export type MLPActivation = 'identity' | 'logistic' | 'tanh' | 'relu';
export type MLPSolver = 'adam' | 'sgd';
export type MLPLearningRateSchedule = 'constant' | 'invscaling' | 'adaptive';
export type MLPOutActivation = 'identity' | 'logistic' | 'softmax';
export type MLPLossKind = 'log' | 'binaryLog' | 'squared';
export interface MLPProps {
    hiddenLayerSizes?: number[];
    activation?: MLPActivation;
    solver?: MLPSolver;
    alpha?: number;
    batchSize?: number | 'auto';
    learningRate?: MLPLearningRateSchedule;
    learningRateInit?: number;
    powerT?: number;
    maxIter?: number;
    shuffle?: boolean;
    randomState?: number;
    tol?: number;
    momentum?: number;
    nesterovsMomentum?: boolean;
    earlyStopping?: boolean;
    validationFraction?: number;
    beta1?: number;
    beta2?: number;
    epsilon?: number;
    nIterNoChange?: number;
}
/** Resolved, validated hyper-parameters (exactly the getParams() shape). */
export interface MLPParams {
    hiddenLayerSizes: number[];
    activation: MLPActivation;
    solver: MLPSolver;
    alpha: number;
    batchSize: number | 'auto';
    learningRate: MLPLearningRateSchedule;
    learningRateInit: number;
    powerT: number;
    maxIter: number;
    shuffle: boolean;
    randomState: number | undefined;
    tol: number;
    momentum: number;
    nesterovsMomentum: boolean;
    earlyStopping: boolean;
    validationFraction: number;
    beta1: number;
    beta2: number;
    epsilon: number;
    nIterNoChange: number;
}
/** Resolve props to the canonical parameter set with sklearn's defaults. */
export declare function resolveMLPProps(props?: MLPProps): MLPParams;
export declare function validateMLPFitInput(X: number[][], y: ArrayLike<unknown>): void;
export interface MLPWeights {
    /** coefs[l][i][j]: weight from unit i of layer l to unit j of layer l+1. */
    coefs: number[][][];
    intercepts: number[][];
}
/**
 * sklearn `_init_coef`: uniform in [-bound, bound] with
 * bound = sqrt(factor / (fanIn + fanOut)), factor 2 for logistic, else 6.
 * Weights are drawn row-major, then the biases — one RNG stream for the
 * whole net, so a fixed randomState fully determines the init.
 */
export declare function initMLPWeights(layerUnits: number[], activation: MLPActivation, rand: () => number): MLPWeights;
/**
 * Forward pass returning the activations of every layer
 * (activations[0] === X, activations[last] = network output).
 */
export declare function forwardMLP(X: number[][], weights: MLPWeights, activation: MLPActivation, outActivation: MLPOutActivation): number[][][];
/** Network output only (last layer of `forwardMLP`). */
export declare function predictMLPOutput(X: number[][], weights: MLPWeights, activation: MLPActivation, outActivation: MLPOutActivation): number[][];
export interface MLPTrainConfig {
    outActivation: MLPOutActivation;
    lossKind: MLPLossKind;
    /**
     * Validation score for early stopping (higher is better): accuracy for
     * the classifier, R² for the regressor.
     */
    validationScore: (weights: MLPWeights, XVal: number[][], YVal: number[][]) => number;
}
export interface MLPTrainResult {
    weights: MLPWeights;
    lossCurve: number[];
    nIter: number;
    /** Best training loss (null when earlyStopping is on). */
    bestLoss: number | null;
    /** Per-epoch validation scores (empty when earlyStopping is off). */
    validationScores: number[];
    bestValidationScore: number | null;
}
/**
 * Minibatch training loop shared by both MLPs — a faithful port of sklearn's
 * `_fit_stochastic`, except that the early-stopping split is a plain
 * deterministic TAIL split (last validationFraction of the rows) instead of
 * sklearn's shuffled/stratified `train_test_split`.
 *
 * `Y` is the already-encoded target matrix (one-hot / 0-1 column / raw
 * column), one row per sample of `X`.
 */
export declare function trainMLP(X: number[][], Y: number[][], params: MLPParams, config: MLPTrainConfig): MLPTrainResult;