void RunNetworkTraining(ITrainableNetwork network, TestConfiguration config, double learningParam, double numEpochs)
{
double err;
for (int i = 0; i < numEpochs; i++)
{
// measure the error before the learning, on the whole training set
err = 0.0;
// TODO randomize the order of training samples
foreach (var trainSample in config.TrainingSet)
{
err += network.GetNetworkTotalError(trainSample.Item1, trainSample.Item2);
}
Console.WriteLine("Step {0}, total network error: {1}", i, err);
// train all samples
foreach (var trainSample in config.TrainingSet)
{
network.Train(trainSample.Item1, trainSample.Item2, learningParam);
}
}
// measure final error of the network
err = 0.0;
foreach (var trainSample in config.TrainingSet)
{
err += network.GetNetworkTotalError(trainSample.Item1, trainSample.Item2);
}
Console.WriteLine("Final total network error: {0}", err);
}
//for the sake of simplicity we crossover two NNs with identical structure and only adjust the weights and biases
public ITrainableNetwork Crossover(ITrainableNetwork other)
{
var network = new TrainableNetwork(this.InputsCount, this.HiddenLayersCounts, this.OutputsCount);
for (int layerIndex = 0; layerIndex < this.HiddenLayersCount + 1; layerIndex++)
{
for (int neuronIndex = 0; neuronIndex < this.Weights[layerIndex].Count; neuronIndex++)
{
for (int weightIndex = 0; weightIndex < this.Weights[layerIndex][neuronIndex].Count; weightIndex++)
{
var copyWeights = Network.Random.NextBool() ? this.Weights : other.Weights;
network.Weights[layerIndex][neuronIndex][weightIndex] = copyWeights[layerIndex][neuronIndex][weightIndex];
}
var copyBiases = Network.Random.NextBool() ? this.Biases : other.Biases;
network.Biases[layerIndex][neuronIndex] = copyBiases[layerIndex][neuronIndex];
}
}
return(network);
}
