diff --git a/src_code_for_reproducibility/chat_utils/__pycache__/apply_template.cpython-312.pyc b/src_code_for_reproducibility/chat_utils/__pycache__/apply_template.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dbc484a917abce606fd373a1d1333cef4561a4ae Binary files /dev/null and b/src_code_for_reproducibility/chat_utils/__pycache__/apply_template.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/chat_utils/__pycache__/chat_turn.cpython-312.pyc b/src_code_for_reproducibility/chat_utils/__pycache__/chat_turn.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c06bb7fc79560eca9567828218cde894ac2ed1d7 Binary files /dev/null and b/src_code_for_reproducibility/chat_utils/__pycache__/chat_turn.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/docs/source/conf.py b/src_code_for_reproducibility/docs/source/conf.py new file mode 100644 index 0000000000000000000000000000000000000000..5c7512678928b6b7580c812cd62d1c22df9945ba --- /dev/null +++ b/src_code_for_reproducibility/docs/source/conf.py @@ -0,0 +1,48 @@ +# Configuration file for the Sphinx documentation builder. +import os +import sys +sys.path.insert(0, os.path.abspath('../..')) + +# -- Project information ----------------------------------------------------- +project = 'llm_negotiation' +copyright = '2023, Your Name' +author = 'Your Name' + +# -- General configuration --------------------------------------------------- +extensions = [ + 'sphinx.ext.autodoc', + 'sphinx.ext.viewcode', + 'sphinx.ext.napoleon', + 'sphinx.ext.autosummary', + 'sphinx.ext.intersphinx', + 'sphinx.ext.mathjax', + 'sphinxcontrib.mermaid', + 'sphinx_rtd_theme', +] + +templates_path = ['_templates'] +exclude_patterns = [] + +# -- Options for HTML output ------------------------------------------------- +html_theme = 'sphinx_rtd_theme' +html_static_path = ['_static'] + +# -- Napoleon settings ------------------------------------------------------- +napoleon_google_docstring = True +napoleon_numpy_docstring = False +napoleon_include_init_with_doc = True +napoleon_include_private_with_doc = False +napoleon_include_special_with_doc = True +napoleon_use_admonition_for_examples = False +napoleon_use_admonition_for_notes = False +napoleon_use_admonition_for_references = False +napoleon_use_ivar = False +napoleon_use_param = True +napoleon_use_rtype = True +napoleon_preprocess_types = False +napoleon_type_aliases = None +napoleon_attr_annotations = True + +# -- Path setup -------------------------------------------------------------- +# Make sure the project's modules can be found by Sphinx +sys.path.insert(0, os.path.abspath('../../src')) \ No newline at end of file diff --git a/src_code_for_reproducibility/docs/source/contributing.rst b/src_code_for_reproducibility/docs/source/contributing.rst new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/src_code_for_reproducibility/docs/source/environments.rst b/src_code_for_reproducibility/docs/source/environments.rst new file mode 100644 index 0000000000000000000000000000000000000000..fa2fc4fbe9c68edfb5b1726a19a0bdf133c3f879 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/environments.rst @@ -0,0 +1,35 @@ +================= +MARL Environments +================= + +This section provides detailed documentation for the multi-agent negotiation environments included in the library. + +Each environment follows the standard interface described in :doc:`../environments` but has its own unique game rules, +dynamics, and implementation details. + +.. toctree:: + :maxdepth: 2 + :caption: Available Environments: + + environments/ipd + environments/diplomacy + environments/dond + +Overview +-------- + +The library currently includes the following environments: + +1. **Iterated Prisoner's Dilemma (IPD)**: A classic game theory problem where two agents repeatedly decide whether to cooperate or defect, with different payoffs based on their joint actions. + +2. **Diplomacy**: An adaptation of the board game Diplomacy, where seven European powers compete for control of supply centers through strategic moves and alliances. + +3. **Deal or No Deal (DOND)**: A negotiation environment based on `the paper Deal or No Deal? End-to-End Learning for Negotiation Dialogues `_ in which agents negotiate over the distribution of a set of prizes. + +Each environment documentation includes: + +- Game rules and background +- Implementation details +- API reference +- Example usage +- Advanced features and customization options \ No newline at end of file diff --git a/src_code_for_reproducibility/docs/source/index.rst b/src_code_for_reproducibility/docs/source/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..cdc1b79539342a9c95ca0cdd9219bce74a7b2c8a --- /dev/null +++ b/src_code_for_reproducibility/docs/source/index.rst @@ -0,0 +1,22 @@ +Welcome to LLM Negotiation's documentation! +=========================================== +This library is a collection of tools for training and evaluating LLM-based agents in multi-agent environments. It is designed to be easy to use and extend. + +.. toctree:: + :maxdepth: 3 + :caption: Contents: + + installation + marl_standard + environments + launch + usage + modules + contributing + +Indices and tables +================== + +* :ref:`genindex` +* :ref:`modindex` +* :ref:`search` \ No newline at end of file diff --git a/src_code_for_reproducibility/docs/source/installation.rst b/src_code_for_reproducibility/docs/source/installation.rst new file mode 100644 index 0000000000000000000000000000000000000000..b148f25d92fd8308e9695f7c17c2b91fb0c9a2c6 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/installation.rst @@ -0,0 +1,10 @@ +Installation +=========== + +To install the package, run: + +.. code-block:: bash + + git clone https://github.com/yourusername/llm_negotiation.git + cd llm_negotiation + pip install -e . \ No newline at end of file diff --git a/src_code_for_reproducibility/docs/source/launch.rst b/src_code_for_reproducibility/docs/source/launch.rst new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/src_code_for_reproducibility/docs/source/marl_standard.rst b/src_code_for_reproducibility/docs/source/marl_standard.rst new file mode 100644 index 0000000000000000000000000000000000000000..b5ea5529892c611b34255645ec68537a236754cf --- /dev/null +++ b/src_code_for_reproducibility/docs/source/marl_standard.rst @@ -0,0 +1,141 @@ +================= +Abstract Standard for Multi-Agent Negotiation Environments +================= + +Multi-Agent Negotiation Environments require more features than gymnasium environments in order to be used as interfaces in general game running code. +The two fundamental differences between gymnasium environments and Multi-Agent Negotiation Environments are: + +1. Response from the LLM is a text action, not a discrete action. Therefore, appropriate parsing of the text is required. The model may need to be run multiple times to get the full action. + This is why we introduce the `AgentHandler` class, which is responsible for parsing the LLM's response. +2. The environment needs to be able to handle multi-agent interactions. + This is why we introduce the `NegotiationEnvironment` class, which is responsible for handling the multi-agent interactions. +3. MARL environments are complex to describe. In different contexts, the same environment may be described differently. Therefore, both the environement and the agent handlers are + responsible for describing a particular trajectory. This information is given by the `get_log_info` method. +4. There might be a lot of overlap between the neural networks used by each agent. For instance, the same model may be used for all agents. This motivates a requirement for a + policy identifier for each agent. + +Taking inspiration from the `gymnasium `_ library, we introduce a new standard for Multi-Agent Negotiation Environments. + +Our standard is based on the following features: + +Environments are of the form: + +.. code-block:: python + + class MarlEnvironment(): + + def __init__(self): + """Initialize the environment.""" + pass + + def reset(self): + """Reset the environment to an initial state and return the initial observation. + Returns: + observation (dict): A dictionary where keys are agent identifiers and values are observations. + """ + # (...) + return observation + + def step(self, actions): + """Take a step in the environment using the provided actions. + + Args: + actions (dict): A dictionary where keys are agent identifiers and values are actions. + + Returns: + observations (dict): A dictionary where keys are agent identifiers and values are observations. + reward (dict): A dictionary where keys are agent identifiers and values are rewards. + done (bool): Whether the episode has ended. + info (dict): Additional information about the environment. + """ + # (...) + return observations, done, info + + def get_log_info(self): + """Get additional information about the environment. This information is used to log the game. + Returns: + log_info (dict): Information about the environment required to log the game. + """ + # (...) + return log_info + + def render(self): + """Render the current state of the environment.""" + pass + + def close(self): + """Perform any necessary cleanup.""" + pass + + + class AgentState(): + + def __init__(self): + """Initialize the agent state.""" + pass + + def step(self, observation_from_env, policy_output=None): + """Update the agent state based on the observation and action. + The action is the output of the LLM. + """ + + Args: + observation_from_env (dict): The observation of the environment. + policy_output : The output of the policy. + + Returns: + policy_id (str): The policy identifier. + policy_input (dict): The input to the policy. + action : The official action to be sent to the environment. + done (bool): Whether the LLM action is ready to be sent to the environment. + info (dict): Additional information about the agent. + """ + # (...) + return policy_id, policy_input, action, done, info + + def get_log_info(self): + """Get information about the agent required to log a trajectory. + Returns: + log_info (dict): Information about the agent required to log a trajectory. + """ + # (...) + return log_info + + def render(self): + """Render the current state of the environment.""" + pass + + def close(self): + """Perform any necessary cleanup.""" + pass + + +Implicitely, the keys of the `observations` in the `step` method of the `MarlEnvironment` interface represent the set of agents from which an action is expected at the current step. The next step should only expect actions from the agents in the `observations` dictionary. + +As you can see, both classes have a `get_log_info` method. This method is used to log the game. It returns a dictionary with keys being the agent identifiers and values being the information to log. The reason we need this is because the environment and the agent handler may need to log different information. It makes it easier to log from the perspective of each agent. The core environment class should not need to know about the details of the agent handler. + + + +Running Environments in Parallel +-------------------------------- +This standard allows the use of the `run_batched_matches` function (TODO: link) to run environments in an efficient way. The core idea is to batch the policy calls for all agents in the environment. + +.. note:: + The ``run_batched_matches`` function allows you to run multiple negotiation games, or "matches," in parallel. + After each environment is initialized, the function continuously loops over all active matches and checks which agents + are still pending actions. Each agent's logic can require multiple calls to the policy (e.g., an LLM) before an action + becomes "ready" to be sent to the environment. (For instance, an agent might need multiple policy calls before having a string which can be parsed into a valid action.) While an agent is waiting for a policy output, these calls for all agents across all matches are grouped together by unique policy identifier and processed in batch for efficiency. This is the core functionality of the ``run_batched_matches`` function. + + Only once all actions from the required agents at a given step for an environment are ready does the function make a single ``env.step(...)`` call; this ensures + every match moves forward in lockstep for all its active agents. As soon as an environment signals it is done, the function + retrieves logged information from both the environment and the agent states before removing this match from the active set. + + If there are more matches waiting to be processed, they are then started one by one to maintain the specified degree of parallelism. + This batching approach provides an efficient mechanism to handle multi-agent or multi-policy environments, ensuring minimal + overhead and a clear, unified flow for stepping through matches. + +Here is a diagram that shows how the `run_batched_matches` function works at a high level: + +.. image:: media/runbatch.png + :alt: Alternate text for the image + :width: 1000px diff --git a/src_code_for_reproducibility/docs/source/modules.rst b/src_code_for_reproducibility/docs/source/modules.rst new file mode 100644 index 0000000000000000000000000000000000000000..e9ff8ac1a89c7bd18e69d633121f5a4022ac6fdf --- /dev/null +++ b/src_code_for_reproducibility/docs/source/modules.rst @@ -0,0 +1,7 @@ +src +=== + +.. toctree:: + :maxdepth: 4 + + src diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_agent.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_agent.rst new file mode 100644 index 0000000000000000000000000000000000000000..8fab765a9c7e749bd446533fdddb5fa5b55e6635 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_agent.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_agent module +======================================== + +.. automodule:: src.environments.dond.dond_agent + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_game.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_game.rst new file mode 100644 index 0000000000000000000000000000000000000000..d0e595aad169a5a8456f83afe5029e7475d7c9e7 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_game.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_game module +======================================= + +.. automodule:: src.environments.dond.dond_game + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_log_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_log_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..cf96327d1bcbc7f0f8785804a49a6975eef889c2 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_log_funcs.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_log\_funcs module +============================================= + +.. automodule:: src.environments.dond.dond_log_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_player.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_player.rst new file mode 100644 index 0000000000000000000000000000000000000000..bab97f1009eb2d5c4e387ac6a83982a51e33c9e3 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_player.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_agent module +========================================= + +.. automodule:: src.environments.dond.dond_agent + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_return_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_return_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..e8084f8f9a291efe75e032183ebd5eeab58f5e41 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_return_funcs.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_return\_funcs module +================================================ + +.. automodule:: src.environments.dond.dond_return_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_statistics_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_statistics_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..4c4d3764f9333b8a2975069160f40952f324a1a8 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_statistics_funcs.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_statistics\_funcs module +==================================================== + +.. automodule:: src.environments.dond.dond_statistics_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.dond_training_data_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_training_data_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..cf31d696a3ed580e24f3c5dffd6f7a2851d16320 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.dond_training_data_funcs.rst @@ -0,0 +1,7 @@ +src.environments.dond.dond\_training\_data\_funcs module +======================================================== + +.. automodule:: src.environments.dond.dond_training_data_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.dond.rst b/src_code_for_reproducibility/docs/source/src.environments.dond.rst new file mode 100644 index 0000000000000000000000000000000000000000..8462de2bdb96d31b5628cfd2942131e13e4e9dc3 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.dond.rst @@ -0,0 +1,19 @@ +src.environments.dond package +============================= + +.. automodule:: src.environments.dond + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.environments.dond.dond_agent + src.environments.dond.dond_game + src.environments.dond.dond_log_funcs + src.environments.dond.dond_statistics_funcs + src.environments.dond.dond_training_data_funcs diff --git a/src_code_for_reproducibility/docs/source/src.environments.env_imports.rst b/src_code_for_reproducibility/docs/source/src.environments.env_imports.rst new file mode 100644 index 0000000000000000000000000000000000000000..4354ba27eee9f0e0fa3f4f0e5d9131c256a4be57 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.env_imports.rst @@ -0,0 +1,7 @@ +src.environments.env\_imports module +==================================== + +.. automodule:: src.environments.env_imports + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.environment_imports.rst b/src_code_for_reproducibility/docs/source/src.environments.environment_imports.rst new file mode 100644 index 0000000000000000000000000000000000000000..d22c53e31cd1c7c064955900c19f34ac51c7006f --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.environment_imports.rst @@ -0,0 +1,7 @@ +src.environments.environment\_imports module +============================================ + +.. automodule:: src.environments.environment_imports + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_agent.rst b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_agent.rst new file mode 100644 index 0000000000000000000000000000000000000000..4845b371089c529493f70de77ceaee0b7500571b --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_agent.rst @@ -0,0 +1,7 @@ +src.environments.ipd.ipd\_agent module +====================================== + +.. automodule:: src.environments.ipd.ipd_agent + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_game.rst b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_game.rst new file mode 100644 index 0000000000000000000000000000000000000000..ede471ef9675c780410189fcf63df0c1a05496d0 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_game.rst @@ -0,0 +1,7 @@ +src.environments.ipd.ipd\_game module +===================================== + +.. automodule:: src.environments.ipd.ipd_game + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_log_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_log_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..edec187f4876cdf653ae4f91035f43bc877a7d40 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_log_funcs.rst @@ -0,0 +1,7 @@ +src.environments.ipd.ipd\_log\_funcs module +=========================================== + +.. automodule:: src.environments.ipd.ipd_log_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_statistics_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_statistics_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..5f54afac07c4d477067ef4c2bf5d883b236cf5fc --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_statistics_funcs.rst @@ -0,0 +1,7 @@ +src.environments.ipd.ipd\_statistics\_funcs module +================================================== + +.. automodule:: src.environments.ipd.ipd_statistics_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_training_data_funcs.rst b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_training_data_funcs.rst new file mode 100644 index 0000000000000000000000000000000000000000..8e4cecea10e25644ff677416823069cd65b500c5 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.ipd.ipd_training_data_funcs.rst @@ -0,0 +1,7 @@ +src.environments.ipd.ipd\_training\_data\_funcs module +====================================================== + +.. automodule:: src.environments.ipd.ipd_training_data_funcs + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.environments.ipd.rst b/src_code_for_reproducibility/docs/source/src.environments.ipd.rst new file mode 100644 index 0000000000000000000000000000000000000000..af26091b3a87dee4d6993f0ae09bdb1c380a130e --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.ipd.rst @@ -0,0 +1,19 @@ +src.environments.ipd package +============================ + +.. automodule:: src.environments.ipd + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.environments.ipd.ipd_agent + src.environments.ipd.ipd_game + src.environments.ipd.ipd_log_funcs + src.environments.ipd.ipd_statistics_funcs + src.environments.ipd.ipd_training_data_funcs diff --git a/src_code_for_reproducibility/docs/source/src.environments.rst b/src_code_for_reproducibility/docs/source/src.environments.rst new file mode 100644 index 0000000000000000000000000000000000000000..221ed1c07ebea145cd23bc06c6474d34b1d8a33e --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.environments.rst @@ -0,0 +1,25 @@ +src.environments package +======================== + +.. automodule:: src.environments + :members: + :undoc-members: + :show-inheritance: + +Subpackages +----------- + +.. toctree:: + :maxdepth: 4 + + src.environments.dond + src.environments.ipd + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.environments.env_imports + src.environments.environment_imports diff --git a/src_code_for_reproducibility/docs/source/src.experiments.arithmetic_test.rst b/src_code_for_reproducibility/docs/source/src.experiments.arithmetic_test.rst new file mode 100644 index 0000000000000000000000000000000000000000..68e0f5da020aee80cc8895ca650a6067317f4bcd --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.experiments.arithmetic_test.rst @@ -0,0 +1,7 @@ +src.experiments.arithmetic\_test module +======================================= + +.. automodule:: src.experiments.arithmetic_test + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.experiments.dond_run_train.rst b/src_code_for_reproducibility/docs/source/src.experiments.dond_run_train.rst new file mode 100644 index 0000000000000000000000000000000000000000..6c94e4bc508836338d5d6393858d403e746b5d2d --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.experiments.dond_run_train.rst @@ -0,0 +1,7 @@ +src.experiments.dond\_run\_train module +======================================= + +.. automodule:: src.experiments.dond_run_train + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.experiments.generate_and_train.rst b/src_code_for_reproducibility/docs/source/src.experiments.generate_and_train.rst new file mode 100644 index 0000000000000000000000000000000000000000..d0d0a0ccdf7839f6c192107d1ad91af4aaadda7b --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.experiments.generate_and_train.rst @@ -0,0 +1,7 @@ +src.experiments.generate\_and\_train module +=========================================== + +.. automodule:: src.experiments.generate_and_train + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.experiments.last_completion.rst b/src_code_for_reproducibility/docs/source/src.experiments.last_completion.rst new file mode 100644 index 0000000000000000000000000000000000000000..1b868ee566283d662a51387046bc070a131f5222 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.experiments.last_completion.rst @@ -0,0 +1,7 @@ +src.experiments.last\_completion module +======================================= + +.. automodule:: src.experiments.last_completion + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.experiments.rst b/src_code_for_reproducibility/docs/source/src.experiments.rst new file mode 100644 index 0000000000000000000000000000000000000000..90f61ff53afa0281b9e6b82188eb2df30b81eb07 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.experiments.rst @@ -0,0 +1,17 @@ +src.experiments package +======================= + +.. automodule:: src.experiments + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.experiments.arithmetic_test + src.experiments.generate_and_train + src.experiments.last_completion diff --git a/src_code_for_reproducibility/docs/source/src.generation.rst b/src_code_for_reproducibility/docs/source/src.generation.rst new file mode 100644 index 0000000000000000000000000000000000000000..14bb2b1364da7067aed5c37e3c77d091d20f011b --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.generation.rst @@ -0,0 +1,15 @@ +src.generation package +====================== + +.. automodule:: src.generation + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.generation.run_games diff --git a/src_code_for_reproducibility/docs/source/src.generation.run_games.rst b/src_code_for_reproducibility/docs/source/src.generation.run_games.rst new file mode 100644 index 0000000000000000000000000000000000000000..dbf42d3f821df187cdd4a8bb9d093839ce6b608a --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.generation.run_games.rst @@ -0,0 +1,7 @@ +src.generation.run\_games module +================================ + +.. automodule:: src.generation.run_games + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.dummy_hf_agent.rst b/src_code_for_reproducibility/docs/source/src.models.dummy_hf_agent.rst new file mode 100644 index 0000000000000000000000000000000000000000..937900b392b98e7f01968d496ae0c350a836d632 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.dummy_hf_agent.rst @@ -0,0 +1,7 @@ +src.models.dummy\_hf\_agent module +================================== + +.. automodule:: src.models.dummy_llm_agent + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.dummy_local_llm.rst b/src_code_for_reproducibility/docs/source/src.models.dummy_local_llm.rst new file mode 100644 index 0000000000000000000000000000000000000000..13b40bd388e445fa60a3c3fc2e089ad89c452dbd --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.dummy_local_llm.rst @@ -0,0 +1,7 @@ +src.models.dummy\_local\_llm module +=================================== + +.. automodule:: src.models.dummy_local_llm + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.hf_agent.rst b/src_code_for_reproducibility/docs/source/src.models.hf_agent.rst new file mode 100644 index 0000000000000000000000000000000000000000..e96abf4e21b1523cd016c6172eed0ddf51b4954a --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.hf_agent.rst @@ -0,0 +1,7 @@ +src.models.hf\_agent module +=========================== + +.. automodule:: src.models.hf_agent + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.local_llm.rst b/src_code_for_reproducibility/docs/source/src.models.local_llm.rst new file mode 100644 index 0000000000000000000000000000000000000000..5c2eebb05e64919d1915eeb63dc18f5e9a36eb2c --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.local_llm.rst @@ -0,0 +1,7 @@ +src.models.local\_llm module +============================ + +.. automodule:: src.models.local_llm + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.new_local_llm.rst b/src_code_for_reproducibility/docs/source/src.models.new_local_llm.rst new file mode 100644 index 0000000000000000000000000000000000000000..d65981b7a6bbf99327fb467a63644bbe98d137f4 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.new_local_llm.rst @@ -0,0 +1,7 @@ +src.models.new\_local\_llm module +================================= + +.. automodule:: src.models.new_local_llm + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.oai_agent.rst b/src_code_for_reproducibility/docs/source/src.models.oai_agent.rst new file mode 100644 index 0000000000000000000000000000000000000000..bd7da8bb6cfc8426bd9725b0ccaf58cfe91eee93 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.oai_agent.rst @@ -0,0 +1,7 @@ +src.models.oai\_agent module +============================ + +.. automodule:: src.models.oai_agent + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.rst b/src_code_for_reproducibility/docs/source/src.models.rst new file mode 100644 index 0000000000000000000000000000000000000000..d03983340a5b0317354d1895df709277d5a4baed --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.rst @@ -0,0 +1,20 @@ +src.models package +================== + +.. automodule:: src.models + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.models.dummy_local_llm + src.models.local_llm + src.models.new_local_llm + src.models.server_llm + src.models.updatable_worker + src.models.vllm_worker_wrap diff --git a/src_code_for_reproducibility/docs/source/src.models.server_llm.rst b/src_code_for_reproducibility/docs/source/src.models.server_llm.rst new file mode 100644 index 0000000000000000000000000000000000000000..6b832a40dafe00b5e56b4060dc07e4dda0a03bad --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.server_llm.rst @@ -0,0 +1,7 @@ +src.models.server\_llm module +============================= + +.. automodule:: src.models.server_llm + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.updatable_worker.rst b/src_code_for_reproducibility/docs/source/src.models.updatable_worker.rst new file mode 100644 index 0000000000000000000000000000000000000000..ee05dfbe7dd407eed4e275f525c04ca6f2c857ae --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.updatable_worker.rst @@ -0,0 +1,7 @@ +src.models.updatable\_worker module +=================================== + +.. automodule:: src.models.updatable_worker + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.models.vllm_worker_wrap.rst b/src_code_for_reproducibility/docs/source/src.models.vllm_worker_wrap.rst new file mode 100644 index 0000000000000000000000000000000000000000..466f755e064534683462b4259876d874678290ed --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.models.vllm_worker_wrap.rst @@ -0,0 +1,7 @@ +src.models.vllm\_worker\_wrap module +==================================== + +.. automodule:: src.models.vllm_worker_wrap + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.rst b/src_code_for_reproducibility/docs/source/src.rst new file mode 100644 index 0000000000000000000000000000000000000000..d2dcff7e18f979e933a27467f4893f3ad5372a88 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.rst @@ -0,0 +1,28 @@ +src package +=========== + +.. automodule:: src + :members: + :undoc-members: + :show-inheritance: + +Subpackages +----------- + +.. toctree:: + :maxdepth: 4 + + src.environments + src.experiments + src.generation + src.models + src.training + src.utils + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.run diff --git a/src_code_for_reproducibility/docs/source/src.run.rst b/src_code_for_reproducibility/docs/source/src.run.rst new file mode 100644 index 0000000000000000000000000000000000000000..379615bab324e654401411db23058b5eec87e277 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.run.rst @@ -0,0 +1,7 @@ +src.run module +============== + +.. automodule:: src.run + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.training.ppo_train.rst b/src_code_for_reproducibility/docs/source/src.training.ppo_train.rst new file mode 100644 index 0000000000000000000000000000000000000000..98f9e085b16ddcfcdecf4237345cb4c1b36b8b80 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.training.ppo_train.rst @@ -0,0 +1,7 @@ +src.training.ppo\_train module +============================== + +.. automodule:: src.training.ppo_train + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.training.ppo_train_value_head.rst b/src_code_for_reproducibility/docs/source/src.training.ppo_train_value_head.rst new file mode 100644 index 0000000000000000000000000000000000000000..a8d6e526eacfce47fac83a7e4617e04db56ecee2 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.training.ppo_train_value_head.rst @@ -0,0 +1,7 @@ +src.training.ppo\_train\_value\_head module +=========================================== + +.. automodule:: src.training.ppo_train_value_head + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.training.reinforce_training.rst b/src_code_for_reproducibility/docs/source/src.training.reinforce_training.rst new file mode 100644 index 0000000000000000000000000000000000000000..5daf4b7250022f523242d6239d0921f362df6d24 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.training.reinforce_training.rst @@ -0,0 +1,7 @@ +src.training.reinforce\_training module +======================================= + +.. automodule:: src.training.reinforce_training + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.training.rl_convs_processing.rst b/src_code_for_reproducibility/docs/source/src.training.rl_convs_processing.rst new file mode 100644 index 0000000000000000000000000000000000000000..cf5db1aa0cb6d010fc70f86c341467ba5e9b485e --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.training.rl_convs_processing.rst @@ -0,0 +1,7 @@ +src.training.rl\_convs\_processing module +========================================= + +.. automodule:: src.training.rl_convs_processing + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.training.rst b/src_code_for_reproducibility/docs/source/src.training.rst new file mode 100644 index 0000000000000000000000000000000000000000..50539fcda2bffa46a72eb48874a7532bf296ff27 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.training.rst @@ -0,0 +1,19 @@ +src.training package +==================== + +.. automodule:: src.training + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.training.ppo_train + src.training.ppo_train_value_head + src.training.reinforce_training + src.training.rl_convs_processing + src.training.train_main diff --git a/src_code_for_reproducibility/docs/source/src.training.train_main.rst b/src_code_for_reproducibility/docs/source/src.training.train_main.rst new file mode 100644 index 0000000000000000000000000000000000000000..838f9ac1cfb2cecd48b2e3d241e7a80c7b86fe25 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.training.train_main.rst @@ -0,0 +1,7 @@ +src.training.train\_main module +=============================== + +.. automodule:: src.training.train_main + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.common_imports.rst b/src_code_for_reproducibility/docs/source/src.utils.common_imports.rst new file mode 100644 index 0000000000000000000000000000000000000000..681ca0651b9298cad189c2b3e7812ba3d31e669f --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.common_imports.rst @@ -0,0 +1,7 @@ +src.utils.common\_imports module +================================ + +.. automodule:: src.utils.common_imports + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.export_ppo_training_set.rst b/src_code_for_reproducibility/docs/source/src.utils.export_ppo_training_set.rst new file mode 100644 index 0000000000000000000000000000000000000000..d05da7df25cb21d8abece3f920c1fb4b88f34d0f --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.export_ppo_training_set.rst @@ -0,0 +1,7 @@ +src.utils.export\_ppo\_training\_set module +=========================================== + +.. automodule:: src.utils.export_ppo_training_set + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.extra_stats.rst b/src_code_for_reproducibility/docs/source/src.utils.extra_stats.rst new file mode 100644 index 0000000000000000000000000000000000000000..7c3e1618d4ae00fc7fc021bf6b96178618321c82 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.extra_stats.rst @@ -0,0 +1,7 @@ +src.utils.extra\_stats module +============================= + +.. automodule:: src.utils.extra_stats + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.inherit_args.rst b/src_code_for_reproducibility/docs/source/src.utils.inherit_args.rst new file mode 100644 index 0000000000000000000000000000000000000000..72d8fc0ff082f4e6d37a852f3dd18a23f58f036d --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.inherit_args.rst @@ -0,0 +1,7 @@ +src.utils.inherit\_args module +============================== + +.. automodule:: src.utils.inherit_args + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.log_gpu_usage.rst b/src_code_for_reproducibility/docs/source/src.utils.log_gpu_usage.rst new file mode 100644 index 0000000000000000000000000000000000000000..44b83082b6eb027ef402603e034c712ccc2cbfcc --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.log_gpu_usage.rst @@ -0,0 +1,7 @@ +src.utils.log\_gpu\_usage module +================================ + +.. automodule:: src.utils.log_gpu_usage + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.log_statistics.rst b/src_code_for_reproducibility/docs/source/src.utils.log_statistics.rst new file mode 100644 index 0000000000000000000000000000000000000000..fde4c1553a35b932ba59aef8d29ceb90b20df48c --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.log_statistics.rst @@ -0,0 +1,7 @@ +src.utils.log\_statistics module +================================ + +.. automodule:: src.utils.log_statistics + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.model_to_cpu.rst b/src_code_for_reproducibility/docs/source/src.utils.model_to_cpu.rst new file mode 100644 index 0000000000000000000000000000000000000000..e98c6dee2e4eb9bd18ec1cb36d6b00740daffb90 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.model_to_cpu.rst @@ -0,0 +1,7 @@ +src.utils.model\_to\_cpu module +=============================== + +.. automodule:: src.utils.model_to_cpu + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.parallel_shuffle.rst b/src_code_for_reproducibility/docs/source/src.utils.parallel_shuffle.rst new file mode 100644 index 0000000000000000000000000000000000000000..3ea9301dcc1ccb245e1c565a135447348244cd65 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.parallel_shuffle.rst @@ -0,0 +1,7 @@ +src.utils.parallel\_shuffle module +================================== + +.. automodule:: src.utils.parallel_shuffle + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.quick_stats.rst b/src_code_for_reproducibility/docs/source/src.utils.quick_stats.rst new file mode 100644 index 0000000000000000000000000000000000000000..76a504eeda8c9b608e1b331cd932fd2f26bffafb --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.quick_stats.rst @@ -0,0 +1,7 @@ +src.utils.quick\_stats module +============================= + +.. automodule:: src.utils.quick_stats + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/docs/source/src.utils.rst b/src_code_for_reproducibility/docs/source/src.utils.rst new file mode 100644 index 0000000000000000000000000000000000000000..4f5cb352cc9ec645c968d0ae99798d47c018c750 --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.rst @@ -0,0 +1,24 @@ +src.utils package +================= + +.. automodule:: src.utils + :members: + :undoc-members: + :show-inheritance: + +Submodules +---------- + +.. toctree:: + :maxdepth: 4 + + src.utils.common_imports + src.utils.export_ppo_training_set + src.utils.extra_stats + src.utils.inherit_args + src.utils.log_gpu_usage + src.utils.log_statistics + src.utils.model_to_cpu + src.utils.parallel_shuffle + src.utils.quick_stats + src.utils.update_start_epoch diff --git a/src_code_for_reproducibility/docs/source/src.utils.update_start_epoch.rst b/src_code_for_reproducibility/docs/source/src.utils.update_start_epoch.rst new file mode 100644 index 0000000000000000000000000000000000000000..72cbad9bd09e056213a2e4cd00a6ba624be333cb --- /dev/null +++ b/src_code_for_reproducibility/docs/source/src.utils.update_start_epoch.rst @@ -0,0 +1,7 @@ +src.utils.update\_start\_epoch module +===================================== + +.. automodule:: src.utils.update_start_epoch + :members: + :undoc-members: + :show-inheritance: diff --git a/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_agent.py b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..5ff402e23224fc7961d9e6796c40daaf2ab4bbaa --- /dev/null +++ b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_agent.py @@ -0,0 +1,259 @@ +from