olmam.java

利用Java实现的神经网络工具箱

/*
 * $RCSfile: OLMAM.java,v $
 * $Revision: 1.11 $
 * $Date: 2005/05/05 02:03:48 $
 *
 * NeuralNetworkToolkit
 * Copyright (C) 2004 Universidade de Brasília
 *
 * This file is part of NeuralNetworkToolkit.
 *
 * NeuralNetworkToolkit 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 2 of the License, or
 * (at your option) any later version.
 *
 * NeuralNetworkToolkit 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 NeuralNetworkToolkit; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA - 02111-1307 - USA.
 */
package neuralnetworktoolkit.methods.gradientbased.quasinewton.lm;

import neuralnetworktoolkit.StatisticalResults;
import neuralnetworktoolkit.math.NeuralMath;
import neuralnetworktoolkit.methods.TrainingParameters;
import neuralnetworktoolkit.neuralnetwork.INeuralNetwork;

/**
 * 
 * 
 * @version $Revision: 1.11 $ - $Date: 2005/05/05 02:03:48 $
 * 
 * @author <a href="mailto:[e-mail]">Hugo Iver V. Gonçalves</a>
 * @author <a href="mailto:[e-mail]">Rodrigo C. M. Coimbra</a>
 */
public class OLMAM extends ModifiedLevenbergMarquardt {
	
	/* (non-Javadoc)
	 * @see neuralnetworktoolkit.methods.ITrainingMethod#train(neuralnetworktoolkit.INeuralNetwork, TrainingParameters)
	 */
	public StatisticalResults train(INeuralNetwork neuralNetwork, TrainingParameters parameters) {
		LevenbergMarquardtParameters param;
		double[] diagonal;

		param = (LevenbergMarquardtParameters) parameters;
		// TODO Erase comented instrumentation code.
		boolean firstUpdate = false;
		numberOfSynapses = neuralNetwork.numberOfSynapses();
		errorGoal = param.getError();
		lmKind = param.getLmKind();
		maximumNumberOfIterations = param.getMaxIterations();

		previousDeltaW = new double[neuralNetwork.numberOfSynapses()][1];

		inicio = System.currentTimeMillis();
		do {
			
			if (this.errorDiminished) {
				//System.out.println("> inicio");

				//System.out.println("> totalErrorEnergy");
				jtXjAndJtXerrorAndError = calculateJacobianTxJacobianAndJacobianTxError(
						neuralNetwork, param.getInputs(), param.getOutputs(),
						50);
				//System.out.println("> jtXjAndJtXerror");
				totalErrorEnergy = jtXjAndJtXerrorAndError.getError();
				jacobianTxJacobian = jtXjAndJtXerrorAndError.getJtXj();
				//System.out.println("> jacobianTxJacobian");
				jacobianTxError = jtXjAndJtXerrorAndError.getJtXerror();
				//System.out.println("> jacobianTxError");
			}
			/*
			 * System.out.println("---------------------------------------------------------------------->");
			 * System.out.println("--->JacobianT X Jacobian");
			 * neuralMath.printMatrix(jacobianTxJacobian);
			 */
			//System.out.println("--->JacobianT X Error");
			//neuralMath.printMatrix(jacobianTxError);
			if (lmKind == LevenbergMarquardt.COMPLETE_LM) {
				diagonal = NeuralMath.diagonalArray(jacobianTxJacobian);
				diagonal = NeuralMath.constantTimesArray(lambda, diagonal);
				hessian = NeuralMath.matrixSumDiagonal(jacobianTxJacobian,
						diagonal);

			} else {
				hessian = NeuralMath.matrixSumConstantDiagonal(lambda,
						jacobianTxJacobian);

