public void EncogXORHelloWorld()
{
// 3. Source code from the new book "Artificial Intelligence for Humans", https://github.com/jeffheaton/aifh/tree/26b07c42a3870277fe201356467bce7b4213b604
var network = new BasicNetwork();
network.AddLayer(new BasicLayer(null, true, 2));
network.AddLayer(new BasicLayer(new ActivationSigmoid(), true, 5));
network.AddLayer(new BasicLayer(new ActivationSigmoid(), false, 1));
network.FinalizeStructure();
network.Reset();
var trainingData = BasicData.ConvertArrays(XOR_INPUT, XOR_IDEAL);
// train the neural network
var train = new ResilientPropagation(network, trainingData);
var epoch = 1;
do
{
train.Iteration();
Console.WriteLine("Epoch #" + epoch + " Error:" + train.LastError);
epoch++;
} while (train.LastError > 0.01);
// test the neural network
Console.WriteLine("Neural Network Results:");
for (var i = 0; i < XOR_INPUT.Length; i++)
{
var output = network.ComputeRegression(XOR_INPUT[i]);
Console.WriteLine(string.Join(",", XOR_INPUT[i])
+ ", actual=" + string.Join(",", output)
+ ",ideal=" + string.Join(",", XOR_IDEAL[i]));
}
}
/// <summary>
/// The entry point for this example. If you would like to make this example
/// stand alone, then add to its own project and rename to Main.
/// </summary>
/// <param name="args">Not used.</param>
public static void ExampleMain(string[] args)
{
var network = new BasicNetwork();
network.AddLayer(new BasicLayer(null, true, 2));
network.AddLayer(new BasicLayer(new ActivationSigmoid(), true, 5));
network.AddLayer(new BasicLayer(new ActivationSigmoid(), false, 1));
network.FinalizeStructure();
network.Reset();
var trainingData = BasicData.ConvertArrays(XOR_INPUT, XOR_IDEAL);
// train the neural network
var train = new BackPropagation(network, trainingData, 0.7, 0.9);
var epoch = 1;
do
{
train.Iteration();
Console.WriteLine("Epoch #" + epoch + " Error:" + train.LastError);
epoch++;
} while (train.LastError > 0.01);
// test the neural network
Console.WriteLine("Neural Network Results:");
for (var i = 0; i < XOR_INPUT.Length; i++)
{
var output = network.ComputeRegression(XOR_INPUT[i]);
Console.WriteLine(string.Join(",", XOR_INPUT[i])
+ ", actual=" + string.Join(",", output)
+ ",ideal=" + string.Join(",", XOR_IDEAL[i]));
}
}
public void Process()
{
Console.WriteLine("Please wait, reading MNIST training data.");
var dir = AppDomain.CurrentDomain.BaseDirectory;
var trainingReader = LearnDigitsBackprop.LoadMNIST(dir, true, MNIST_DEPTH);
var validationReader = LearnDigitsBackprop.LoadMNIST(dir, false, MNIST_DEPTH);
Console.WriteLine("Training set size: " + trainingReader.NumImages);
Console.WriteLine("Validation set size: " + validationReader.NumImages);
var inputCount = trainingReader.Data[0].Input.Length;
var outputCount = trainingReader.Data[0].Ideal.Length;
var network = new BasicNetwork();
network.AddLayer(new BasicLayer(null, true, inputCount));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 100));
network.AddLayer(new DropoutLayer(new ActivationReLU(), true, 50, 0.5));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 25));
network.AddLayer(new BasicLayer(new ActivationSoftMax(), false, outputCount));
network.FinalizeStructure();
network.Reset();
// train the neural network
Console.WriteLine("Training neural network.");
var train = new BackPropagation(network, trainingReader.Data, 1e-4, 0.9);
train.L1 = 0;
train.L2 = 1e-11;
PerformIterationsClassifyEarlyStop(train, network, validationReader.Data, 5);
}
/// <inheritdoc />
public void FinalizeStructure(BasicNetwork theOwner, int theLayerIndex, TempStructureCounts counts)