typing import Dict, List, Tuple, Optional, Any +import copy + +class DiplomacyAgent: + """Agent handler for Diplomacy game that follows the MARL standard. + + This class is responsible for parsing LLM output into valid Diplomacy orders, + managing the agent state, and providing information for logging. + """ + + def __init__(self, policy_id: str, power_name: str, random_valid_move=False): + """Initialize the agent handler for a power in the Diplomacy game. + + Args: + power_name: The name of the power this agent controls (e.g., 'FRANCE', 'ENGLAND') + policy_id: The identifier for the policy this agent uses + random_valid_move: If True, will select random valid moves instead of using LLM (default: False) + """ + self.policy_id = policy_id + self.power_name = power_name + self.orders = [] + self.wait = True + self.processing_state = "WAITING_FOR_ORDERS" + self.parsed_orders = [] + self.order_status = {} + self.message_history = [] + self.random_valid_move = random_valid_move + + def step(self, observation_from_env, policy_output=None): + """Update the agent state based on the observation and LLM output. + + Args: + observation_from_env: The observation from the environment + policy_output: The output from the LLM + + Returns: + policy_id: The policy identifier + policy_input: The input to the policy + action: The official action to be sent to the environment + done: Whether the LLM action is ready to be sent to the environment + info: Additional information about the agent + """ + info = {} + + # If random_valid_move is enabled, select random valid moves + if self.random_valid_move: + valid_orders = self._select_random_valid_moves(observation_from_env) + self.orders = valid_orders + self.wait = False + action = { + "orders": valid_orders, + "wait": False + } + return self.policy_id, {}, action, True, info + + # If no policy output, this is the initial step - prepare prompt + if policy_output is None: + # Create initial prompt for the LLM + phase = observation_from_env.get('phase', '') + units = observation_from_env.get('units', {}).get(self.power_name, []) + centers = observation_from_env.get('centers', {}).get(self.power_name, []) + orderable_locations = observation_from_env.get('orderable_locations', {}) + + prompt = self._create_prompt(phase, units, centers, orderable_locations) + + return self.policy_id, {"prompt": prompt}, None, False, info + + # Process the LLM output to extract orders + success, parsed_orders = self._parse_llm_output(policy_output) + self.parsed_orders = parsed_orders + + if not success: + # Need more information from LLM + clarification_prompt = self._create_clarification_prompt(policy_output, parsed_orders) + return self.policy_id, {"prompt": clarification_prompt}, None, False, info + + # Validate if the orders are valid for the current phase + valid_orders = self._validate_orders(parsed_orders, observation_from_env) + + if valid_orders: + # Orders are valid, prepare action for environment + self.orders = valid_orders + self.wait = False + action = { + "orders": valid_orders, + "wait": False + } + return self.policy_id, {}, action, True, info + else: + # Orders are invalid, ask for new ones + error_prompt = self._create_error_prompt(parsed_orders, observation_from_env) + return self.policy_id, {"prompt": error_prompt}, None, False, info + + def _create_prompt(self, phase, units, centers, orderable_locations): + """Create the initial prompt for the LLM. + + Args: + phase: The current game phase + units: List of units controlled by this power + centers: List of supply centers controlled by this power + orderable_locations: List of locations where orders can be issued + + Returns: + A prompt string for the LLM + """ + prompt = f"You are playing as {self.power_name} in Diplomacy. The current phase is {phase}.\n\n" + prompt += f"Your units: {', '.join(units)}\n" + prompt += f"Your supply centers: {', '.join(centers)}\n" + prompt += f"Locations you can order: {', '.join(orderable_locations)}\n\n" + + if phase.endswith('M'): # Movement phase + prompt += "Please provide orders for your units in the form:\n" + prompt += "- A LON H (hold)\n" + prompt += "- F NTH - NWY (move)\n" + prompt += "- A WAL S F LON (support)\n" + prompt += "- F NWG C A NWY - EDI (convoy)\n" + elif phase.endswith('R'): # Retreat phase + prompt += "Please provide retreat orders for your dislodged units:\n" + prompt += "- A PAR R MAR (retreat to MAR)\n" + prompt += "- A PAR D (disband)\n" + elif phase.endswith('A'): # Adjustment phase + if len(units) < len(centers): + prompt += "You can build units. Please provide build orders:\n" + prompt += "- A PAR B (build army in PAR)\n" + prompt += "- F BRE B (build fleet in BRE)\n" + prompt += "- WAIVE (waive a build)\n" + elif len(units) > len(centers): + prompt += "You must remove units. Please provide disbandment orders:\n" + prompt += "- A PAR D (disband army in PAR)\n" + prompt += "- F BRE D (disband fleet in BRE)\n" + + prompt += "\nProvide your orders as a list, one per line." + return prompt + + def _parse_llm_output(self, llm_output): + """Parse the LLM output to extract orders. + + Args: + llm_output: The raw output from the LLM + + Returns: + success: Whether parsing was successful + parsed_orders: List of parsed orders + """ + # Simple parsing for now - extract lines that look like orders + lines = llm_output.strip().split('\n') + orders = [] + + for line in lines: + # Remove list markers, hyphens, etc. + line = line.strip('- *•').strip() + + # Skip empty lines and lines that don't look like orders + if not line or line.startswith('I ') or line.startswith('Let\'s'): + continue + + # Check if it looks like a Diplomacy order + if (' H' in line or ' -' in line or ' S ' in line or ' C ' in line or + ' R ' in line or ' D' in line or ' B' in line or line == 'WAIVE'): + orders.append(line) + + return len(orders) > 0, orders + + def _validate_orders(self, orders, observation): + """Validate if the orders are valid for the current phase. + + Args: + orders: List of orders to validate + observation: Current observation from the environment + + Returns: + List of valid orders or None if invalid + """ + # For simplicity, we'll assume all parsed orders are valid + # In a real implementation, we would use the game's validation logic + return orders + + def _create_clarification_prompt(self, previous_output, parsed_orders): + """Create a prompt asking for clarification when orders couldn't be parsed. + + Args: + previous_output: The previous LLM output + parsed_orders: Any orders that were successfully parsed + + Returns: + A prompt string for the LLM + """ + prompt = f"I couldn't fully understand your orders for {self.power_name}. " + + if parsed_orders: + prompt += f"I understood these orders:\n" + for order in parsed_orders: + prompt += f"- {order}\n" + + prompt += "\nPlease provide clear, valid Diplomacy orders in the format:\n" + prompt += "- A LON H\n- F NTH - NWY\n- etc.\n" + return prompt + + def _create_error_prompt(self, invalid_orders, observation): + """Create a prompt when orders are invalid. + + Args: + invalid_orders: The invalid orders + observation: Current observation from the environment + + Returns: + A prompt string for the LLM + """ + prompt = f"The following orders for {self.power_name} are invalid:\n" + for order in invalid_orders: + prompt += f"- {order}\n" + + prompt += "\nPlease provide valid orders for your units." + return prompt + + def get_log_info(self): + """Get information about the agent required to log a trajectory. + + Returns: + log_info: Information about the agent required to log a trajectory. + """ + return { + "power_name": self.power_name, + "orders": self.orders, + "wait": self.wait, + "parsing_state": self.processing_state, + "message_history": self.message_history + } + + def render(self): + """Render the current state of the agent.""" + print(f"Power: {self.power_name}") + print(f"Orders: {self.orders}") + print(f"Wait: {self.wait}") + + def close(self): + """Perform any necessary cleanup.""" + pass + + def _select_random_valid_moves(self, observation): + """Select random valid moves for all units. + + Args: + observation: Current observation from the environment + + Returns: + List of valid orders + """ + import random + + possible_orders = observation.get('possible_orders', {}) + valid_orders = [] + + # For each location with possible orders, select one randomly + for location, orders in possible_orders.items(): + if orders: # If there are any possible orders for this location + valid_orders.append(random.choice(orders)) + + return valid_orders \ No newline at end of file diff --git a/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_env.py b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_env.py new file mode 100644 index 0000000000000000000000000000000000000000..9b72612c43f2535d353b0157ce72a9b79c23cbb3 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_env.py @@ -0,0 +1,230 @@ +from typing import Dict, List, Tuple, Optional, Any +from diplomacy import Game +import random + +class DiplomacyEnv: + """Multi-Agent Reinforcement Learning environment for Diplomacy. + + This class wraps the Diplomacy game engine to provide an interface + compliant with the MARL standard. + """ + + def __init__(self, random_seed=None, map_name="standard", game_id=None, rules=None, max_steps=50): + """Initialize the Diplomacy environment. + + Args: + map_name: The name of the map to use (default: "standard") + game_id: Optional game ID + rules: Optional rules to apply to the game + max_steps: Maximum number of steps before forcing game end (default: 10) + """ + self.random_seed = random_seed + self.map_name = map_name + self.game_id = game_id + self.rules = rules or [] + self.game = None + self.active_powers = [] + self.render_mode = None + self.max_steps = max_steps + self.current_steps = 0 + + def reset(self): + """Reset the environment to an initial state and return the initial observation. + + Returns: + observation: A dictionary where keys are agent identifiers and values are observations. + """ + # Initialize a new game + self.game = Game(game_id=self.game_id, map_name=self.map_name) + + # Apply rules + for rule in self.rules: + self.game.add_rule(rule) + + # Determine active powers (not eliminated) + self.active_powers = [name for name, power in self.game.powers.items() + if not power.is_eliminated()] + + # Reset step counter + self.current_steps = 0 + + # Create initial observations for all powers + observations = {} + for power_name in self.active_powers: + observations[power_name] = self._create_observation(power_name) + + return observations + + def step(self, actions): + """Take a step in the environment using the provided actions. + + Args: + actions: A dictionary where keys are agent identifiers and values are actions. + + Returns: + observations: A dictionary where keys are agent identifiers and values are observations. + done: Whether the episode has ended. + info: Additional information about the environment. + """ + print(f"stepping {self.current_steps}") + self.current_steps += 1 + # Apply actions (orders) for each power + for power_name, action in actions.items(): + if power_name in self.active_powers: + orders = action.get("orders", []) + wait = action.get("wait", True) + + # Set orders for the power + if orders: + self.game.set_orders(power_name, orders) + + # Set wait flag + self.game.set_wait(power_name, wait) + + # Check if all active powers are ready to proceed + if self.game.does_not_wait(): + # Process the current phase + self.game.process() + + + # Update active powers list after processing + self.active_powers = [name for name, power in self.game.powers.items() + if not power.is_eliminated()] + + # Create observations for all active powers + observations = {} + for power_name in self.active_powers: + observations[power_name] = self._create_observation(power_name) + + # Check if the game is done (either naturally or due to max steps) + done = self.game.is_game_done or self.current_steps >= self.max_steps + + # Create info dict + info = { + "phase": self.game.get_current_phase(), + "active_powers": self.active_powers, + "centers": self.game.get_centers(), + "units": self.game.get_units(), + "current_steps": self.current_steps, + "max_steps_reached": self.current_steps >= self.max_steps + } + + return observations, done, info + + def _create_observation(self, power_name): + """Create observation for a specific power. + + Args: + power_name: The name of the power + + Returns: + An observation dictionary + """ + observation = { + "phase": self.game.get_current_phase(), + "units": self.game.get_units(), + "centers": self.game.get_centers(), + "orderable_locations": self.game.get_orderable_locations(power_name), + "order_status": self.game.get_order_status(power_name), + "possible_orders": self._get_possible_orders_for_power(power_name) + } + return observation + + def _get_possible_orders_for_power(self, power_name): + """Get all possible orders for a power's units. + + Args: + power_name: The name of the power + + Returns: + A dictionary mapping units to their possible orders + """ + all_possible_orders = self.game.get_all_possible_orders() + + # Filter for only the locations where this power has units + power_units = self.game.get_units(power_name) + power_unit_locations = [unit[2:] for unit in power_units] + + # For retreat phases, include retreating units + if self.game.phase_type == 'R': + power = self.game.get_power(power_name) + power_unit_locations.extend([unit[2:] for unit in power.retreats]) + + # For adjustment phases, include buildable locations + elif self.game.phase_type == 'A': + power = self.game.get_power(power_name) + # If we have more centers than units, we can build + if len(power.centers) > len(power.units): + buildable_sites = self.game._build_sites(power) + power_unit_locations.extend(buildable_sites) + # If we have more units than centers, we need to remove + elif len(power.units) > len(power.centers): + # All units are candidates for removal + pass + + # Filter the possible orders to only those for this power's units/locations + power_possible_orders = {} + for loc, orders in all_possible_orders.items(): + if loc[:3] in power_unit_locations: + power_possible_orders[loc] = orders + + return power_possible_orders + + def get_log_info(self): + """Get additional information about the environment for logging. + + Returns: + log_info: Information about the environment required to log the game. + """ + if not self.game: + return {} + + return { + "game_id": self.game.game_id, + "phase": self.game.get_current_phase(), + "map_name": self.game.map_name, + "centers": self.game.get_centers(), + "units": self.game.get_units(), + "powers": {name: { + "units": power.units, + "centers": power.centers, + "is_eliminated": power.is_eliminated(), + "order_status": self.game.get_order_status(name) + } for name, power in self.game.powers.items()}, + "orders": self.game.get_orders(), + "active_powers": self.active_powers, + "is_game_done": self.game.is_game_done, + "outcome": self.game.outcome if self.game.is_game_done else None + } + + def render(self, mode='human'): + """Render the current state of the environment. + + Args: + mode: The rendering mode ('human', 'svg', etc.) + + Returns: + The rendered image if applicable + """ + self.render_mode = mode + if self.game: + if mode == 'human': + # Just print basic game state + print(f"Game: {self.game.game_id}") + print(f"Phase: {self.game.get_current_phase()}") + print(f"Active Powers: {self.active_powers}") + print("Supply Centers:") + for power_name, centers in self.game.get_centers().items(): + print(f" {power_name}: {centers}") + print("Units:") + for power_name, units in self.game.get_units().items(): + print(f" {power_name}: {units}") + return None + elif mode == 'svg': + # Return SVG representation + return self.game.render(output_format='svg') + return None + + def close(self): + """Perform any necessary cleanup.""" + self.game = None \ No newline at end of file diff --git a/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_logging.py b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_logging.py new file mode 100644 index 0000000000000000000000000000000000000000..3f60e82d9738c116c0b8b8d3f7818eddebb18fa2 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_logging.py @@ -0,0 +1,360 @@ +import os +import json +from utils.common_imports import * + + + +def diplomacy_log_match( + path, + agents_log_info, + env_log_info, + metrics_func=None, + metrics_func_args=None + ): + """ + Logs the Diplomacy game data and generates HTML visualizations using the get_log_info methods. + + Args: + path (str): Base path to save the data. + agents_log_info (list): List of agent information dictionaries containing the get_log_info results. + env_log_info (dict): Environment information from its get_log_info method. + metrics_func (str, optional): Name of the function to calculate metrics. + metrics_func_args (dict, optional): Arguments for the metrics function. + """ + # Create directory structure + os.makedirs(path, exist_ok=True) + + # Save the environment log info + env_log_path = os.path.join(path, "env_log.json") + with open(env_log_path, "w") as f: + json.dump(env_log_info, f, indent=4, default=_json_serialize) + + # Process each agent's log info + for agent_log in agents_log_info: + power_name = agent_log["power_name"] + + # Define paths for raw data and statistics subfolders + power_path = os.path.join(path, power_name) + raw_data_path = os.path.join(power_path, "raw_data") + statistics_path = os.path.join(power_path, "statistics") + + # Ensure directories exist + os.makedirs(raw_data_path, exist_ok=True) + os.makedirs(statistics_path, exist_ok=True) + + # Determine the next available file number for raw data + raw_files = os.listdir(raw_data_path) + raw_numbers = [int(f.split('_')[-1].split('.')[0]) for f in raw_files if f.startswith("log_")] + next_raw_number = max(raw_numbers, default=0) + 1 + raw_file = os.path.join(raw_data_path, f"log_{next_raw_number}.json") + + # Save agent log info + with open(raw_file, "w") as f: + json.dump(agent_log, f, indent=4, default=_json_serialize) + + # Log metrics if a metrics function is provided + if metrics_func: + metrics_files = os.listdir(statistics_path) + metrics_numbers = [int(f.split('_')[-1].split('.')[0]) for f in metrics_files if f.startswith("metrics_")] + next_metrics_number = max(metrics_numbers, default=0) + 1 + metrics_file = os.path.join(statistics_path, f"metrics_{next_metrics_number}.json") + + metrics = globals()[metrics_func](agent_log, info, **metrics_func_args) + with open(metrics_file, "w") as f: + json.dump(metrics, f, indent=4) + + # Generate the HTML visualization + html_content = generate_diplomacy_html(agents_log_info, env_log_info) + + # Ensure the html directory exists + html_path = os.path.join(path, "html") + os.makedirs(html_path, exist_ok=True) + + # Determine the next available file number for HTML + html_files = os.listdir(html_path) + html_numbers = [int(f.split('_')[-1].split('.')[0]) for f in html_files if f.startswith("game_summary_")] + next_html_number = max(html_numbers, default=0) + 1 + html_file = os.path.join(html_path, f"game_summary_{next_html_number}.html") + + # Save the HTML content to a file + with open(html_file, "w") as f: + f.write(html_content) + +def generate_diplomacy_html(agent_infos, env_info): + """ + Generate HTML visualization for a Diplomacy game. + + Args: + agent_infos (list): List of agent information dictionaries from get_log_info. + env_info (dict): Environment information from get_log_info. + + Returns: + str: HTML content for the game visualization. + """ + # Extract game information + game_id = env_info.get("game_id", "Unknown") + phase = env_info.get("phase", "Unknown") + map_name = env_info.get("map_name", "standard") + is_game_done = env_info.get("is_game_done", False) + outcome = env_info.get("outcome", []) + + centers = env_info.get("centers", {}) + units = env_info.get("units", {}) + + # HTML head and style + html_content = """ + + + + + + Diplomacy Game {game_id} + + + +
+

