AlphaZero-Style Chess & Go Engine

介紹

AlphaZero-style chess trained from self-play on multi GPU nodes: native C++ search with batched TensorRT inference, Python training. 3,251 Elo at 100k searches against fixed-node Stockfish 13 (paper, arXiv:2609.37447); 3,338 with a follow-up attention network.

翻譯準備中

取自儲存庫 README

README

正在載入完整 README。 以下為已儲存的摘錄。

How strong can an AlphaZero-style chess system become under a very limited compute budget when the entire learning loop is engineered for efficiency? This project trained chess models from random initialization through self-play, without human games or pretrained chess weights, while sharing eight RTX 4070 SUPER GPUs among self-play, learning, and evaluation.