I investigated two deep reinforcement learning methods on the Atari version of Ms. Pacman–Deep Q-Learning (DQL) and Proximal Policy Optimization (PPO). Of the two models trained, DQL performed significantly better and was able to learn novel strategies, while the PPO model was unable to learn from the environment effectively, even after hours of training. We conclude that DQL performed better in this relatively simple environment with shorter training, and that further research is needed with greater computational power and time to evaluate the true effectiveness of PPO.

Github Repo: https://github.com/noazlee/cs320pacmanDRL
Youtube Video: https://www.youtube.com/watch?v=zFr55abPhQE