boardgame-rl

Various board games and reinforcement learning agent trained via Q-learning. The agent can be trained for various episodes and played against. The code is meant to be flexible enough so that the agent can learn to play any board game that implements the Game abstract base class.

Motivation

This project was completed in order to learn the basics of implementing Q-learning with a simple look up table for Q-values. Inspiration was taken from Reinforcement Learning: An Introduction

Built with

numpy

Run instructions

python main.py -game tictactoe

Test instructions

python -m unittest discover -v

Supported Board Games

Tic Tac Toe
Connect Four
Chomp

Tic Tac Toe

Connect Four

Alternate state representation of full board vs. 3 column state. Significantly reduces memory with little cost to agent performance.

Chomp

While implementing the game, I encountered some interesting problems. For Chomp, it is especially important to know which rectangles of chocolate were eaten at the same time.

For example, although

0 1 1
0 1 1
P * *

and

0 1 1
0 1 1
P * *

are the same, they could have been arrived at in different ways, not necessarily temporally but in terms of which blocks were eaten together.

0|1 1
0|1 1
  - -
P * *

and

0|1 1
- - -
0|1 1
  - -
P * *

Implement Your Own Custom Game

Just add your game class to the game module and make sure it is a subclass of the Game abstract base class. Then just pass in your game to the rl agent.

jaehunro/boardgame-rl