chess-value-net

A small MLP value network for chess, trained on human games and self-play, and continually updated as it plays. It scores a position from the side-to-move's perspective in [-1, 1]. Used inside a negamax + alpha-beta search engine.

• Network version: 2
• Games trained on: 1,503
• Input: 773-dim board features (see chessai/encoding.py)

Not a standalone transformers model -- load with the project's ValueNet.load().