/// <summary>
/// Считаем дельта-отклонение, пересчитываем веса нейронов
/// </summary>
/// <param name="answer"></param>
private double CalculateDeviation(NeuralNetworkDomain neuralNet, double answer)
{
//Считаем значение средней квадратичной ошибки
var outNeuron = neuralNet.Neurons.Single(n => n.NeuronType == NeuronType.OutputNeuronType);
neuralNet.MSEcounter++;
neuralNet.ErrorMSE += Math.Pow((answer - outNeuron.DataOut), 2);
//Считаем дельта-отклонение для слоя вывода
outNeuron.DeltaDeviation = DeltaOutput(answer, outNeuron.DataOut);
CalculateHiddenLayerDeltaDeviation(neuralNet, NeuronType.SecondLayerHiddenNeuronType);
CalculateHiddenLayerDeltaDeviation(neuralNet, NeuronType.FirstLayerHiddenNeuronType);
//Считаем изменение веса синапсов слоя ввода
foreach (var neuron in neuralNet.Neurons.Where(n => n.NeuronType == NeuronType.InputNeuronType))
{
//градиент для градиентного спуска
double grad = 0;
foreach (var synapse in neuron.OutputSynapses)
{
var targetDeviation = synapse.OutputNeuron.DeltaDeviation;
grad = targetDeviation * neuron.DataOut;
synapse.DeltaWeight = neuralNet.LearningSpeed * grad + neuralNet.Moment * synapse.DeltaWeight;
synapse.Weight += synapse.DeltaWeight;
}
}
return(neuralNet.ErrorMSE / neuralNet.MSEcounter);
}
public CalibrationResultDomain Calibrate(NeuralNetworkDomain neuralNet, double answer)
{
Calculate(neuralNet);
var error = CalculateDeviation(neuralNet, answer);
return(new CalibrationResultDomain
{
Error = error,
Result = neuralNet.Neurons.Single(n => n.NeuronType == NeuronType.OutputNeuronType).DataOut
});
}
/// <summary>
/// Считаем выходное значение проходом вперёд
/// </summary>
/// <param name="neuralNet"></param>
private void Calculate(NeuralNetworkDomain neuralNet)
{
//Обсчитываем скрытый слой и слои вывода
var neurons = neuralNet.Neurons
.Where(n => n.NeuronType != NeuronType.InputNeuronType)
.OrderBy(n => n.NeuronType);
foreach (var neuron in neurons)
{
neuron.DataIn = 0;
foreach (var synapse in neuron.InputSynapses)
{
neuron.DataIn += synapse.InputNeuron.DataOut * synapse.Weight;
}
neuron.DataOut = Sigmoid(neuron.DataIn);
}
}
/// <summary>
/// Считаем дельта-отклонение для указанного слоя скрытых нейронов и изменяем веса синапсов
/// </summary>
/// <param name="hiddenLayerType"></param>
private void CalculateHiddenLayerDeltaDeviation(NeuralNetworkDomain neuralNet, NeuronType hiddenLayerType)
{
foreach (var neuron in neuralNet.Neurons.Where(n => n.NeuronType == hiddenLayerType))
{
//сумма произведения всех исходящих весов и дельта-отклонения нейрона, с которым связан синапс
double sum = 0;
//градиент для градиентного спуска
double grad = 0;
foreach (var synapse in neuron.OutputSynapses)
{
var targetDeviation = synapse.OutputNeuron.DeltaDeviation;
sum += targetDeviation * synapse.Weight;
grad = targetDeviation * neuron.DataOut;
synapse.DeltaWeight = neuralNet.LearningSpeed * grad + neuralNet.Moment * synapse.DeltaWeight;
synapse.Weight += synapse.DeltaWeight;
}
neuron.DeltaDeviation = SigmoidDiff(neuron.DataOut) * sum;
}
}
public double Reckon(NeuralNetworkDomain neuralNet)
{
Calculate(neuralNet);
return(neuralNet.Neurons.Single(n => n.NeuronType == NeuronType.OutputNeuronType).DataOut);
}
