/// <summary>
/// Returns tuple containing 1) percentage of weight values that differ, and 2) the absolute difference in weight values of dna2 as a percentage of the aggregated absolute weight values of dna1
/// </summary>
public static System.Tuple <float, double> CompareWeights(Dna dna1, Dna dna2)
{
if (!Dna.TopologiesEqual(dna1, dna2))
{
throw new System.ArgumentException("Can't compare dna weights of different topologies!");
}
if (dna1.WeightsAndBiases.Count != dna2.WeightsAndBiases.Count) // should always be false if topologies are the same. but just in case...
{
throw new System.ArgumentException("Weights arrays not equal???");
}
double accumulatedAbsoluteDifference = 0;
int nWeightsDiffering = 0;
for (int i = 0; i < dna1.WeightsAndBiases.Count; i++)
{
double absoluteDifference = System.Math.Abs(dna1.WeightsAndBiases[i] - dna2.WeightsAndBiases[i]);
accumulatedAbsoluteDifference += absoluteDifference;
if (absoluteDifference > 0)
{
nWeightsDiffering++;
}
}
double dna1TotalWeight = dna1.WeightsAndBiases.Aggregate(0.0, (total, weight) => total + System.Math.Abs(weight));
double weightDifferenceAsProportion = accumulatedAbsoluteDifference / dna1TotalWeight;
return(new System.Tuple <float, double>((float)nWeightsDiffering / (float)dna1.WeightsAndBiases.Count, weightDifferenceAsProportion));
}
public static void DebugDnaDiff(Dna dna1, Dna dna2, string relation)
{
var comparison = CompareWeights(dna1, dna2);
string log = System.String.Format("{0} | {1:P2} of weights are different, {2:P2} absolute value difference", relation, comparison.Item1, comparison.Item2);
Debug.Log(log);
}
public void AddDna(Dna newDna, bool enforceDistinctive)
{
Dna addition = enforceDistinctive && GenePool.Any(dna => dna.Equals(newDna)) ?
Dna.CloneAndMutate(newDna, newDna.Heritage, species.MutationSeverity, species.ActivationMutationSeverity) :
newDna;
GenePool.Add(addition);
Composition.TryGetValue(addition.Heritage, out int count);
Composition[addition.Heritage] = count + 1;
}
private static List <Dna> GenerateRandomDna(CarSpecies species, int number)
{
return(Enumerable.Range(0, number).Select((_) =>
Dna.GenerateRandomDnaEncoding(
species.Inputs,
species.HiddenLayersNeuronCount,
CarSpecies.Outputs,
species.OutputLayerActivation,
species.HeterogeneousHiddenActivation
)
).ToList());
}
public static Dna SelectRandomBasedOnFitness(List <Dna> parentPool, Dna excluding = null)
{
List <Dna> candidates = excluding == null ? parentPool : parentPool.Where(p => p != excluding).ToList();
float totalFitness = candidates.Aggregate(0f, (total, candidate) => total + candidate.RawFitnessRating); // TODO: optimise this
List <KeyValuePair <Dna, float> > candidateChances = candidates.ConvertAll(c => new KeyValuePair <Dna, float>(c, c.RawFitnessRating / totalFitness));
float diceRoll = Random.Range(0f, 1f);
float cumulative = 0f;
for (int i = 0; i < candidateChances.Count; i++)
{
cumulative += candidateChances[i].Value;
if (diceRoll < cumulative)
{
return(candidateChances[i].Key);
}
}
Debug.LogWarning("Failed to choose new random parent by fitness...");
return(candidates[Random.Range(0, candidates.Count)]);
}
public static Generation FromPrevious(Generation previous, int spawnLocationIndex)
{
CarSpecies species = previous.species;
List <Dna> previousGenePool = previous.GenePool;
