T. L. Paine, A. Cowie, Z. Wang, B. Piot, and N. de Freitas. Acme: A research",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 126,
+ 599,
+ 492,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 599,
+ 492,
+ 613
+ ],
+ "score": 1.0,
+ "content": "framework for distributed reinforcement learning. arXiv preprint arXiv:2006.00979, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "[31] H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “Other-Play” for zero-shot coordination. In",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 126,
+ 625,
+ 381,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 625,
+ 381,
+ 639
+ ],
+ "score": 1.0,
+ "content": "International Conference on Machine Learning (ICML), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 507,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 507,
+ 654
+ ],
+ "score": 1.0,
+ "content": "[32] E. Hughes, J. Z. Leibo, M. G. Phillips, K. Tuyls, E. A. Duéñez-Guzmán, A. G. Castañeda,",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 126,
+ 651,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 651,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": "I. Dunning, T. Zhu, K. R. McKee, R. Koster, H. Roff, and T. Graepel. Inequity aversion",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 127,
+ 663,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 663,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "improves cooperation in intertemporal social dilemmas. In Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 127,
+ 674,
+ 232,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 674,
+ 232,
+ 686
+ ],
+ "score": 1.0,
+ "content": "Systems (NeurIPS), 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 703
+ ],
+ "score": 1.0,
+ "content": "[33] B. Ibarz, J. Leike, T. Pohlen, G. Irving, S. Legg, and D. Amodei. Reward learning from human",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 714
+ ],
+ "score": 1.0,
+ "content": "preferences and demonstrations in Atari. In Neural Information Processing Systems (NeurIPS),",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 126,
+ 709,
+ 155,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 709,
+ 155,
+ 725
+ ],
+ "score": 1.0,
+ "content": "2018.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 25.5
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 751
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "score": 1.0,
+ "content": "11",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 40,
+ 507,
+ 727
+ ],
+ "lines": [],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 73,
+ 507,
+ 725
+ ],
+ "lines_deleted": true
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 59,
+ 507,
+ 725
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "[34] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 82,
+ 507,
+ 97
+ ],
+ "score": 1.0,
+ "content": "N. C. Rabinowitz, A. S. Morcos, A. Ruderman, et al. Human-level performance in 3D mul-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 95,
+ 507,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 95,
+ 507,
+ 109
+ ],
+ "score": 1.0,
+ "content": "tiplayer games with population-based reinforcement learning. Science, 364(6443):859–865,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 104,
+ 154,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 104,
+ 154,
+ 119
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 120,
+ 507,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 120,
+ 507,
+ 134
+ ],
+ "score": 1.0,
+ "content": "[35] S. Javdani, S. S. Srinivasa, and J. A. Bagnell. Shared autonomy via hindsight optimization.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 127,
+ 131,
+ 277,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 131,
+ 277,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Robotics Science and Systems, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 147,
+ 507,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 507,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[36] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, and J. B. Tenenbaum. Co-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 157,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 157,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "ordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 127,
+ 168,
+ 371,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 168,
+ 371,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Conference of the Cognitive Science Society (CogSci), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 183,
+ 506,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 183,
+ 506,
+ 197
+ ],
+ "score": 1.0,
+ "content": "[37] G. Klien, D. D. Woods, J. M. Bradshaw, R. R. Hoffman, and P. J. Feltovich. Ten challenges for",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 127,
+ 195,
+ 505,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 195,
+ 505,
+ 209
+ ],
+ "score": 1.0,
+ "content": "making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 127,
+ 207,
+ 197,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 207,
+ 197,
+ 218
+ ],
+ "score": 1.0,
+ "content": "(6):91–95, 2004.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 236
+ ],
+ "score": 1.0,
+ "content": "[38] P. Knott, M. Carroll, S. Devlin, K. Ciosek, K. Hofmann, A. Dragan, and R. Shah. Evaluating",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 126,
+ 231,
+ 506,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 231,
+ 506,
+ 245
+ ],
+ "score": 1.0,
+ "content": "the robustness of collaborative agents. In International Conference on Autonomous Agents and",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 127,
+ 244,
+ 276,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 244,
+ 276,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Multiagent Systems (AAMAS), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "[39] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 127,
+ 269,
+ 506,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 269,
+ 506,
+ 282
+ ],
+ "score": 1.0,
+ "content": "T. Graepel. A unified game-theoretic approach to multiagent reinforcement learning. In Neural",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 127,
+ 280,
+ 317,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 280,
+ 317,
+ 294
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems (NIPS), 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 107,
+ 296,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 296,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "[40] A. Lerer and A. Peysakhovich. Learning existing social conventions via observationally",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 126,
