public void UpdateWeights()
{
for (int i = 0; i < Neurons.Length; i++)
{
// Update each input link's weight
for (int k = 0; k < Neurons[i].Inputs.Length; k++)
{
Synapse link = Neurons[i].Inputs[k];
// Add to the input link weight a percentage (the learning rate) of the
// derived error value (computed previously) multiplied by the input value
// We also add a percentage (momentum value) of the last change, to help overcome
// settling at local minima
double weightChange = Neurons[i].ErrorDelta * link.Input.Value;
switch (LearningMethod)
{
case LearningType.SGDMomentum:
link.LastChange = (LearningMomentum * link.LastChange) + (LearningRate * weightChange) - (2 * WeightDecay * link.Weight);
link.Weight += link.LastChange;
break;
case LearningType.Adagrad:
link.LastChange += weightChange * weightChange;
link.Weight += LearningRate * weightChange / (Math.Sqrt(link.LastChange) + 1e-4);
break;
default:
throw new NotSupportedException("Unsupported learning method");
}
}
}
}
public void ConnectForwardFully(Layer nextLayer, bool clearConnections)
{
if (nextLayer == null)
{
throw new ArgumentNullException("nextLayer");
}
// Adjust each neuron in the next layer's input connection lists to have room for our new connections
// Expand by the number of neurons in our layer plus one for the bias node
for (int i = 0; i < nextLayer.Neurons.Length; i++)
{
if (clearConnections)
{
nextLayer.Neurons[i].Inputs = new Synapse[Neurons.Length + 1];
}
else
{
Array.Resize(ref nextLayer.Neurons[i].Inputs, nextLayer.Neurons[i].Inputs.Length + Neurons.Length + 1);
}
}
for (int i = 0; i < Neurons.Length; i++)
{
// Resize source neuron's output connection list to have room for our new connections
int beginIndex = Neurons[i].Outputs.Length;
if (clearConnections)
{
Neurons[i].Outputs = new Synapse[nextLayer.Neurons.Length];
beginIndex = 0;
}
else
{
Array.Resize(ref Neurons[i].Outputs, Neurons[i].Outputs.Length + nextLayer.Neurons.Length);
}
//Synapse[] outputList = new Synapse[Neurons[i].Outputs.Length + nextLayer.Neurons.Length];
//Array.Copy(Neurons[i].Outputs, 0, outputList, 0, Neurons[i].Outputs.Length);
//Neurons[i].Outputs = outputList;
for (int k = 0; k < nextLayer.Neurons.Length; k++)
{
Synapse link = new Synapse(Neurons[i], nextLayer.Neurons[k], Neuron.GetRandomWeight());
Neurons[i].Outputs[beginIndex + k] = link;
nextLayer.Neurons[k].Inputs[nextLayer.Neurons[k].Inputs.Length - 1 - (Neurons.Length - i)] = link;
}
}
// Connect bias node
Array.Resize(ref Bias.Outputs, nextLayer.Neurons.Length);
for (int i = 0; i < nextLayer.Neurons.Length; i++)
{
nextLayer.Neurons[i].Inputs[nextLayer.Neurons[i].Inputs.Length - 1] = Bias.Outputs[i]
= new Synapse(Bias, nextLayer.Neurons[i], Neuron.GetRandomWeight());
}
}
public void ComputeError()
{
for (int i = 0; i < Neurons.Length; i++)
{
double errorSum = 0.0;
for (int k = 0; k < Neurons[i].Outputs.Length; k++)
{
Synapse link = Neurons[i].Outputs[k];
errorSum += link.Output.ErrorDelta * link.Weight;
}
Neurons[i].ErrorDelta = Neurons[i].ErrorDerivFunction(Neurons[i].Value) * errorSum;
}
}
public void ComputeError2(int[] outputIndexes)
{
for (int i = 0; i < Neurons.Length; i++)
{
double errorSum = 0.0;
for (int k = 0; k < outputIndexes.Length; k++)
{
int outputSyn = outputIndexes[k];
Synapse link = Neurons[i].Outputs[outputSyn];
errorSum += link.Output.ErrorDelta * link.Weight;
}
Neurons[i].ErrorDelta = errorSum;
}
}
