alan-cooney / rllib-multi-agent

RLlib Multi Agent Example

RLlib Multi Agent Example

An example to demonstrate the speed & simplicity of RlLib. Each example trains a model, outputting results to .results and videos of the last iteration to .videos.

Setup

This repository includes a devcontianer for one-click setup. You can also install manually (pip install .).

Models

Cartpole (Single agent example)

Cartpole is included as a comparison against multi-agent examples.

python ./src/singleAgent.py

This shows the performance of RlLib with tuned hyperparameters, solving Cartpole. Key stats are as follows (for hitting a reward of 990):

• Iterations: 42 (4000 steps per iteration, so c. 50 episodes by the end)
• Total time: 50 seconds (6-core CPU with GPU disabled)

Pistonball (Multi-agent example)

python ./src/multiAgent.py

This takes around 100 iterations to get to a positive score (10 minutes on a 6-core CPU machine with 1 GPU).