import { Rng } from './policy';
export type PlayerRef = number | string;
/**
 * The forward model the search drives. Implement these over your engine:
 * `applyAction` is the transition function, `evaluate` is the leaf heuristic
 * (higher = better for `perspective`), and `determinize` samples the hidden
 * state so the search isn't allowed to peek at information a player shouldn't
 * have. `orderActions` is an optional breadth control (good move ordering lets
 * `maxBranch` prune to the promising moves).
 */
export interface SearchModel<TState, TAction> {
    legalActions(state: TState): TAction[];
    applyAction(state: TState, action: TAction): TState;
    isTerminal(state: TState): boolean;
    currentPlayer(state: TState): PlayerRef;
    /** Position value from `perspective`'s point of view (higher is better). */
    evaluate(state: TState, perspective: PlayerRef): number;
    /** Sample a concrete world consistent with `perspective`'s knowledge. */
    determinize?(state: TState, perspective: PlayerRef, rng: Rng): TState;
    /** Reorder actions best-first; the search expands only the first `maxBranch`. */
    orderActions?(state: TState, actions: TAction[]): TAction[];
}
export interface SearchOptions {
    /** Plies to look ahead, including the root action itself (>= 1). */
    depth: number;
    perspective: PlayerRef;
    rng?: Rng;
    /** Worlds to sample for imperfect-information averaging (default 1). */
    determinizations?: number;
    /** Cap candidates expanded per node (default unlimited). */
    maxBranch?: number;
}
export interface ScoredAction<TAction> {
    readonly action: TAction;
    readonly value: number;
}
/**
 * Value every root action by simulating it forward. For each action we average
 * its minimax value across `determinizations` sampled worlds. Returns one entry
 * per (ordered, breadth-capped) root action; the caller turns these values into
 * a choice (e.g. skill-scaled softmax via {@link chooseActionIndex}).
 */
export declare function searchActionValues<TState, TAction>(rootState: TState, model: SearchModel<TState, TAction>, options: SearchOptions): ScoredAction<TAction>[];