public void Connect(Neuron from, Neuron to, double weight)
{
Synapse s = new Synapse();
s.Axon = from;
s.Dentrite = to;
s.Weight = weight;
synapses.Add(s);
}
public void Evaluate()
{
foreach (Neuron outNeuron in outputNeurons)
{
double value = 0.0;
foreach (Neuron inNeuron in inputNeurons)
{
Synapse s = GetSynapse(inNeuron, outNeuron);
value += s.Weight * inNeuron.Value;
}
outNeuron.Value = value;
}
}
public void Train(TrainingData[] data)
{
// set weights random
Random r = new Random(0);
foreach (Synapse s in synapses)
{
s.Weight = r.NextDouble() * 2 - 1.0; // value between -1.0 and 1.0
}
// minimize the error
double learningRate = 0.01;
double precision = 0.01;
double lastError;
double currentError = double.MaxValue;
do
{
lastError = currentError;
currentError = 0.0;
foreach (Synapse s in synapses)
{
s.dW = 0.0;
}
// for each training point...
foreach (TrainingData d in data)
{
// for each output neuron...
for (int j = 0; j < outputNeurons.Count; j++)
{
// calculate Yj from inputNeurons and weights
outputNeurons[j].Value = 0.0;
for (int i = 0; i < inputNeurons.Count; i++)
{
Synapse s = GetSynapse(inputNeurons[i], outputNeurons[j]);
outputNeurons[j].Value += s.Weight * d.X[i];
}
// determine error contribution from this output node and training point
currentError += Math.Pow(d.T[j] - outputNeurons[j].Value, 2.0);
// determine weight gradient for each synapse
for (int i = 0; i < inputNeurons.Count; i++)
{
Synapse s = GetSynapse(inputNeurons[i], outputNeurons[j]);
s.dW += (d.T[j] - outputNeurons[j].Value) * d.X[i];
}
}
}
// update error for number of training points
currentError /= data.Length;
// adjust weights
foreach (Synapse s in synapses)
{
s.Weight += learningRate * s.dW;
}
}while (Math.Abs(currentError - lastError) > precision);
}
