public NNManager()
{
const int nbOfInputs = 2;
const int nbOfOutputs = 1;
const int sizeDataSet = 10000;
var sizeLayers = new int[0];
//sizeLayers[0] = 2;
var Rand = new Random(Guid.NewGuid().GetHashCode());
var nn = new NNNeuralNetwork(nbOfInputs, nbOfOutputs, sizeLayers);
nn.Initialise();
double[][] dataset = new double[sizeDataSet][];
double[][] expected = new double[sizeDataSet][];
for (int i = 0; i < sizeDataSet; i++)
{
dataset[i] = new double[nbOfInputs];
dataset[i][0] = Rand.Next(2);
dataset[i][1] = Rand.Next(2);
expected[i] = new double[1];
expected[i][0] = dataset[i][0] + dataset[i][1];
if (expected[i][0] == 2)
{
expected[i][0] = 1;
}
}
var Trainer = new NNTrainer();
Trainer.Train(nn, dataset, expected, 0.5, 15);
}
private void UpdateWeights(NNNeuralNetwork nn, double[] nnInputs, double learningRate)
{
for (int i = 0; i < nn.nbOfHiddenLayers + 1; i++)
{
NNLayer Layer = nn.Layers[i];
double[] inputs;
// Récupération des inputs du layer
if (i != 0)
{
inputs = new double[nn.Layers[i - 1].nbOfNeurons];
for (int j = 0; j < nn.Layers[i - 1].nbOfNeurons; j++)
{
inputs[j] = nn.Layers[i - 1].Neurons[j].output;
}
}
else
{
inputs = nnInputs;
}
// Mise à jour des poids de chaque neurone du layer
Layer.Neurons.ForEach(Neuron =>
{
for (int j = 0; j < inputs.Length; j++)
{
Neuron.weights[j] += learningRate * Neuron.delta * inputs[j];
}
Neuron.w0 += learningRate * Neuron.delta;
});
}
}
private int BackwardPropagateError(NNNeuralNetwork nn, double[] expected)
{
if (nn.nbOfOutputs == expected.Length)
{
// Parcours inverse des layers
for (int i = nn.nbOfHiddenLayers; i >= 0; i--)
{
NNLayer Layer = nn.Layers[i];
double[] errors = new double[Layer.nbOfNeurons];
// Si ce n'est pas le dernier layer
if (i != nn.nbOfHiddenLayers)
{
// Parcours des neurones du layer courant
// Calcul de l'erreur
for (int j = 0; j < Layer.nbOfNeurons; j++)
{
var error = 0.0;
// Parcours des neurones du layers précédent
nn.Layers[i + 1].Neurons.ForEach(Neuron => {
error += Neuron.weights[j] * Neuron.delta;
});
errors[i] = error;
}
}
// Si c'est le dernier layer
else
{
// Parcours des neurones du layer courant
// Calcul de l'erreur
for (int j = 0; j < Layer.nbOfNeurons; j++)
{
errors[j] = (expected[j] - Layer.Neurons[j].output);
}
}
// Parcours des neurones du layer courant
// Calcul de delta
for (int j = 0; j < Layer.nbOfNeurons; j++)
{
NNNeuron Neuron = Layer.Neurons[j];
Neuron.delta = errors[j] * NNNeuron.TransfertDerivative(Neuron.output);
}
}
return(0);
}
else
{
return(1);
}
}
public int Train(NNNeuralNetwork nn, double[][] dataset, double[][] expectedResults, double learningRate, int nbEpoch)
{
/*
* if (dataset.GetLength(0) != nn.nbOfInputs || expectedResults.GetLength(0) != nn.nbOfOutputs)
* {
* return -1;
* }
*/
for (int i = 0; i < nbEpoch; i++)
{
double sumError = 0.0;
for (int j = 0; j < dataset.Length; j++)
{
double[] inputs = dataset[j];
double[] outputs = nn.ProcessInputs(ref inputs);
double[] expected = expectedResults[j];
/*
* var strBuilder = new StringBuilder();
* strBuilder.Append("Line=");
* strBuilder.Append(j);
* strBuilder.Append(" : inputs=");
* for (int k = 0; k < inputs.Length-1; k++)
* {
* strBuilder.Append(inputs[k]);
* strBuilder.Append(",");
* }
* strBuilder.Append(inputs[inputs.Length-1]);
*
* strBuilder.Append(" ; outputs=");
* for (int k = 0; k < outputs.Length - 1; k++)
* {
* strBuilder.Append(outputs[k]);
* strBuilder.Append(",");
* }
* strBuilder.Append(outputs[outputs.Length - 1]);
*
* strBuilder.Append(" ; expected=");
* for (int k = 0; k < expected.Length - 1; k++)
* {
* strBuilder.Append(expected[k]);
* strBuilder.Append(",");
* }
* strBuilder.Append(expected[expected.Length - 1]);
*
* Console.WriteLine(strBuilder.ToString());
*/
for (int k = 0; k < outputs.Length; k++)
{
sumError += (expected[k] - outputs[k]);
}
BackwardPropagateError(nn, expected);
UpdateWeights(nn, inputs, learningRate);
}
Console.WriteLine(">Epoch={0}, learningRate={1}, error={2}", i, learningRate, sumError);
Console.WriteLine("---------------------------------------");
}
return(0);
}