An AI that plays chess and is hopefully better than me
GIF demonstrating human play, AI is not complete yet (see below)
Written in Java 17 using the Maven project management tool. Currently, the chess program itself is complete and I have written up a neural network-based agent with stochastic training. Currently it doesn’t train very well and doesn’t implement techniques like decaying learning rate, but the functionality is there and I’ll be working on improving the strategy over time.
This project implements a chess game that can be played by two humans, two AIs, or humans vs. AI. Furthermore, the game can also be played in a ‘headless’ state by using the command line to enter in moves if there is a human in the game (otherwise the game will run to completion).
There are currently 3 different AIs created for this game:
RandomAgentworks by simply picking a random, legal move and it performs as poorly as you’d expect.
GreedyAgentpicks the move that gets it the most points by capturing a piece, assuming it’s legal and randomly chooses between ties. Although usually beating
RandomAgentin AI vs. AI games, it only looks one move into the future and doesn’t understand the concept of a protected piece, so will happily sack its queen for a pawn.
NeuralAgentfeeds in the board as a 3-dimensional input (height, width, one-hot encoded piece) through a Convolutional Neural Network (CNN) and outputs the scalar evaluation of the position. It does this evaluation for every possible position it can reach given a legal move and picks the move with the highest evaluation. However, the training for this model isn’t as effective as I would like it to be, so I will be working to improve this in the future.
To play the game with graphics enabled (not headless), use the
--graphics flag. To change the humans/AIs playing the game, currently you will have to change that in
git clone https://github.com/ben9583/chess-ai.git chess-ai && cd chess-ai mvn build