private void Mutate(int individualIndex)//a helper method which calls itself recursively to make mutation truly probabalistic
{
if (RandHolder.NextDouble() < mutationChance)
{
individuals[individualIndex].Mutate();
Mutate(individualIndex);
}
}
public IGeneticIndividual TrainGeneration(int numGenerations)//execute a number of iterations of the algorithm creating more fit solutions. Return most fit individual
{
for (int i = 0; i < numGenerations; i++)
{
generationCount++;
//create elite array
IGeneticIndividual[] eliteIndividuals = new IGeneticIndividual[numToSave];
for (int j = 0; j < numToSave; j++)
{
eliteIndividuals[j] = individuals[j];
}
//create 2 dim array of parents(1) for each child(0)
IGeneticIndividual[,] reproductionGroups = new IGeneticIndividual[numChildrenNeeded, numParents]; //uses tournament selection process to fill array
for (int childIter = 0; childIter < reproductionGroups.GetLength(0); childIter++) //iterate once per child
{
for (int parentIter = 0; parentIter < reproductionGroups.GetLength(1); parentIter++) //iterate once per parent. This will
{
IGeneticIndividual[] tournamentSubset = new IGeneticIndividual[tournamentSize];
for (int touranmentIter = 0; touranmentIter < tournamentSubset.Length; touranmentIter++)
{
tournamentSubset[touranmentIter] = individuals[(int)(RandHolder.NextDouble() * populationSize)];//pick a random individual
}
Array.Sort(tournamentSubset);
reproductionGroups[childIter, parentIter] = tournamentSubset[0];
}
}
//generate children
IGeneticIndividual[] newChildren = new IGeneticIndividual[numChildrenNeeded];//create empty array to hold newly generated children
for (int childIter = 0; childIter < newChildren.Length; childIter++)
{
IGeneticIndividual[] reproductionGroup = new IGeneticIndividual[numParents];//fill the single dimensioned array, reproductionGroup, with its corresponding individuals from the 2 dimensional array, reproductionGroups.
for (int parentIter = 0; parentIter < reproductionGroup.Length; parentIter++)
{
reproductionGroup[parentIter] = reproductionGroups[childIter, parentIter];
}
newChildren[childIter] = reproductionGroup[0].Reproduce(reproductionGroup, numCrossoverPoints, 1)[0];//generate a single child with reproductionGroup as its parents. Method called by reproductionGroup[0] for convenience.
}
//mutate over children array
for (int mutationIter = 0; mutationIter < populationSize; mutationIter++)
{
Mutate(mutationIter);
}
//combine elite and children array and copy into individuals array
for (int eliteIter = 0; eliteIter < eliteIndividuals.Length; eliteIter++)//copy the high fitness individuals that were saved from the previous generation
{
individuals[eliteIter] = eliteIndividuals[eliteIter];
}
for (int childIter = 0; childIter < newChildren.Length; childIter++)//copy the newly generated children after the elite individuals
{
individuals[eliteIndividuals.Length + childIter] = newChildren[childIter];
}
FitnessSort(); //order the individuals by fitness
}
return(individuals[0]); //the highest fitness
}
void RandomizeWeights() //randomizes biases and weights
{
for (int i = 0; i < weights.GetLength(0); i++)
{
for (int j = 0; j < weights.GetLength(1); j++)
{
weights[i, j] = RandHolder.NextDouble() * 2 - 1;//from [-1, 1)
}
}
}
public void Mutate()
{
int indexToMutate = (int)(RandHolder.NextDouble() * genes.Length);
if (RandHolder.NextDouble() > 0.5)
{
genes[indexToMutate] = 1;
}
else
{
genes[indexToMutate] = 0;
}
}
public IGeneticIndividual[] Reproduce(IGeneticIndividual[] IParents, int numCrossoverPoints, int numChildren)
{
int numParents = IParents.Length;//convenient variable to have
//convert array of type IGeneticIndividual to type TestGeneticIndividual
TestGeneticIndividual[] parents = new TestGeneticIndividual[numParents];
for (int i = 0; i < numParents; i++)
{
parents[i] = IParents[i] as TestGeneticIndividual;
}
TestGeneticIndividual[] children = new TestGeneticIndividual[numChildren]; //variable to hold generated children. Will be output by method