{
_owner = theOwner;
_layerIndex = theLayerIndex;
var prevLayer = _layerIndex > 0 ? _owner.Layers[_layerIndex - 1] : null;
var nextLayer = _layerIndex < _owner.Layers.Count - 1 ? _owner.Layers[_layerIndex + 1] : null;
var tc = TotalCount;
counts.AddNeuronCount(tc);
if (prevLayer != null)
{
counts.AddWeightCount(Count*prevLayer.TotalCount);
}
int weightIndex, layerIndex;
if (theLayerIndex == _owner.Layers.Count - 1)
{
weightIndex = 0;
layerIndex = 0;
}
else
{
weightIndex = nextLayer.WeightIndex
+ TotalCount*nextLayer.Count;
layerIndex = nextLayer.NeuronIndex + nextLayer.TotalCount;
}
_neuronIndex = layerIndex;
_weightIndex = weightIndex;
}
/// <inheritdoc />
public void finalizeStructure(BasicNetwork theOwner, int theLayerIndex, TempStructureCounts counts)
{
FinalizeStructure(theOwner, theLayerIndex, counts);
var prevLayer = LayerIndex > 0 ? Owner.Layers[LayerIndex - 1] : null;
var nextLayer = LayerIndex < Owner.Layers.Count - 1 ? Owner.Layers[LayerIndex + 1] : null;
if (prevLayer == null)
{
throw new AIFHError("Conv2DLayer must have a previous layer (cannot be used as the input layer).");
}
var inColumns = prevLayer.DimensionCounts[0];
var inRows = prevLayer.DimensionCounts[1];
_inDepth = prevLayer.DimensionCounts[2];
_outColumns = Math.Floor((double)(inColumns + Padding * 2 - FilterRows) / Stride + 1);
_outRows = Math.Floor((double)(inRows + Padding * 2 - FilterColumns) / Stride + 1);
}
/// <inheritdoc />
public void finalizeStructure(BasicNetwork theOwner, int theLayerIndex, TempStructureCounts counts)
{
FinalizeStructure(theOwner, theLayerIndex, counts);
var prevLayer = LayerIndex > 0 ? Owner.Layers[LayerIndex - 1] : null;
var nextLayer = LayerIndex < Owner.Layers.Count - 1 ? Owner.Layers[LayerIndex + 1] : null;
if (prevLayer == null)
{
throw new AIFHError("Conv2DLayer must have a previous layer (cannot be used as the input layer).");
}
var inColumns = prevLayer.DimensionCounts[0];
var inRows = prevLayer.DimensionCounts[1];
_inDepth = prevLayer.DimensionCounts[2];
_outColumns = Math.Floor((double) (inColumns + Padding*2 - FilterRows)/Stride + 1);
_outRows = Math.Floor((double) (inRows + Padding*2 - FilterColumns)/Stride + 1);
}
/// <inheritdoc />
public void FinalizeStructure(BasicNetwork theOwner, int theLayerIndex, TempStructureCounts counts)
{
throw new NotImplementedException();
}
// calculate the forecast the next barChange for the day p_dateWeekDays[p_iRebalance]
// double target = p_barChanges[iRebalance + 1]; // so target is the p_iRebalance+1 day %change; so the last index that can be used in training is p_barChanges[p_iRebalance] as output
double GetMemberForecast(NNInputDesc p_nnInputDesc, int p_nNeurons, int p_maxEpoch, int p_iRebalance, int p_lookbackWindowSize, int[] p_barChangeLookbackDaysInds, double p_outputOutlierThreshold, double p_inputOutlierClipInSD, double p_inputNormalizationBoost, double p_outputNormalizationBoost, int p_notNNStrategyType, double[] p_dateWeekDays, double[] p_barChanges, out double p_trainError)
{
int barChangeDim = p_barChangeLookbackDaysInds == null ? 0 : p_barChangeLookbackDaysInds.Length;
p_trainError = Double.MaxValue;
StrongAssert.True(p_dateWeekDays.Length == p_barChanges.Length);
bool isUseDateWeekDays = (p_nnInputDesc & NNInputDesc.WeekDays) != 0;
bool isUseBarChange = (p_nnInputDesc & NNInputDesc.BarChange) != 0;
bool isUseEurChanges = (p_nnInputDesc & NNInputDesc.EurChange) != 0;
int inputDim = 0; // can be 3 or 5+1+1
int dateWeekDaysInd = 0;
int barChangeInd = 0;
