Atari Pong with Deep Reinforcement Learning Project

This repository contains code for the Atari Pong with Deep Reinforcement Learning using a CNN-based Deep Q-Network project. Below is a quick overview of the repo layout:

• config: This folder contains config files which specify the training parameters and output directories for model training.
• core: This folder contains the core components of the deep reinforcement models. It contains a general DQN class and torch models for the q_network, which sits on top of it. This directory contains the bulk of the code of this repo.
• env_req: This folder contains environment.yml and environment_cuda.yml files denoting the configuration of the virtual environment required to run the modules of this project.
• refs: This folder contains reference materials (i.e. academic papers) discussing the deep reinforcement learning techniques used in this project.
• results: This folder holds all the results saved to disk during training including eval score plots, model weights, episode recordings etc.
• utils: This folder contains additional modules that act as utility functions to the code contained in core/.
• run.py: This is the main driver script of the project and is used to train the RL models i.e. python run.py --config={config_name}.

This project leveraged materials from Stanford University's Reinforcement Learning (XCS234) course, with many modifications made to:

• Improve the overall code structure, clarity, simplicity, documentation, and format
• Address potential issues in the way the original repo was configured
• Improve runtime efficiency by switching to a torch based replay buffer
• Add support for a double DQN loss function
• Add support for prioritized experience replay (PER)
• Add support for periodically recording episodes of gameplay