Summary
This repository contains implementations of DQN and C51 on the classic Atari game called Pacman. We used the environment provided by OpenAI Gym. The goal of this project was to reproduce results first introduced in this paper and become familiarized with the simple single network Deep RL algorithms. We borrowed and adapted the code from repositories mentioned in the References section.
Team:
- Ronak Mehta (DQN)
- Manmeet Singh (Categorical DQN)
Folder Structure:
assets: contains static artifacts, primarily plots of results from logs
colabs: notebooks with execution code and tensorboard visualzations
logs: a detailed list of metrics logged per frame to calculate performance
Models: saved models in h5 and pickle formats
training_videos: video captures of episodic training saved every n videos
Presentation.pdf
Results Summary Todo: Add images under assets folder
Code References:
- https://github.com/MichaelSpencerA/atari-breakout-dqn
- https://github.com/flyyufelix/C51-DDQN-Keras
- https://github.com/Jash-R/ReinforcmentLearning-For-MsPacman (original starter code which fails to converge and contains bugs)