/*
 * Copyright (C) 2015 Edward Raff
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see .
 */
package jsat.math.optimization.stochastic;

import java.util.Arrays;
import jsat.linear.IndexValue;
import jsat.linear.Vec;

/**
 * The Rprop algorithm provides adaptive learning rates using only first order
 * information. Rprop works best with the true gradient, and may not work well
 * when using stochastic gradients.
*
* See: Riedmiller, M., & Braun, H. (1993). A direct adaptive method for * faster backpropagation learning: the RPROP algorithm. IEEE International * Conference on Neural Networks, 1(3), 586–591. doi:10.1109/ICNN.1993.298623 * * @author Edward Raff */ public class Rprop implements GradientUpdater { private double eta_pos = 1.2; private double eta_neg = 0.5; private double eta_start = 0.1; private double eta_max = 50; private double eta_min = 1e-6; /** * holds what would be w^(t-1) */ private double[] prev_w; private double[] prev_grad; private double[] cur_eta; private double prev_grad_bias; private double cur_eta_bias; private double prev_bias; /** * Creates a new Rprop instance for gradient updating */ public Rprop() { } /** * Copy constructor * @param toCopy the object to copy */ public Rprop(Rprop toCopy) { if (toCopy.prev_grad != null) this.prev_grad = Arrays.copyOf(toCopy.prev_grad, toCopy.prev_grad.length); if (toCopy.cur_eta != null) this.cur_eta = Arrays.copyOf(toCopy.cur_eta, toCopy.cur_eta.length); if (toCopy.prev_w != null) this.prev_w = Arrays.copyOf(toCopy.prev_w, toCopy.prev_w.length); this.prev_grad_bias = toCopy.prev_grad_bias; this.cur_eta_bias = toCopy.cur_eta_bias; this.prev_bias = toCopy.prev_bias; } @Override public void update(Vec w, Vec grad, double eta) { update(w, grad, eta, 0, 0); } @Override public double update(Vec w, Vec grad, double eta, double bias, double biasGrad) { for(IndexValue iv : grad) { final int i = iv.getIndex(); final double g_i = iv.getValue(); final double g_prev = prev_grad[i]; final double w_i = w.get(i); prev_grad[i] = g_i; final double sign_g_i = Math.signum(g_i); final double sign_g_prev = Math.signum(g_prev); if(sign_g_i == 0 || sign_g_prev == 0) { double eta_i = cur_eta[i]; w.increment(i, -sign_g_i*eta_i*eta); } else if(sign_g_i == sign_g_prev) { double eta_i = cur_eta[i] = Math.min(cur_eta[i]*eta_pos, eta_max); w.increment(i, -sign_g_i*eta_i*eta); } else//not equal, sign change { double eta_i = cur_eta[i] = Math.max(cur_eta[i]*eta_neg, eta_min); w.increment(i, -prev_w[i]*eta_i*eta); prev_grad[i] = 0; } prev_w[i] = w_i; } //and again for the bias term if(bias != 0 && biasGrad != 0) { double toRet; final double g_i = biasGrad; final double g_prev = prev_grad_bias; final double w_i = bias; prev_grad_bias = g_i; final double sign_g_i = Math.signum(g_i); final double sign_g_prev = Math.signum(g_prev); if(sign_g_i == 0 || sign_g_prev == 0) { double eta_i = cur_eta_bias; toRet = sign_g_i*eta_i; } else if(sign_g_i == sign_g_prev) { double eta_i = cur_eta_bias = Math.min(cur_eta_bias*eta_pos, eta_max); toRet = sign_g_i*eta_i; } else//not equal, sign change { double eta_i = cur_eta_bias = Math.max(cur_eta_bias*eta_neg, eta_min); prev_grad_bias = 0; toRet = -prev_bias*eta_i; } prev_bias = w_i; return toRet*eta; } return 0; } @Override public void setup(int d) { prev_grad = new double[d]; cur_eta = new double[d]; Arrays.fill(cur_eta, eta_start); prev_w = new double[d]; cur_eta_bias = eta_start; prev_grad_bias = 0; prev_bias = 0; } @Override public Rprop clone() { return new Rprop(this); } }

Ads help maintain this website.