private void crossover(Chromosone chromosoneA, Chromosone chromosoneB, CROSSOVERTYPE type)
{
if (!shouldCombine)
{
return;
}
switch (type)
{
case CROSSOVERTYPE.DUAL:
int pointA = NeuralNet.Random.Next(0, chromosoneA.Genes.Count);
int pointB = NeuralNet.Random.Next(0, chromosoneA.Genes.Count);
int temp = pointA;
pointA = pointA < pointB ? pointA : pointB;
pointB = temp > pointB ? temp : pointB; // assert A is smaller than B
var copyOfA = new List <Gene>(chromosoneA.Genes);
var copyOfB = new List <Gene>(chromosoneB.Genes);
chromosoneA.Genes.Clear();
chromosoneB.Genes.Clear();
chromosoneA.Genes.AddRange(new List <Gene>(copyOfA.GetRange(0, pointA)));
chromosoneA.Genes.AddRange(new List <Gene>(copyOfB.GetRange(pointA, pointB - pointA)));
chromosoneA.Genes.AddRange(new List <Gene>(copyOfA.GetRange(pointB, copyOfA.Count - pointB)));
chromosoneB.Genes.AddRange(new List <Gene>(copyOfB.GetRange(0, pointA)));
chromosoneB.Genes.AddRange(new List <Gene>(copyOfA.GetRange(pointA, pointB - pointA)));
chromosoneB.Genes.AddRange(new List <Gene>(copyOfB.GetRange(pointB, copyOfB.Count - pointB)));
break;
default:
crossover(chromosoneA, chromosoneB, CROSSOVERTYPE.DUAL); // Should never happen
break;
}
}
private void mutate(Chromosone chrom)
{
chrom.Genes.ForEach(gene => {
if (!shouldMutate)
{
return;
}
gene.generateRandom();
});
}
public void Train(List <DataSet> dataSets, int numEpochs, int numGenerations)
{
List <Chromosone> bestInEpoch = new List <Chromosone>();
for (int epoch = 0; epoch < numEpochs; epoch++)
{
string debugStr = "";
debugStr += ("**********NEW EPOCH (" + epoch + ")*************\n");
//population.individuals.ForEach(chrom => debugStr += chrom.Genes.Count + ", ");
Debug.Log(debugStr);
// Generate genes
population.regenerate();
for (int generation = 0; generation < numGenerations; generation++)
{
debugStr = "";
debugStr += ("**********NEW GENERATION (" + generation + ")*************\n");
//population.individuals.ForEach(chrom => debugStr += chrom.Genes.Count + ", ");
Debug.Log(debugStr);
population.individuals.ForEach(chrom => chrom.resetFitness());
dataSets.ForEach(ds => {
for (int individual = 0; individual < population.individuals.Count; individual++)
{
// Set weights to genes
updateWeights(population.individuals[individual].getWeights());
ForwardPropagate(ds.Values);
// Calculate fitness
// TODO: Only calculates last dataset technically, should be nested, dataset => evolve => new dataset
population.individuals[individual].calculateFitness(OutputLayer.Select(neuron => neuron.Value).ToArray(), ds.Targets, dataSets.Count);
}
// Evolve
population.evolve();
});
}
// Save best individual
bestInEpoch.Add(Population.getBest(population.individuals, false));
string outputStr = "";
var descendingPopulation = new List <Chromosone>(new List <Chromosone>(population.individuals).OrderByDescending(chrom => chrom.Fitness));
descendingPopulation.ForEach(chrom => outputStr += chrom.Fitness + ", ");
Debug.Log(outputStr + "\nBest in epoch: " + bestInEpoch.Last().Fitness);
}
// Get best individual from the epochs
Chromosone bestOverall = Population.getBest(bestInEpoch, false);
updateWeights(bestOverall.getWeights());
Debug.Log("Best overall: " + bestOverall.Fitness);
}
public void evolve()
{
Population childPopulation = new Population(0, 0, mutationRate);
while (childPopulation.individuals.Count < individuals.Count)
{
// Selection
Chromosone chromA = Selection(SELECTIONTYPE.TOURNAMENT);
Chromosone chromB = Selection(SELECTIONTYPE.TOURNAMENT);
// Cross over
crossover(chromA, chromB, CROSSOVERTYPE.DUAL);
// Mutations
mutate(chromA);
mutate(chromB);
// Add new individuals
childPopulation.individuals.Add(chromA);
childPopulation.individuals.Add(chromB);
}
individuals = new List <Chromosone>(childPopulation.individuals); // NOPE SHALLOW COPY!
}
public Chromosone(Chromosone other)
{
Genes = new List <Gene>(other.Genes);
Fitness = other.Fitness;
}