Crate treant_gumbel 

Gumbel MuZero search: policy improvement by planning with Gumbel.

Implements the algorithm from Danihelka et al., “Policy improvement by planning with Gumbel” (ICLR 2022).

Key features:

• Gumbel-Top-k sampling at the root for action selection
• Sequential Halving for optimal simulation budget allocation
• PUCT selection for tree traversal below the root
• Improved policy output — a better training target than visit counts

Design

Gumbel search is fundamentally different from standard MCTS at the root level: instead of UCT/PUCT selection, it samples Gumbel noise, selects top-m actions, then uses Sequential Halving to allocate simulations. Below the root, standard PUCT guides tree traversal. This produces monotonically improving policies — more simulations always help.

The crate reuses treant::GameState so any game implemented for the core MCTS crate works with Gumbel search.

Example

use treant::GameState;
use treant_gumbel::{GumbelSearch, GumbelConfig, GumbelEvaluator};

let mut search = GumbelSearch::new(Eval, GumbelConfig::default());
let result = search.search(&MyGame, 100);
println!("Best move: {:?}", result.best_move);

Structs

GumbelConfig

Configuration for Gumbel search.

GumbelSearch

Gumbel MCTS search engine.

MoveStats

Per-move statistics from Gumbel search.

SearchResult

Result of a Gumbel search.

Traits

GumbelEvaluator

Evaluator providing policy logits and value estimates.