int eurChangeInd = 0;
if (isUseDateWeekDays)
{
//inputVectorDim = inputVectorDim + size(p_dateWeekDays, 2); %1 dim or 5 dimension if p_dateWeekDays is 5 dimensional
//currBarChangeInd = currBarChangeInd + size(p_dateWeekDays, 2);
//eurChangeInd = eurChangeInd + size(p_dateWeekDays, 2);
inputDim += 1;
barChangeInd += 1;
eurChangeInd += 1;
}
if (isUseBarChange)
{
inputDim += barChangeDim;
eurChangeInd += barChangeDim;
}
if (isUseEurChanges)
{
inputDim++;
}
// extract original data. We have to exclude outliers and normalize later
double[][] nnInput = new double[p_lookbackWindowSize][];
double[][] nnOutput = new double[p_lookbackWindowSize][];
int inputIdx = p_iRebalance - p_lookbackWindowSize; // p_barChanges[p_iRebalance] can be used as a last output for the day p_iRebalance-1
for (int i = 0; i < p_lookbackWindowSize; i++, inputIdx++)
{
nnInput[i] = new double[inputDim];
if (isUseDateWeekDays)
nnInput[i][dateWeekDaysInd] = p_dateWeekDays[inputIdx]; // forecast the barChange for the day p_dateWeekDays[p_iRebalance]
if (isUseBarChange)
{
// nnInput[i][barChangeInd] = p_barChanges[inputIdx]; // forecast the barChange for the day p_dateWeekDays[p_iRebalance]
for (int j = 0; j < barChangeDim; j++)
{
int barChangeLookbackInd = p_barChangeLookbackDaysInds[j];
nnInput[i][barChangeInd + j] = p_barChanges[inputIdx - barChangeLookbackInd];
}
}
nnOutput[i] = new double[1];
nnOutput[i][0] = p_barChanges[inputIdx + 1]; // we want to forecast the next day
}
// Exclude outliers if requisted
EliminateOutliers(p_outputOutlierThreshold, ref nnInput, ref nnOutput);
int nTrainingSamples = nnInput.GetLength(0);
// normalize target and input?
// 2bins, output is normalized??
//for (int i = 0; i < nTrainingSamples; i++)
//{
// //nnInput[i][currBarChangeInd] = Math.Sign(nnInput[i][0]);
// nnOutput[i][0] = Math.Sign(nnOutput[i][0]);
//}
// inputOutlierClipInSD: only use it for CurrChange (continous), not dayOfTheWeek (discrete)
//Utils.StrongAssert(nnInput[0].Length == 1); // we would like to normalize only the first input: currChange, not the dayOfTheWeek
//double nnInputMean = nnInput.Select(r => r[0]).Average();
//double nnInputSD = nnInput.Select(r => r[0]).StandardDeviation();
//for (int i = 0; i < nnInput.Length; i++)
//{
// if (nnInput[i][0] < nnInputMean - nnInputSD * p_inputOutlierClipInSD)
// {
// nnInput[i][0] = nnInputMean - nnInputSD * p_inputOutlierClipInSD;
// }
// else if (nnInput[i][0] > nnInputMean + nnInputSD * p_inputOutlierClipInSD)
// {
// nnInput[i][0] = nnInputMean + nnInputSD * p_inputOutlierClipInSD;
// }
//}
// A. Normalize nnInput
RangeNormalizationParams dateWeekDaysNormParam = new RangeNormalizationParams();
double[] barChangeNormParam = new double[barChangeDim]; //I should fill Double.NaN;
if (isUseDateWeekDays)
dateWeekDaysNormParam = NeuralSnifferUtil.NormalizeIntoUnitRange(nnInput, dateWeekDaysInd, p_inputNormalizationBoost);
if (isUseBarChange)
{
//barChangeNormParam = Utils.NormalizeWithoutMovingCenterStd(nnInput, barChangeInd, p_inputNormalizationBoost);
for (int j = 0; j < barChangeDim; j++)
{
barChangeNormParam[j] = NeuralSnifferUtil.NormalizeWithoutMovingCenterStd(nnInput, barChangeInd + j, p_inputNormalizationBoost);
}
}
// B. Normalize nnInput
double outputMultiplier = NeuralSnifferUtil.NormalizeWithoutMovingCenterStd(nnOutput, 0, p_outputNormalizationBoost);
// C. generate testInput now and normalize
double[] testInput = new double[inputDim];
if (isUseDateWeekDays)
{
testInput[dateWeekDaysInd] = p_dateWeekDays[inputIdx]; // forecast the barChange for the day p_dateWeekDays[p_iRebalance]
