midpoint commit of neural network rewrite. I realy need a disributed version control system!

Author: magicindian

src/aima/learning/data/rabbiteyes.csv

@@ -0,0 +1,70 @@
+15,21.66
+15,22.75
+15,22.3
+18,31.25
+28,44.79
+29,40.55
+37,50.25
+37,46.88
+44,52.33
+50,63.47
+50,61.13
+60,81
+61,73.09
+64,79.09
+65,79.51
+65,65.31
+72,71.9
+75,86.1
+75,94.6
+82,92.5
+85,105
+91,101.7
+91,102.9
+97,110
+98,104.3
+125,134.9
+142, 130.68
+147, 155.3
+147,152.2
+150,144.5
+159,142.15
+165,139.81
+183,153.22
+192,145.72
+195,161.1
+218,174.18
+218,173.03
+219,173.54
+224,178.86
+225,177.68
+227,173.73
+232,159.98
+232,161.29
+237,187.07
+246,176.13
+258,183.4
+276,186.26
+285,189.66
+300,186.09
+301,186.7
+305,186.8
+312,195.1
+317,216.41
+338,203.23
+347,188.38
+354,189.7
+357,195.31
+375,202.63
+394,224.82
+513,203.3
+535,209.7
+554,233.9
+591,234.7
+648,244.3
+660,231
+705,242.4
+723,230.77
+756,242.57
+768,232.12
+860,246.7

src/aima/learning/data/rabbiteyes.txt

@@ -0,0 +1,4 @@
+parameter 1 = age in days
+parameter 2 = weights of eyelenses in milligrams
+
+source = Neural Networks A Comprehensive Foundation - Simon Haykin

src/aima/learning/neural/BackPropLearning.java

@@ -0,0 +1,51 @@
+package aima.learning.neural;
+
+public class BackPropLearning implements NNTrainingScheme {
+	private final double learningRate;
+	private final double momentum;
+
+	private final Layer hiddenLayer;
+	private final Layer outputLayer;
+
+	public BackPropLearning(FeedForwardNeuralNetwork network,
+			double learningRate, double momentum) {
+
+		this.hiddenLayer = network.getHiddenLayer();
+		this.outputLayer = network.getOutputLayer();
+		this.learningRate = learningRate;
+		this.momentum = momentum;
+	}
+
+	public Vector processInput(FeedForwardNeuralNetwork network, Vector input) {
+
+		hiddenLayer.feedForward(input);
+		outputLayer.feedForward(hiddenLayer.getLastActivationValues());
+		return outputLayer.getLastActivationValues();
+	}
+
+	public void processError(FeedForwardNeuralNetwork network, Vector error) {
+		// TODO calculate total error somewhere
+		// create Sensitivity Matrices
+		outputLayer.sensitivityMatrixFromErrorMatrix(error);
+		hiddenLayer.sensitivityMatrixFromSucceedingLayer(network
+				.getOutputLayer());
+
+		// calculate weight Updates
+		outputLayer.calculateWeightUpdates(hiddenLayer
+				.getLastActivationValues(), learningRate, momentum);
+		hiddenLayer.calculateWeightUpdates(hiddenLayer.getLastInputValues(),
+				learningRate, momentum);
+
+		// calculate Bias Updates
+		outputLayer.calculateBiasUpdates(learningRate, momentum);
+		hiddenLayer.calculateBiasUpdates(learningRate, momentum);
+
+		// update weightsAndBiases
+		outputLayer.updateWeights();
+		outputLayer.updateBiases();
+
+		hiddenLayer.updateWeights();
+		hiddenLayer.updateBiases();
+
+	}
+}

src/aima/learning/neural/FeedForwardNeuralNetwork.java

@@ -0,0 +1,107 @@
+package aima.learning.neural;
+
+import aima.learning.framework.DataSet;
+import aima.util.Matrix;
+
+public class FeedForwardNeuralNetwork implements FunctionApproximator {
+
+	private final Layer hiddenLayer;
+	private final Layer outputLayer;
+
+	private final double learningRate, momentum;
+
+	private final NNTrainingScheme trainingScheme;
+
+	/*
+	 * constructor to be used for non testing code for now assume that config
+	 * contains learning rate, momentum parameter, and number of epochs. change
+	 * this later to accomodate varied learning schemes like early stopping
+	 */
+	public FeedForwardNeuralNetwork(NNConfig config) {
+
+		learningRate = config.getParameterAsDouble("learning_rate");
+		momentum = config.getParameterAsDouble("momentum");
+
+		int numberOfInputNeurons = config
+				.getParameterAsInteger("number_of_inputs");
+		int numberOfHiddenNeurons = config
+				.getParameterAsInteger("number_of_hidden_neurons");
+		int numberOfOutputNeurons = config
+				.getParameterAsInteger("number_of_outputs");
+
+		double lowerLimitForWeights = config
+				.getParameterAsDouble("lower_limit_weights");
+		double upperLimitForWeights = config
+				.getParameterAsDouble("upper_limit_weights");
+
+		hiddenLayer = new Layer(numberOfHiddenNeurons, numberOfInputNeurons,
+				lowerLimitForWeights, upperLimitForWeights,
+				new LogSigActivationFunction());
+
+		outputLayer = new Layer(numberOfOutputNeurons, numberOfHiddenNeurons,
+				lowerLimitForWeights, upperLimitForWeights,
+				new PureLinearActivationFunction());
+		trainingScheme = new BackPropLearning(this, learningRate, momentum);
+
+	}
+
+	/*
+	 * ONLY for testing to set up a network with known weights in future use to
+	 * deserialize networks after adding variables for pen weightupdate,
+	 * lastnput etc
+	 */
+	public FeedForwardNeuralNetwork(Matrix hiddenLayerWeights,
+			Vector hiddenLayerBias, Matrix outputLayerWeights,
+			Vector outputLayerBias, double learningRate, double momentum) {
+		this.learningRate = learningRate;
+		this.momentum = momentum;
+		hiddenLayer = new Layer(hiddenLayerWeights, hiddenLayerBias,
+				new LogSigActivationFunction());
+		outputLayer = new Layer(outputLayerWeights, outputLayerBias,
+				new PureLinearActivationFunction());
+		trainingScheme = new BackPropLearning(this, learningRate, momentum);
+	}
+
+	public void processError(Vector error) {
+
+		trainingScheme.processError(this, error);
+
+	}
+
+	public Vector processInput(Vector input) {
+		return trainingScheme.processInput(this, input);
+	}
+
+	public void testOn(DataSet ds) {
+		// TODO Auto-generated method stub
+	}
+
+	public Matrix getHiddenLayerWeights() {
+
+		return hiddenLayer.getWeightMatrix();
+	}
+
+	public Vector getHiddenLayerBias() {
+
+		return hiddenLayer.getBiasVector();
+	}
+
+	public Matrix getOutputLayerWeights() {
+
+		return outputLayer.getWeightMatrix();
+	}
+
+	public Vector getOutputLayerBias() {
+
+		return outputLayer.getBiasVector();
+	}
+
+	public Layer getHiddenLayer() {
+		return hiddenLayer;
+	}
+
+	public Layer getOutputLayer() {
+		return outputLayer;
+	}
+
+}

src/aima/learning/neural/FunctionApproximator.java

@@ -0,0 +1,14 @@
+package aima.learning.neural;
+
+public interface FunctionApproximator {
+	/*
+	 * accepts input pattern and processe it returning an output value
+	 */
+	Vector processInput(Vector input);
+
+	/*
+	 * accept an error and change the parameters to accomodate it
+	 */
+	void processError(Vector error);
+
+}