This is a fast, root-parallel C++ implementation of the Expert Iteration algorithm proposed by Anthony et al. The implementation is generic in a choice of a Markov decision process as well as your choice of an apprentice.
The current version of the program presents a UCI interface to be used with Lichess as a chess bot. The apprentice used in the chess bot is a trivial one, which evaluates every state to 0 and performs no training. With the trivial apprentice, the Expert Iteration algorithm degenerates to plain Monte-Carlo Tree Search. Even with pure MCTS, the program has managed to beat a human player :) A better apprentice model is in progress.
You will need to add libtorch as a submodule under include/ and build it with CMake.1 You need to have SConstruct installed. The following incanation builds and runs the program on macOS and Linux.
git clone https://github.com/jaykru/mcts-chess
cd mcts-chess
scons && ./run.sh
The main executable presents a UCI chess interface. You can play manually with this, but it's recommended that you instead hook it up with lichess-bot. Some tweaking to lichess-bot is required to make it tolerant of long thinking time when using high iteration counts for the tree search.
Copyright Jay Kruer 2023. You probably won't want to use the code (yet) but contact me if you do. I haven't decided on a license yet.
Footnotes
-
I'd like to make this less manual and kludgy in the future, but this is a hobby project for now... ↩