NeuralSnifferUtil.RenormalizeIntoUnitRange(testInput, dateWeekDaysInd, dateWeekDaysNormParam);
}
if (isUseBarChange)
{
//testInput[barChangeInd] = p_barChanges[inputIdx]; // forecast the barChange for the day p_dateWeekDays[p_iRebalance]
for (int j = 0; j < barChangeDim; j++)
{
int barChangeLookbackInd = p_barChangeLookbackDaysInds[j];
testInput[barChangeInd + j] = p_barChanges[inputIdx - barChangeLookbackInd];
NeuralSnifferUtil.RenormalizeWithoutMovingCenterStd(testInput, barChangeInd + j, barChangeNormParam[j]);
}
}
if (p_notNNStrategyType == 1)
{
// Buy&hold
return 1;
}
if (p_notNNStrategyType == 2 || p_notNNStrategyType == 3) // 2 = MR, 3 = FT
{
// deterministic MR and FT
double returnInUpDays = p_notNNStrategyType == 2 ? -1 : 1;
if (p_barChanges[p_iRebalance] < 0)
return -1 * returnInUpDays;
else if (p_barChanges[p_iRebalance] > 0)
return returnInUpDays;
else
return 0;
}
if (p_notNNStrategyType == 4) // Naive learner, 1 bins
{
return NeuralSnifferUtil.NaiveLearner1DForecast(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { Double.PositiveInfinity }, testInput[0]);
}
if (p_notNNStrategyType == 5) // Naive learner, 2 bins
{
return NeuralSnifferUtil.NaiveLearner1DForecast(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { 0.0, Double.PositiveInfinity }, testInput[0]);
}
if (p_notNNStrategyType == 6) // Naive learner, 4 bins
{
double inputStd = nnInput.Select(r => r[0]).ToArray().StandardDeviation();
return NeuralSnifferUtil.NaiveLearner1DForecast(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { -0.6717, 0.0, 0.6717, Double.PositiveInfinity }.ScalarProduct(inputStd), testInput[0]);
//%4 bins with equal distribution: %25% of the samples are under -0.6717 std away
//range4Bins = [-0.6717 0 0.6717 inf];
//tick4Bins = [2*-0.6717 -0.6717 0 0.6717];
//%6 bins with equal distribution: 16.6%: -0.9678; 33.3%: -0.429; 50%: 0,
//range6Bins = [-0.96787 -0.429 0 0.429 0.96787 inf];
//tick6Bins = [(-0.96787-0.96787+0.429) -0.96787 -0.429 0 0.429 0.96787];
//%10 bins with equal distribution: 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%: -1.2801/-0.8391/-0.5216/-0.2513/0/0.2513/0.5216/0.8391/1.2801
//range10Bins = [-1.2801 -0.8391 -0.5216 -0.2513 0 0.2513 0.5216 0.8391 1.2801 inf];
//tick10Bins = [-1.2801-(1.2801-0.8391) -1.2801 -0.8391 -0.5216 -0.2513 0 0.2513 0.5216 0.8391 1.2801];
//%20 bins with equal distribution: 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%:
//range20Bins = [-1.6445 -1.2801 -1.0343 -0.8391 -0.6717 -0.5216 -0.3828 -0.2513 -0.1244 0 0.1244 0.2513 0.3828 0.5216 0.6717 0.8391 1.0343 1.2801 1.6445 inf];
//tick20Bins = [-1.6445-(1.6445-1.2801) -1.6445 -1.2801 -1.0343 -0.8391 -0.6717 -0.5216 -0.3828 -0.2513 -0.1244 0 0.1244 0.2513 0.3828 0.5216 0.6717 0.8391 1.0343 1.2801 1.6445];
}
if (p_notNNStrategyType != -1)
{
StrongAssert.Fail("Unrecognized p_notNNStrategyType");
}
// Matlab emulation network
// consider the newFF() in Matlab:
// 1. The input is not a layer; no activation function, no bias
// 2. The middle layer has a bias, and tansig transfer function
// 3. The output is a layer; having a bias (I checked); but has Linear activation (in the default case); in the Matlab book, there are examples with tansig ouptput layers too
// Jeff use a similar one
// I've been using a linear activation function on the output layer, and sigmoid or htan on the input and hidden lately for my prediction nets,
// and getting lower error rates than a uniform activation function. (uniform: using the same on every layer)
