A Retrieval-based Chess Engine

Most modern chess-engines utilize complex value estimations, heuristics, tree-search algorithms and table-bases. We propose a simple retrieval-based chess-engine that utilizes transcripts from billions of public games. We use a pretrained embedder to embed chess positions into 64-dimensional vectors, create a large vector database of 100M positions, then train ChessBERT, a BERT-like model which learns a mapping from chess position to optimal move when given k retrieved similar positions in context.