chikazemori / deepreinforcementlearning-applied-to-atari

Deep Reinforcement Learning Applied to Atari MsPacman

• Created a tabular reinforcement learning environment.
• Applied Q-learning in the environment created.
• Applied Soft Actor Critic to Atari MsPacman.
• Applied Deep Q-Network to Atari MsPacman.

Packages

Python Version: 3.8.3

Main Packages used:

• opencv-python==4.5.1.48
• torch==1.8.1
• torchvision==0.9.1

Requirements: <pip install -r requirements.txt>

Original Data

The original data was from OpenAI Gym.

Examples of Results

Best game by SAC:

Pacman gets stuck in a corner and stops moving.

Best game by DQN:

Pacman moves more smoothly and does not stop moving.

Conclusion

• DQN outperforms SAC in Atari MsPacman, which suggests that DQN performs better than SAC when the environment has complex rules and uncertainty.

Specifications

Jupyter Notebooks

All jupyter notebooks are available as ipynb.

• Task_1: Create a tabular reinforcement learning environment
• Task_2: Apply Q-learning to the environment created in Task_1
• Task_3: Apply SAC to Atari MsPacman
• Task_4: Apply DQN to Atari MsPacman

Folders

• Code: Contains all of the codes implemented

Files

• Report: The report for the whole project