schroederdewitt/rec-maddpg

Decentralized deep multi-agent reinforcement learning in physical environments.

Recurrent Multi-Agent Deep Deterministic Policy Gradient (Rec-MADDPG)

This is the code for implementing the Rec-MADDPG algorithm presented in my MSc Dissertation "Communication and Cooperation in Decentralized Multi-AgentReinforcement Learning". It is configured to be run in conjunction with environments from the Multi-Agent Particle Environments (MPE).

Installation

• Install requirements with pip install -r requirements.txt
• Install my adaption of the Multi-Agent Particle environment: https://github.com/nicoring/multiagent-particle-envs

Usage

• cd into the maddpg directory
• Run the code with python trainer.py --scenario SCENARIO_NAME
• python trainer.py --help gives a description of all the available command line options.
• The code stores the success rates and returns as well as the policies of the agents.

Code Structure

This repository contains the code for MADDPG and Rec-MADDPG in the maddpg directory, which contains the following files:

• trainer.py which is the main file to run and contains the training logic.
• agent.py contains the code for MADDPG and Rec-MADDPG agents.
• models.py contains the code for the actor and policy networks.
• memory.py contains the replay buffer code.
• distribitions.py contains the code for the KL-divergence between Gumbel-Softmax distributions
• Additionally, there are multiple run scripts.