BasicNetwork network = new BasicNetwork();
//network.Logic = new Encog.Neural.Networks.Logic.FeedforwardLogic(); // the default is SimpleRecurrentLogic; but FeedforwardLogic() is faster, simpler
network.AddLayer(new BasicLayer(new ActivationLinear(), false, inputDim)); // doesn't matter what is here. nor the act.function, neither the bias are used
network.AddLayer(new BasicLayer(new ActivationTANH(), true, p_nNeurons));
network.AddLayer(new BasicLayer(new ActivationLinear(), true, 1));
network.FinalizeStructure(); // after this, the Layers.BiasWeight and Synapses.Weights are zero
//Utils.Input1DVisualizer(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { 0.0, Double.MaxValue }, new double[] { });
//Utils.Input1DVisualizer(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { 0.5, 1.5, 2.5, 3.5, 4.5}, new double[] { });
double[] agyTestForecast = new double[1];
double[] agyTestTrainError = new double[1];
ResilientPropagation train = null;
for (int i = 0; i < agyTestForecast.Length; i++)
{
network.Reset();
//network.Reset(new Random(123)); // randomizes Layers.BiasWeight and Synapses.Weights; if initialweights are left zero; they will be zeroWeights after Resilient and Backprog training. Only the last biasWeight will be non-zero. means: the output will be the same regardless of the input; Bad.
// train the neural network
var trainingData = BasicData.ConvertArrays(nnInput, nnOutput);
train = new ResilientPropagation(network, trainingData); // err remained=0.99; with 100 random samples: accumulatedValue=0.69, 0.09, 0.57; up/down ratio stayed at 0.52 in all cases (ratio is stable)
//train.NumThreads = 1; // default is 0; that means Encog can determine how many; (but I specificaly want 1 threads train, because I run 4 backtests in parallel; so I use the CPUs anyway)
int epoch = 1;
do
{
train.Iteration();
//Utils.Logger.Info("Epoch #" + epoch + " Error:" + train.Error);
epoch++;
} while ((epoch <= p_maxEpoch) && (train.LastError > 0.001)); // epoch = 5000?
//var inputData = new BasicData(testInput);
var ouput = network.ComputeRegression(testInput);
//double[] ouput = outputData.Data;
NeuralSnifferUtil.DenormalizeWithoutMovingCenterStd(ouput, 0, outputMultiplier);
//double forecast = outputData[0];
double forecast = ouput[0];
//Utils.Logger.Info(@"Real%change: " + p_barChanges[p_iRebalance] * 100 + "%, Network forecast: " + outputData.ToString() + "%");
agyTestForecast[i] = forecast;
agyTestTrainError[i] = train.LastError;
}
//NNVisualizer(network, Enumerable.Range(0, 11).Select(r => new double[] { ((double)r / 10.0 - 0.5) * 2 * 0.001}).ToArray(), AppDomain.CurrentDomain.BaseDirectory + "NNchart.csv");
//NNVisualizer(network, Enumerable.Range(0, 101).Select(r => new double[] { ((double)r / 100.0 - 0.5) * 2 * 0.10 * inputMultiplier }).ToArray(), AppDomain.CurrentDomain.BaseDirectory + "NNchart.csv"); // -10% to +10%
p_trainError = agyTestTrainError[0];
return agyTestForecast[0];
}
public void Process()
{
IList<BasicData> trainingData = LoadSunspots();
BasicNetwork network = new BasicNetwork();
network.AddLayer(new BasicLayer(null, true, this.INPUT_WINDOW));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 50));
network.AddLayer(new BasicLayer(new ActivationLinear(), false, 1));
network.FinalizeStructure();
network.Reset();
BackPropagation train = new BackPropagation(network, trainingData, 1e-9, 0.5);
train.BatchSize = 0;
PerformIterations(train, 100000, 650, true);
Query(network, trainingData);
}
/// <summary>
/// Run the example.
