b7leung / reinforcement-learning-connect-four

AI that learns to play connect four, using reinforcement learning methods (Q-Learning and Monte-Carlo policy iteration).

An AI that learns to play connect four, using reinforcement learning. For details, please refer to the report.

• connect_four.py: The code for the connect four game mechanics, including managing the board state, updating it with a move input, and checking for win conditions (ie, if the current board states has a vertical, horizontal, or diagonal four in a row, as well as if the board state is a tie)

• q_learning.py: The code for learning the Q function for the Connect Four AI. Contains the code for generating episodes with self-play. Also contains code for the main training loop (ie, implementing the MC-TS and Q-Learning algorithm) and testing the trained Connect Four AI.

• run.py: Used to actually start the RL experiments.

• Plots.py: A jupyter notebook used to plot the results.

Note: For the code to be runnable, the matplotlib, numpy, pandas, and tqdm python packages are required. Also, clone the github repository from https://github.com/codebox/connect4. The code from this repo is used solely for evaluating my trained RL Connect Four agent at test time, as a black box. It is not used for the training, or anything reinforcement learning related (all the things related to RL is implemented by me).