Generation TNG = new Generation(species, previous.GenerationNumber + 1, spawnLocationIndex);
// Create and add fresh dna into next gen
int nNew = Mathf.RoundToInt(species.GenerationSize * species.NewDnaRate);
if (nNew > 0)
{
TNG.AddMultipleDna(GenerateRandomDna(species, nNew), false);
}
// Preserve elites
var previousGenerationOrderedByFitness = previousGenePool.OrderByDescending((dna => dna.RawFitnessRating));
int nElites = Mathf.RoundToInt(species.GenerationSize * species.ProportionUnchanged);
if (nElites > 0)
{
TNG.AddMultipleDna(previousGenerationOrderedByFitness.Take(nElites).Select(dna => Dna.Clone(dna)), true);
}
// Add mutated versions of elites
int nMutatedElites = Mathf.RoundToInt(species.GenerationSize * species.ProportionMutatantsOfUnchanged);
if (((species.GenerationSize - (nElites + nMutatedElites + nNew)) % 2 == 1))
{
nMutatedElites++; // make sure remaining spaces for offspring is an even number
}
for (int i = 0; i < nMutatedElites; i++)
{
Dna randomElite = Darwin.SelectRandomBasedOnFitness(
previousGenerationOrderedByFitness.Take(Mathf.RoundToInt(species.GenerationSize * 0.2f)).ToList()
);
Dna mutatedElite = Dna.CloneAndMutate(randomElite, DnaHeritage.MutatedElite, species.MutationSeverity, species.ActivationMutationSeverity);
TNG.AddDna(mutatedElite, true);
}
// Populate the rest with offspring of previous
int nMutatedOffspring = 0;
int freeSpacesForOffspring = species.GenerationSize - (nElites + nMutatedElites + nNew);
int targetNumMutatedOffspring = Mathf.RoundToInt(freeSpacesForOffspring * species.OffspringMutationProbability);
for (int i = 0; i < Mathf.RoundToInt(freeSpacesForOffspring / 2); i++)
{
Dna parent1 = Darwin.SelectRandomBasedOnFitness(previousGenePool);
Dna parent2 = Darwin.SelectRandomBasedOnFitness(previousGenePool, parent1);
List <Dna> children = new List <Dna>(2);
/* Attempts at crossover may fail if the genomes of the 2 parents are too similar. If for example:
* p1 = 1,2,3,4
* p2 = 1,2,3,5
* then no matter how we try to cross them, the children will end up as clones of the parents. To mitigate
* this we try a few times and if we consistently fail, then we mutate the dna as a fallback
*/
bool crossoverFailed = false;
for (int crossoverAttempt = 0; crossoverAttempt < 5; crossoverAttempt++)
{
children = Dna.CreateOffspring(
parent1,
parent2,
species.CrossoverPasses,
species.IncludeActivationCrossover
);
crossoverFailed = parent1.Equals(children[0]) || parent1.Equals(children[1]) || parent2.Equals(children[0]) || parent2.Equals(children[1]) || children[0].Equals(children[1]);
if (!crossoverFailed)
{
break;
}
}
if (crossoverFailed)
{
Debug.LogWarning("Crossover failed after several attempts - selected parent genomes are likely too similar. Falling back to mutation");
}
if (crossoverFailed || (nMutatedOffspring <= targetNumMutatedOffspring && Random.Range(0f, 1f) < species.OffspringMutationProbability))
{
TNG.AddMultipleDna(children.ConvertAll(c => Dna.CloneAndMutate(c, DnaHeritage.MutatedOffspring, species.MutationSeverity, species.ActivationMutationSeverity)), true);
nMutatedOffspring += 2;
}
else
{
TNG.AddMultipleDna(children, true);
}
}
DnaUtils.DebugGenerationDiff(previousGenePool, TNG.GenePool);
return(TNG);
}
public Identicle(Dna dna, List <Dna> clonesInPrevGen)
{
NextGenDna = dna;
ClonesInPrevGen = clonesInPrevGen;
}