/// </summary>
public void Process()
{
// read the iris data from the resources
var assembly = Assembly.GetExecutingAssembly();
var res = assembly.GetManifestResourceStream("AIFH_Vol3.Resources.iris.csv");
// did we fail to read the resouce
if (res == null)
{
Console.WriteLine("Can't read iris data from embedded resources.");
return;
}
// load the data
var istream = new StreamReader(res);
var ds = DataSet.Load(istream);
istream.Close();
// The following ranges are setup for the Iris data set. If you wish to normalize other files you will
// need to modify the below function calls other files.
ds.NormalizeRange(0, 0, 1);
ds.NormalizeRange(1, 0, 1);
ds.NormalizeRange(2, 0, 1);
ds.NormalizeRange(3, 0, 1);
var species = ds.EncodeOneOfN(4);
var trainingData = ds.ExtractSupervised(0, 4, 4, 3);
var network = new BasicNetwork();
network.AddLayer(new BasicLayer(null, true, 4));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 20));
network.AddLayer(new BasicLayer(new ActivationSoftMax(), false, 3));
network.FinalizeStructure();
network.Reset();
var train = new BackPropagation(network, trainingData, 0.001, 0.9);
PerformIterations(train, 100000, 0.02, true);
QueryOneOfN(network, trainingData, species);
}
/// <inheritdoc />
public void FinalizeStructure(BasicNetwork theOwner, int theLayerIndex, TempStructureCounts counts)
{
throw new NotImplementedException();
}
/// <summary>
/// Run the example.
/// </summary>
public void Process()
{
// read the iris data from the resources
var assembly = Assembly.GetExecutingAssembly();
var res = assembly.GetManifestResourceStream("AIFH_Vol3.Resources.auto-mpg.data.csv");
// did we fail to read the resouce
if (res == null)
{
Console.WriteLine("Can't read auto MPG data from embedded resources.");
return;
}
// load the data
var istream = new StreamReader(res);
var ds = DataSet.Load(istream);
istream.Close();
// The following ranges are setup for the Auto MPG data set. If you wish to normalize other files you will
// need to modify the below function calls other files.
// First remove some columns that we will not use:
ds.DeleteColumn(8); // Car name
ds.DeleteColumn(7); // Car origin
ds.DeleteColumn(6); // Year
ds.DeleteUnknowns();
ds.NormalizeZScore(1);
ds.NormalizeZScore(2);
ds.NormalizeZScore(3);
ds.NormalizeZScore(4);
ds.NormalizeZScore(5);
var trainingData = ds.ExtractSupervised(1, 4, 0, 1);
var splitList = DataUtil.Split(trainingData, 0.75);
trainingData = splitList[0];
var validationData = splitList[1];
Console.WriteLine("Size of dataset: " + ds.Count);
Console.WriteLine("Size of training set: " + trainingData.Count);
Console.WriteLine("Size of validation set: " + validationData.Count);
var inputCount = trainingData[0].Input.Length;
var network = new BasicNetwork();
network.AddLayer(new BasicLayer(null, true, inputCount));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 50));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 25));
network.AddLayer(new BasicLayer(new ActivationReLU(), true, 5));
network.AddLayer(new BasicLayer(new ActivationLinear(), false, 1));
network.FinalizeStructure();
network.Reset();
var train = new BackPropagation(network, trainingData, 0.000001, 0.9);
PerformIterationsEarlyStop(train, network, validationData, 20, new ErrorCalculationMSE());
Query(network, validationData);
}
internal void TestEncogSimilarThatIsUsedByVirtualBroker()
{
int p_nNeurons = 5;
int p_lookbackWindowSize = 200;
int p_maxEpoch = 5000;
int inputDim = 0; // can be 3 or 5+1+1
double[] testInput = new double[inputDim];
// extract original data. We have to exclude outliers and normalize later