+ 306,
+ 505,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 306,
+ 505,
+ 320
+ ],
+ "score": 1.0,
+ "content": "augmented self-play. In AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 126,
+ 316,
+ 185,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 316,
+ 185,
+ 330
+ ],
+ "score": 1.0,
+ "content": "(AIES), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 330,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 330,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "[41] E. Lockhart, N. Burch, N. Bard, S. Borgeaud, T. Eccles, L. Smaira, and R. Smith. Human-agent",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 125,
+ 342,
+ 416,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 342,
+ 416,
+ 357
+ ],
+ "score": 1.0,
+ "content": "cooperation in bridge bidding. arXiv preprint arXiv:2011.14124, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 357,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 357,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "[42] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch. Multi-agent actor-critic",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 126,
+ 368,
+ 505,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 368,
+ 505,
+ 383
+ ],
+ "score": 1.0,
+ "content": "for mixed cooperative-competitive environments. In Neural Information Processing Systems",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 126,
+ 380,
+ 186,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 380,
+ 186,
+ 394
+ ],
+ "score": 1.0,
+ "content": "(NIPS), 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 408
+ ],
+ "score": 1.0,
+ "content": "[43] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. Learning to learn to communicate. In",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 126,
+ 406,
+ 352,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 406,
+ 352,
+ 419
+ ],
+ "score": 1.0,
+ "content": "ICML Adaptive & Multitask Learning Workshop, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 507,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 507,
+ 435
+ ],
+ "score": 1.0,
+ "content": "[44] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. On the interaction between super-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 127,
+ 431,
+ 505,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 431,
+ 505,
+ 446
+ ],
+ "score": 1.0,
+ "content": "vision and self-play in emergent communication. In International Conference on Learning",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 443,
+ 254,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 443,
+ 254,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Representations (ICLR), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "[45] A. Lupu, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 126,
+ 468,
+ 434,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 468,
+ 434,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "[46] A. Mahajan, T. Rashid, M. Samvelyan, and S. Whiteson. Maven: Multi-agent variational",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 126,
+ 495,
+ 421,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 495,
+ 421,
+ 508
+ ],
+ "score": 1.0,
+ "content": "exploration. In Neural Information Processing Systems (NeurIPS), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 509,
+ 507,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 509,
+ 507,
+ 525
+ ],
+ "score": 1.0,
+ "content": "[47] K. R. McKee, I. Gemp, B. McWilliams, E. A. Duèñez-Guzmán, E. Hughes, and J. Z. Leibo.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 127,
+ 522,
+ 507,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 522,
+ 507,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Social diversity and social preferences in mixed-motive reinforcement learning. In International",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 532,
+ 435,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 532,
+ 435,
+ 545
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 507,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 507,
+ 562
+ ],
+ "score": 1.0,
+ "content": "[48] K. R. McKee, X. Bai, and S. Fiske. Understanding human impressions of artificial intelligence.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 126,
+ 557,
+ 196,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 557,
+ 196,
+ 572
+ ],
+ "score": 1.0,
+ "content": "PsyArXiv, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "[49] K. R. McKee, E. Hughes, T. O. Zhu, M. J. Chadwick, R. Koster, A. G. Castaneda, C. Beattie,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 127,
+ 584,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 584,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "T. Graepel, M. Botvinick, and J. Z. Leibo. Deep reinforcement learning models the emergent",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 128,
+ 596,
+ 425,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 596,
+ 425,
+ 609
+ ],
+ "score": 1.0,
+ "content": "dynamics of human cooperation. arXiv preprint arXiv:2103.04982, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "[50] K. R. McKee, J. Z. Leibo, C. Beattie, and R. Everett. Quantifying environment and population",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 127,
+ 621,
+ 482,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 621,
+ 482,
+ 635
+ ],
+ "score": 1.0,
+ "content": "diversity in multi-agent reinforcement learning. arXiv preprint arXiv:2102.08370, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "[51] F. Melo and A. Sardinha. Ad hoc teamwork by learning teammates’ task. International",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 127,
+ 648,
+ 441,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 648,
+ 441,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 507,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 507,
+ 677
+ ],
+ "score": 1.0,
+ "content": "[52] B. Mutlu, A. Terrell, and C.-M. Huang. Coordination mechanisms in human-robot collaboration.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 126,
+ 673,
+ 343,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 673,
+ 343,
+ 687
+ ],
+ "score": 1.0,
+ "content": "In HRI Collaborative Manipulation Workshop, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "[53] S. Nikolaidis and J. Shah. Human-robot cross-training: Computational formulation, modeling",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 699,
+ 507,
+ 715
+ ],
+ "score": 1.0,
+ "content": "and evaluation of a human team training strategy. In ACM/IEEE International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 127,
+ 711,
