rogue-gym-agents-cog19

This repository contains codes for exprimens in COG2019 Rogue-Gym: A New Challenge for Generalization in Reinforcement Learning.

Watch agents

Here I show gifs of agents appeared in Section 6.4 in the paper.

Overfitted Agent

CNN + 10 training seeds

Generalized Agent

ResNet + L2 regularization + 40 training seeds

Setup

1. Install pipenv

pip3 install pipenv -U --user

2. (Optional) Modify Pipfile For example, if you want to use PyTorch 1.0.0, specify torch = '==1.0' in Pipfile.

3. Create the virtual env

pipenv --site-package --three
pipenv install

Usage

All hyper paremters are at env.py and you need to edit the file to change the experiment setting.

Train agents

• PPO with nature CNN

pipenv run python agents/ppo_naturecnn.py train

• PPO with impala CNN

pipenv run python agents/ppo_impalacnn.py train

• PPO with β-VAE feature extractor(β is hard coded in the file)

pipenv run python agents/vae_ppo.py train

Evaluate agents

pipenv run python agents/eval_seeds.py --logdir=$YOUR_LOD_DIR