private reinforcementNet[] PickBestPopulation()
{
reinforcementNet[] newPopulation = new reinforcementNet[initialPopulation];
for (int i = 0; i < bestAgentSelection; i++)
{
newPopulation[naturallySelected] = population[i].InitialiseCopy(controller.LAYERS, controller.NEURONS);
newPopulation[naturallySelected].fitness = 0;
naturallySelected++;
int f = Mathf.RoundToInt(population[i].fitness * 10);
for (int c = 0; c < f; c++)
{
genePool.Add(i);
}
}
for (int i = 0; i < worstAgentSelection; i++)
{
int last = population.Length - 1;
last -= i;
int f = Mathf.RoundToInt(population[last].fitness * 10);
for (int c = 0; c < f; c++)
{
genePool.Add(last);
}
}
return(newPopulation);
}
public reinforcementNet InitialiseCopy(int hiddenLayerCount, int hiddenNeuronCount)
{
reinforcementNet n = new reinforcementNet();
List <Matrix <float> > newWeights = new List <Matrix <float> >();
for (int i = 0; i < this.weights.Count; i++)
{
Matrix <float> currentWeight = Matrix <float> .Build.Dense(weights[i].RowCount, weights[i].ColumnCount);
for (int x = 0; x < currentWeight.RowCount; x++)
{
for (int y = 0; y < currentWeight.ColumnCount; y++)
{
currentWeight[x, y] = weights[i][x, y];
}
}
newWeights.Add(currentWeight);
}
List <float> newBiases = new List <float>();
newBiases.AddRange(biases);
n.weights = newWeights;
n.biases = newBiases;
n.InitialiseHidden(hiddenLayerCount, hiddenNeuronCount);
return(n);
}
//later set starting index to end because all nets up to that point have been
//created with other nets
private void FillPopulationWithRandomValues(reinforcementNet[] newPopulation, int startingIndex)
{
while (startingIndex < initialPopulation)
{
newPopulation[startingIndex] = new reinforcementNet();
newPopulation[startingIndex].Initialise(controller.LAYERS, controller.NEURONS);
startingIndex++;
}
}
private void FillPopulationWithRandomValues(reinforcementNet[] newPopulation, reinforcement2Child[] newPopulation2, int startingIndex)
{
while (startingIndex < initialPopulation)
{
newPopulation[startingIndex] = new reinforcementNet();
newPopulation[startingIndex].Initialise(controller.LAYERS, controller.NEURONS);
//second person
newPopulation2[startingIndex] = new reinforcement2Child();
newPopulation2[startingIndex].Initialise2(player2.LAYERS, player2.NEURONS);
startingIndex++;
}
}
public void Death(float fitness, reinforcementNet network)
{
if (currentGenome < population.Length - 1)
{
population[currentGenome].fitness = fitness;
Debug.Log("death first layer");
currentGenome++;
Debug.Log("death active");
ResetToCurrentGenome();
}
else
{
RePopulate();
}
}
//this helps improve the performance of current genome
public void Death(float fitness, reinforcementNet network)
{
if (currentGenome < population.Length - 1)
{
//set the previous fitness from the death of the car controller
//fitness then stored in dec order
population[currentGenome].fitness = fitness;
currentGenome++;
//reset car with different net but saved fitness
ResetToCurrentGenome();
}
else
{
//repopulate current generation
RePopulate();
}
}
private void SortPopulation()
{
//i might rework this
for (int i = 0; i < population.Length; i++)
{
for (int j = i; j < population.Length; j++)
{
if (population[i].fitness < population[j].fitness)
{
//order from high to low fitness
reinforcementNet temp = population[i];
population[i] = population[j];
population[j] = temp;
}
}
}
}
private reinforcementNet[] PickBestPopulation()
{
reinforcementNet[] newPopulation = new reinforcementNet[initialPopulation];
//diverse gene pool to make higher chance of top network over lower network
for (int i = 0; i < bestAgentSelection; i++)
{
//copy is there because without ypu would be changine the refernce population i
//passing over the best networks
//increments number of filled nets
//being added to new pop
newPopulation[naturallySelected] = population[i].InitialiseCopy(controller.LAYERS, controller.NEURONS);
newPopulation[naturallySelected].fitness = 0;
//the remaining non selected have rabdomized values
naturallySelected++;
//take fitness, multiply by ten and add to gene pool
//ie survival of the fittest
int f = Mathf.RoundToInt(population[i].fitness * 10);
for (int c = 0; c < f; c++)
{
//adding indexes to reference neural net
//example newpopulation[genepool[i]] == neural net
genePool.Add(i);
}
}
for (int i = 0; i < worstAgentSelection; i++)
{
//loop from bottom of array
int last = population.Length - 1;
last -= i;
int f = Mathf.RoundToInt(population[last].fitness * 8);
for (int c = 0; c < f; c++)
{
genePool.Add(last);
}
}
return(newPopulation);
}
private void SortPopulation()
{
