This is the initial release of the ChessSL supervised learning chess engine with Monte Carlo Tree Search (MCTS).
Implemented features
- ResNetCBAM model
- BatchMCTS algorithm
- A minimal implementation for the Universal Chess Interface (UCI) protocol
As of 23/Nov/2024, the engine has achieved a rating of 2154 in bullet mode and 2113 in blitz mode on Lichess, which is better than 90.3% and 94.7% players.
Issues
- Our neural network is very weak, compared to the Lc0's network.
- Our batch MCTS algorithm can be improved
- Lacking an "arena" mechanism to compare two version of the engine
- The UCI protocol is only good enough for the lichess API
This is probably the final release as I have invested too much time into this project. Nevertheless, I had a lot of fun and the final result is satisfactory.
Comment after a year
- After some thinking, I figured the whole program basically relies on the value network and MCTS. The policy network is trained very badly for some reason. If I have time I may take a look at it.