/// <summary>
/// Initializes the internal neural network.
/// </summary>
/// <param name="state">The state.</param>
private void InitializeQNetwork(State state)
{
if (File.Exists(NetworkPath))
{
Q = new QNetwork(NetworkPath);
}
else
{
var inputLength = state.ToArray().Length;
Q = new QNetwork(p =>
{
p.TrainingError = Parameters.QNetwork.TrainingError;
p.MaxIterationPerTrainging = Parameters.QNetwork.MaxIterationPerTrainging;
// Input Layer
p.Layers.Add(new LayerParameters(null, false, inputLength));
// Hidden Layers
for (int i = 0; i < Parameters.QNetwork.HiddenLayersCount; i++)
{
p.Layers.Add(new LayerParameters(new ActivationTANH(), true, Parameters.QNetwork.NeuronCountForHiddenLayers));
}
// Output Layer
p.Layers.Add(new LayerParameters(new ActivationTANH(), false, 3));
});
Q.OnTrainingEpochComplete += (e, a) =>
{
a.AgentId = this.Parameters.Id;
OnTrainingEpochComplete?.Invoke(e, a);
};
}
}
/// <summary>
/// Trains the neural network with the passed in training set.
/// Receive a list of tuple, where each tuple represent
/// Tuple = (State, ExpectedValueForBuy, ExpectedValueForSell, ExpectedValueForWait)
/// </summary>
/// <param name="trainingSet">The training set.</param>
public void Train(IList <Tuple <State, double[]> > trainingSet)
{
var trainingData = new List <IMLDataPair>();
foreach (var sample in trainingSet)
{
var flattenState = sample.Item1.ToArray();
var actuals = new BasicMLData(flattenState);
var ideals = new BasicMLData(sample.Item2);
trainingData.Add(new BasicMLDataPair(actuals, ideals));
}
IMLDataSet dataSet = new BasicMLDataSet(trainingData);
//IMLTrain train = new Backpropagation(NeuralNetwork, dataSet, Parameters.LearningRate, Parameters.LearningMomemtum);
IMLTrain train = new ResilientPropagation(NeuralNetwork, dataSet);
int epoch = 1;
do
{
train.Iteration();
epoch++;
OnTrainingEpochComplete?.Invoke(this, new OnTrainingEpochCompleteArgs()
{
Epoch = epoch,
Error = train.Error
});
} while (train.Error > Parameters.TrainingError && epoch < Parameters.MaxIterationPerTrainging);
//foreach (var item in dataSet)
//{
// var output = NeuralNetwork.Compute(item.Input);
// Console.WriteLine("output: {0} - {1} - {2} | ideal: {3} - {4} - {5}", output[0], output[1], output[2], item.Ideal[0], item.Ideal[1], item.Ideal[2]);
//}
}