import { ClassifierBase } from '../base';
import { Params } from '../base/estimator';
import { MLPProps } from './mlpBase';
export interface MLPClassifierProps extends MLPProps {
}
/**
 * Multi-layer perceptron classifier trained with minibatch backpropagation
 * (sklearn `MLPClassifier` parity for the 'adam' and 'sgd' solvers; the
 * lbfgs solver is not implemented).
 *
 * Binary problems use a SINGLE logistic output unit with binary
 * cross-entropy, exactly like sklearn's LabelBinarizer encoding; multiclass
 * problems use a softmax output layer with categorical cross-entropy.
 * The L2 penalty `alpha` applies to weights only, never biases.
 *
 * Early stopping (`earlyStopping: true`) holds out the LAST
 * `validationFraction` of the rows as a plain deterministic split (sklearn
 * uses a shuffled, stratified split for classifiers — documented deviation)
 * and restores the weights with the best validation accuracy.
 */
export declare class MLPClassifier extends ClassifierBase {
    private params;
    private fitted;
    private classesList;
    private coefsState;
    private interceptsState;
    private lossCurveState;
    private nIterState;
    constructor(props?: MLPClassifierProps);
    getParams(): Params;
    /** Sorted class labels seen during fit. */
    get classes(): number[];
    /** Mean training loss recorded at the end of each epoch. */
    get lossCurve(): number[];
    /** Number of epochs actually run by the last fit. */
    get nIter(): number;
    fit(trainX: number[][], trainY: number[]): void;
    /**
     * Class-membership probabilities, columns ordered by the sorted class
     * labels (`classes`); each row sums to 1.
     */
    predictProba(testX: number[][]): number[][];
    predict(testX: number[][]): number[];
    private rawOutput;
}