/// <summary>
/// Get the error of the network for a single Input
/// </summary>
/// <param name="td"></param>
/// <returns></returns>
public double GetTotalError(TrainingData td)
{
double error = 0;
//get network Output
double[] output = CalculateOutput(td);
for (int k = 0; k < output.Length; k++)
{
double diff = output[k] - td.Output[k];
error += 0.5 * (diff * diff);
}
return(error / output.Length);
}
public void Train(TrainingData td)
{
//Calculate Output to fill layer outputs
CalculateOutput(td);
//Calculate deltas
for (int j = 0; j < Layers[LastLayer].Neurons.Length; j++)
{
Layers[LastLayer].Neurons[j].Delta = GetOutputNeuronDelta(Layers[LastLayer].Output[j], td.Output[j]);
}
for (int i = LastHiddenLayer; i >= 0; i--)
{
for (int j = 0; j < Layers[i].Neurons.Length; j++)
{
Layers[i].Neurons[j].Delta = GetHiddenNeuronDelta(j, Layers[i].Output[j], Layers[i + 1]);
}
}
//Adjust Weights
for (int i = LastLayer; i > 0; i--)
{
//Output and middle layers use the previous layer's Output
Layer l = Layers[i];
for (int j = 0; j < l.Neurons.Length; j++)
{
double lrateDt = LRate * l.Neurons[j].Delta;
Layers[i].Neurons[j].Bias += lrateDt;
for (int k = 0; k < l.Neurons[j].Weights.Length; k++)
{
Layers[i].Neurons[j].Weights[k] += lrateDt * Layers[i - 1].Output[k];
}
}
}
//The first layer uses the Input from the training data, so do that separately
//to avoid an excess if in the inner loops that would slow things down
for (int j = 0; j < Layers[0].Neurons.Length; j++)
{
double lrateDt = LRate * Layers[0].Neurons[j].Delta;
Layers[0].Neurons[j].Bias += lrateDt;
for (int k = 0; k < Layers[0].Neurons[j].Weights.Length; k++)
{
Layers[0].Neurons[j].Weights[k] += lrateDt * td.Input[k];
}
}
}
/// <summary>
/// Calculate Output of the network for a given Input
/// </summary>
/// <param name="td">Training data object to get Input from</param>
/// <returns>Output of the Neural Network</returns>
public double[] CalculateOutput(TrainingData td)
{
return(CalculateOutput(td.Input));
}