Game Information

+
+
+

Game Details

+

Game ID: {game_id}

+

Phase: {phase}

+

Map: {map_name}

+

Status: {status}

+
+
+

Supply Centers

+
+ """.format( + game_id=game_id, + phase=phase, + map_name=map_name, + status="Completed" if is_game_done else "Active" + ) + + # Add supply center information + for power, power_centers in centers.items(): + html_content += f""" +
+ {power}: {len(power_centers)} +
+ """ + + html_content += """ +
+
+
+ """ + + # Add outcome if game is done + if is_game_done and outcome: + winners = outcome[1:] if len(outcome) > 1 else ["Draw"] + html_content += f""" +
+

Game Outcome

+

Winners: {', '.join(winners)}

+
+ """ + + html_content += """ +
+
+ """ + + # Add each power's information + for agent_log in agent_infos: + power_name = agent_log["power_name"] + power_class = power_name.lower() + orders = agent_log.get("orders", []) + message_history = agent_log.get("message_history", []) + + html_content += f""" +
+
{power_name}
+ +
+

Units

+
    + """ + + # Add units information + power_units = units.get(power_name, []) + for unit in power_units: + html_content += f"
  • {unit}
    • " + + html_content += """ +
+
+ +
+
Final Orders
+
    + """ + + # Add orders + for order in orders: + html_content += f"
  • {order}
    • " + + html_content += """ +
+
+ """ + + # Add message history + for message in message_history: + if isinstance(message, dict): + # Skip system messages or handle differently + if message.get("role") == "system": + continue + + role = message.get("role", "unknown") + content = message.get("content", "") + + role_class = "user" if role == "user" else "assistant" + role_display = "Environment" if role == "user" else f"LLM ({power_name})" + + # Escape HTML characters in content + content = content.replace("<", "<").replace(">", ">").replace("\n", "
") + + html_content += f""" +
+
{role_display}
+

{content}

+
+ """ + elif isinstance(message, str): + # Simple string messages (may be used in some implementations) + html_content += f""" +
+

