private void SetWeights(NetHolder holder)
{
holder.NeuralNet.SetInputWeights(holder.Genome.Dna);
}
private void IncrementFitness(NetHolder holder, double fitness)
{
holder.Genome.Fitness += fitness;
}
private void TeachXor()
{
var population = new List<NetHolder>();
var genes = new List<Genome>();
var populationCount = 100;
var numberOfGenerations = 5000;
var trainingSet = 1000;
var initialFitness = 0;
for (int i = 0; i < populationCount; i++)
{
var neuralNet = new NeuralNet(2, 1, 2, 6);
var genome = new Genome(neuralNet.GetTotalInputs());
genome.Fitness = initialFitness; // Set initial fitness
var holder = new NetHolder() {Genome = genome, NeuralNet = neuralNet};
this.SetWeights(holder);
population.Add(holder);
genes.Add(genome);
}
var algorithm = new GeneticAlgorithm(genes, 0.1, 0.7, 0.3, (int) (populationCount * 0.1));
var random = new Random();
var inputHolders = new List<InputHolder>();
var maxValue = 1;
trainingSet = 4;
inputHolders.Add(new InputHolder() {X = 1, Y = 1});
inputHolders.Add(new InputHolder() {X = 1, Y = 0});
inputHolders.Add(new InputHolder() {X = 0, Y = 1});
inputHolders.Add(new InputHolder() {X = 0, Y = 0});
for (int i = 0; i < numberOfGenerations; i++)
{
Console.WriteLine("Generation {0}", i);
int generationCorrect = 0;
double maxError = double.MinValue;
double minError = double.MaxValue;
double averageError = 0;
int errorIteration = 0;
double minFitness = double.MaxValue;
double maxFitness = double.MinValue;
double averageFitness = 0;
int fitnessIteration = 0;
foreach (var netHolder in population)
{
netHolder.Genome.Fitness = initialFitness;
int totalCorrect = 0;
foreach (var input in inputHolders)
{
var list =
netHolder.NeuralNet.Update(new List<double>() {input.X, input.Y});
var output = list[0];
var expectedOutput = (int) input.X ^ (int) input.Y;
var error = Math.Abs(output - expectedOutput);
averageError = (averageError * errorIteration + error) /
++errorIteration;
if (netHolder.MaxError > error)
{
netHolder.MaxError = error;
}
lock (this)
{
if (error > maxError)
{
maxError = error;
}
if (error < minError)
{
minError = error;
}
}
var bonus = 100.0;
var coef = 1.0;
if (error > 0)
{
coef = (maxValue - error) / (maxValue);
}
else
{
totalCorrect++;
}
bonus = bonus * coef;
double fitnessPenalty = bonus / trainingSet;
netHolder.Genome.Fitness += fitnessPenalty;
/* Console.WriteLine("{0} -> output: {1}, expected: {2}, error: {3}", j, output, expectedOutput,
Math.Round(error, 2));*/
}
//netHolder.Genome.Fitness += totalCorrect;
generationCorrect += totalCorrect;
//Console.WriteLine("{0} -> Fitness: {1}", j, population[j].Genome.Fitness);
lock (this)
{
if (netHolder.Genome.Fitness > maxFitness)
{
maxFitness = netHolder.Genome.Fitness;
}
if (netHolder.Genome.Fitness < minFitness)
{
minFitness = netHolder.Genome.Fitness;
}
averageFitness = (averageFitness * fitnessIteration +
netHolder.Genome.Fitness) /
++fitnessIteration;
}
}
if (generationCorrect / populationCount == trainingSet)
{
Console.WriteLine("Found solution");
for (int j = 0; j < 1000; j++)
{
int x = random.Next(1);
int y = random.Next(1);
var z = population[0].NeuralNet.Update(new List<double>() {x, y})[0];
if ((x ^ y) != z)
{
}
}
}
Console.WriteLine("Fitness: min: {0} max: {1}, avg: {2}", Math.Round(minFitness, 2),
Math.Round(maxFitness, 2), Math.Round(averageFitness, 2));
Console.WriteLine("Error: min: {0} max: {1}, avg: {2}, corr: {3}", Math.Round(minError, 2),
Math.Round(maxError, 2), Math.Round(averageError, 2), generationCorrect);
var newPopulation = algorithm.Epoch();
for (int j = 0; j < newPopulation.Count; j++)
{
population[j].Genome = newPopulation[j];
}
foreach (var netHolder in population)
{
this.SetWeights(netHolder);
}
}
NetHolder best = population[0];
foreach (var netHolder in population)
{
if (netHolder.Genome.Fitness > best.Genome.Fitness)
{
best = netHolder;
}
}
best.NeuralNet.Update(0, 0);
}
private double TrainNetwork(int initialFitness, List<InputHolder> inputHolders, int maxValue, NetHolder netHolder, double averageError, int errorIteration, ref double maxError, ref double minError, ref int generationCorrect, ref double maxFitness, ref double minFitness, ref double averageFitness, int fitnessIteration)
{
netHolder.Genome.Fitness = initialFitness;
int totalCorrect = 0;
foreach (var input in inputHolders)
{
var list = netHolder.NeuralNet.Update(new List<double>() {input.X, input.Y});
var output = list[0] * maxValue;
var expectedOutput = input.X + input.Y;
var error = Math.Abs(output - expectedOutput);
