A collaboration of four started during my time at the Recurse Center.
We trained a neural network on 16.9M chess games (FICS database) to detect all legal moves a player can take at each given state.
The biggest challenge in building a chess AI is to model the number of states a player can end up in over the course of the game. That is, it is nearly impossible to fully grasp the decision tree in each given state. Our current plan is to adopt a supervised learning approach and model the decision making process for finding the best move by the Monte Carlo Tree Search.
This is still somewhat on-going, but for technical details, please see the Github repo above.