public QLearner(List <Action> actions, int numHidden, int numInputs, int seed)
{
this.actions = actions;
network = new BackPropNeuralNet(numInputs, numHidden, actions.Count);
int numWeights = (numInputs * numHidden) + (numHidden * actions.Count) + (numHidden + actions.Count);
List <double> weights = new List <double>();
for (int i = 0; i < numWeights; i++)
{
weights.Add(Random.Range(-1.0f, 1.0f));
}
network.SetWeights(weights.ToArray());
}
public Detect(BackPropNeuralNet bnn)
{
InitializeComponent();
DoubleBuffered = true;
Dock = DockStyle.Fill;
blobFinder = new BlobFinder();
this.bnn = bnn;
this.panel1.Paint += new PaintEventHandler(this.panel1_Paint);
this.panel1.MouseDown += new MouseEventHandler(this.panel1_MouseDown);
this.panel1.MouseMove += new MouseEventHandler(this.panel1_MouseMove);
this.panel1.MouseUp += new MouseEventHandler(this.panel1_MouseUp);
}
public void Setup(IEnumerable <IFunctor> functors, StockHistorySet stockHistorySet)
{
Fns = functors.ToArray();
m_StockHistorySet = stockHistorySet;
if (this.Fns == null)
{
throw new ApplicationException();
}
if (!Fns.Any())
{
throw new ApplicationException();
}
int numInput = Fns.Length;
int numHidden = 8;
int numOutput = 1;
int numWeights = (numInput * numHidden) + (numHidden * numOutput) + (numHidden + numOutput);
DebugWrite("Creating a " + numInput + "-input, " + numHidden + "-hidden, " + numOutput + "-output neural network");
DebugWrite("Using hard-coded tanh function for hidden layer activation");
DebugWrite("Using hard-coded log-sigmoid function for output layer activation");
m_Bnn = new BackPropNeuralNet(numInput, numHidden, numOutput);
DebugWrite("\nGenerating random initial weights and bias values");
double[] initWeights = new double[numWeights];
for (int i = 0; i < initWeights.Length; ++i)
{
initWeights[i] = (rnd.NextDouble() - 0.5d) * 1.0d;
}
DebugWrite("Loading neural network initial weights and biases into neural network");
m_Bnn.SetWeights(initWeights);
double learnRate = 0.2; // learning rate - controls the maginitude of the increase in the change in weights.
double momentum = 0.1; // momentum - to discourage oscillation.
DebugWrite("Setting learning rate = " + learnRate.ToString("F2") + " and momentum = " + momentum.ToString("F2"));
int maxEpochs = 8000000;
double errorThresh = 0.01;
DebugWrite("\nSetting max epochs = " + maxEpochs + " and error threshold = " + errorThresh.ToString("F6"));
// Train
int epoch = 0;
double error = double.MaxValue;
DebugWrite("\nBeginning training using back-propagation\n");
int stocksCount = stockHistorySet.AllStockHistories.Count();;
while (epoch < maxEpochs) // train
{
int stockNum = rnd.Next(0, stocksCount);
var stock = stockHistorySet.AllStockHistories[stockNum];
double realValue = stock.Closes[Today + DaysInFuture] / stock.Closes[Today];
double updateValue = Math.Tanh(realValue);
double predictedOutput = m_Bnn.ComputeOutputs(ComputeInputs(stock))[0];
double predictedValue = ATanh(predictedOutput) * stock.Closes[Today];
m_Bnn.UpdateWeights(new[] { updateValue }, learnRate, momentum);
++epoch;
if (epoch % 20000 == 0)
{
error = GetAverageError(true);
if (error < errorThresh)
{
DebugWrite("Found weights and bias values that meet the error criterion at epoch " + epoch);
break;
}
DebugWrite("epoch = " + epoch);
DebugWrite(" error = " + error + "\n");
}
} // train loop
double[] finalWeights = m_Bnn.GetWeights();
DebugWrite("");
DebugWrite("Final neural network weights and bias values are:");
Helpers.ShowVector(finalWeights, 5, 8, true);
}
public State()
{
NeuralNetwork = new BackPropNeuralNet(NumberOfInputs, NumberOfHidden, NumberOfOutputs);
}
public QLearner(int eValue, double alpha, double gamma, double eta, List <Action> actions, List <int> hiddenSizes, int numInputs, int seed)
{
this.actions = actions;
//hiddenSizes.Add(1);
network = new BackPropNeuralNet(numInputs + actions.Count, hiddenSizes[0], 1);
}
public void Setup(IEnumerable<IFunctor> functors, StockHistorySet stockHistorySet)
{
Fns = functors.ToArray();
m_StockHistorySet = stockHistorySet;
if (this.Fns == null) throw new ApplicationException();
if (!Fns.Any()) throw new ApplicationException();
int numInput = Fns.Length;
int numHidden = 8;
int numOutput = 1;
int numWeights = (numInput * numHidden) + (numHidden * numOutput) + (numHidden + numOutput);
DebugWrite("Creating a " + numInput + "-input, " + numHidden + "-hidden, " + numOutput + "-output neural network");
DebugWrite("Using hard-coded tanh function for hidden layer activation");
DebugWrite("Using hard-coded log-sigmoid function for output layer activation");
m_Bnn = new BackPropNeuralNet(numInput, numHidden, numOutput);
DebugWrite("\nGenerating random initial weights and bias values");
double[] initWeights = new double[numWeights];
for (int i = 0; i < initWeights.Length; ++i)
initWeights[i] = (rnd.NextDouble() - 0.5d) * 1.0d;
DebugWrite("Loading neural network initial weights and biases into neural network");
m_Bnn.SetWeights(initWeights);
double learnRate = 0.2; // learning rate - controls the maginitude of the increase in the change in weights.
double momentum = 0.1; // momentum - to discourage oscillation.
DebugWrite("Setting learning rate = " + learnRate.ToString("F2") + " and momentum = " + momentum.ToString("F2"));
int maxEpochs = 8000000;
double errorThresh = 0.01;
DebugWrite("\nSetting max epochs = " + maxEpochs + " and error threshold = " + errorThresh.ToString("F6"));
// Train
int epoch = 0;
double error = double.MaxValue;
DebugWrite("\nBeginning training using back-propagation\n");
int stocksCount = stockHistorySet.AllStockHistories.Count(); ;
while (epoch < maxEpochs) // train
{
int stockNum = rnd.Next(0, stocksCount);
var stock = stockHistorySet.AllStockHistories[stockNum];
double realValue = stock.Closes[Today + DaysInFuture] / stock.Closes[Today];
double updateValue = Math.Tanh(realValue);
double predictedOutput = m_Bnn.ComputeOutputs(ComputeInputs(stock))[0];
double predictedValue = ATanh(predictedOutput) * stock.Closes[Today];
m_Bnn.UpdateWeights(new[] { updateValue }, learnRate, momentum);
++epoch;
if (epoch % 20000 == 0)
{
error = GetAverageError(true);
if (error < errorThresh)
{
DebugWrite("Found weights and bias values that meet the error criterion at epoch " + epoch);
break;
}
DebugWrite("epoch = " + epoch);
DebugWrite(" error = " + error + "\n");
}
} // train loop
double[] finalWeights = m_Bnn.GetWeights();
DebugWrite("");
DebugWrite("Final neural network weights and bias values are:");
Helpers.ShowVector(finalWeights, 5, 8, true);
}
