yashbhutwala / pong-ai

Deep Q-Learning Networks vs. Policy Gradient Learning in OpenAI Gym's Pong Environment

Pong AI

This is the final project for CSCI 379 (Intro to AI & CogSci, Spring 2017).

Team Reveries:

• Yash Bhutwala
• Matt McNally
• Kenny Rader
• John Simmons

Problem Statement

Given a set a pixels from a game of Pong, a mechanism for measuring wins and losses, and a hard-coded opponent agent, could we create an agent that could potentially beat a human opponent, and what would be the best way to do that?

Our approach tests Deep Q-Learning Networks (DQN) against Policy Gradient (PG) learning in order to see which algorithm and architecture learns the best.

How to run the programs:

Deep Q-Learning (DQN)

Our DQN agent can be ran from the ./dqn directory. You can run it by using the command:

python main.py --env_name=Pong-v0 --is_train=True --display=True

This will run the program on the Pong environment with Training Mode and Rendering turned on.

Policy Gradient (PG)

Our PG agent can be ran from the ./pg directory. You can run it by using the command:

python3 yashPong.py

This will run the agent in Training Mode and Rendering turned on by default, though this can be changed in code.

Acknowledgements:

The code for our DQN approach is modified from existing code from devsisters. The original repository can be found here

Likewise, the code for our PG approach is modified from existing code from Dr. Andrej Karpathy. The original code can be found here