/// <summary>
/// Main constructor for evolution. Requires a configuration instance
/// </summary>
/// <param name="config"></param>
public Evolution(EvolutionConfig config)
{
// Setup new lists
genomes = new List<Genome>();
species = new List<Species>();
// Reset basic variables
evaluationIndex = 0;
speciesFitnessSum = 0;
generationNumber = 0;
topScore = 0;
// Set the configuration
evoConfig = config;
// Create new random genomes
for (int i = 0; i < config.PopulationSize; i++)
genomes.Add(new Genome(config));
// Mutate each genenome, add add them to a species
foreach (Genome g in genomes) {
g.Mutate();
AddGenomeToSpecies(g);
}
// Setup netwrok
evaluationNetwork = new NeuralNetwork(genomes[0]);
}
/// <summary>
/// After the network has been used, we need to evaluate the network, assign a fitness score and move on.
/// </summary>
/// <param name="fitnessScore"></param>
public void EvaluateNetwork(double fitnessScore)
{
// Update the fitness of this genome
genomes[evaluationIndex++].Fitness = fitnessScore;
// When the last genome was evaluated, setup the next generation
if (evaluationIndex >= genomes.Count()) {
NextGeneration();
onGenerationEnded(EventArgs.Empty);
}
else // Setup next network
evaluationNetwork = new NeuralNetwork(genomes[evaluationIndex]);
}
/// <summary>
/// This constructor creates a new generation from a previous generation
/// </summary>
/// <param name="evoConfig"></param>
/// <param name="previousGeneration"></param>
public void NextGeneration()
{
// Clear the list of genomes (we want to update them)
genomes.Clear();
List<Species> nextSpecies = new List<Species>(species);
foreach (Species s in species) {
// 'Prune' the species
s.KeepTopHalf();
// Remove old 'dead' species that have not made any progress in a while
s.UpdateStaleness();
if (s.Staleness >= evoConfig.StaleSpecies && s.TopFitness < species.Max().TopFitness)
nextSpecies.Remove(s);
}
// Update species
species = nextSpecies;
nextSpecies = new List<Species>(species);
UpdateSpeciesFitnessSum();
// Remove species that are not going to be bred because they have low fitness sums
foreach (Species s in species) {
if (evoConfig.PopulationSize * s.AverageFitness / speciesFitnessSum < 1 && s.TopFitness < species.Max().TopFitness)
nextSpecies.Remove(s);
}
// Update species
species = nextSpecies;
UpdateSpeciesFitnessSum();
// Breed new children
List<Genome> children = new List<Genome>();
foreach (Species s in species) {
// Estimate how many children should be bred per species
int breedCount = (int) (evoConfig.PopulationSize * s.AverageFitness / speciesFitnessSum);
// Make new children
for (int i = 0; i < breedCount; i++)
children.Add(Genome.BreedChild(evoConfig, s));
// Keep the best from each species
s.KeepTop(1);
// Add them to the list of genomes
foreach(Genome g in s.genomes)
genomes.Add(g);
}
// Find the top score
topScore = genomes.Max(g => g.Fitness);
// Since we only estimated how many children should be bred per species, we can't be sure we actually fit the number required (decimals)
// So we need to add more children by randomly selecting species and breeding them again
while (children.Count < evoConfig.PopulationSize)
children.Add(Genome.BreedChild(evoConfig,
species.ElementAt(
evoConfig.rng.Next(species.Count()))));
// Assign each child a species (most likely within its same species, but sometimes they get adopted)
foreach (Genome g in children) {
AddGenomeToSpecies(g);
// Add to global genome list
genomes.Add(g);
}
// Reset evaluation index;
evaluationIndex = 0;
// Setup netwrok
evaluationNetwork = new NeuralNetwork(genomes[0]);
// Increment generation counter
generationNumber++;
}