averageError = (averageError * errorIteration + error) / ++errorIteration;
if (netHolder.MaxError < error)
{
netHolder.MaxError = error;
}
lock (this)
{
if (error > maxError)
{
maxError = error;
}
if (error < minError)
{
minError = error;
}
}
var bonus = 100.0;
var coef = 1.0;
if (error > 1)
{
coef = (maxValue - error) / (maxValue);
}
else
{
totalCorrect++;
}
bonus = bonus * Math.Pow(coef, 3);
double fitnessPenalty = bonus / inputHolders.Count;
netHolder.Genome.Fitness += fitnessPenalty;
/* Console.WriteLine("{0} -> output: {1}, expected: {2}, error: {3}", j, output, expectedOutput,
Math.Round(error, 2));*/
}
netHolder.Genome.Fitness += totalCorrect;
generationCorrect += totalCorrect;
//Console.WriteLine("{0} -> Fitness: {1}", j, population[j].Genome.Fitness);
lock (this)
{
if (netHolder.Genome.Fitness > maxFitness)
{
maxFitness = netHolder.Genome.Fitness;
}
if (netHolder.Genome.Fitness < minFitness)
{
minFitness = netHolder.Genome.Fitness;
}
averageFitness = (averageFitness * fitnessIteration + netHolder.Genome.Fitness) /
++fitnessIteration;
}
return averageError;
}
private void TeachSum()
{
var population = new List<NetHolder>();
var genes = new List<Genome>();
var populationCount = 100;
var numberOfGenerations = 200;
var trainingSet = 100;
var initialFitness = 0;
for (int i = 0; i < populationCount; i++)
{
var neuralNet = new NeuralNet(2, 1, 3, 6);
var genome = new Genome(neuralNet.GetTotalInputs());
genome.Fitness = initialFitness; // Set initial fitness
var holder = new NetHolder(){Genome = genome, NeuralNet = neuralNet};
this.SetWeights(holder);
population.Add(holder);
genes.Add(genome);
}
NetHolder bestEver = new NetHolder()
{
Genome = new Genome(new List<double>()) {Fitness = double.MinValue},
NeuralNet = new NeuralNet(2, 1, 3, 6)
};
var algorithm = new GeneticAlgorithm(genes, 0.1, 0.7, 0.3, (int) (populationCount * 0.1));
var random = new Random();
var inputHolders = new List<InputHolder>();
var maxNumber = 5;
var maxValue = maxNumber * 2;
inputHolders.Add(new InputHolder() {X = 0, Y = 1});
inputHolders.Add(new InputHolder() {X = 5, Y = 5});
trainingSet = inputHolders.Count;
for (int i = 0; i < numberOfGenerations; i++)
{
Console.WriteLine("Generation {0}", i);
int generationCorrect = 0;
double maxError = double.MinValue;
double minError = double.MaxValue;
double averageError = 0;
int errorIteration = 0;
double minFitness = double.MaxValue;
double maxFitness = double.MinValue;
double averageFitness = 0;
int fitnessIteration = 0;
foreach (var netHolder in population)
{
averageError = TrainNetwork(initialFitness, inputHolders, maxValue, netHolder,
averageError, errorIteration, ref maxError, ref minError,
ref generationCorrect, ref maxFitness, ref minFitness,
ref averageFitness, fitnessIteration);
}
foreach (var holder in population)
{
if (holder.MaxError < maxError)
{
var errorBonus = (maxError - holder.MaxError) * 10;
holder.Genome.Fitness += errorBonus;
}
if (holder.Genome.Fitness > bestEver.Genome.Fitness)
{
bestEver.Genome = holder.Genome;
bestEver.Genome.Fitness = holder.Genome.Fitness;
bestEver.MaxError = holder.MaxError;
this.SetWeights(bestEver);
}
}
Console.WriteLine("Fitness: min: {0} max: {1}, avg: {2}", Math.Round(minFitness, 2),
Math.Round(maxFitness, 2), Math.Round(averageFitness, 2));
Console.WriteLine("Error: min: {0} max: {1}, avg: {2}, corr: {3}", Math.Round(minError, 2),
Math.Round(maxError, 2), Math.Round(averageError, 2), generationCorrect);
Console.WriteLine("Best: fitness {0}, max error {1}%", Math.Round(bestEver.Genome.Fitness, 2),
Math.Round(bestEver.MaxError / maxValue * 100, 2));
var newPopulation = algorithm.Epoch();
for (int j = 0; j < newPopulation.Count; j++)
{
population[j].Genome = newPopulation[j];
population[j].MaxError = 0;
}
foreach (var netHolder in population)
{
this.SetWeights(netHolder);
}
}
Console.WriteLine("Completed. Testing best");
for (int i = 0; i < 10; i++)
{
double avg = 0d;
int iter = 0;
double maxError = 0;
double minError =0;
int generationCorrect=0;
double maxFitness=0;
double minFitness=0;
double averageFitness=0;
int fitnessIteration=0;
this.TrainNetwork(initialFitness, inputHolders, maxValue, bestEver, avg, iter, ref maxError, ref minError,
ref generationCorrect, ref maxFitness, ref minFitness, ref averageFitness,
fitnessIteration);
var x = random.Next(maxNumber);
var y = random.Next(maxNumber);
var z = bestEver.NeuralNet.Update(new List<double> {x, y})[0] * maxValue;
Console.WriteLine("{0} + {1} = {2}", x, y, Math.Round(z, 2));
}
}