Struct Ismcts 

pub struct Ismcts { /* private fields */ }

Single-observer information-set MCTS for hidden-information games.

Standard Mcts cheats at hidden information: it searches the one true state, so it sees opponents’ secret cards. ISMCTS instead searches from a player’s information set. Each simulation asks the game for a fresh determinization (Determinize::determinize) — a full world consistent with what the searcher can see, but with the hidden parts resampled — and runs one UCT iteration in that world. Averaging over many sampled worlds yields a move that is good on expectation without ever peeking.

A single tree is shared across determinizations. Because different worlds offer different legal moves, selection uses an availability count (how many simulations a move was legal for) in the UCB denominator, and only moves legal in the current world are considered. Values are backed up as a vector, one entry per seat, and each node selects to maximize the mover’s own entry (max^n), so this handles three- and four-player games, not just two-player zero-sum. Rollouts are uniform-random. All randomness comes from an internal seeded generator, so a run is reproducible.

Implementations

impl Ismcts

pub const fn new(iterations: u32, seed: u64) -> Self

Creates a search running iterations simulations per move, seeded from seed. Uses the standard UCT exploration constant (sqrt 2).

pub const fn with_exploration(self, exploration: f64) -> Self

Sets the UCT exploration constant (higher explores more).

pub fn evaluate<G>( &mut self, game: &G, state: &G::State, player: PlayerId, ) -> f64

where G: Determinize, G::State: Clone, G::Action: Clone,

Estimates player’s value of state as the mean rollout reward over sampled worlds, searching from player’s information set.

Trait Implementations

impl<G> Bot<G> for Ismcts

where G: Determinize, G::State: Clone, G::Action: Clone,

fn choose( &mut self, game: &G, state: &G::State, player: PlayerId, ) -> Option<G::Action>

Returns the action to play for player in state, or None if there is nothing to do (no legal actions, e.g. a terminal state).

impl<G> RankedBot<G> for Ismcts

where G: Determinize, G::State: Clone, G::Action: Clone,

fn rank( &mut self, game: &G, state: &G::State, player: PlayerId, ) -> Vec<(G::Action, f64)>

Ranks the searcher’s moves by visit share (the ISMCTS policy). Scores are the fraction of simulations spent on each move and sum to 1.