Learning to Communicate with Deep Multi-Agent Reinforcement Learning

This is a PyTorch implementation of the original Lua code release.

Overview

This codebase implements two approaches to learning discrete communication protocols for playing collaborative games: Reinforced Inter-Agent Learning (RIAL), in which agents learn a factorized deep Q-learning policy across game actions and messages, and Differentiable Inter-Agent Learning (DIAL), in which the message vectors are directly learned by backpropagating errors through a noisy communication channel during training, and discretized to binary vectors during test time. While RIAL and DIAL share the same individual network architecture, one would expect learning to be more efficient under DIAL, which directly backpropagates downstream errors during training, a fact that is verified in comparing the performance of the two approaches.

Execution

$ virtualenv .venv
$ source .venv/bin/activate
$ pip install -r requirements.txt
$ python main.py -c config/switch_3_dial.json

Results for switch game

This chart was generated by plotting an exponentially-weighted average across 20 trials for each curve.

More info

More generally, main.py takes multiple arguments:

Configuration

JSON configuration files passed to main.py should consist of the following key-value pairs:

Visualizing results

You can use analyze_results.py to graph results output by main.py. This script will plot the average results across all csv files per path specified after -r. Further, -a can take an alpha value to plot results as exponentially-weighted moving averages, and -l takes an optional list of labels corresponding to the paths.

$ python util/analyze_results -r <paths to results> -a <weight for EWMA>

Bibtex

@inproceedings{foerster2016learning,
    title={Learning to communicate with deep multi-agent reinforcement learning},
    author={Foerster, Jakob and Assael, Yannis M and de Freitas, Nando and Whiteson, Shimon},
    booktitle={Advances in Neural Information Processing Systems},
    pages={2137--2145},
    year={2016} 
}

License

Code licensed under the Apache License v2.0