+ 288,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 711,
+ 288,
+ 724
+ ],
+ "score": 1.0,
+ "content": "Human-Robot Interaction (HRI), 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 25.5
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 751
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 59,
+ 507,
+ 725
+ ],
+ "lines": [],
+ "index": 25.5,
+ "bbox_fs": [
+ 104,
+ 72,
+ 507,
+ 724
+ ],
+ "lines_deleted": true
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 45,
+ 507,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "[54] OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 125,
+ 82,
+ 507,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 82,
+ 507,
+ 99
+ ],
+ "score": 1.0,
+ "content": "M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 94,
+ 506,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 94,
+ 506,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Q. Yuan, W. Zaremba, and L. Zhang. Solving Rubik’s Cube with a robot hand. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 105,
+ 232,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 105,
+ 232,
+ 118
+ ],
+ "score": 1.0,
+ "content": "arXiv:1910.07113, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 119,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 119,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "[55] E. Peer, L. Brandimarte, S. Samat, and A. Acquisti. Beyond the Turk: Alternative platforms for",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 127,
+ 132,
+ 507,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 132,
+ 507,
+ 146
+ ],
+ "score": 1.0,
+ "content": "crowdsourcing behavioral research. Journal of Experimental Social Psychology, 70:153–163,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 127,
+ 142,
+ 155,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 142,
+ 155,
+ 157
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 158,
+ 505,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 158,
+ 505,
+ 170
+ ],
+ "score": 1.0,
+ "content": "[56] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 126,
+ 168,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 168,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "control with dynamics randomization. In IEEE International Conference on Robotics and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 125,
+ 180,
+ 236,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 180,
+ 236,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Automation (ICRA), 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 195,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 195,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "[57] P. M. Pilarski, A. Butcher, M. Johanson, M. M. Botvinick, A. Bolt, and A. S. Parker. Learned",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 126,
+ 207,
+ 507,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 207,
+ 507,
+ 221
+ ],
+ "score": 1.0,
+ "content": "human-agent decision-making, communication and joint action in a virtual reality environment.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 126,
+ 216,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 216,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "In Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM),",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 126,
+ 227,
+ 155,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 227,
+ 155,
+ 243
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": "[58] D. A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 127,
+ 256,
+ 289,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 256,
+ 289,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Neural Computation, 3(1):88–97, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 285
+ ],
+ "score": 1.0,
+ "content": "[59] S. Russell. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 127,
+ 279,
+ 155,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 279,
+ 155,
+ 295
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": "[60] D. Sadigh, N. Landolfi, S. S. Sastry, S. A. Seshia, and A. D. Dragan. Planning for cars that",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 127,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "coordinate with people: leveraging effects on human actions for planning and active information",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 126,
+ 319,
+ 459,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 319,
+ 459,
+ 332
+ ],
+ "score": 1.0,
+ "content": "gathering over human internal state. Autonomous Robots, 42(7):1405–1426, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "[61] N. Schurr, J. Marecki, M. Tambe, and P. Scerri. Towards flexible coordination of human-agent",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 126,
+ 345,
+ 348,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 345,
+ 348,
+ 358
+ ],
+ "score": 1.0,
+ "content": "teams. Multiagent and Grid Systems, 1(1):3–16, 2005.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "[62] A. Shih, A. Sawhney, J. Kondic, S. Ermon, and D. Sadigh. On the critical role of conventions in",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 127,
+ 372,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 372,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "adaptive human-AI collaboration. In International Conference on Learning Representations",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 126,
+ 382,
+ 187,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 382,
+ 187,
+ 397
+ ],
+ "score": 1.0,
+ "content": "(ICLR), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": "[63] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 126,
+ 409,
+ 505,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 409,
+ 505,
+ 423
+ ],
+ "score": 1.0,
+ "content": "M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 125,
+ 418,
+ 507,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 418,
+ 507,
+ 435
+ ],
+ "score": 1.0,
+ "content": "D. Hassabis. Mastering the game of go without human knowledge. Nature, 550(7676):354–359,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 126,
+ 428,
+ 156,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 428,
+ 156,
+ 446
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "[64] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 127,
+ 457,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 457,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 125,
+ 466,
+ 507,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 466,
+ 507,
+ 484
+ ],
+ "score": 1.0,
+ "content": "learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 127,
+ 480,
+ 202,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 480,
