/// <summary>
/// Tworzy nową instancję klasy Perceptron.
/// </summary>
/// <param name="inputsNumber">Liczba inputów jaką perceptron może przyjąć.</param>
/// <param name="activationFunction">Funkcja aktywacyjna dla tego perceptrona.</param>
/// <param name="addBias">Wskazuje na to czy bias ma być dodawany.</param>
/// <param name="weights">Wagi jakie perceptron otrzyma na początku.</param>
public Perceptron(byte inputsNumber, ActivationFunctionHandler activationFunction, bool addBias, float[] weights)
{
this.inputsNumber = inputsNumber;
this.activationFunction = activationFunction;
this.addBias = addBias;
LearningRate = 0.001f;
this.weights = weights;
}
/// <summary>
/// Tworzy nową instancję klasy Perceptron.
/// </summary>
/// <param name="inputsNumber">Liczba inputów jaką perceptron może przyjąć.</param>
/// <param name="activationFunction">Funkcja aktywacyjna dla tego perceptrona.</param>
/// <param name="addBias">Wskazuje na to czy bias ma być dodawany.</param>
public Perceptron(byte inputsNumber, ActivationFunctionHandler activationFunction, bool addBias)
{
this.inputsNumber = inputsNumber;
this.activationFunction = activationFunction;
this.addBias = addBias;
LearningRate = 0.001f;
weights = new float[inputsNumber];
for (byte i = 0; i < weights.Length; i++)
{
weights[i] = RandomGenerator.GetFloat(-1, 1);
}
}
void Backpropagation()
{
int outputLayer = layers.Count - 1;
int numberOfHiddenLayers = layers.Count > 2 ? layers.Count - 2 : 0;
//Output layer
for (int i = 0; i < layers[outputLayer].GetNeurons().Count; i++) //Iterate the neurons in the output layer
//Calculate the error for the neuron by subtracting the actual output from the desired output of this output neuron
{
double error = desiredOutputs[i][trainingIndex] - layers[outputLayer].GetNeurons()[i].GetOutput();
//Calculate the errorGradient for the neuron (used for the errorGradientSum in the hidden layer to follow)
layers[outputLayer].GetNeurons()[i].SetErrorGradient(ActivationFunctionHandler.TriggerDerativeFunction(layers[outputLayer].GetNeurons()[i].GetActivationFunction(),
layers[outputLayer].GetNeurons()[i].GetOutput()) * error);
//Update the neuron's weights
for (int j = 0; j < layers[outputLayer].GetNeurons()[i].GetWeights().Count; j++)
{
layers[outputLayer].GetNeurons()[i].GetWeights()[j] += alpha * layers[outputLayer].GetNeurons()[i].GetInputs()[j] * error;
}
//Update the neuron's bias
layers[outputLayer].GetNeurons()[i].SetBias(alpha * -1 * layers[outputLayer].GetNeurons()[i].GetErrorGradient());
}
//Hidden layer
for (int i = numberOfHiddenLayers; i > 1; i--) //Iterate the hidden layers
{
for (int j = 0; j < layers[i].GetNeurons().Count; j++) //Iterate the layer's neurons
//Calculate the errorGradientSum for the previous layer
{
double errorGradientSum = 0;
for (int k = 0; k < layers[i + 1].GetNeurons().Count; k++)
{
errorGradientSum += layers[i + 1].GetNeurons()[k].GetErrorGradient() * layers[i + 1].GetNeurons()[k].GetWeights()[j];
}
//Calculate the errorGradient for the neuron (used for the errorGradientSum in the hidden layer to follow)
layers[i].GetNeurons()[j].SetErrorGradient(ActivationFunctionHandler.TriggerDerativeFunction(layers[i].GetNeurons()[j].GetActivationFunction(),
layers[i].GetNeurons()[j].GetOutput()) * errorGradientSum);
//Update the neuron's weights
for (int k = 0; k < layers[i].GetNeurons()[j].GetWeights().Count; k++)
{
layers[i].GetNeurons()[j].GetWeights()[k] += alpha * layers[i].GetNeurons()[j].GetInputs()[k] * layers[i].GetNeurons()[j].GetErrorGradient();
}
//Update the neuron's bias
layers[i].GetNeurons()[j].SetBias(alpha * -1 * layers[i].GetNeurons()[j].GetErrorGradient());
}
}
}
/// <summary>
/// Tworzy nową 3 warstwową sieć neuronową z określonym rozmiarem każdej warstwy i określonymi funkcjami aktywacyjnymi.
/// </summary>
/// <param name="inputsLayerSize">Ilość inputów.</param>
/// <param name="hiddenLayerSize">Ilość ukrytych neuronów.</param>
/// <param name="outputsLayerSize">Ilość outputów.</param>
/// <param name="hiddenLayerActivationFunction">Funkcja aktywacyjna dla neuronów ukrytej warstwy.</param>
/// <param name="outputLayerActivationFunction">Funkcja aktywacyjna dla neuronów warstwy outputów.</param>
/// <param name="hiddenLayerDerivativeFunction">Pochodna dla funckcji aktywacyjnej warstwy ukrytej.</param>
/// <param name="outputLayerDerivativeFunction">Pochodna dla funckcji aktywacyjnej warstwy outputów.</param>
public SimpleNeuralNetwork(int inputsLayerSize, int hiddenLayerSize, int outputsLayerSize, ActivationFunctionHandler hiddenLayerActivationFunction, ActivationFunctionHandler outputLayerActivationFunction, FloatOperationHandler hiddenLayerDerivativeFunction, FloatOperationHandler outputLayerDerivativeFunction)
{
InputsLayerSize = inputsLayerSize;
HiddenLayerSize = hiddenLayerSize;
OutputsLayerSize = outputsLayerSize;
HiddenLayerActivationFunction = hiddenLayerActivationFunction;
OutputLayerActivationFunction = outputLayerActivationFunction;
HiddenLayerDerivativeFunction = hiddenLayerDerivativeFunction;
OutputLayerDerivativeFunction = outputLayerDerivativeFunction;
weightsIh = new Matrix(HiddenLayerSize, InputsLayerSize);
weightsHo = new Matrix(OutputsLayerSize, HiddenLayerSize);
weightsIh.Randomize(-1, 1);
weightsHo.Randomize(-1, 1);
biasH = new Matrix(HiddenLayerSize, 1);
biasO = new Matrix(OutputsLayerSize, 1);
biasH.Randomize(-1, 1);
biasO.Randomize(-1, 1);
}
