AlphaZero‑style micro‑chess engine with beam‑search inference
It provides scripts to download elite chess games, prepare data, and run a lightweight AlphaZero‑inspired inference engine that combines a 15M‑parameter GPT‑2 policy with a depth‑3 beam search using the PUCT formula. The pipeline lets researchers evaluate the model against Stockfish and benchmark prediction accuracy. Targeted at ML researchers and hobbyists building small‑scale chess AI without massive resources. It stands out by offering an end‑to‑end workflow and HPC‑ready batch scripts, more complete than typical toy chess bots.
View on GitHub →LIF18/AlphaZero-Micro-Chess-TIF360