This repo contains the source code to reproduce the results in the paper A Closer Look at Invalid Action Masking in Policy Gradient Algorithms.

Our experiments use docker containers to run and Weight and Biases (https://www.wandb.com/) to record the experiments, so the first step is to register a wandb account and get an API key, which we refer to as YOUR_WANDB_KEY

# build the docker container
docker build -t invalid_action_masking:latest -f sharedmemory.Dockerfile .
# build docker run commands. replace `{YOUR_WANDB_KEY}` with your own
WANDB_KEY={YOUR_WANDB_KEY} python docker.py > docker.sh
# run experiments (96 in total)
# if you have limited computational resources, consider not running all of them at a time.
# in addition, notice the commands have --cpuset-cpus="0", --cpuset-cpus="1" for different runs
# to make sure each container is only using one core. By default I assume your machine has 40 cores,
# but feel free to modify the `cores` variable in `docker.py`
bash docker.sh

Record your wandb username, which we will refer to as YOUR_WANDB_ENTITY

cd plots
WANDB_ENTITY={YOUR_WANDB_ENTITY} python episode_reward.py
WANDB_ENTITY={YOUR_WANDB_ENTITY} python approx_kl.py

These command should reproduce the PDFs in plots that are attached to the repo.

Although it would be possible, it would require a significant amount of effort to properly log metrics and redo the plotting, so at this time we would not have intructions to do reproduction without WANDB. Note that it is possible to use wandb locally by following https://docs.wandb.com/self-hosted/local.

We have tested these scripts to reproduce but it is possible that there is a bug and maybe we are assuming something specific regarding the environment. If you couldn't reproduce our results, please file an issue and we will address it as soon as the double-blind review is over.