public EA(int population, int numInputs, LayerParams[] networkStructure)
{
this.Population = new IndividualStruct[population];
for (int i = 0; i < population; i++)
{
Population[i] = new IndividualStruct(numInputs);
for (int j = 0; j < networkStructure.Length; j++)
{
this.Population[i].Individual.AddLayer(networkStructure[j].F, networkStructure[j].G, networkStructure[j].NumNeurons);
}
Population[i].Individual.BuildNetwork();
}
this.RNG = new Random();
}
private Network TournamentSelection(IndividualStruct[] population, int k, double chanceWin)
{
IndividualStruct[] tournament = new IndividualStruct[k];
for (int i = 0; i < k; i++)
{
tournament[i] = population[this.RNG.Next(population.Length)];
}
if (chanceWin > this.RNG.NextDouble())
{
return(tournament[this.RNG.Next(k)].Individual);
}
else
{
return(tournament.OrderBy(f => f.Fitness).ElementAt(0).Individual);
}
}
public void EvolutionaryLoop(DataStruct[] dataset, int generations, int numChildren, double crossoverRate, double mutationRate, int kTournament, double chanceWin)
{
for (int i = 0; i < generations; i++)
{
double bestFitness = int.MaxValue;
double averageFitness = 0;
IndividualStruct[] newPopulation = new IndividualStruct[this.Population.Length];
for (int j = 0; j < this.Population.Length; j++)
{
this.Population[j].Fitness = Fitness(dataset, this.Population[j]);
if (this.Population[j].Fitness < bestFitness)
{
bestFitness = this.Population[j].Fitness;
}
averageFitness += this.Population[j].Fitness;
}
if (i != generations - 1)
{
var survivors = this.Population.OrderBy(x => x.Fitness).ThenBy(x => x.FromGeneration);
for (int c = 0; c < numChildren; c++)
{
newPopulation[c] = new IndividualStruct(this.Crossover(this.TournamentSelection(this.Population, kTournament, chanceWin), this.TournamentSelection(this.Population, kTournament, chanceWin), crossoverRate), i + 1);
}
for (int a = numChildren; a < this.Population.Length; a++)
{
newPopulation[a] = survivors.ElementAt(a - numChildren);
}
for (int ind = 0; ind < newPopulation.Length; ind++)
{
if (this.RNG.NextDouble() < mutationRate)
{
this.Mutation(newPopulation[ind].Individual);
newPopulation[ind].NumMutations += 1;
}
}
this.Population = newPopulation;
}
Console.WriteLine($"Generation: {i + 1}");
Console.WriteLine($"Best fitness: {bestFitness}");
Console.WriteLine($"Average fitness: {averageFitness / this.Population.Length}");
Console.WriteLine("-----------------------------------");
}
}
public double Fitness(DataStruct[] dataset, IndividualStruct individual)
{
//TODO avoid duplicate code from the supervised learning implementation in network (Train())
double accumulatedError = 0;
foreach (var example in dataset)
{
double[] result = individual.Individual.Inference(example.Input);
double[] delta = individual.Individual.Compare(example.Output, result);
double[] MSE = new double[delta.Length];
for (int iDelta = 0; iDelta < delta.Length; iDelta++)
{
MSE[iDelta] = Math.Pow(delta[iDelta], 2);
}
accumulatedError += MSE.Sum() / MSE.Length;
}
return(accumulatedError);
}