{message}

+
+ """ + + html_content += """ +
+ """ + + html_content += """ +
+ + + """ + + return html_content + +def _json_serialize(obj): + """ + A helper function to convert non-JSON-serializable objects + (like OrderResult) into strings or dicts. + """ + # Check for the specific object types you know are problematic + if obj.__class__.__name__ == "OrderResult": + # Return a string representation or a dict + return str(obj) + + # Fallback: attempt to convert anything else to string + return str(obj) \ No newline at end of file diff --git a/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_logging_for_training.py b/src_code_for_reproducibility/markov_games/diplomacy/diplomacy_logging_for_training.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/src_code_for_reproducibility/markov_games/ipd/Ipd_hard_coded_agents.py b/src_code_for_reproducibility/markov_games/ipd/Ipd_hard_coded_agents.py new file mode 100644 index 0000000000000000000000000000000000000000..a974bddc69c1a3002ce5d84aac868f59bb731900 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/ipd/Ipd_hard_coded_agents.py @@ -0,0 +1,72 @@ +from dataclasses import dataclass +from typing import Any, Tuple + +from mllm.markov_games.ipd.ipd_agent import IPDAgent +from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn + + +@dataclass +class AlwaysCooperateIPDAgent(IPDAgent): + async def act(self, observation) -> Tuple[Any, AgentActLog]: + """ + Always plays the cooperate action, ignoring observation. + Returns the configured cooperate_string so the simulation parses it as "C". + """ + + action = self.cooperate_string + + # Log a minimal, structured chat turn for consistency with other agents + turn_text = f"Playing cooperate: {action}" + self.state.chat_history.append( + ChatTurn( + agent_id=self.agent_id, + role="assistant", + content=turn_text, + is_state_end=True, + ) + ) + + act_log = AgentActLog( + chat_turns=[self.state.chat_history[-1]], + info=None, + ) + + # Advance internal counters similar to IPDAgent semantics + self.state.chat_counter = len(self.state.chat_history) + self.state.round_nb = observation.round_nb + + return action, act_log + + +@dataclass +class AlwaysDefectIPDAgent(IPDAgent): + async def act(self, observation) -> Tuple[Any, AgentActLog]: + """ + Always plays the defect action, ignoring observation. + Returns the configured defect_string so the simulation parses it as "D". + """ + + action = self.defect_string + + # Log a minimal, structured chat turn for consistency with other agents + turn_text = f"Playing defect: {action}" + self.state.chat_history.append( + ChatTurn( + agent_id=self.agent_id, + role="assistant", + content=turn_text, + is_state_end=True, + ) + ) + + act_log = AgentActLog( + chat_turns=[self.state.chat_history[-1]], + info=None, + ) + + # Advance internal counters similar to IPDAgent semantics + self.state.chat_counter = len(self.state.chat_history) + self.state.round_nb = observation.round_nb + + return action, act_log + diff --git a/src_code_for_reproducibility/markov_games/ipd/__init__.py b/src_code_for_reproducibility/markov_games/ipd/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c7f2388f6380fd3a54a2c80d1f1f77ae1d1fd4c8 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/ipd/__init__.py @@ -0,0 +1,7 @@ +from .Ipd_hard_coded_agents import AlwaysCooperateIPDAgent, AlwaysDefectIPDAgent + +__all__ = [ + "AlwaysCooperateIPDAgent", + "AlwaysDefectIPDAgent", +] + diff --git a/src_code_for_reproducibility/markov_games/ipd/__pycache__/Ipd_hard_coded_agents.cpython-312.pyc b/src_code_for_reproducibility/markov_games/ipd/__pycache__/Ipd_hard_coded_agents.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..354fbab18d57425054c5e83b8a8c467d967ddccc Binary files /dev/null and b/src_code_for_reproducibility/markov_games/ipd/__pycache__/Ipd_hard_coded_agents.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/ipd/__pycache__/__init__.cpython-312.pyc b/src_code_for_reproducibility/markov_games/ipd/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2ad972161150c1335d840f5d86f3a662a30f5826 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/ipd/__pycache__/__init__.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e1154ff84f7e1316b3bb49a684c6e5c789276cb1 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c559efc356f340201bb70a3558c4b9abe6a5c7df Binary files /dev/null and b/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_statistics.cpython-312.pyc b/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_statistics.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0bf55aa8dc1407f9fa71222cef9dd1ef44f6be6d Binary files /dev/null and b/src_code_for_reproducibility/markov_games/ipd/__pycache__/ipd_statistics.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/ipd/ipd_agent.py b/src_code_for_reproducibility/markov_games/ipd/ipd_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..3a6f64b542dcc9ee7e114e617bde9cc1181ea301 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/ipd/ipd_agent.py @@ -0,0 +1,115 @@ +import copy +import json +import random +import re +from collections.abc import Callable +from copy import deepcopy +from dataclasses import dataclass, field +from typing import Any, Dict, List, Optional, Tuple, Union + +from mllm.markov_games.agent import Agent +from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn + + +@dataclass +class IPDAgentState: + """ + TOWRITE + """ + + nb_retries: int + round_nb: int + chat_counter: int + chat_history: List[ChatTurn] + + +@dataclass +class IPDAgent(Agent): + seed: int + agent_id: str + agent_name: str + policy: Callable[[List[Dict]], str] + intro_prompt: str # Introduction prompt explaining the game rules + goal_prompt: str # Prompt explaining the agent's goal + strategy_prompt: str # Prompt suggesting a strategy to the agent + max_errors: int # Maximum number of errors allowed before default action + allow_reasoning: bool # Whether to allow reasoning in the response + max_reasoning_chars: int # Maximum number of characters for reasoning + cooperate_string: str # string parsed as playing cooperate by simulation + defect_string: str # string parsed as playing defect by simulation + + def __post_init__(self): + self.state = IPDAgentState( + nb_retries=0, round_nb=0, chat_counter=0, chat_history=[] + ) + + async def act(self, observation) -> Tuple[Any, AgentActLog]: + """ + TOWRITE + """ + + action = None + action_is_ready = False + round_nb = observation.round_nb + + # If it's the first round, we need to send the intro prompt + if round_nb == 0 and self.state.chat_counter == 0: + self.state.chat_history.append( + ChatTurn( + agent_id=self.agent_id, + role="user", + content=self.intro_prompt, + is_state_end=True, + ) + ) + + # If new round + if round_nb > self.state.round_nb: + coagent_action = observation.last_coagent_move + user_message = f"Last round, the other agent played {coagent_action}." + self.state.chat_history.append( + ChatTurn( + agent_id=self.agent_id, + role="user", + content=user_message, + is_state_end=True, + ) + ) + + # If not new round, try to get valid action from policy + output_chat_turn: ChatTurn = await self.policy( + state=self.state.chat_history, + agent_id=self.agent_id, + regex=f"({self.cooperate_string}|{self.defect_string})", + ) + self.state.chat_history.append(output_chat_turn) + action = output_chat_turn.content + + agent_step_log = AgentActLog( + chat_turns=self.state.chat_history[self.state.chat_counter :], info=None + ) + self.state.chat_counter = len(self.state.chat_history) + self.state.round_nb = round_nb + + return action, agent_step_log + + def get_safe_copy(self): + """ + Return a safe copy of the agent. + """ + agent_copy = copy.copy(self) + agent_copy.state = copy.deepcopy(self.state) + return agent_copy + + def reset(self): + self.state = IPDAgentState() + raise NotImplementedError + + def render(self): + pass + + def close(self): + pass + + def get_agent_info(self): + pass diff --git a/src_code_for_reproducibility/markov_games/ipd/ipd_simulation.py b/src_code_for_reproducibility/markov_games/ipd/ipd_simulation.py new file mode 100644 index 0000000000000000000000000000000000000000..238c7d319bd6e679284d2636aeffa194662a664b --- /dev/null +++ b/src_code_for_reproducibility/markov_games/ipd/ipd_simulation.py @@ -0,0 +1,162 @@ +import copy +import random +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple + +import numpy as np + +from mllm.markov_games.markov_game import Simulation +from mllm.markov_games.rollout_tree import SimulationStepLog +from mllm.utils.get_coagent_id import get_coagent_id + + +@dataclass +class IPDState: + """ + State of the Iterated Prisoner's Dilemma game. + """ + + round_nb: int = 0 + done: bool = False + last_moves: Dict[str, str] | None = None + + +@dataclass +class IPDObs: + """ + Observation in Iterated Prisoner's Dilemma game. + """ + + round_nb: int + last_coagent_move: str | None + + +class IPD(Simulation): + """ + Iterated Prisoner's Dilemma simulation following the standard. + + In each round of the game, two agents simultaneously choose to either cooperate (C) or defect (D). + The payoffs are as follows: + - If both cooperate: Both receive the "reward" (usually 3 points) + - If both defect: Both receive the "punishment" (usually 1 point) + - If one cooperates and one defects: The defector receives the "temptation" (usually 5 points) + and the cooperator receives the "sucker" payoff (usually 0 points) + + The game is played for a specified number of rounds. + """ + + def __init__( + self, + agent_ids: List[str], + agent_names: List[str], + seed: int, + rounds_per_game: int, + reward: float, # Both cooperate + punishment: float, # Both defect + temptation: float, # Defector's reward when other cooperates + sucker: float, # Cooperator's reward when other defects + cooperate_actions: List[str], + defect_actions: List[str], + ): + self.agent_ids = agent_ids + self.agent_names = agent_names + self.seed = seed + self.rounds_per_game = rounds_per_game + self.reward = reward + self.punishment = punishment + self.temptation = temptation + self.sucker = sucker + self.cooperate_actions = cooperate_actions + self.defect_actions = defect_actions + self.state = IPDState() + + def step(self, actions: Dict[str, str]) -> Tuple[bool, SimulationStepLog]: + """ + Take a step in the environment using the provided actions. + Here, the observations are just the states of the game. + + Args: + actions (dict): A dictionary where keys are agent identifiers and values are actions ('C' or 'D'). + + Returns: + observations (dict): A dictionary where keys are agent identifiers and values are observations. + done (bool): Whether the episode has ended. + info (dict): Additional information about the environment. + """ + + # Calculate rewards using payoff matrix + agent0_action = actions[self.agent_ids[0]] + agent1_action = actions[self.agent_ids[1]] + + # Normalize actions to standard cooperate/defect/gibberish format + def normalize_action(action): + if action in self.cooperate_actions: + return "C" + elif action in self.defect_actions: + return "D" + else: + return "D" + + norm_action0 = normalize_action(agent0_action) + norm_action1 = normalize_action(agent1_action) + + payoffs = { + ("C", "C"): [self.reward, self.reward], + ("C", "D"): [self.sucker, self.temptation], + ("D", "C"): [self.temptation, self.sucker], + ("D", "D"): [self.punishment, self.punishment], + } + + round_rewards = { + self.agent_ids[0]: payoffs[(norm_action0, norm_action1)][0], + self.agent_ids[1]: payoffs[(norm_action0, norm_action1)][1], + } + + # Update game state + self.state.round_nb += 1 + self.state.last_moves = copy.deepcopy(actions) + done = self.state.round_nb >= self.rounds_per_game + step_log = SimulationStepLog( + rewards=round_rewards, + info={ + "actions": { + self.agent_ids[0]: norm_action0, + self.agent_ids[1]: norm_action1, + } + }, + ) + + return done, step_log + + def get_obs(self): + """Returns all agent observations in dict + Returns: + observations + """ + observations = {} + for agent_id in self.agent_ids: + observations[agent_id] = self.get_obs_agent(agent_id) + return observations + + def get_obs_agent(self, agent_id): + """Returns observation for agent_id""" + if self.state.last_moves != None: + other_id = get_coagent_id(self.agent_ids, agent_id) + last_coagent_move = self.state.last_moves[other_id] + else: + last_coagent_move = None + obs = IPDObs(round_nb=self.state.round_nb, last_coagent_move=last_coagent_move) + return obs + + def reset(self): + """Returns initial observations and states""" + self.state = IPDState() + return self.get_obs() + + def get_safe_copy(self): + """ + Return a safe copy of the simulation. + """ + simulation_copy = copy.copy(self) + simulation_copy.state = copy.deepcopy(self.state) + return simulation_copy diff --git a/src_code_for_reproducibility/markov_games/ipd/ipd_statistics.py b/src_code_for_reproducibility/markov_games/ipd/ipd_statistics.py new file mode 100644 index 0000000000000000000000000000000000000000..8740fda6bc2550c92aef27ed9fbe7bc945be42ca --- /dev/null +++ b/src_code_for_reproducibility/markov_games/ipd/ipd_statistics.py @@ -0,0 +1,18 @@ +from __future__ import annotations + +from typing import Dict, Callable, List, Tuple + +from mllm.markov_games.rollout_tree import SimulationStepLog + + +def avg_reward(sl: SimulationStepLog) -> List[Tuple[str, float]]: + for aid in sl.rewards.keys(): + if "buffer" in str(aid) and "live" not in str(aid): + return None + # One value per agent at each step + rewards_dict = {f"reward-{aid}": float(v) for aid, v in (sl.rewards or {}).items()} + return [(key, value) for key, value in rewards_dict.items() if