Pong RL

A variant of Andrej Karpathy's Reinforcement Learning implementation of a Pong agent. Check out the original implementation.

Here, we implement PPO and PG with the option to use GAE. We are using PyTorch because we aren't as cracked as the man. We also make a custon pong-rl gym that's compatible with PyGame. This way we control either paddle with a human or AI, allowing us to do adverserial RL.

Prerequisites

• Python 3.9
• numpy 1.23.5
• gym[atari, accept-rom-license]
• torch
• wandb

Installation

conda create -n pong-rl python==3.9
pip install numpy==1.23.5 "gym[atari, accept-rom-license]" pygame torch wandb

Then install that foundryml repo with pip install -e . in the root foundryml directory.