for (int i = 0; i < population.Length; i++)
{
for (int j = i; j < population.Length; j++)
{
if (population[i].fitness < population[j].fitness)
{
reinforcementNet temp = population[i];
population[i] = population[j];
population[j] = temp;
}
if (population2[i].fitness < population2[j].fitness)
{
reinforcement2Child temp2 = population2[i];
population2[i] = population2[j];
population2[j] = temp2;
}
}
}
}
public void ResetWithNetwork(reinforcementNet net)
{
network = net;
Reset();
}
private void Awake()
{
startPosition = transform.position;
startRotation = transform.eulerAngles;
network = GetComponent <reinforcementNet>();
}
private void Crossover(reinforcementNet[] newPopulation, reinforcement2Child[] newPopulation2)
{
for (int i = 0; i < numberToCrossover; i += 2)
{
int AIndex = i;
int BIndex = i + 1;
if (genePool.Count >= 1)
{
for (int l = 0; l < 100; l++)
{
AIndex = genePool[Random.Range(0, genePool.Count)];
BIndex = genePool[Random.Range(0, genePool.Count)];
if (AIndex != BIndex)
{
break;
}
}
}
reinforcementNet Child1 = new reinforcementNet();
reinforcementNet Child2 = new reinforcementNet();
//
reinforcement2Child Child1_A = new reinforcement2Child();
reinforcement2Child Child2_A = new reinforcement2Child();
Child1.Initialise(controller.LAYERS, controller.NEURONS);
Child2.Initialise(controller.LAYERS, controller.NEURONS);
Child1_A.Initialise2(player2.LAYERS, player2.NEURONS);
Child2_A.Initialise2(player2.LAYERS, player2.NEURONS);
Child1.fitness = 0;
Child2.fitness = 0;
Child1_A.fitness = 0;
Child2_A.fitness = 0;
for (int w = 0; w < Child1.weights.Count; w++)
{
if (Random.Range(0.0f, 1.0f) < 0.5f)
{
Child1.weights[w] = population[AIndex].weights[w];
Child2.weights[w] = population[BIndex].weights[w];
//second person
Child2_A.weights[w] = population2[AIndex].weights[w];
Child1_A.weights[w] = population2[BIndex].weights[w];
}
else
{
Child2.weights[w] = population[AIndex].weights[w];
Child1.weights[w] = population[BIndex].weights[w];
//second person
Child1_A.weights[w] = population2[AIndex].weights[w];
Child2_A.weights[w] = population2[BIndex].weights[w];
}
}
for (int w = 0; w < Child1.biases.Count; w++)
{
if (Random.Range(0.0f, 1.0f) < 0.5f)
{
Child1.biases[w] = population[AIndex].biases[w];
Child2.biases[w] = population[BIndex].biases[w];
//second person
Child2_A.biases[w] = population2[AIndex].biases[w];
Child1_A.biases[w] = population2[BIndex].biases[w];
}
else
{
Child2.biases[w] = population[AIndex].biases[w];
Child1.biases[w] = population[BIndex].biases[w];
//second person
Child1_A.biases[w] = population2[AIndex].biases[w];
Child2_A.biases[w] = population2[BIndex].biases[w];
}
}
newPopulation[naturallySelected] = Child1;
newPopulation2[naturallySelected] = Child1_A;
naturallySelected++;
newPopulation[naturallySelected] = Child2;
newPopulation2[naturallySelected] = Child2_A;
naturallySelected++;
}
}
private void Crossover(reinforcementNet[] newPopulation)
{
//cross over based on gene pool
//increment by two
for (int i = 0; i < numberToCrossover; i += 2)
{
//first parent index
int AIndex = i;
//second parent
int BIndex = i + 1;
if (genePool.Count >= 1)
{
for (int l = 0; l < 100; l++)
{
//find random elemnt in gene pool
AIndex = genePool[Random.Range(0, genePool.Count)];
BIndex = genePool[Random.Range(0, genePool.Count)];
//ensure parents are not the same
//we found what we looking for break
if (AIndex != BIndex)
{
break;
}
}
}
reinforcementNet Child1 = new reinforcementNet();
reinforcementNet Child2 = new reinforcementNet();
Child1.Initialise(controller.LAYERS, controller.NEURONS);
Child2.Initialise(controller.LAYERS, controller.NEURONS);
Child1.fitness = 0;
Child2.fitness = 0;
//should be more subtle in practice to much transfer of info
for (int w = 0; w < Child1.weights.Count; w++)
{
if (Random.Range(0.0f, 1.0f) < 0.5f)
{
//taje weights from parents
Child1.weights[w] = population[AIndex].weights[w];
Child2.weights[w] = population[BIndex].weights[w];
}
else
{
//random chance for kids
Child2.weights[w] = population[AIndex].weights[w];
Child1.weights[w] = population[BIndex].weights[w];
}
}
//indv values instead of matices
for (int w = 0; w < Child1.biases.Count; w++)
{
if (Random.Range(0.0f, 1.0f) < 0.5f)
{
Child1.biases[w] = population[AIndex].biases[w];
Child2.biases[w] = population[BIndex].biases[w];
}
else
{
Child2.biases[w] = population[AIndex].biases[w];
Child1.biases[w] = population[BIndex].biases[w];
}
}
//add to new population
newPopulation[naturallySelected] = Child1;
naturallySelected++;
//add to new population
newPopulation[naturallySelected] = Child2;
naturallySelected++;
}
}
