Are there any attempts at writing Stockfish MCTS?

Question

As of time of writing, there is one major traditional engine using Monte Carlo Tree Search for its search algorithm - Komodo. The MCTS version of Komodo started out weak but gained strength quickly, and Komodo MCTS recently defeated Komodo AB (i.e. Komodo using the more traditional alpha-beta pruning) in the TCEC Cup 3. That indicates that MCTS is at least comparable to standard AB search.

Are there any attempts at converting Stockfish to MCTS search?

Answer 1

This is difficult to answer definitively as there is obviously no way to say that there isn't someone somewhere working on an MCTS version of Stockfish. However, as of the time of writing, there exists no official attempt to create an MCTS version of Stockfish.

While some forks of Stockfish, such as ShashChess and SugaR have MCTS-like modes, they differ from what you are looking for:

• One of these modes applies a simple neural network to determine R in Late Move Reductions in Stockfish's alpha-beta search.
• The alternative collects MCTS-like information for each root move by collecting the evaluation of every node visited in a standard alpha-beta search. This is not really MCTS as it only collects additional statistics on the same alpha-beta search function.

Komodo MCTS, which does use its own handcrafted evaluation function and a full-fledged MCTS search, is only competitive with the standard version due to the fact that its implementation of MCTS applies alpha-beta rollouts: i.e. the alpha-beta search used in standard Komodo is used to determine the expected win % for leaf nodes, and also as the policy to determine which nodes should be preferentially expanded.

The exact methods by which Komodo MCTS converts its alpha-beta score to an expected win probability, as well as the details in how alpha-beta searches are used to determine which node MCTS should expand, are all proprietary. As such, it's unlikely that implementing MCTS in Stockfish is anything close to trivial.