value is not None] + +stat_functs: list[Callable[[SimulationStepLog], List[Tuple[str, float]]]] = [ + avg_reward, +] \ No newline at end of file diff --git a/src_code_for_reproducibility/markov_games/negotiation/README.md b/src_code_for_reproducibility/markov_games/negotiation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..c8ebadee705971c5331924ed1b9d53c7e5f69770 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/README.md @@ -0,0 +1,40 @@ +## Negotiation Games: core mechanics and variants + +This family of games feature two agents who, in each round, may briefly communicate and then simultaneously propose how to split a fixed resource (most commonly 10 coins). Rewards are the amount kept multiplied by an agent’s per-unit value. The starting speaker alternates deterministically across rounds. + +Communication is optional and variant-dependent: some settings encourage rich messaging to share private information, while others remove messaging entirely to focus on allocation behavior. + +Proportional splitting is used when the two proposals exceed the available total: allocations are scaled proportionally rather than discarded. This preserves a useful learning signal even when agents over-claim. + +### Variants (in increasing difficulty) + +- No‑Press Split + - Single item type (coins) + - No communication; agents go straight to making split proposals, with the starting player alternating deterministically. + - Motivation: mirrors no‑communication setups (e.g., Advantage Alignment) while keeping the split decision nontrivial. + - Deterministic Mode: values are fixed and public: one agent values coins at 10, the other at 1 (alternates each round). + - Stochastic Mode: values are random and uncorrelated. + +- Trust-and-Split RPS (TAS-RPS) + - Single item type (coins) + - Each round, a rock–paper–scissors hand draw creates a strong asymmetry: the winner’s per-coin value is 10, the loser’s is 1. + - Each agent initially sees only their own hand and must communicate to coordinate an optimal split. + - Motivation: enforce large value disparity so one’s own value reveals little about the other’s (avoiding ceiling effects) and incentivize meaningful communication. + +- Trust-and-Split (TAS) + - Single item type (coins); each round, each agent’s per-coin value is independently sampled in a broad range (e.g., 1–20). + - Each agent observes only their own value; they may use short messages to share and negotiate. + - Motivation: a simple blend that tests whether agents learn to exchange private information and coordinate proportional, value-aware splits. + +- Deal-or-No-Deal (DOND) + - Introduced in [Deal or No Deal? End-to-End Learning for Negotiation Dialogues](https://arxiv.org/pdf/1706.05125) + - Multiple item types (typically "books", "hats" and "balls") with limited stocks; each agent has its own per-type values. + - A deal pays out only if both proposals exactly agree and respect the stock; otherwise no deal (zero reward) that round. + - Motivation: a known benchmark closer to real-world bargaining, where both parties must explicitly agree. + + + + + + + diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/dond_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/dond_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5c3cdd472992d134e05776b55eeeaea51ef72feb Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/dond_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/dond_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/dond_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f30a76c150da8a74c1987b2d712c714b6b3bea49 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/dond_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c5e0ddf5cac0cf4f5b9909e743fbaee6940fc4ab Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_hard_coded_policies.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_hard_coded_policies.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7d0e8f66077980f9dca7955a12d324c195ff89ce Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_hard_coded_policies.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6cea4e60166d85502b70f5d5ab56eaa4707eb7e9 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/nego_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/negotiation_statistics.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/negotiation_statistics.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ebb546838e912f12d62abfe2b22a71ce2adbd98e Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/negotiation_statistics.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/no_press_nego_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/no_press_nego_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad86d92929bd6c38a1aeba2fa3d8ad619cab45df Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/no_press_nego_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/no_press_nego_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/no_press_nego_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0c9443951c86d71f7d7150be5007aee9061220f2 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/no_press_nego_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c8c381a73774bad0361fbe12560a32c645ca13dc Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_rps_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_rps_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9228dcd110476851547e76bac7ba6f795c991a0a Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_rps_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_rps_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_rps_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a3e363d683aca4bb5e1f4ed73801043a87b64d50 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_rps_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simple_agent.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simple_agent.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..708db5895faf46b2e0eba53b1fe8567e7b17cde1 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simple_agent.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simple_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simple_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..19528e275e9b4653c58fd3f011cbc819b5351403 Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simple_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simulation.cpython-312.pyc b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simulation.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..67cba7cc63da526ed267724c423a113da94355ea Binary files /dev/null and b/src_code_for_reproducibility/markov_games/negotiation/__pycache__/tas_simulation.cpython-312.pyc differ diff --git a/src_code_for_reproducibility/markov_games/negotiation/dond_agent.py b/src_code_for_reproducibility/markov_games/negotiation/dond_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..958756f16e1ab0e5b348d2ae7a37777be2a0ad21 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/dond_agent.py @@ -0,0 +1,61 @@ +import copy +import re +from collections.abc import Callable +from dataclasses import dataclass +from typing import Any, Dict, List, Tuple + +from mllm.markov_games.agent import Agent +from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn +from mllm.markov_games.negotiation.dond_simulation import ( + DealNoDealObs, +) +from mllm.markov_games.negotiation.nego_simulation import Split +from mllm.markov_games.negotiation.nego_agent import NegotiationAgent, NegotiationAgentState + +class DealNoDealAgent(NegotiationAgent): + def __init__( + self, + *args, + **kwargs, + ): + super().__init__(*args, **kwargs) + self.intro_prompt = ( + "You are {agent_id}. You are playing an iterated game. " + "At each round, you and other agent will try to distribute among yourselves items of types {item_types}. " + "You only know how much you value each item type, but not the other agent's values. " + "You can communicate with the other agent by sending up to {quota_messages_per_agent_per_round} short messages per round. " + "Each round, after exchanging messages, you and the other agent will submit a private proposal. " + "A deal is accepted only if both proposals match exactly and are within stock; otherwise no deal (0 points for both at that round). " + "The values of the items of the other agent at the previous round are revealed to you after each round. " + "Your goal is: {goal}." + ) + self.new_round_prompt = ("New round {round_nb}. Items: {stock}. Your values: {values}. ") + self.last_round_prompt = ("Last round, other agent's values: {previous_values_coagent}. ") + self.send_split_prompt = ("Respond with ... where you propose how many items of each type you want to keep.") + + def get_message_regex(self, observation: DealNoDealObs) -> str: + return r"[\s\S]{0,400}" + + def get_split_regex(self, observation: DealNoDealObs) -> str: + parts = [] + for t in observation.item_types: + s = int(observation.quantities.get(t, 0)) + allowed = "|".join(str(k) for k in range(0, s + 1)) + rng = f"({allowed})" + parts.append(fr"<{t}>{rng}") + items_block = "".join(parts) + return fr"({items_block})" + + def get_split_action(self, policy_output: str, observation: DealNoDealObs) -> Split: + import re as _re + allocations: Dict[str, int] = {} + for t in observation.item_types: + m = _re.search(fr"<{t}>([0-9]+)", policy_output) + if m: + allocations[t] = int(m.group(1)) + else: + allocations[t] = 0 + return Split(items_given_to_self=allocations) + + + diff --git a/src_code_for_reproducibility/markov_games/negotiation/dond_simulation.py b/src_code_for_reproducibility/markov_games/negotiation/dond_simulation.py new file mode 100644 index 0000000000000000000000000000000000000000..a27d6ce2cf7e31a0cddd341db39ae7898b086115 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/dond_simulation.py @@ -0,0 +1,153 @@ +import copy +from dataclasses import dataclass +from typing import Any, Dict, List, Tuple + +from numpy.random import default_rng + +from mllm.markov_games.rollout_tree import SimulationStepLog +from mllm.markov_games.negotiation.nego_simulation import Split, NegotiationState, NegotiationObs, NegotiationSimulation +from mllm.utils.get_coagent_id import get_coagent_id + + +AgentId = str + + +@dataclass +class DealNoDealState(NegotiationState): + item_types: List[str] + values: Dict[AgentId, Dict[str, int]] + +@dataclass +class DealNoDealObs(NegotiationObs): + my_values: Dict[str, int] + item_types: List[str] + previous_values_coagent: Dict[str, int] | None + + +def random_partition_integer(rng, total: int, parts: int) -> List[int]: + if parts <= 0: + return [] + if total <= 0: + return [0 for _ in range(parts)] + cuts = sorted(rng.integers(0, total + 1, size=parts - 1).tolist()) + vals = [] + prev = 0 + for c in cuts + [total]: + vals.append(c - prev) + prev = c + return vals + +class DealNoDealSimulation(NegotiationSimulation): + + def __init__( + self, + item_types: List[str] = ["books", "hats", "balls"], + *args, + **kwargs, + ): + super().__init__(item_types=item_types, *args, **kwargs) + self.reset() + + def _other(self, agent_id: AgentId) -> AgentId: + return get_coagent_id(self.agent_ids, agent_id) + + def _sample_stock(self) -> Dict[str, int]: + # total items between 5 and 7 + total_items = int(self.rng.integers(5, 8)) + # nonnegative per-type counts summing to total_items + parts = random_partition_integer(self.rng, total_items, len(self.item_types)) + # allow zeros per type + return {t: int(c) for t, c in zip(self.item_types, parts)} + + def _sample_values_pair(self) -> Dict[AgentId, Dict[str, int]]: + # Each agent has integer non-negative values that sum to 10 + # Each item type valued by at least one agent + # Some item type valued by both agents + while True: + vals_a = random_partition_integer(self.rng, 10, len(self.item_types)) + vals_b = random_partition_integer(self.rng, 10, len(self.item_types)) + a = {t: int(v) for t, v in zip(self.item_types, vals_a)} + b = {t: int(v) for t, v in zip(self.item_types, vals_b)} + # each item valued by at least one + ok1 = all((a[t] > 0) or (b[t] > 0) for t in self.item_types) + # some item valued by both + ok2 = any((a[t] > 0) and (b[t] > 0) for t in self.item_types) + if ok1 and ok2: + return {self.agent_ids[0]: a, self.agent_ids[1]: b} + + def _is_valid_allocation(self, allocation: Dict[str, int], stock: Dict[str, int]) -> bool: + for t in self.item_types: + v = allocation.get(t) + if v is None: + return False + if not isinstance(v, int): + return False + if v < 0 or v > int(stock.get(t, 0)): + return False + return True + + def set_new_round_of_variant(self): + # Keep same values, resample stock + self.state.quantities = self._sample_stock() + + def get_info_of_variant(self, state: NegotiationState, actions: Dict[AgentId, Any]) -> Dict[str, Any]: + return { + "quantities": copy.deepcopy(state.quantities), + "values": copy.deepcopy(state.values), + 'splits': copy.deepcopy(state.splits), + } + + def get_rewards(self, splits: Dict[AgentId, Split]) -> Dict[AgentId, float]: + """ + Returns the rewards for each agent. + """ + split_a = splits[self.agent_ids[0]].items_given_to_self + split_b = splits[self.agent_ids[1]].items_given_to_self + rewards = {self.agent_ids[0]: 0, self.agent_ids[1]: 0} + for t in self.item_types: + # If not complementary, return 0! + if not split_a[t] + split_b[t] == self.state.quantities[t]: + return {self.agent_ids[0]: 0, self.agent_ids[1]: 0} + rewards[self.agent_ids[0]] += split_a[t] * self.state.values[self.agent_ids[0]][t] + rewards[self.agent_ids[1]] += split_b[t] * self.state.values[self.agent_ids[1]][t] + return rewards + + def