Maia Chess is a neural network chess engine designed to play like a human. Over the summer of 2022 I worked on research at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) intending to understand how Maia Chess works and improve its prediction of human moves. I primarily focused on writing the input and output moves to Portable Game Notation (PGN) files so they could be more easily visualized, but I also did experiments optimizing hyperparameters and implementing a GAN framework.
Maia Chess is trained on positions from players of a specified rating (in this case 1100) from the website Lichess. During testing it receives the previous eight moves as an input, and outputs the probability that each move is played. I converted that information to a PGN so the strengths and weaknesses of the model could be more easily understood.
The model was able to attain a first-choice accuracy of 50 percent. Subsequent experiments focused on optimizing the model's hyperparameters and attempting to implement a GAN framework.
