/// <summary>
/// Updates own synapse weights
/// </summary>
/// <param name="ownDelta">The delta values of this network</param>
/// <param name="previousLayerValues">The values of the previous layer</param>
/// <param name="learnParameter">The resolution to learn at</param>
public void UpdateValues(double[] ownDelta, double[] previousLayerValues, double learnParameter)
{
//Calculate and add weight deltas
for (int i = 0; i < ownDelta.Length; i++)
{
for (int j = 0; j < previousLayerValues.Length; j++)
{
_neurons[i][j] += NeuronMath.GetWeightDelta(learnParameter, ownDelta[i], previousLayerValues[j]);
}
}
double biasDelta = 0;
//Calculate bias delta
for (int i = 0; i < NeuronCount; i++)
{
biasDelta += ownDelta[i] * _neurons[i][_neurons[i].Length - 1];
}
//Update connection weights to bias
for (int i = 0; i < NeuronCount; i++)
{
_neurons[i][_neurons[i].Length - 1] *= NeuronMath.GetWeightDelta(learnParameter, biasDelta, Bias);
}
}
/// <summary>
/// Returns the neuron delta values for an output layer
/// </summary>
/// <param name="isValues">The values the neurons in this layer have</param>
/// <param name="wantValues">The values the neurons in this layer should have</param>
/// <returns>The delta values for the neurons</returns>
public double[] GetDeltaOutputLayer(double[] isValues, double[] wantValues)
{
//Create result array
var result = new double[NeuronCount];
//Calculate delta values
for (int i = 0; i < result.Length; i++)
{
result[i] = NeuronMath.GetDeltaOutputLayer(Derivative(isValues[i]), isValues[i], wantValues[i]);
}
return(result);
}
/// <summary>
/// Returns the neuron delta values for an output layer
/// </summary>
/// <param name="isValues">The values the neurons in this layer have</param>
/// <param name="calcValues">The calc values of the next layer</param>
/// <returns>The delta values for the neurons</returns>
public double[] GetDeltaHiddenLayer(double[] isValues, double[] calcValues)
{
//Create result array
var result = new double[calcValues.Length];
//Calculate delta values
for (int i = 0; i < calcValues.Length; i++)
{
result[i] = NeuronMath.GetDeltaHiddenLayer(isValues[i], calcValues[i]);
}
return(result);
}
/// <summary>
/// Returns the derivative of the activation function in the layer
/// </summary>
/// <param name="value">The value to calculate</param>
/// <returns>The result</returns>
private double Derivative(double value)
{
//Check own type
switch (Type)
{
//Case Sigmoid:
case LayerType.Sigmoid:
//Return SigmoidDerivative of value
return(NeuronMath.SigmoidDerivative(value));
//This should not happen
default:
throw new Exception("This should not happen !");
}
}
/// <summary>
/// Claculates the neuron values for a forward pass in this layer
/// </summary>
/// <param name="input">The values of the previous layer</param>
/// <returns>The processed values</returns>
public double[] ForwardPass(double[] input)
{
//Check input size
if (input.Count() != _neurons[0].Length - 1)
{
throw new ArgumentException("Invalid input array size !");
}
//Add bias to input
Array.Resize(ref input, input.Length + 1);
input[input.Length - 1] = Bias;
var result = new double[NeuronCount];
//Calculate neuron values
for (var i = 0; i < NeuronCount; i++)
{
result[i] = ActivationFunction(NeuronMath.MatrixMultiply(_neurons[i], input));
}
return(result);
}
