An RL agent that learns to play Tic Tac Toe through value iteration. Inspired by Sutton and Barto's, "Reinforcement Learning - An Introduction". Can train an agent through self play, and then test that policy against a human. A human can directly play against an agent.

1. Install requirements pip install -r requirements.txt
2. Run the program, python main.py

This program was designed using OOP principles. A description of files is below.

• main.py - main program - run this from your terminal python main.py.
• board_state.py (class) - manages the state of a TicTacToe game.
• agent.py (class) - an autonomous agent that will make decisions based on a policy and learn through value iteration.
• human.py (class) - helper class so that a player may interact with the game.
• game.py (static class) - play the game, determines if a game is over, and prints board state to the console.

Right now, an agent learns through TD(0) learning and value iteration. In order to acheive an optimal policy the agent must have been in each state multiple times. Although the state space of tic tac toe is relitively small compared to other games, through my experience this approx 200k iterations to train an optimal agent. Other approaches such as q-learning, DNN's and even non RL approaches are encouraged, such as game theory or evolutionary approaches. Forks, & merge-requests are encouraged.

See LICENSE, if used for academic purposes, please cite this repository.