Nostradamus - UCI Chess Engine
Author: David Carteau
The Nostradamus UCI chess engine uses language models inspired techniques to play chess:
unlike other engines, it does not rely on a traditional search tree to find the optimal sequence of moves instead, it uses a network architecture based on Transformers to predict the next best move given a specific position.
The engine is intended for educational purposes and is currently very weak (see below).
unlike other engines, it does not rely on a traditional search tree to find the optimal sequence of moves instead, it uses a network architecture based on Transformers to predict the next best move given a specific position.
The engine is intended for educational purposes and is currently very weak (see below).
Nostradamus 0.4 what's new?
How it works
Current version (v0.4):
takes a fenstring as input
transforms it into a sequence of 64 tokens (pieces on the board)
iteratively predicts a sequence of 3 new tokens (fr, to and up) which, when combined, give the move to play (i.e. from square, to square and promotion, if any)
transforms this sequence back into a UCI-formatted move (i.e. a7a8q)
Current model's architecture has:
256 dimensions/features per token
8 layers (Transformers blocks)
8 heads per layer
3,751,045 total parameters
Comments
Post a Comment