search.mcts

search.mcts

AlphaZero-style MCTS with chance-node handling for backgammon.

Classes

Name	Description
Analysis	Full MCTS analysis of a position, for display/debugging.
Candidate	A single candidate move from MCTS analysis.
MCTS	Monte Carlo Tree Search with PUCT selection.
MCTSNode	A node in the MCTS tree. Only decision and terminal nodes are stored.

Analysis

search.mcts.Analysis(candidates, root_value, num_simulations)

Full MCTS analysis of a position, for display/debugging.

Candidate

search.mcts.Candidate(action, visits, visit_prob, prior, q_value)

A single candidate move from MCTS analysis.

MCTS

search.mcts.MCTS(
    network,
    num_simulations=100,
    c_puct=1.5,
    virtual_loss_count=1,
    dirichlet_alpha=0.0,
    noise_eps=0.25,
)

Monte Carlo Tree Search with PUCT selection.

Methods

Name	Description
analyze	Run MCTS and return a full Analysis with per-candidate stats.
search	Run MCTS and return (action -> visit proportion, visit entropy).
search_with_value	Run MCTS and return (action_probs, visit_entropy, root_q_value).

analyze

search.mcts.MCTS.analyze(state)

Run MCTS and return a full Analysis with per-candidate stats.

search

search.mcts.MCTS.search(state)

Run MCTS and return (action -> visit proportion, visit entropy).

search_with_value

search.mcts.MCTS.search_with_value(state)

Run MCTS and return (action_probs, visit_entropy, root_q_value).

MCTSNode

search.mcts.MCTSNode(state, parent=None, parent_action=None, prior=0.0)

A node in the MCTS tree. Only decision and terminal nodes are stored.

Uses lazy child creation: when a node is expanded (evaluated by the network), only the prior probabilities are stored. Child nodes are created on-demand when first selected by PUCT, avoiding expensive state cloning for actions that are never visited.

Functions

Name	Description
select_action	Sample an action from the visit distribution.

select_action

search.mcts.select_action(action_probs, temperature=1.0)

Sample an action from the visit distribution.