double[][] nnInput = new double[p_lookbackWindowSize][];
double[][] nnOutput = new double[p_lookbackWindowSize][];
// Matlab emulation network
// consider the newFF() in Matlab:
// 1. The input is not a layer; no activation function, no bias
// 2. The middle layer has a bias, and tansig transfer function
// 3. The output is a layer; having a bias (I checked); but has Linear activation (in the default case); in the Matlab book, there are examples with tansig ouptput layers too
// Jeff use a similar one
// I've been using a linear activation function on the output layer, and sigmoid or htan on the input and hidden lately for my prediction nets,
// and getting lower error rates than a uniform activation function. (uniform: using the same on every layer)
BasicNetwork network = new BasicNetwork();
//var ffPattern = new FeedForwardPattern(); maybe I don't need that Logic settings
//network.Logic = new Encog.Neural.Networks.Logic.FeedforwardLogic(); // the default is SimpleRecurrentLogic; but FeedforwardLogic() is faster, simpler
network.AddLayer(new BasicLayer(new ActivationLinear(), false, inputDim)); // doesn't matter what is here. nor the act.function, neither the bias are used
network.AddLayer(new BasicLayer(new ActivationTANH(), true, p_nNeurons));
network.AddLayer(new BasicLayer(new ActivationLinear(), true, 1));
//network.Structure.FinalizeStructure(); // after this, the Layers.BiasWeight and Synapses.Weights are zero
network.FinalizeStructure();
//Utils.Input1DVisualizer(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { 0.0, Double.MaxValue }, new double[] { });
//Utils.Input1DVisualizer(nnInput.Select(r => r[0]).ToArray(), nnOutput.Select(r => r[0]).ToArray(), new double[] { 0.5, 1.5, 2.5, 3.5, 4.5}, new double[] { });
double[] agyTestForecast = new double[1];
double[] agyTestTrainError = new double[1];
ResilientPropagation train = null;
for (int i = 0; i < agyTestForecast.Length; i++)
{
network.Reset();
//network.Reset(new Random((int)DateTime.Now.Ticks).Next());
//network.Reset(new Random(123)); // randomizes Layers.BiasWeight and Synapses.Weights; if initialweights are left zero; they will be zeroWeights after Resilient and Backprog training. Only the last biasWeight will be non-zero. means: the output will be the same regardless of the input; Bad.
// train the neural network
var trainingData = BasicData.ConvertArrays(nnInput, nnOutput);
train = new ResilientPropagation(network, trainingData);
//INeuralDataSet trainingSet = new BasicNeuralDataSet(nnInput, nnOutput);
//train = new ResilientPropagation(network, trainingSet); // err remained=0.99; with 100 random samples: accumulatedValue=0.69, 0.09, 0.57; up/down ratio stayed at 0.52 in all cases (ratio is stable)
//train.NumThreads = 1; // default is 0; that means Encog can determine how many; (but I specificaly want 1 threads train, because I run 4 backtests in parallel; so I use the CPUs anyway)
int epoch = 1;
do
{
train.Iteration();
//Utils.Logger.Info("Epoch #" + epoch + " Error:" + train.Error);
epoch++;
} while ((epoch <= p_maxEpoch) && (train.LastError > 0.001)); // epoch = 5000?
//INeuralData inputData = new BasicNeuralData(testInput);
//INeuralData outputData = network.Compute(inputData);
//double[] ouput = outputData.Data;
//NeuralSnifferUtil.DenormalizeWithoutMovingCenterStd(ouput, 0, outputMultiplier);
//double forecast = outputData[0];
////Utils.Logger.Info(@"Real%change: " + p_barChanges[p_iRebalance] * 100 + "%, Network forecast: " + outputData.ToString() + "%");
//agyTestForecast[i] = forecast;
//agyTestTrainError[i] = train.Error;
}
}