private void SpawnCreatures(BitArray newGenome)
{
GameObject creaturesParent = new GameObject("creatures");
_parents.Add(creaturesParent);
for (int i = 0; i < CreatureAmount; i++)
{
CreatureBehaviour creature = Instantiate(Resources.Load("Prefabs/creature") as GameObject).GetComponent <CreatureBehaviour>();
if (newGenome != null)
{
newGenome = GeneticAlgorithm.Mutation(newGenome);
creature.Genome.SetGenome(newGenome);
}
else
{
creature.Genome.BaseMass = Random.Range(1f, 10f);
creature.Genome.VisionLength = 2f;
}
Vector2 screenBounds = new Vector2(Width, Height);
creature.transform.position = Camera.main.ScreenToWorldPoint(new Vector2(Random.Range(0f, screenBounds.x), Random.Range(0f, screenBounds.y)));
creature.transform.Rotate(0, 0, Random.Range(0, 360));
creature.transform.SetParent(creaturesParent.transform, true);
_creatures.Add(creature);
}
}
private static CreatureBehaviour ChooseWinner(CreatureBehaviour[] candidates)
{
float random = Random.value;
CreatureBehaviour winner = null;
candidates = candidates.OrderBy(x => x.AccumulatedNormalizedFitness).ToArray(); // Make sure the list is ordered right (high norm fitness = better score)
//foreach (CreatureBehaviour candidate in candidates.Reverse()) // Wikipedia doesn't mention reversing the list, but if we don't then we will always pick the latest one in the list (with the highest accNormFitness)
// if (candidate.AccumulatedNormalizedFitness >= random)
// winner = candidate;
foreach (CreatureBehaviour candidate in candidates) // Isn't this a better option? This way if random = 0.9, the closest one below 0.9 would be taken instead of the ones above 0.9 (higher accumulated fitness, the worse you are)
{
if (candidate.AccumulatedNormalizedFitness <= random)
{
winner = candidate;
}
}
// If we have no winner, take the best one (fallback)
// TODO: I don't know why we won't always take the best one? Is it to introduce more diversity? Is mutation not enough for this?
if (!winner)
{
winner = candidates.First();
}
Debug.Log($"Winner fitness: {winner.Fitness}, Best fitness: {candidates.Max(x => x.Fitness)}");
return(winner);
}
public void OnCollisionEnter2D(Collision2D col)
{
CreatureBehaviour creature = col.collider.gameObject.GetComponent <CreatureBehaviour>();
if (creature != null)
{
creature.Die();
}
}
public static CreatureBehaviour[] GetParents(List <CreatureBehaviour> candidates)
{
CreatureBehaviour parent1 = CalculateWinner(candidates);
candidates.Remove(parent1);
CreatureBehaviour parent2 = CalculateWinner(candidates);
return(new CreatureBehaviour[] { parent1, parent2 });
}
// Selection
// 1. Normalize each fitness (normFitness = fitness / sum(allFitnesses)), so that the sum of all fitnesses is 1
// 2. Sort population desc fitness
// 3. Accumulate normalized fitnesses (accFitness = normFitness + allPreviousNormFitnesses)
// ex: accFitness1 would be (normFitness1), accFitness2 would be (normFitness2 + normFitness1)
// accFitness3 would be (normFitness3 + normFitness2 + normFitness1) ...
// 4. Choose random number R between 0 and 1
// 5. Winner is the last one whose accFitness >= R
// https://en.wikipedia.org/wiki/Selection_(genetic_algorithm)
public static CreatureBehaviour CalculateWinner(List <CreatureBehaviour> candidates)
{
// Calculate total fitness of all candidates summed together
float totalFitnesses = candidates.Select(x => x.Fitness).Sum();
// Normalize fitnesses
candidates.ForEach(x => x.NormalizedFitness = NormalizeFitness(x.Fitness, totalFitnesses));
// Sort by descending fitness
candidates = candidates.OrderByDescending(x => x.Fitness).ToList();
// Accumulate normalized fitnesses
foreach (CreatureBehaviour candidate in candidates)
{
int index = candidates.IndexOf(candidate);
float[] previousFitnesses = candidates.GetRange(0, index).Select(x => x.NormalizedFitness).ToArray();
candidate.AccumulatedNormalizedFitness = AccumulateNormalizedFitness(candidate.NormalizedFitness, previousFitnesses);
}
// Choose winner
CreatureBehaviour winner = ChooseWinner(candidates.ToArray());
return(winner);
}