get_obs(self): + return {agent_id: self.get_obs_agent(agent_id) for agent_id in self.agent_ids} + + def get_obs_agent(self, agent_id): + other_id = self._other(agent_id) + obs = DealNoDealObs( + round_nb=self.state.round_nb, + last_message=self.state.last_message, + current_agent=self.state.current_agent, + quantities=copy.deepcopy(self.state.quantities), + value=0.0, # unused in DOND + other_agent_split=None, # not meaningful until split + split_phase=self.state.split_phase, + quota_messages_per_agent_per_round=self.quota_messages_per_agent_per_round, + my_values=copy.deepcopy(self.state.values[agent_id]), + item_types=list(self.item_types), + previous_values_coagent=copy.deepcopy(self.state.values.get(other_id, {})), + ) + return obs + + def reset(self): + start_agent = self.agent_ids[self._starting_agent_index] + stock = self._sample_stock() + values = self._sample_values_pair() + self.state = DealNoDealState( + round_nb=0, + last_message="", + current_agent=start_agent, + quantities=stock, + values=values, + previous_values=None, + splits={aid: None for aid in self.agent_ids}, + nb_messages_sent={aid: 0 for aid in self.agent_ids}, + split_phase=False, + item_types=list(self.item_types), + ) + return self.get_obs() + + diff --git a/src_code_for_reproducibility/markov_games/negotiation/nego_hard_coded_policies.py b/src_code_for_reproducibility/markov_games/negotiation/nego_hard_coded_policies.py new file mode 100644 index 0000000000000000000000000000000000000000..2b5c191e15ef6b0abada72b1b6ba3a4c59421fdf --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/nego_hard_coded_policies.py @@ -0,0 +1,64 @@ +import asyncio +from typing import Optional +from mllm.markov_games.negotiation.nego_agent import NegotiationAgent +from mllm.markov_games.negotiation.no_press_nego_agent import NoPressAgent +from mllm.markov_games.negotiation.no_press_nego_simulation import NoPressObs +from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn +from mllm.markov_games.negotiation.nego_simulation import Split +from typing import Any, Tuple + +class HardCodedNegoWelfareMaximizingPolicy(NoPressAgent): + async def act(self, observation: NoPressObs) -> Tuple[Any, AgentActLog]: + """ + Policy that gives all of the items to the agent who values them more. + If the items are equally valued, give them to the agent who values them more. + """ + quantities = observation.quantities + my_values = observation.value + other_values = observation.other_value + + items_given_to_self = {} + for item, qty in quantities.items(): + my_v = float(my_values.get(item, 0)) + other_v = float(other_values.get(item, 0)) + if my_v == other_v: + items_given_to_self[item] = int(qty) / 2 + else: + items_given_to_self[item] = int(qty if my_v > other_v else 0) + + action = Split(items_given_to_self=items_given_to_self) + act_log = AgentActLog( + chat_turns=[ + ChatTurn( + agent_id=self.agent_id, + role="assistant", + content="Using welfare-maximizing split (all to higher-value agent).", + is_state_end=True, + ) + ], + info=None, + ) + return action, act_log + +class HardCodedNegoGreedyPolicy(NoPressAgent): + async def act(self, observation: NoPressObs) -> Tuple[Any, AgentActLog]: + """ + Always gives itself all of the items. + """ + quantities = observation.quantities + items_given_to_self = {item: int(qty) for item, qty in quantities.items()} + + action = Split(items_given_to_self=items_given_to_self) + act_log = AgentActLog( + chat_turns=[ + ChatTurn( + agent_id=self.agent_id, + role="assistant", + content="Using greedy split (keep all items).", + is_state_end=True, + ) + ], + info=None, + ) + return action, act_log + diff --git a/src_code_for_reproducibility/markov_games/negotiation/no_press_nego_agent.py b/src_code_for_reproducibility/markov_games/negotiation/no_press_nego_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..c0a62f2465ac8b34dc09cbc003dcc663b170ffe7 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/no_press_nego_agent.py @@ -0,0 +1,94 @@ +from typing import Any, Dict, List, Tuple + +from mllm.markov_games.negotiation.nego_agent import ( + NegotiationAgent, + NegotiationAgentState, +) +from mllm.markov_games.negotiation.nego_simulation import Split +from mllm.markov_games.negotiation.no_press_nego_simulation import NoPressObs +from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn + + +class NoPressAgent(NegotiationAgent): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # No communication in this variant + self.intro_prompt = ( + "Welcome to an iterated game. You are {agent}. The other agent is {other_agent}.\n" + "Setup:\n" + "1. The game consists of multiple independent rounds.\n" + "2. In each round, there are multiple items to split between the two agents.\n" + "3. Both agents are assigned a per-item value between 1 and 20 (inclusive) in each round.\n" + "4. You can observe per-item values of both agents.\n" + "5. Because assignments are random, both agents are equally likely to have same expected per-item value.\n" + "\n" + "Protocol:\n" + "1. Both agents simultaneously propose the amount of each item they will keep.\n" + "2. If the total sum of proposals is less than or equal to the item quantity, both agents receive their proposed amounts.\n" + "3. If the total sum of proposals exceeds the item quantity, they are allocated proportionally.\n" + "4. Your points for the round = (amount you receive per item) x (your per-item value for that round), added across all items.\n" + "5. Points are accumulated across rounds.\n" + "Your goal: {goal}\n" + ) + self.new_round_prompt = ( + "A New Round Begins\n" + "The items to split are {quantities}.\n" + "Your per-item values are {value} and {other_agent}'s per-item values are {other_value}." + ) + self.last_round_prompt = ( + "Last Round Summary:\n" + " - Items to split: {last_quantities}\n" + " - Your per-item values: {last_value_agent}\n" + " - {other_agent}'s per-item values: {last_value_coagent}\n" + " - You proposed: {last_split_agent}\n" + " - You earned: {last_points_agent} points\n" + " - {other_agent} proposed: {last_split_coagent}\n" + " - {other_agent} earned: {last_points_coagent} points\n" + " - Round Complete.\n" + ) + self.send_split_prompt = "Submit Your Proposal\n" "Respond as {proposal_style}" + + def get_message_regex(self, observation: NoPressObs) -> str: + return r"^$" # No messages allowed + + def get_split_regex(self, observation: NoPressObs) -> str: + items = list(observation.quantities.keys()) + # Accept both singular and plural forms + item_pattern = "|".join( + [f"{item[:-1]}s?" if item.endswith("s") else f"{item}s?" for item in items] + ) + regex = rf"(?i)Proposal:\s*((?:\s*(?P(10|[0-9]))\s*(?P{item_pattern})\s*,?)+)" + return regex + + def get_split_action(self, policy_output: str, observation: NoPressObs) -> Split: + items = list(observation.quantities.keys()) + import re as _re + + split_regex = self.get_split_regex(observation) + items_given_to_self = {item: 0 for item in items} + m = _re.match(split_regex, policy_output.strip()) + if m: + # Find all (number, item) pairs + item_pattern = "|".join( + [ + f"{item[:-1]}s?" if item.endswith("s") else f"{item}s?" + for item in items + ] + ) + inner_regex = rf"(?i)(10|[0-9])\s*({item_pattern})" + + def normalize_item_name(item_str): + for orig in items: + if item_str.lower() == orig.lower(): + return orig + if orig.endswith("s") and item_str.lower() == orig[:-1].lower(): + return orig + if ( + not orig.endswith("s") + and item_str.lower() == orig.lower() + "s" + ): + return orig + + for num, item in _re.findall(inner_regex, m.group(1)): + items_given_to_self[normalize_item_name(item)] = int(num) + return Split(items_given_to_self=items_given_to_self) diff --git a/src_code_for_reproducibility/markov_games/negotiation/tas_agent.py b/src_code_for_reproducibility/markov_games/negotiation/tas_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..002160873969ab7292f0f62a091e12ec376022c6 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/tas_agent.py @@ -0,0 +1,108 @@ +from mllm.markov_games.negotiation.nego_agent import NegotiationAgent +from mllm.markov_games.negotiation.nego_simulation import Split +from mllm.markov_games.negotiation.tas_simulation import TrustAndSplitObs + + +class TrustAndSplitAgent(NegotiationAgent): + def __init__(self, num_message_chars, *args, **kwargs): + self.num_message_chars = num_message_chars + super().__init__(*args, **kwargs) + self.intro_prompt = ( + "Welcome to an iterated game. You are {agent}. The other agent is {other_agent}.\n" + "Setup:\n" + "1. The game has multiple independent rounds.\n" + "2. In each round, there are multiple items to split between the two agents.\n" + "3. Both agents are assigned a per-item value between 1 and 20 (inclusive) in each round.\n" + "4. You can only observe your own per-item values.\n" + "5. Because assignments are random, both agents are equally likely to have same expected per-item value.\n" + "\n" + "Protocol:\n" + "1. At the start of the round, one agent begins the conversation. The starting role alternates each round.\n" + "2. Agents exchange a short chat ({quota_messages_per_agent_per_round} messages per round per agent) to negotiate how to split the item.\n" + " - Use this chat to communicate your private per-item value to make informed proposals.\n" + "3. After the chat, both agents simultaneously propose the amount of each item they will keep.\n" + "4. If the total sum of proposals is less than or equal to the item quantity, both agents receive their proposed amounts.\n" + "5. If the total sum of proposals exceeds the item quantity, they are allocated proportionally.\n" + "6. Your points for the round = (amount you receive per item) x (your per-item value for that round), added across all items.\n" + "7. Points are accumulated across rounds.\n" + "Your goal: {goal}\n" + ) + self.new_round_prompt = ( + "A New Round Begins\n" + "The items to split are {quantities}.\n" + "Your per-item values are {value}." + ) + self.last_round_prompt = ( + "Last Round Summary:\n" + " - Items to split: {last_quantities}\n" + " - Your per-item values: {last_value_agent}\n" + " - {other_agent}'s per-item values: {last_value_coagent}\n" + " - You proposed: {last_split_agent}\n" + " - You earned: {last_points_agent} points\n" + " - {other_agent} proposed: {last_split_coagent}\n" + " - {other_agent} earned: {last_points_coagent} points\n" + " - Round Complete.\n" + ) + self.send_split_prompt = ( + "Message quota is finished for this round.\n" + "{other_agent} has finalized their proposal.\n" + "Submit your finalization now\n" + "Respond with {proposal_style2}" + ) + # self.wait_for_message_prompt = "Wait for {other_agent} to send a message..." + self.wait_for_message_prompt = "" + self.last_message_prompt = "{other_agent} said: {last_message}" + # self.send_message_prompt = ( + # f"Send your message now (max {self.num_message_chars} chars)." + # ) + self.send_message_prompt = f"Send your message now in ... (<={self.num_message_chars} chars)." + + def get_message_regex(self, observation: TrustAndSplitObs) -> str: + return rf"[\s\S]{{0,{self.num_message_chars}}}" + + # def get_message_regex(self, observation: TrustAndSplitObs) -> str: + # return rf"(?s).{{0,{self.num_message_chars}}}" + + def get_split_regex(self, observation: TrustAndSplitObs) -> str: + items = list(observation.quantities.keys()) + # Accept both singular and plural forms + item_pattern = "|".join( + [f"{item[:-1]}s?" if item.endswith("s") else f"{item}s?" for item in items] + ) + regex = rf"(?i) ?((?:\s*(?P(10|[0-9]))\s*(?P{item_pattern})\s*,?)+) ?" + return regex + + def get_split_action( + self, policy_output: str, observation: TrustAndSplitObs + ) -> Split: + items = list(observation.quantities.keys()) + import re as _re + + split_regex = self.get_split_regex(observation) + items_given_to_self = {item: 0 for item in items} + m = _re.match(split_regex, policy_output.strip()) + if m: + # Find all (number, item) pairs + item_pattern = "|".join( + [ + f"{item[:-1]}s?" if item.endswith("s") else f"{item}s?" + for item in items + ] + ) + inner_regex = rf"(?i)(10|[0-9])\s*({item_pattern})" + + def normalize_item_name(item_str): + for orig in items: + if item_str.lower() == orig.lower(): + return orig + if orig.endswith("s") and item_str.lower() == orig[:-1].lower(): + return orig + if ( + not orig.endswith("s") + and item_str.lower() == orig.lower() + "s" + ): + return orig + + for num, item in _re.findall(inner_regex, m.group(1)): + items_given_to_self[normalize_item_name(item)] = int(num) + return Split(items_given_to_self=items_given_to_self) diff --git a/src_code_for_reproducibility/markov_games/negotiation/tas_rps_agent.py b/src_code_for_reproducibility/markov_games/negotiation/tas_rps_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..e711c2a65d336e4d9b991c68662069e96b4dfee8 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/tas_rps_agent.py @@ -0,0 +1,118 @@ +import copy +from collections.abc import Callable +from dataclasses import dataclass +from typing import Any, Dict, List, Tuple + +from mllm.markov_games.agent import Agent +from mllm.markov_games.negotiation.nego_agent import ( + Message, + NegotiationAgent, + NegotiationAgentState, + Split, +) +from mllm.markov_games.negotiation.tas_rps_simulation import TrustAndSplitRPSObs +from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn + + +class TrustAndSplitRPSAgent(NegotiationAgent): + def __init__( + self, + num_message_chars: int, + message_start_end_format: bool = False, + proposal_start_end_format: bool = False, + *args, + **kwargs, + ): + self.num_message_chars = num_message_chars + self.message_start_end_format = message_start_end_format + self.proposal_start_end_format = proposal_start_end_format + super().