A developer combines Theano and deep learning to build a chess AI called Deep Pink. The approach trains a 3-layer neural network (2048 units wide, ~10M parameters) on 100M chess games from FICS, learning a position evaluation function purely from game data — no chess rules encoded explicitly. The model uses triplet-based training with a soft inequality objective and is plugged into a negamax search with alpha-beta pruning. When benchmarked against Sunfish (a pure Python chess engine), Deep Pink wins roughly 1/3 of games. Key insight: evaluation function speed matters as much as accuracy, since faster functions allow deeper tree search. Potential improvements include better search algorithms, harder training examples, and GPU-accelerated inference.
Sort: