public void addTrainingRow(TrainingRow tr)
{
this.rows.Add(tr);
}
public void addTrainingRow(TrainingRow tr)
{
this.rows.Add(tr);
}
public double trainNetwork(TrainingTemplate trainingTemplate, AriesNeuroNet.Neurons.Neuron neuron, int maxGenerations, ErrorHistory errorProg)
{
int stableLimit = trainingTemplate.rows.Count;
int stableGenerations = 0;
int currentGeneration = 0;
double error = 0;
List <double> errorHistory = new List <double>();
//We innitialize the flags
bool adjustedWeights = false;
bool stableGenFlag = true;
bool genLimitFlag = true;
bool templateFlag = false;
//Step 1 initialize the neurons to randomize weights
neuron.randomizeWeights();
/*
* Possible breaking mecanism
* if(flag) {
* Console.Writeline(adequate message)
* return error
* which breaks the function
* }
* */
for (int currentRow = 0; currentRow < trainingTemplate.rows.Count; currentRow++)
{
// I extract the current row
TrainingRow row = trainingTemplate.rows[currentRow];
do
{
// I begin a new generation
//Console.WriteLine("========================================================================");
//Console.WriteLine("Begining Generation: " + currentGeneration);
//Console.WriteLine("Current row: " + currentRow);
// I reset the adjutedWeights flag
adjustedWeights = false;
// I set the inputs
neuron.setInputValues(row.inputs);
// I fire the neuron
neuron.fireNeuron();
// I get the expected output out of the template
double expectedOutput = row.outputs[0];
// I get the real output fromt he neuron
double realOutput = neuron.output.weightedReading;
Console.WriteLine("Output is " + realOutput);
// I calculate the error
error = expectedOutput - realOutput;
//Console.WriteLine("Error is " + error);
// I make a decision based on the error
if (error == 0)
{
//Console.WriteLine("I have not found an error");
stableGenerations++;
// I set the flag so that I exit the while
adjustedWeights = false;
}
else
{
//Console.WriteLine("I found an error");
//Console.WriteLine("The error is " + error);
// I reset the stable generations counter
stableGenerations = 0;
// These are for debugging purposes
//List<double> oldWeights = new List<double>();
//List<double> newWeights = new List<double>();
// I mark that I needed to adjust the weights.
adjustedWeights = true;
//Do the heavy duty processing
foreach (NeuronPort input in neuron.inputs)
{
//oldWeights.Add(input.weight);
input.weight += input.reading * learningRate * error; // To do finish this
//newWeights.Add(input.weight);
}
Console.WriteLine("I corrected with " + (learningRate * error));
// I log the error to history
errorHistory.Add(error);
// I publish the old weights
/*
* Console.WriteLine("Old weights: ");
* foreach (double weight in oldWeights)
* {
* Console.Write(weight + " ");
* }
*
* // I publish the new weights
* Console.WriteLine("New weights: " + newWeights);
* foreach (double weight in newWeights)
* {
* Console.Write(weight + " ");
* }
* */
}
// I mark that I've finished these generation
currentGeneration++;
//I check the conditions
stableGenFlag = stableGenerations < stableLimit;
genLimitFlag = currentGeneration < maxGenerations;
templateFlag = currentRow < trainingTemplate.rows.Count;
//the breaking conditions
if (!stableGenFlag)
{
//Console.WriteLine("Ended due to stable limit gen") ;
return(error);
}
if (!genLimitFlag)
{
//Console.WriteLine("Ended due to limit of gens to train");
return(error);
}
} while (adjustedWeights);
//maybe not necesary ???
if (!stableGenFlag)
{
return(error);
}
if (!genLimitFlag)
{
return(error);
}
}
errorProg.errorPoints = errorHistory;
return(error);
}
/// <summary>
/// trains the netwrok though back propagation
/// </summary>
/// <param name="trainingTemplate"></param>
/// <param name="inputLayer">the input layer of the network</param>
/// <param name="outputLayer">the output layer of the network</param>
/// <param name="hiddenLayer">the hidden layer of the network</param>
/// <param name="maxGenerations"></param>
/// <param name="errorProg"></param>
/// <returns></returns>
public double trainNetwork(TrainingTemplate trainingTemplate, Layer inputLayer, Layer outputLayer, Layer hiddenLayer, int maxGenerations, double acceptableError, ErrorHistory errorProg)
{
// Step 0 innitialize
int stableLimit = trainingTemplate.rows.Count;
int stableGenerations = 0;
int currentGeneration = 0;
double error = 0;
List <double> errorHistory = new List <double>();
bool adjustedWeights = false;
bool stableGenFlag = true;
bool genLimitFlag = true;
bool templateFlag = false;
foreach (Neuron neuron in inputLayer.neurons)
{
neuron.randomizeWeights();
}
foreach (Neuron neuron in hiddenLayer.neurons)
{
neuron.randomizeWeights();
}
//Should I set the output neuron's weights to 1 ?
for (int currentRow = 0; currentRow < trainingTemplate.rows.Count; currentRow++)
{
// I extract the current row
TrainingRow row = trainingTemplate.rows[currentRow];
do
{
// I reset the adjutedWeights flag
adjustedWeights = false;
//extract the inputs and distribute them
List <double> templateInputs = row.inputs;
// Need to reverse them as to pop them
templateInputs.Reverse();
//Fire the neurons
foreach (Neuron neuron in inputLayer.neurons)
{
foreach (NeuronPort input in neuron.inputs)
{
// We get the last input from the template and check if it's not null if it's null we put 0
double var_input = 0;
if (templateInputs.Count > 0)
{
var_input = templateInputs.Last();
templateInputs.RemoveAt(templateInputs.Count - 1);
}
input.reading = var_input;
}
neuron.fireNeuron();
}
// fire the hidden layer
foreach (Neuron neuron in hiddenLayer.neurons)
{
neuron.fireNeuron();
}
//Fire the output layer
outputLayer.neurons[0].fireNeuron();
error = outputLayer.neurons[0].output.weightedReading - trainingTemplate.rows[0].outputs[0];
if (error > acceptableError)
{
//Process the error here
Neuron outputNeuron = outputLayer.neurons[0];
//First of calculate the deltaSigma for output layer
outputNeuron.nodeDelta = (-1 * error) * outputNeuron.fireRule.fireNeuronDerivative(outputNeuron.inputs, outputNeuron.bias);
// and hidden layer (no input layer)
foreach (Neuron neuron in hiddenLayer.neurons)
{
neuron.nodeDelta = outputNeuron.nodeDelta * (neuron.fireRule.fireNeuronDerivative(neuron.inputs, neuron.bias) + neuron.output.weight);
// need to add sum of weights going out
}
//EXperimental !!!
foreach (Neuron neuron in inputLayer.neurons)
{
neuron.nodeDelta = outputNeuron.nodeDelta * (neuron.fireRule.fireNeuronDerivative(neuron.inputs, neuron.bias) + neuron.output.weight);
// need to add sum of weights going out
}
//Calculate the gradient
//For the output
outputNeuron.output.gradient = outputNeuron.output.weightedReading * outputNeuron.nodeDelta;
foreach (Neuron neuron in hiddenLayer.neurons)
{
neuron.output.gradient = neuron.output.weightedReading * outputNeuron.nodeDelta;
}
//Experimental!!!
foreach (Neuron neuron in inputLayer.neurons)
{
neuron.output.gradient = neuron.output.weightedReading * outputNeuron.nodeDelta;
}
//Calculate the new weights
//For the output layer
outputNeuron.deltaWeight = learningRate * outputNeuron.output.gradient + moementum * outputNeuron.deltaWeight;
outputNeuron.output.weight += outputNeuron.deltaWeight;
//for the hidden layer
foreach (Neuron neuron in hiddenLayer.neurons)
{
neuron.deltaWeight = learningRate * neuron.output.gradient + moementum * neuron.deltaWeight;
neuron.output.weight += neuron.deltaWeight;
}
//For the input layer ?
foreach (Neuron neuron in inputLayer.neurons)
{
neuron.deltaWeight = learningRate * neuron.output.gradient + moementum * neuron.deltaWeight;
neuron.output.weight += neuron.deltaWeight;
}
}
// I mark that I've finished these generation
currentGeneration++;
//I check the conditions
stableGenFlag = stableGenerations < stableLimit;
genLimitFlag = currentGeneration < maxGenerations;
templateFlag = currentRow < trainingTemplate.rows.Count;
//the breaking conditions
if (!stableGenFlag)
{
//Console.WriteLine("Ended due to stable limit gen") ;
return(error);
}
if (!genLimitFlag)
{
//Console.WriteLine("Ended due to limit of gens to train");
return(error);
}
} while (error > acceptableError);
}
;
//IAppDomainSetup still need to return the error history
return(error);
}