/// <summary>
/// Tell how different are 2 genomes. This is used in speciation.
/// </summary>
/// <returns>A normalized float.</returns>
public static float GeneticDistance(NEATConfig config, Genome genome1, Genome genome2)
{
var distance = 0f;
var comparisonInfo = new GenomeComparisonInfo(genome1, genome2);
var genesCountInBiggest = Mathf.Max(genome1.Genes.Count, genome2.Genes.Count);
var disjointDist = comparisonInfo.Disjoint.Count * config.speciesCompatibility.disjointImpact;
var excessDist = comparisonInfo.Excess.Count * config.speciesCompatibility.excessImpact;
distance += disjointDist / genesCountInBiggest;
distance += excessDist / genesCountInBiggest;
float totalWeightDif = 0;
foreach (var gene in comparisonInfo.Matching)
{
var weight1 = Mathf.Abs(genome1.GetGene(gene).Weight);
var weight2 = Mathf.Abs(genome2.GetGene(gene).Weight);
var diff = Mathf.Abs(weight1 - weight2) / Mathf.Max(weight1, weight2);
totalWeightDif += diff;
}
distance += totalWeightDif / comparisonInfo.Matching.Count * config.speciesCompatibility.weightMeanDifCoef;
distance /= config.speciesCompatibility.TotalImpactSum;
return(distance);
}
/// <summary>
/// Feed inputs into the neural network.
/// Starting from the output Neurons, solve all necessary neurons,
/// including recurent sequences.
/// The input and bias neurons are considered solved.
/// </summary>
/// <param name="inputs"> The given inner values of the Input layer. </param>
public float[] FeedNeuralNetwork(NEATConfig config, float[] inputs)
{
var inputCount = InputCount;
if (inputs.Length != inputCount)
{
throw new System.Exception(string.Format("Invalid number of inputs, {0} given {1} expected",
inputs.Length, inputCount));
}
int i = 0;
foreach (var inputNeuron in InputNeurons)
{
inputNeuron.InnerValue = inputs[i++];
}
solvedNeurons.Clear();
foreach (var outNeuron in OutputNeurons)
{
FeedNeuralNetworkRecursively(config.activationFunction, outNeuron, solvedNeurons);
}
return(CollectNetworkOutputs());
}
/// <summary>
/// Groups all the genomes in Species, so that they can
/// develop their own structures, without being 'distracted'
/// by other local maximums.
/// </summary>
public void Speciate(NEATConfig config, List <Genome> genomes)
{
foreach (var genome in genomes)
{
var compatibleSpecies = GetGenomeSpecies(config, SpeciesList, genome);
if (compatibleSpecies == null)
{
var newSpecies = new Species();
newSpecies.NewGenomes.Add(genome);
SpeciesList.Add(newSpecies);
}
else
{
compatibleSpecies.NewGenomes.Add(genome);
}
}
foreach (var speciesI in SpeciesList)
{
speciesI.Genomes = speciesI.NewGenomes;
speciesI.NewGenomes = new List <Genome>();
speciesI.Generation++;
}
EliminateEmptySpecies();
AssignSpeciesToGenomes(SpeciesList);
}
public static void MutateConnection(NEATConfig config, Genome genome)
{
foreach (var gene in genome.Genes)
{
gene.Mutate(config);
}
}
/// <summary>
/// Mutates the given genome.
/// First, try to mutate the structure. Only one
/// structural mutation can happen at a time.
/// Then, mutate the genome's weights, each
/// genome has a chance to be modified.
/// </summary>
/// <param name="gradient">The mutation chances are multiplied by this gradient.</param>
/// <returns>The same genome, but mutated.</returns>
public static Genome Mutate(NEATConfig config, Genome genome, float gradient = 1f)
{
float rand = Random.Range(0, 1f);
// Mutate structure. It's a weighted random choice.
if (rand <= config.structMutation.connectionAddProb * gradient)
{
MutateAddGene(config, genome);
}
else if ((rand -= config.structMutation.connectionAddProb) <= config.structMutation.connectionDelProb * gradient)
{
MutateDelGene(genome);
}
else if ((rand -= config.structMutation.connectionDelProb) <= config.structMutation.nodeAddProb * gradient)
{
MutateAddNode(genome);
}
else if ((rand -= config.structMutation.nodeAddProb) <= config.structMutation.nodeDelProb * gradient)
{
MutateDelNode(genome);
}
// Mutate weight values.
foreach (var gene in genome.Genes)
{
if (Random.Range(0, 1.0f) <= config.weightMutation.prob * gradient)
{
gene.Mutate(config);
}
}
return(genome);
}
public static Gene MutateAddGene(NEATConfig config, Genome genome)
{
List <Neuron> possibleStartingNodes = new List <Neuron>(genome.Neurons);
while (possibleStartingNodes.Count > 0)
{
Neuron startingNode = possibleStartingNodes[Random.Range(0, possibleStartingNodes.Count)];
var alreadyConnectedNodes = startingNode.OutNeurons;
Neuron[] possibleEndNodes = genome.Neurons.Where(x =>
x.Type != ENeuronType.Input &&
x.Type != ENeuronType.Bias &&
!alreadyConnectedNodes.Contains(x)
).ToArray();
if (possibleEndNodes.Length != 0)
{
Neuron endNeuron = possibleEndNodes[Random.Range(0, possibleEndNodes.Length)];
var newConnection = new Gene(
startingNode,
endNeuron,
(config == null) ? 0 : config.NewRandomWeight());
startingNode.OutGenes.Add(newConnection);
endNeuron.InGenes.Add(newConnection);
genome.Genes.Add(newConnection);
return(newConnection);
}
else
{
possibleStartingNodes.Remove(startingNode);
}
}
return(null);
}
/// <summary>
/// Get N genomes from this species.
