benchmark of MROCS algorithm inspired by MADDPG and my previous work ( wheeler ), benchmark for new framework rewheeler ( comming soon ) on multi agent environment Tennis from Unity
Project Details
- state_size=24, action_size=2 as default UnityML - Tennis environment provides
- however for critic we used state_size=48, action_size=4, as MROCS approach ( update of MADDPG algo )
- 2 players environment, used shared actor + critic for controlling all arms
- Policy Gradients used, namely MROCS algorithm ( DDPG + MADDPG ideas/implementation and updates )
- How to install :
- install anaconda : https://www.anaconda.com/download/
- then follow :
conda install -y pytorch -c pytorch pip install unityagents- then replicate my results by running MROCS.ipynb
- and you can download my pretrained weights : https://github.com/rezer0dai/UnityMLEnvs/tree/master/models/mrocs/benchmark
- drop those checkpoints in main dir, and load agents cells, avoid retraining, and skip to cell which loads those weights and try EXPLORER vs TARGET network
- environment itself, download and copy to ./data/ folder in this project : https://s3-us-west-1.amazonaws.com/udacity-drlnd/P3/Tennis/Tennis_Linux.zip