Code accompanying the paper
The Emergence of Adversarial Communication in Multi-Agent Reinforcement Learning
Jan Blumenkamp, Amanda Prorok
(University of Cambridge)
arXiv: 2008.02616.
The five minute video presentation for CoRL 2020:
Supplementary video material:
Clone the repository, change directory into its root and run:
pip install -e .
This will install the package and all requirements. It will also set up the entry points we are referring to later in these instructions.
Generally, training is performed for the policies and for the interpreter. We first explain the three policy training steps (cooperative, self-interested, and re-adaptation) for all three experiments (coverage, split coverage and path planning) and then for the interpreters.
The policy training follows this scheme:
train_policy [experiment] -t [total time steps in millions]
continue_policy [cooperative checkpoint path] -t [total time steps] -e [experiment] -o self_interested
continue_policy [self-interested checkpoint path] -t [total time steps] -e [experiment] -o re_adapt
where experiment is one of {coverage, coverage_split, path_planning}, -t is the total number of time steps at which the experiment is to be terminated (note that this is not per call, but total time steps, so if a policy is trained with train_policy -t 20 and then continued with continue_policy -t 20 it will terminate immediately) and -o is a config option (one of {self_interested, re_adapt} as can be found in the alternative_config key in each of the config files in config).
When running each experiment, Ray will print the trial name to the terminal, which looks something like MultiPPO_coverage_f4dc4_00000. By default, Ray will create the directory ~/ray_results/MultiPPO in which the trial with the given name can be found with its checkpoint. continue_policy expects the path to one of such checkpoints, for example ~/ray_results/MultiPPO/MultiPPO_coverage_f4dc4_00000/checkpoint_440. The first continue_policy expects the checkpoint generated in the first train_policy call and the second continue_policy the checkpoint generated in the first continue_policy call. You should take note of each experiment's checkpoint path.
train_policy coverage -t 20
continue_policy [cooperative checkpoint path] -t 60 -e coverage -o self_interested
continue_policy [adversarial checkpoint path] -t 80 -e coverage -o re_adapt
train_policy coverage_split -t 3
continue_policy [cooperative checkpoint path] -t 20 -e coverage_split -o self_interested
continue_policy [adversarial checkpoint path] -t 30 -e coverage_split -o re_adapt
train_policy path_planning -t 20
continue_policy [cooperative checkpoint path] -t 60 -e path_planning -o self_interested
continue_policy [adversarial checkpoint path] -t 80 -e path_planning -o re_adapt
We provide three methods for evaluation:
evaluate_coop: Evaluate cooperative only performance while disabling self-interested agents with and without communication among cooperative agents.evaluate_adv: Evaluate cooperative and self-interested agents with and without communication between cooperative and self-interested agents (cooperative agents can always communicate to each other).evaluate_random: Run a random policy that visits random neighboring (preferably uncovered) cells.
The evaluation is run as
evaluate_{coop, adv} [checkpoint path] [result path] --trials 100
evaluate_random [result path] --trials 100
for 100 evaluation runs with different seeds. The resulting file is a Pandas dataframe containing the rewards for all agents at every time step. It can be processed and visualized by running evaluate_plot [pickled data path].
Additionally, a checkpoint can be rolled out and rendered for a randomly generated environment with evaluate_serve [checkpoint_path] --seed 0.
If you use any part of this code in your research, please cite our paper:
@article{blumenkamp2020adversarial,
title={The Emergence of Adversarial Communication in Multi-Agent Reinforcement Learning},
author={Blumenkamp, Jan and Prorok, Amanda},
journal={Conference on Robot Learning (CoRL)},
year={2020}
}