/// The best genome is simply copied.
/// A part of the genomes are soft copied. Soft, means that the mutation
/// is less than normal.
/// The rest, are a result of a crossover between weighted random chosen
/// genomes. The random is weighted, based on the fitness. The parents can't
/// be the same.
/// </summary>
public Genome[] GetNGenomesForNextGeneration(NEATConfig config, int n)
{
var result = new Genome[n];
Genomes = Genomes.OrderByDescending(x => x.Fitness).ToList();
for (int i = 0; i < n; i++)
{
if (i == 0)
{
result[i] = new Genome(Genomes[0]);
}
else if (i <= config.partOfGenomesToCopyForNextGenerations * n || Genomes.Count == 1)
{
result[i] = new Genome(FitnessWeightedRandomChoice(Genomes)).Mutate(config, 0.8f);
}
else
{
var parent1 = FitnessWeightedRandomChoice(Genomes, config.weightedRandomGradient);
var parent2 = FitnessWeightedRandomChoice(Genomes.Where(x => x != parent1), config.weightedRandomGradient);
result[i] = Genome.Crossover(config, parent1, parent2).Mutate(config);
}
}
return(result);
}
public Genome(NEATConfig config, PackedGenome packedGenome)
{
SpeciesId = null;
Fitness = 0;
CreateNetwork(config.inputCount,
config.outputCount,
packedGenome.genes.Select(x => new Gene(x))
);
}
/// <summary>
/// There is a config.WeightMutateNewValProb chance to
/// reinitialize the Weight.
/// </summary>
public void Mutate(NEATConfig config)
{
if (Random.Range(0, 1.0f) < config.weightMutation.NewValProb)
{
Weight = config.NewRandomWeight();
}
else
{
Weight += config.GetAWeightMutation();
}
}
public Population(int genomeCount, int inCount, int outCount, NEATConfig config)
{
SpeciesCtrl = new SpeciesControl();
GenomeCount = genomeCount;
InputCount = inCount;
OutputCount = outCount;
Config = config;
Genomes = new List <Genome>(GenomeCount);
Populate();
}
public static Genome LoadGenome(NEATConfig config, string filePath)
{
if (!File.Exists(filePath))
{
throw new Exception(filePath + " doesn't exist.");
}
var bf = new BinaryFormatter();
FileStream file = File.Open(filePath, FileMode.Open);
var packedGenome = (PackedGenome)bf.Deserialize(file);
file.Close();
return(new Genome(config, packedGenome));
}
/// <summary>
/// Create a new offspring by combining the 2 parents' genes.
/// The Matching genes are already randomly chosen from parent1 or p2.
/// </summary>
/// <returns>A fresh Genome (not mutated yet).</returns>
public static Genome Crossover(NEATConfig config, Genome parent1, Genome parent2)
{
var fittestParent = (parent1.Fitness > parent2.Fitness) ? parent1 : parent2;
var genesToInherit = new List <Gene>(fittestParent.Genes.Count);
var comparisonInfo = new GenomeComparisonInfo(parent1, parent2);
// The matching genes are chosen randomly from target1 or target2
genesToInherit.AddRange(comparisonInfo.Matching.Select(x => new Gene(x)));
// Add the disjoint genes from the fittests.
genesToInherit.AddRange(comparisonInfo.Disjoint.Where(x => x.Value == fittestParent)
.Select(x => new Gene(x.Key)));
// Add the excess genes from the fittests.
genesToInherit.AddRange(comparisonInfo.Excess.Where(x => x.Value == fittestParent)
.Select(x => new Gene(x.Key)));
return(new Genome(fittestParent, genesToInherit));
}
private string _mutationLogs = ""; // Mutation Logs
public Evaluator(NEATConfig config, Genome startingGenome, InnovationCounter nodeInnovation, InnovationCounter connectionInnovation)
{
_config = config;
GenerationNumber = 0;
_nodeInnovation = nodeInnovation;
_connectionInnovation = connectionInnovation;
_genomes = new List <Genome>();
for (int i = 0; i < _config.populationSize; ++i)
{
_genomes.Add(Genome.GenomeRandomWeights(startingGenome, _r));
}
_nextEvaluationGenomes = new List <Genome>();
_species = new List <Species>();
_genomesSpecies = new Dictionary <Genome, Species>();
BestFitness = -1f;
BestGenome = null;
}
/// <summary>
/// Determines the species in which the given genome is located.
/// Two genomes are in the same species if their genetic distance
/// is less than the given compatibility treshold.
/// </summary>
private Species GetGenomeSpecies(NEATConfig config, List <Species> species, Genome genome)
{
if (species == null)
{
return(null);
}
foreach (var speciesI in species)
{
if (speciesI.Genomes.Count == 0)
{
continue;
}
var threshold = config.speciesCompatibility.threshold;
var geneticDist = Genome.GeneticDistance(config, genome, speciesI.Representative);
if (geneticDist <= threshold)
{
return(speciesI);
}
}
return(null);
}
public Genome Mutate(NEATConfig config, float gradient = 1f)
{
return(Mutator.Mutate(config, this, gradient));
}
public bool DidntImproveInLastGenerations(NEATConfig config)
{
return(GenerationsSinceFitnesImprovement >= config.maxStagnation);
}