/// <summary>
/// Строит модель искусственного нейрона с заданными параметрами.
/// </summary>
/// <param name="synapticConnectionAmount">Количество синаптических связей (размерность вектора входных данных)ю</param>
/// <param name="threshold">Порог, увеличивающий выходное значение функции активации.</param>
/// <param name="learningSpeed">Коэффициент скорости обучения. Число в полуинтервале (0; double.MaxValue]. Рекомендуемое значение: 1.</param>
/// <returns>Модель искусственного нейрона с сигмоидальной функцией активацией (гиперболический тангенс)</returns>
public INeuron Build(int synapticConnectionAmount, double threshold, double learningSpeed)
{
var neuron = new SigmoidNeuron(synapticConnectionAmount, threshold, learningSpeed);
var dispersion = 1.0 / Math.Sqrt(synapticConnectionAmount);
for (var i = 0; i < neuron.SynapticWeights.Count; i++)
{
neuron.SynapticWeights[i] = MathHelper.GetRandomValue(0, dispersion);
}
return(neuron);
}
public void TestSigmoidalFunction()
{
NeuronInput inputA = new NeuronInput(0.67, 1.5);
NeuronInput inputB = new NeuronInput(0.5, 1.0);
NeuronInput inputC = new NeuronInput(0.8, 0.8);
NeuronInput[] inputs = new NeuronInput[3];
inputs[0] = inputA;
inputs[1] = inputB;
inputs[2] = inputC;
SigmoidNeuron neuron = new SigmoidNeuron(inputs, 3.0);
double output = neuron.SigmoidalFunction();
double expected = 0.99420529989699;
double delta = expected - output;
Assert.AreEqual(output, expected, delta);
}
public void Backprop()
{
FeedForward();
// go through the output layer
for (int i = 0; i < LayerNeuronCounts[NumLayers - 1]; ++i)
{
// for each neuron in the output layer calculate the difference between the output and the expected output
SigmoidNeuron tempNeuron = NetLayers[NumLayers - 1].Nodes[i];
tempNeuron.Delta = CostDerivertive(Outputs[i], ExpectedOutputs[i]) * SigmoidPrime(tempNeuron.z);
tempNeuron.NablaBias += tempNeuron.Delta;
// then go through each neuron connected to this neuron.
for (int j = 0; j < tempNeuron.Inputs.Length; ++j)
{
// set the change required in the weight.
tempNeuron.NablaWeights[j] += tempNeuron.Delta * tempNeuron.Inputs[j].Activation;
// and set it's required change property
tempNeuron.Inputs[j].Delta = (tempNeuron.Weights[j] * tempNeuron.Delta) * SigmoidPrime(tempNeuron.Inputs[j].z);
}
}
// go through the remaning layers and nodes in those layers
for (int i = NumLayers - 2; i > 0; --i)
{
for (int j = 0; j < LayerNeuronCounts[i]; ++j)
{
NetLayers[i].Nodes[j].NablaBias += NetLayers[i].Nodes[j].Delta;
// for each node connected to this node set the required change in weight and it's required change property.
for (int k = 0; k < NetLayers[i].Nodes[j].NablaWeights.Length; ++k)
{
NetLayers[i].Nodes[j].NablaWeights[k] += NetLayers[i].Nodes[j].Delta * NetLayers[i].Nodes[j].Inputs[k].Activation;
NetLayers[i].Nodes[j].Inputs[k].Delta = (NetLayers[i].Nodes[j].Weights[k] * NetLayers[i].Nodes[j].Delta) * SigmoidPrime(NetLayers[i].Nodes[j].Inputs[k].z);
}
}
}
}
public void FeedForward()
{
for (int i = 0; i < LayerNeuronCounts[0]; ++i)
{
NetLayers[0].Nodes[i].Activation = Inputs[i];
}
// Go through each layer that isn't the input layer.
for (int i = 1; i < NumLayers; ++i)
{
// Go through each node in this layer and calculate it's activation value.
for (int j = 0; j < LayerNeuronCounts[i]; ++j)
{
SigmoidNeuron tempNeuron = NetLayers[i].Nodes[j];
tempNeuron.Activation = Sigmoid(tempNeuron.GetZ());
}
}
// Update the outputs from the output layer.
for (int i = 0; i < LayerNeuronCounts[NumLayers - 1]; ++i)
{
Outputs[i] = NetLayers[NumLayers - 1].Nodes[i].Activation;
}
}
