Configuration Options

Every configurable method in GameState and MCTS, in one place.

Conventions: "Required" means no default -- you must implement it. All other methods have defaults you can override. Types shown are return types.

GameState methods

Method	Type	Default	Purpose	Tutorial
`current_player()`	`Self::Player`	required	Identifies whose turn it is	2
`available_moves()`	`Self::MoveList`	required	Legal moves from this state; empty = terminal	2
`make_move(mov)`	`()`	required	Apply a move, mutating state in place	2
`max_children(visits)`	`usize`	`usize::MAX`	Max children to expand at this visit count; override for progressive widening	Progressive widening
`terminal_value()`	`Option<ProvenValue>`	`None`	Classify terminal state for MCTS-Solver (current player's perspective)	4
`terminal_score()`	`Option<i32>`	`None`	Exact minimax score of terminal state for Score-Bounded MCTS	4
`chance_outcomes()`	`Option<Vec<(Move, f64)>>`	`None`	Chance event outcomes with probabilities (must sum to 1.0)	5

MCTS associated types

Type	Bounds	Purpose	Tutorial
`State`	`GameState + Send + Sync + 'static`	The game state type	2
`Eval`	`Evaluator<Self> + Send + 'static`	State and move evaluator	2
`TreePolicy`	`TreePolicy<Self> + Send + 'static`	Child selection policy	2
`NodeData`	`Default + Sync + Send + 'static`	Custom per-node data (accessible via `SearchHandle`)	7
`TranspositionTable`	`TranspositionTable<Self> + Send + 'static`	Transposition table (use `()` for none)	7
`ExtraThreadData`	`'static`	Custom per-thread data (accessible via `SearchHandle`)	7

MCTS methods

Search control

Method	Type	Default	Purpose	Tutorial
`virtual_loss()`	`i64`	`0`	Bias subtracted during descent, added back during backprop; discourages thread collision	Parallel MCTS
`fpu_value()`	`f64`	`f64::INFINITY`	First Play Urgency: value for unvisited children; `INFINITY` = try all first; finite = let prior guide	6
`visits_before_expansion()`	`u64`	`1`	Visits to a leaf before creating a tree node	7
`node_limit()`	`usize`	`usize::MAX`	Maximum tree nodes; search halts when reached	7
`max_playout_length()`	`usize`	`1_000_000`	Safety cap: panics if a single playout exceeds this depth	7
`max_playout_depth()`	`usize`	`usize::MAX`	Quality knob: forces leaf evaluation when exceeded	7

Selection and output

Method	Type	Default	Purpose	Tutorial
`select_child_after_search(children)`	`&MoveInfo<Self>`	Most-visited (solver/bounds-aware)	Post-search child selection; override for custom behavior	7
`selection_temperature()`	`f64`	`0.0`	Temperature for `best_move()`; 0 = argmax, 1 = proportional to visits	6

Solver and bounds

Method	Type	Default	Purpose	Tutorial
`solver_enabled()`	`bool`	`false`	Enable MCTS-Solver: propagate proven win/loss/draw values	4
`score_bounded_enabled()`	`bool`	`false`	Enable Score-Bounded MCTS: track `[lower, upper]` minimax bounds	4

Stochastic games

Method	Type	Default	Purpose	Tutorial
`closed_loop_chance()`	`bool`	`false`	Give each chance outcome its own tree node (larger tree, better per-outcome stats)	5

Determinism and noise

Method	Type	Default	Purpose	Tutorial
`rng_seed()`	`Option<u64>`	`None`	Seed for deterministic search; each thread gets `seed + thread_id`	7
`dirichlet_noise()`	`Option<(f64, f64)>`	`None`	`(epsilon, alpha)` for root exploration noise: `(1-eps)prior + epsDir(alpha)`	6

Callbacks

Method	Type	Default	Purpose	Tutorial
`on_backpropagation(evaln, handle)`	`()`	no-op	Called for each node on the playout path during backpropagation	7
`cycle_behaviour()`	`CycleBehaviour<Self>`	`PanicWhenCycleDetected` (with TT) / `Ignore` (without TT)	Strategy for graph cycles from transposition tables	7

Interaction matrix

Some options interact with or require others.

Option	Requires	Interacts with
`solver_enabled()`	`GameState::terminal_value()` or `terminal_score()`	`select_child_after_search()` prefers proven wins
`score_bounded_enabled()`	`GameState::terminal_score()` or `terminal_value()`	Converged bounds set proven values when solver is also active
`closed_loop_chance()`	`GameState::chance_outcomes()`	Increases tree size; interacts with `node_limit()`
`dirichlet_noise()`	`TreePolicy::apply_dirichlet_noise()` (non-trivial impl)	Only effective with `AlphaGoPolicy` or custom policy with `f64` priors
`fpu_value()` (finite)	Nothing	Best combined with `AlphaGoPolicy` and neural network priors
`virtual_loss()`	Nothing	Only meaningful with parallel search (`playout_n_parallel`, etc.)
`selection_temperature()`	Nothing	Only affects `best_move()`; `principal_variation()` always uses argmax
`rng_seed()`	Nothing	Seeds policy thread data via `TreePolicy::seed_thread_data()`
`max_children()`	Moves returned in priority order	Works with `TreePolicy::compare_move_evaluations()` for ordering

`treant-gumbel` configuration

The treant-gumbel crate uses GumbelConfig instead of the MCTS trait. See Gumbel Search tutorial and Traits Reference.

Field	Type	Default	Purpose
`m_actions`	`usize`	`16`	Actions to consider after Gumbel-Top-k sampling; higher = broader search
`c_puct`	`f64`	`1.25`	PUCT exploration constant for below-root tree traversal
`max_depth`	`usize`	`200`	Maximum depth per simulation before forcing leaf evaluation
`value_scale`	`f64`	`50.0`	Scale factor mapping Q-values to logit scale; controls Q vs prior balance in the improved policy
`seed`	`u64`	`42`	RNG seed for reproducible Gumbel noise sampling

Tuning guidance:

m_actions: set to the typical number of legal moves or smaller. For games with 200+ moves, 16-32 is typical. For games with fewer than 10 moves, use the move count.
value_scale: higher values make the improved policy sharper (more exploitation). Lower values let the prior dominate. The paper uses 50.0 with Q-values in [-1, 1].
c_puct: same role as in AlphaGoPolicy. Typical range: 1.0-2.5.

GameState methods​

MCTS associated types​

MCTS methods​

Search control​

Selection and output​

Solver and bounds​

Stochastic games​

Determinism and noise​

Callbacks​

Interaction matrix​

treant-gumbel configuration​