DRIML: Deep Reinforcement and InfoMax Learning

Code for Deep Reinforcement and InfoMax Learning (Neurips 2020)

Note: The repo is under construction right now, things will get added progressively to it as code is optimized/cleaned. For now, the parallelized Procgen code is released for rlpyt version of Feb.19 2020, but the goal is to make it compatible for the latest stable version of rlpyt.

Overview of algorithm

Prerequisites

• rlpyt (commit a0f1c3045eac1b12d6305b35200139f9ee2a63cd). Newer commits might throw errors. Goal: rewrite code in latest stable rlpyt version.
• torch. Latest stable release seems to work.

Instructions

• Clone the repo
• Run python main_procgen.py --lambda_LL "0" --lambda_GL "0" --lambda_LG "0" --lambda_GG "1" --experiment-name "test" --env-name "procgen-bigfish-v0.500" \ --n_step-return "7" --nce-batch-size "256" --horizon "10000" --algo "c51" --n-cpus "8" --n-gpus "1" --weight-save-interval "-1" --n_step-nce "-2" \ --frame_stack "3" --nce_loss "InfoNCE_action_loss" --log-interval-steps=1000 --mode "serial", for example. Trains DRIML-randk on 500 Bigfish levels.

To cite:

@inproceedings{mazoure2020deep,
  title={Deep Reinforcement and InfoMax Learning},
  author={Mazoure, Bogdan and Combes, Remi Tachet des and Doan, Thang and Bachman, Philip and Hjelm, R Devon},
  journal={Advances in Neural Information Processing Systems},
  year={2020}
}