wangle-zg/marl__matd3

Code for the RL method MATD3 described in the paper "Reducing Overestimation Bias in Multi-Agent Domains Using Double Centralized Critics"

• Use this code to replicate the results from the paper, for a more readable TF 2.x implementation check out tf2multiagentrl.

Implementation of Multi-Agent TD3

This is the implemetation of MATD3, presented in our paper Reducing Overestimation Bias in Multi-Agent Domains Using Double Centralized Critics. Multi-Agent TD3 is an algorithm for multi-agent reinforcement learning, that combines the improvements of TD3 with MADDPG.

The implementation here is closely based on maddpg from Ryan Lowe / OpenAI, to enable a fair comparision. The environments used are from multiagent-particle-envs from OpenAI.

Requirements

• python == 3.6
• TF == 1.12.0 any 1.x should work
• Gym == 0.10.5 this one is important
• Numpy >= 1.16.2

Example Useage

To start training on simple_crypto, with an MATD3 team of agents and an MADDPG adversary, use

python train.py --scenario simple_speaker_listener --good-policy matd3 --adv-policy maddpg

Reference

If you use our implementation, please also cite our paper with

@misc{ackermann2019reducing,
    title={Reducing Overestimation Bias in Multi-Agent Domains Using Double Centralized Critics},
    author={Johannes Ackermann and Volker Gabler and Takayuki Osa and Masashi Sugiyama},
    year={2019},
    eprint={1910.01465},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}