An implementation of MADDPG

1. Introduction

This is a pytorch implementation of multi-agent deep deterministic policy gradient algorithm.

The experimental environment is a modified version of Waterworld based on MADRL.

2. Environment

The main features (different from MADRL) of the modified Waterworld environment are:

• evaders and poisons now bounce at the wall obeying physical rules
• sizes of the evaders, pursuers and poisons are now the same so that random actions will lead to average rewards around 0.
• need exactly n_coop agents to catch food.

3. Dependency

• pytorch
• visdom
• python==3.6.1 (recommend using the anaconda/miniconda)
• if you need to render the environments, opencv is required

4. Install

• Install MADRL.
• Replace the madrl_environments/pursuit directory with the one in this repo.
• python main.py

if scene rendering is enabled, recommend to install opencv through conda-forge.

5. Results

two agents, cooperation = 2

The two agents need to cooperate to achieve the food for reward 10.

the average

one agent, cooperation = 1

6. TODO

• reproduce the experiments in the paper with competitive environments.