This repo contains a PyTorch implementation of a Trust Region Policy Optimization agent for an environment with a discrete action space.

1. Install conda for Python 2.7.

conda create --name trpo --file requirements/conda_requirements.txt
source activate trpo
pip install -r requirements/pip_requirements.txt

3. Install PyTorch from source at commit eff5b8b.

python run_trpo.py --env=GYM_ENV_ID

where GYM_ENV_ID is the environment ID of the gym environment you which to train on.

A game of Pong as played using the policy model learned from a TRPO agent

Plot of total reward per episode of Pong vs. episode number

OpenAI's Baseline implementation of parallel TRPO in TensorFlow

Ilya Kostrikov's implementation of TRPO for continuous control in PyTorch