private void UpdateWeights(IEnumerator Inputs)
{
for (int i = 0; i < neuronLayers.Count; i++)
{
NeuronLayer processingLayer = neuronLayers[i];
double entry;
int j = 0;
while (Inputs.MoveNext())
{
entry = Inputs.Current.GetType() == typeof(Neuron) ? ((Neuron)Inputs.Current).activation : (double)Inputs.Current;
foreach (Neuron neuron in processingLayer)
{
neuron.weights[j] += _learningRate * neuron.delta * entry;
neuron.weights[j] += _momentum * neuron.prev_delta;
neuron.weights[j] -= _decay * neuron.weights[j];
neuron.bias += _learningRate * neuron.bias_delta * 1;
neuron.bias += _momentum * neuron.bias_prevdelta;
neuron.bias -= _decay * neuron.bias;
neuron.bias_prevdelta = _learningRate * neuron.bias_delta * 1;
}
j++;
}
Inputs = processingLayer.GetEnumerator();
}
}
/// <summary>
/// Backward - Propagation
/// </summary>
/// <param name="ExpectedResult">Expected output of Neural network</param>
private void BackProp(double[] ExpectedResult)
{
for (int i = 0; i < ExpectedResult.Length; i++)
{
Neuron currNeuron = neuronLayers.Last <NeuronLayer>()[i];
currNeuron.delta = (1 + currNeuron.activation) * (1 - currNeuron.activation) * (ExpectedResult[i] - currNeuron.activation);
}
// -2 because the last one processed above
for (int i = neuronLayers.Count - 2; i >= 0; i--)
{
NeuronLayer processingLayer = neuronLayers[i];
Neuron processingNeuron;
NeuronLayer backLayer = neuronLayers[i + 1];
for (int j = 0; j < processingLayer.Count; j++)
{
processingNeuron = processingLayer[j];
double total = 0;
foreach (Neuron backLayerNeuron in backLayer)
{
total += backLayerNeuron.delta * backLayerNeuron.weights[j];
}
processingNeuron.delta = (1 + processingNeuron.activation) * (1 - processingNeuron.activation) * total;
processingNeuron.bias_delta = (1 + processingNeuron.activation) * (1 - processingNeuron.activation) * 1;
}
}
}
public double ErorrCost(double[] ExpectedResult)
{
double errorCostValue = 0;
NeuronLayer outputLayer = neuronLayers.Last <NeuronLayer>();
for (int i = 0; i < ExpectedResult.Length; i++)
{
errorCostValue += Math.Pow(outputLayer[i].activation - ExpectedResult[i], 2);
}
return(errorCostValue);
}