+ 202,
+ 492
+ ],
+ "score": 1.0,
+ "content": "1140–1144, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 507,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 507,
+ 509
+ ],
+ "score": 1.0,
+ "content": "[65] H. F. Song, A. Abdolmaleki, J. T. Springenberg, A. Clark, H. Soyer, J. W. Rae, S. Noury,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 126,
+ 505,
+ 507,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 505,
+ 507,
+ 519
+ ],
+ "score": 1.0,
+ "content": "A. Ahuja, S. Liu, D. Tirumala, N. Heess, D. Belov, M. Riedmiller, and M. M. Botvinick. V-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 125,
+ 515,
+ 507,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 515,
+ 507,
+ 532
+ ],
+ "score": 1.0,
+ "content": "MPO: On-policy maximum a posteriori policy optimization for discrete and continuous control.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 527,
+ 418,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 527,
+ 418,
+ 542
+ ],
+ "score": 1.0,
+ "content": "In International Conference on Learning Representations (ICLR), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 543,
+ 507,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 507,
+ 557
+ ],
+ "score": 1.0,
+ "content": "[66] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Collabo-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 125,
+ 554,
+ 470,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 554,
+ 470,
+ 568
+ ],
+ "score": 1.0,
+ "content": "ration without pre-coordination. In AAAI Conference on Artificial Intelligence, 2010.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "[67] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 128,
+ 582,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 582,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "transferring deep neural networks from simulation to the real world. In IEEE/RSJ International",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 128,
+ 592,
+ 370,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 592,
+ 370,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Conference on Intelligent Robots and Systems (IROS), 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 621
+ ],
+ "score": 1.0,
+ "content": "[68] P. Tylkin, G. Radanovic, and D. C. Parkes. Learning robust helpful behaviors in two-player",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 127,
+ 618,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 618,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "cooperative Atari environments. In International Conference on Autonomous Agents and",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 126,
+ 629,
+ 277,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 629,
+ 277,
+ 643
+ ],
+ "score": 1.0,
+ "content": "Multiagent Systems (AAMAS), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 507,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 507,
+ 658
+ ],
+ "score": 1.0,
+ "content": "[69] O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 127,
+ 655,
+ 507,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 655,
+ 507,
+ 669
+ ],
+ "score": 1.0,
+ "content": "R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre,",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 127,
+ 667,
+ 507,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 667,
+ 507,
+ 681
+ ],
+ "score": 1.0,
+ "content": "T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V. Dalibard,",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 125,
+ 676,
+ 507,
+ 692
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 676,
+ 507,
+ 692
+ ],
+ "score": 1.0,
+ "content": "D. Budden, Y. Sulsky, J. Molloy, T. L. Paine, C. Gulcehre, Z. Wang, T. Pfaff, Y. Wu, R. Ring,",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 127,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "D. Yogatama, D. Wünsch, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, K. Kavukcuoglu,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "D. Hassabis, C. Apps, and D. Silver. Grandmaster level in StarCraft II using multi-agent",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 127,
+ 711,
+ 366,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 711,
+ 366,
+ 723
+ ],
+ "score": 1.0,
+ "content": "reinforcement learning. Nature, 575(7782):350–354, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "13",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 45,
+ 507,
+ 723
+ ],
+ "lines": [],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 72,
+ 507,
+ 723
+ ],
+ "lines_deleted": true
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 169
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "[70] R. E. Wang, S. A. Wu, J. A. Evans, J. B. Tenenbaum, D. C. Parkes, and M. Kleiman-Weiner. Too",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 126,
+ 83,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 83,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "many cooks: Coordinating multi-agent collaboration through inverse planning. In International",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 93,
+ 437,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 93,
+ 437,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 109,
+ 505,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 109,
+ 505,
+ 122
+ ],
+ "score": 1.0,
+ "content": "[71] F. Wu, S. Zilberstein, and X. Chen. Online planning for ad hoc autonomous agent teams. In",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 126,
+ 120,
+ 414,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 120,
+ 414,
+ 132
+ ],
+ "score": 1.0,
+ "content": "International Joint Conference on Artificial Intelligence (IJCAI), 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 135,
+ 505,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 135,
+ 505,
+ 148
+ ],
+ "score": 1.0,
+ "content": "[72] S. Zheng, A. Trott, S. Srinivasa, N. Naik, M. Gruesbeck, D. C. Parkes, and R. Socher. The AI",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 126,
+ 146,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 146,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "economist: Improving equality and productivity with AI-driven tax policies. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 158,
+ 231,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 158,
+ 231,
+ 169
+ ],
+ "score": 1.0,
+ "content": "arXiv:2004.13332, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "14",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 169
+ ],
+ "lines": [],
+ "index": 3.5,
+ "bbox_fs": [
+ 106,
+ 72,
+ 506,
+ 169
+ ],
+ "lines_deleted": true
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file