			}
			//System.out.println("> hessian");
			//System.out.println("--->Hessian = Soma JacobianT X Jacobian +
			// diagonal");
			//NeuralMath.printMatrix(hessian);
			//System.runFinalization();
			inverseHessian = NeuralMath.inverseMatrix(hessian);
			//System.out.println("> inverse");
			//System.out.println("--->Hessian Inverse");
			//NeuralMath.printMatrix(hessian);

			minusLMDeltaW = NeuralMath.matrixProduct(inverseHessian,
					jacobianTxError);

			if ((numberOfIterations > 0) && (firstUpdate)) {
				IGG = NeuralMath.matrixProduct(NeuralMath
						.calculateTransposed(jacobianTxError), minusLMDeltaW)[0][0];
				IGF = NeuralMath.matrixProduct(NeuralMath
						.calculateTransposed(jacobianTxError), previousDeltaW)[0][0];
				IFF = NeuralMath.matrixProduct(NeuralMath
						.calculateTransposed(previousDeltaW), NeuralMath
						.matrixProduct(hessian, previousDeltaW))[0][0];

				epsilon = Math.sqrt(1 - ((IGF * IGF) / (IGG * IFF)));
				deltaP = Math.sqrt(IGG);
				deltaQt = -epsilon * deltaP * Math.sqrt(IGG);

				/*
				 * System.out.println("--->IGG");
				 * NeuralMath.printMatrix(NeuralMath.matrixProduct(NeuralMath.calculateTransposed(jacobianTxError),
				 * minusLMDeltaW)); System.out.println("--->IGF");
				 * NeuralMath.printMatrix(NeuralMath.matrixProduct(NeuralMath.calculateTransposed(jacobianTxError),
				 * previousDeltaW)); System.out.println("--->IFF");
				 * NeuralMath.printMatrix(NeuralMath.matrixProduct(NeuralMath.calculateTransposed(previousDeltaW),
				 * NeuralMath.matrixProduct(hessian, previousDeltaW)));
				 * 
				 * System.out.println("previousDeltaW"+" e deltaQt: " + deltaP);
				 * NeuralMath.printMatrix(previousDeltaW);
				 */

				lagrange2 = (.5) * Math.sqrt((IFF * IGG - IGF * IGF)
						/ (IGG * deltaP * deltaP - deltaQt * deltaQt));
				lagrange1 = (-2 * lagrange2 * deltaQt + IGF) / IGG;

				/*
				 * System.out.println("--->lagrange1: " + lagrange1);
				 * System.out.println("--->lagrange2: " + lagrange2);
				 */

				deltaW = NeuralMath.matrixSum(NeuralMath.constantTimesMatrix(
						-(lagrange1 / (2 * lagrange2)), minusLMDeltaW),
						NeuralMath.constantTimesMatrix(1 / (2 * lagrange2),
								previousDeltaW));

			} else {
				deltaW = NeuralMath.constantTimesMatrix(-1, minusLMDeltaW);

			}
			neuralNetwork.updateWeights(deltaW);

			instantaneousError = calculateTotalError(neuralNetwork, param
					.getInputs(), param.getOutputs());

			if (instantaneousError < totalErrorEnergy
					+ .1
					* (NeuralMath.matrixProduct(NeuralMath
							.calculateTransposed(jacobianTxError), deltaW)[0][0])) {
				//System.out.println("---> - O erro diminuiu "+ lambda);
				lambda = lambda / multiplier;
				previousDeltaW = (double[][]) deltaW.clone();
				firstUpdate = true;
				totalErrorEnergy = instantaneousError;
				this.errorDiminished = true;
			} else {
				//System.out.println("---> + O erro aumentou " + lambda);
				deltaW = NeuralMath.constantTimesMatrix(-1, deltaW);
				neuralNetwork.updateWeights(deltaW);
				lambda = lambda * multiplier;
				this.errorDiminished = false;
			}
			numberOfIterations++;
			if (numberOfIterations % /* (maximumNumberOfIterations/100) */1 == 0) {

				System.out
						.println("Numero de Iteracoes: " + numberOfIterations);
				System.out.println("Erro atual: " + totalErrorEnergy);

			}
			//System.out.println("---------------------------------------------------------------------->");
		} while (((totalErrorEnergy / param.getInputs().length) > errorGoal)
				&& (numberOfIterations < maximumNumberOfIterations)
				&& (goAhead = true));

		fim = System.currentTimeMillis();

		neuralNetwork.setError(totalErrorEnergy / (param.getInputs().length));

		results.setNumberIterations(numberOfIterations);
		numberOfIterations = 0;
		results.setError(totalErrorEnergy / param.getInputs().length);
		results.setTrainingTime((fim - inicio) / 1000);

		return results;
		
	} //train()
	
} //OLMAM