quasinewton.java

利用Java实现的神经网络工具箱

/*
 * $RCSfile: QuasiNewton.java,v $
 * $Revision: 1.15 $
 * $Date: 2005/05/08 02:16:29 $
 *
 * NeuralNetworkToolkit
 * Copyright (C) 2004 Universidade de Bras韑ia
 *
 * 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;

import neuralnetworktoolkit.debugger.mlp.MLPDebugger;
import neuralnetworktoolkit.math.NeuralMath;
import neuralnetworktoolkit.methods.gradientbased.GradientBasedMethod;
import neuralnetworktoolkit.neuralnetwork.INeuralNetwork;

/**
 * An implementation of Quasi-Newton training method.
 * <br><b>Experimental implementation. Not properly tested!</b>
 * 
 * @version 1.0 19 Jun 2004
 * 
 * @author <a href="mailto:[e-mail]">Hugo Iver V. Gon锟絘lves</a>
 * @author <a href="mailto:[e-mail]">Rodrigo C. M. Coimbra</a>
 */
public abstract class QuasiNewton extends GradientBasedMethod {

	/** Maximum number of iterations. */
	public static final int MAX_ITERATIONS = 10000000;
	
	protected MLPDebugger debugger;

	/**
	 * Calculates neuron deltas.
	 * 
	 * @param neuralNetwork Neural network to calculate deltas.
	 */
	public abstract void calculateNeuronDeltas(INeuralNetwork neuralNetwork);

	/**
	 * Calculates the product of transposed jacobian matrix by
	 * jacobian matrix, the product of transposed jacobian matrix
	 * by error vector and the sum of the squared error.
	 * 
	 * @param neuralNetwork   Neural network to calculate the products.
	 * @param instanceSet     Training instances set.
	 * @param outputs         Network expected outputs.
	 * @param subInstanceSize Sub-block size of <code>instanceSet</code>.
	 * 
	 * @return The two products.
	 */
	public JacobianTxJacobianAndJacobianTxErrorAndError calculateJacobianTxJacobianAndJacobianTxError(
			INeuralNetwork neuralNetwork,
			double[][] instanceSet,
			double[][] outputs,
			int subInstanceSize) {

		double[][] subInstanceSet =
			new double[subInstanceSize][instanceSet[0].length];
		double[][] finalInstances;
		int numberOfSynapses = neuralNetwork.numberOfSynapses();
		int index = 0;
		int subInstanceIndex = 0;
		int numberOfBlocks = 0;
		int blockId = 0;
		double[][] jtXj = new double[numberOfSynapses][numberOfSynapses];
		double[][] jtXerror = new double[numberOfSynapses][1];
		double[][] jacobian = new double[subInstanceSize][numberOfSynapses];
		double[][] jacobianT;
		double[][] subError = new double[subInstanceSize][1];
		double error = 0;

		
		//neuralMath.printMatrix(instanceSet);
		//neuralMath.printMatrix(outputs);
		
		numberOfBlocks = (int)instanceSet.length/subInstanceSize;
		System.out.println(">>"+numberOfBlocks);
		
		for (int i = 0; i < numberOfBlocks * subInstanceSize; i++) {
			subInstanceSet[subInstanceIndex] = instanceSet[i];
			subInstanceIndex++;
			if ((i % (subInstanceSize) == (subInstanceSize - 1)) && (i > 0)) {

				for (int m0 = 0; m0 < subInstanceSet.length; m0++) {
					double[] input = subInstanceSet[m0];
					neuralNetwork.inputLayerSetup(input);
					neuralNetwork.propagateInput();
					calculateNeuronDeltas(neuralNetwork);
					//debugger.printNetworkState();
					jacobian[m0] = calculateDerivativesVector(neuralNetwork);
					subError[m0][0] =
						outputs[blockId * subInstanceSize
							+ m0][0]
							- neuralNetwork.retrieveFinalResults()[0];
					error += subError[m0][0]*subError[m0][0];

				}

				//System.out.println("J ----------------------");
				//NeuralMath.printMatrix(jacobian);
				//System.out.println("------------------------");
				
				jtXj =
					NeuralMath.matrixSum(
						jtXj,
						NeuralMath.transposeProduct(jacobian));
				jacobianT = NeuralMath.calculateTransposed(jacobian);
				jtXerror =
					NeuralMath.matrixSum(
						jtXerror,
						NeuralMath.matrixProduct(jacobianT, subError));

				blockId++;
				subInstanceIndex = 0;
			}
			
			

		}

		if ((instanceSet.length - numberOfBlocks * subInstanceSize) > 0) {

			finalInstances =
				new double[instanceSet.length
					- numberOfBlocks * subInstanceSize][instanceSet[0].length];
			for (int i = 0;
				i < (instanceSet.length - numberOfBlocks * subInstanceSize);
				i++) {
				finalInstances[i] =
					instanceSet[numberOfBlocks * subInstanceSize + i];
			}

			jacobian = new double[finalInstances.length][numberOfSynapses];
			subError = new double[finalInstances.length][1];
			for (int m0 = 0; m0 < finalInstances.length; m0++) {
				double[] input = finalInstances[m0];

				neuralNetwork.inputLayerSetup(input);
				neuralNetwork.propagateInput();
				calculateNeuronDeltas(neuralNetwork);
				//debugger.printNetworkState();
				jacobian[m0] = calculateDerivativesVector(neuralNetwork);
				subError[m0][0] =
					outputs[numberOfBlocks * subInstanceSize
						+ m0][0]
						- neuralNetwork.retrieveFinalResults()[0];
				error += subError[m0][0]*subError[m0][0];

			}
			
			//System.out.println("J ----------------------");
			//NeuralMath.printMatrix(jacobian);
			//System.out.println("------------------------");
			
			jtXj =
				NeuralMath.matrixSum(
					jtXj,
					NeuralMath.transposeProduct(jacobian));
			
			jacobianT = NeuralMath.calculateTransposed(jacobian);
			jtXerror =
				NeuralMath.matrixSum(
					jtXerror,
					NeuralMath.matrixProduct(jacobianT, subError));

		}
		
		//System.out.println("Jt x J -----------------");
		//NeuralMath.printMatrix(jtXj);
		//System.out.println("------------------------");
		//System.out.println("Jt x Error -----------------");
		//NeuralMath.printMatrix(jtXerror);
		//System.out.println("------------------------");

		return new JacobianTxJacobianAndJacobianTxErrorAndError(jtXj, jtXerror, error);

	} //calculateJacobianTxJacobian()
	
		
	
} //QuasiNewton