__init__(*args, **kwargs) + self.intro_prompt = ( + "Welcome to an iterated game. You are {agent}. The other agent is {other_agent}.\n" + "\n" + "Setup:\n" + "1. The game has multiple independent rounds.\n" + "2. In each round, there are 10 coins to split between the two agents.\n" + "3. Each agent's per-coin value for that round is determined as follows:\n" + " - Both agents are randomly assigned a rock, paper or scissors hands\n" + " - Rock has the upper hand over scissors, scissors has the upper hand over paper and paper has the upper hand over rock.\n" + " - The agent with the upper hand has a per-coin value of 10.\n" + " - The agent with the lower hand has a per-coin value of 1.\n" + "4. You only see your own hand, but you may communicate it in messages and infer your value based on the other agent's hand.\n" + "5. Over many rounds both agents are equally likely to have the upper and lower hand.\n" + "\n" + "Protocol:\n" + "1. At the start of the round, one agent begins the conversation. The starting role alternates each round.\n" + "2. Agents exchange a short chat ({quota_messages_per_agent_per_round} messages per round per agent) to negotiate how to split the 10 coins.\n" + " - Use this chat to communicate your hand so that both agents can determine their per-coin values.\n" + "3. After the chat, both agents simultaneously propose how many coins they keep.\n" + "4. If the total sum of proposals is less than or equal to 10, both agents receive their proposals.\n" + "5. If the total sum of proposals exceeds 10, the coins are allocated proportionally.\n" + "6. Your points for the round = (coins you receive) x (your per-coin value for that round). \n" + "7. The points are accumulated across rounds.\n" + "Your goal: {goal}\n" + ) + self.new_round_prompt = ( + "A New Round Begins\n" + "Your hand is {hand}. You don't know {other_agent}'s hand yet.\n" + ) + # self.last_round_prompt = ( + # "Last Round Summary:\n" + # " - Your hand: {last_hand_agent}\n" + # " - {other_agent}'s hand: {last_hand_coagent}\n" + # " - Your value per coin: {last_value_agent}\n" + # " - {other_agent}'s value per coin: {last_value_coagent}\n" + # " - You proposed: {last_split_agent} coins\n" + # " - You earned: {last_points_agent} points\n" + # " - {other_agent} proposed: {last_split_coagent} coins\n" + # " - {other_agent} earned: {last_points_coagent} points\n" + # " - Round Complete.\n" + # ) + self.last_round_prompt = "In the previous round, {other_agent} had a {last_hand_value_coagent} hand and proposed {last_split_coagent} coins.\n" + if self.proposal_start_end_format: + self.send_split_prompt = ( + "Submit your proposal\n" + "Respond with <> x <> where x is an integer in [0, 10]." + ) + else: + self.send_split_prompt = ( + "Submit your proposal\n" + "Respond with x where x is an integer in [0, 10]." + ) + self.wait_for_message_prompt = "Wait for {other_agent} to send a message..." + # self.wait_for_message_prompt = "" + self.last_message_prompt = "{other_agent} said: {last_message}" + if self.message_start_end_format: + self.send_message_prompt = f"Send your message now in <>...<> (<={self.num_message_chars} chars)." + else: + self.send_message_prompt = f"Send your message now in ... (<={self.num_message_chars} chars)." + + def get_message_regex(self, observation: TrustAndSplitRPSObs) -> str: + if self.message_start_end_format: + return ( + rf"<>[\s\S]{{0,{self.num_message_chars}}}<>" + ) + else: + return rf"[\s\S]{{0,{self.num_message_chars}}}" + + def get_split_regex(self, observation: TrustAndSplitRPSObs) -> str: + if self.proposal_start_end_format: + return r"<> ?(10|[0-9]) ?<>" + else: + return r" ?(10|[0-9]) ?" + + def get_split_action( + self, policy_output: str, observation: TrustAndSplitRPSObs + ) -> Split: + import re as _re + + if self.proposal_start_end_format: + m = _re.search( + r"<> ?(10|[0-9]) ?<>", policy_output + ) + else: + m = _re.search( + r" ?(10|[0-9]) ?", policy_output + ) + coins_int = int(m.group(1)) if m else int(policy_output) + return Split(items_given_to_self={"coins": coins_int}) diff --git a/src_code_for_reproducibility/markov_games/negotiation/tas_simple_agent.py b/src_code_for_reproducibility/markov_games/negotiation/tas_simple_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..e4439b53a04e8efe4553cb1aa0d85459a6e90c9d --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/tas_simple_agent.py @@ -0,0 +1,90 @@ +from mllm.markov_games.negotiation.nego_agent import NegotiationAgent +from mllm.markov_games.negotiation.nego_simulation import Split +from mllm.markov_games.negotiation.tas_simple_simulation import TrustAndSplitSimpleObs + + +class TrustAndSplitSimpleAgent(NegotiationAgent): + def __init__( + self, + num_message_chars, + message_start_end_format: bool = False, + proposal_start_end_format: bool = False, + *args, + **kwargs, + ): + self.num_message_chars = num_message_chars + self.message_start_end_format = message_start_end_format + self.proposal_start_end_format = proposal_start_end_format + super().__init__(*args, **kwargs) + self.intro_prompt = ( + "Welcome to an iterated game. You are {agent}. The other agent is {other_agent}.\n" + "Setup:\n" + "1. The game has multiple independent rounds.\n" + "2. In each round, there are 10 coins to split between the two agents.\n" + "3. Both agents are assigned a per-coin value between 1 and 10 (inclusive) in each round.\n" + "4. You can only observe your own per-coin value.\n" + "5. Because assignments are random, both agents are equally likely to have same expected per-coin value.\n" + "\n" + "Protocol:\n" + "1. At the start of the round, one agent begins the conversation. The starting role alternates each round.\n" + "2. Agents exchange a short chat ({quota_messages_per_agent_per_round} messages per round per agent) to negotiate how to split the coins.\n" + " - Use this chat to communicate your private per-coin value to make informed proposals.\n" + "3. After the chat, both agents simultaneously propose how many coins they keep.\n" + "4. If the total sum of proposals is less than or equal to 10, both agents receive their proposals.\n" + "5. If the total sum of proposals exceeds 10, the coins are allocated proportionally.\n" + "6. Your points for the round = (coins you receive) x (your per-coin value for that round). \n" + "7. Points are accumulated across rounds.\n" + "Your goal: {goal}\n" + ) + self.new_round_prompt = ( + "A New Round Begins\n" + "Your per-coin value is {value}. You don't know {other_agent}'s value yet.\n" + ) + self.last_round_prompt = "In the previous round, {other_agent} had a {last_value_str_coagent} value and proposed {last_split_coagent} coins.\n" + if self.proposal_start_end_format: + self.send_split_prompt = ( + "Submit your proposal\n" + "Respond with <> x <> where x is an integer in [0, 10]." + ) + else: + self.send_split_prompt = ( + "Submit your proposal\n" + "Respond with x where x is an integer in [0, 10]." + ) + self.wait_for_message_prompt = "Wait for {other_agent} to send a message..." + # self.wait_for_message_prompt = "" + self.last_message_prompt = "{other_agent} said: {last_message}" + if self.message_start_end_format: + self.send_message_prompt = f"Send your message now in <>...<> (<={self.num_message_chars} chars)." + else: + self.send_message_prompt = f"Send your message now in ... (<={self.num_message_chars} chars)." + + def get_message_regex(self, observation: TrustAndSplitSimpleObs) -> str: + if self.message_start_end_format: + return ( + rf"<>[\s\S]{{0,{self.num_message_chars}}}<>" + ) + else: + return rf"[\s\S]{{0,{self.num_message_chars}}}" + + def get_split_regex(self, observation: TrustAndSplitSimpleObs) -> str: + if self.proposal_start_end_format: + return r"<> ?(10|[0-9]) ?<>" + else: + return r" ?(10|[0-9]) ?" + + def get_split_action( + self, policy_output: str, observation: TrustAndSplitSimpleObs + ) -> Split: + import re as _re + + if self.proposal_start_end_format: + m = _re.search( + r"<> ?(10|[0-9]) ?<>", policy_output + ) + else: + m = _re.search( + r" ?(10|[0-9]) ?", policy_output + ) + coins_int = int(m.group(1)) if m else int(policy_output) + return Split(items_given_to_self={"coins": coins_int}) diff --git a/src_code_for_reproducibility/markov_games/negotiation/tas_simple_simulation.py b/src_code_for_reproducibility/markov_games/negotiation/tas_simple_simulation.py new file mode 100644 index 0000000000000000000000000000000000000000..3dbd0c43d73e3f7b18204b62e71d72b2df1d13e6 --- /dev/null +++ b/src_code_for_reproducibility/markov_games/negotiation/tas_simple_simulation.py @@ -0,0 +1,169 @@ +import copy +from collections import defaultdict +from dataclasses import dataclass +from typing import Any, Dict, List, Literal + +from numpy.random import default_rng + +from mllm.markov_games.negotiation.nego_simulation import ( + NegotiationObs, + NegotiationSimulation, + NegotiationState, + Split, + compute_tas_style_rewards, +) + +AgentId = str + + +@dataclass +class TrustAndSplitSimpleState(NegotiationState): + pass + + +@dataclass +class TrustAndSplitSimpleObs(NegotiationObs): + last_value_str_coagent: str | None + + +class TrustAndSplitSimpleSimulation(NegotiationSimulation): + def __init__( + self, + game_type: Literal["10-1-exclusive", "1-to-10"] = "1-to-10", + dist_type: Literal["uniform", "bimodal"] = "uniform", + beta_dist_alpha: float = 0.1, + beta_dist_beta: float = 0.1, + *args, + **kwargs, + ): + self.game_type = game_type + self.dist_type = dist_type + self.beta_dist_alpha = beta_dist_alpha + self.beta_dist_beta = beta_dist_beta + super().__init__(*args, **kwargs) + + def _sample_values(self) -> Dict[AgentId, dict]: + values = {} + while True: + if self.game_type == "10-1-exclusive": + v = int(self.rng.choice([1, 10])) + values[self.agent_ids[0]] = v + values[self.agent_ids[1]] = 10 if v == 1 else 1 + elif self.game_type == "1-to-10": + for aid in self.agent_ids: + if self.dist_type == "uniform": + values[aid] = int(self.rng.integers(1, 11)) + elif self.dist_type == "bimodal": + alpha, beta = self.beta_dist_alpha, self.beta_dist_beta + values[aid] = int(round(self.rng.beta(alpha, beta) * 9) + 1) + if len(set(values.values())) != 1: + break + return values + + def _sample_quantities(self) -> Dict[str, int]: + return {"coins": 10} + + def set_new_round_of_variant(self): + self.state.quantities = self._sample_quantities() + self.state.values = self._sample_values() + self.state.split_phase = False + + def get_info_of_variant( + self, state: NegotiationState, actions: Dict[AgentId, Any] + ) -> Dict[str, Any]: + return { + "quantities": copy.deepcopy(state.quantities), + "values": copy.deepcopy(state.values), + # "previous_values": copy.deepcopy(state.previous_values), + "splits": copy.deepcopy(state.splits), + } + + def get_rewards(self, splits: Dict[AgentId, Split]) -> Dict[AgentId, float]: + return compute_tas_style_rewards( + self.agent_ids, self.state.values, splits, self.state.quantities + ) + + def get_obs(self): + return {agent_id: self.get_obs_agent(agent_id) for agent_id in self.agent_ids} + + def get_obs_agent(self, agent_id): + other_id = self._other(agent_id) + last_value_coagent = ( + None + if self.state.previous_values is None + else self.state.previous_values.get(other_id) + ) + last_points_coagent = ( + None + if self.state.previous_points is None + else round(self.state.previous_points.get(other_id), 1) + ) + last_value_agent = ( + None + if self.state.previous_values is None + else self.state.previous_values.get(agent_id) + ) + last_points_agent = ( + None + if self.state.previous_points is None + else round(self.state.previous_points.get(agent_id), 1) + ) + last_split_coagent = None + last_split_agent = None + if self.state.previous_splits is not None: + last_split_coagent = self.state.previous_splits[ + other_id + ].items_given_to_self["coins"] + last_split_agent = self.state.previous_splits[agent_id].items_given_to_self[ + "coins" + ] + if last_value_agent is None or last_value_coagent is None: + last_value_str_coagent = None + else: + if last_value_coagent > last_value_agent: + last_value_str_coagent = "higher" + elif last_value_coagent < last_value_agent: + last_value_str_coagent = "lower" + else: + raise ValueError("Should not be equal values") + + obs = TrustAndSplitSimpleObs( + round_nb=self.state.round_nb, + last_message=self.state.last_message, + quota_messages_per_agent_per_round=self.quota_messages_per_agent_per_round, + current_agent=self.state.current_agent, + other_agent=self.agent_id_to_name[other_id], + quantities=self.state.quantities, + item_types=self.item_types, + value=self.state.values[agent_id], + split_phase=self.state.split_phase, + last_split_agent=last_split_agent, + last_value_agent=last_value_agent, + last_points_agent=last_points_agent, + last_split_coagent=last_split_coagent, + last_value_coagent=last_value_coagent, + last_points_coagent=last_points_coagent, + last_quantities=self.state.previous_quantities, + last_value_str_coagent=last_value_str_coagent, + ) + return obs + + def reset(self): + start_agent = self.agent_ids[self._starting_agent_index] + quantities = self._sample_quantities() + values = self._sample_values() + self.state = TrustAndSplitSimpleState( + round_nb=0, + last_message="", + current_agent=start_agent, + quantities=quantities, + values=values, + previous_values=None, + splits={aid: None for aid in self.agent_ids}, + nb_messages_sent={aid: 0 for aid in self.agent_ids}, + split_phase=False, + previous_splits=None, + previous_points=None, + previous_quantities=None, + ) + return self.get_obs()