for (int childIter = 0; childIter < numChildren; childIter++) //iterate once for each child to be generated
{
int[] crossoverPoints = new int[numCrossoverPoints];
//fill crossoverPoints array with random ints which are indexes of genes array
for (int i = 0; i < crossoverPoints.Length; i++)
{
crossoverPoints[i] = (int)(RandHolder.NextDouble() * parents[0].genes.Length);
}
int activeParentIndex = 0;
//generate child
children[childIter] = new TestGeneticIndividual(parents[0].genes.Length);
for (int i = 0; i < parents[0].genes.Length; i++)
{
children[childIter].genes[i] = parents[activeParentIndex].genes[i];
for (int iter = 0; iter < numCrossoverPoints; iter++)
{
if (i == crossoverPoints[iter])
{
int temp = (int)(RandHolder.NextDouble() * (parents.Length - 1));//minus one for value exclusion, so between first and second to last index
if (temp != activeParentIndex)
{
activeParentIndex = temp;
}
else
{
activeParentIndex = parents.Length - 1;//minus one because it is max index of array
}
}
}
}
}
//convert array of type TestGeneticIndividual to type IGeneticIndividual
IGeneticIndividual[] IChildren = new IGeneticIndividual[numChildren];
for (int i = 0; i < numChildren; i++)
{
IChildren[i] = children[i] as IGeneticIndividual;
}
return(IChildren);
}
public void Randomize()
{
for (int i = 0; i < genes.Length; i++)
{
if (RandHolder.NextDouble() > 0.5)
{
genes[i] = 1;
}
else
{
genes[i] = 0;
}
}
}
public void Mutate()//set a single random weight to a random value from 0 to 1
{
//layer to mutate//
int i = (int)(RandHolder.NextDouble() * (numHiddenLayers + 1)); //+1 for output
//node to mutate//
int j = -1; //should cause an index out of bounds if not set by subsequent if cases
if (i == 0) //if first hidden layer (same as else case)(same number of nodes as hidden layer so doesn't matter)
{
j = (int)(RandHolder.NextDouble() * hiddenLayerSize);
}
else if (i == numHiddenLayers) //if output layer
{
j = (int)(RandHolder.NextDouble() * numOutputs);
}
else //if any other hidden layer
{
j = (int)(RandHolder.NextDouble() * hiddenLayerSize);
}
//weight to mutate//
int k = -1; //should cause an index out of bounds if not set by subsequent if cases
if (i == 0) //if first hidden layer (same as else case)(same number of nodes as hidden layer so doesn't matter)
{
k = (int)(RandHolder.NextDouble() * (numInputs + 1)); //+1 for bias
}
else if (i == numHiddenLayers) //if output layer same as else case
{
k = (int)(RandHolder.NextDouble() * (hiddenLayerSize) + 1); //+1 for bias
}
else//if any other hidden layer
{
k = (int)(RandHolder.NextDouble() * (hiddenLayerSize) + 1);//+1 for bias
}
//set new value;
double currentValue = Layers[i].Weights[j, k];
double newValue = (RandHolder.NextDouble() + 0.5) * currentValue; //[0.5, 1.5)
if (RandHolder.NextDouble() > 0.5)
{
newValue = -newValue;//(-1.5, -0.5] union [0.5, 1.5)
}
Layers[i].Weights[j, k] = newValue;
}
public IGeneticIndividual[] Reproduce(IGeneticIndividual[] IParents, int numCrossoverPoints, int numChildren) //returns an array of nets given an array of all parents.
// Doesn't necessarily need to be called by one of the parents but that's probably convenient.
//numCrossover points SHOULD be > numParents-1 (to use all parents at least once), but it's fine if not
{
int numParents = IParents.Length;//convenient variable to have
//convert array of type IGeneticIndividual to type NeuralNet
NeuralNet[] parents = new NeuralNet[numParents];
for (int i = 0; i < numParents; i++)
{
parents[i] = IParents[i] as NeuralNet;
}
NeuralNet[] children = new NeuralNet[numChildren]; //variable to hold generated children. Will be output by method
for (int childIter = 0; childIter < numChildren; childIter++) //iterate once for each child to be generated
{
//generate random crossoverPoints
int[,] crossoverPoints = new int[numCrossoverPoints, 3]; //3 because (i, j, k) are (layer, node, weight)
for (int cpIter = 0; cpIter < numCrossoverPoints; cpIter++) //iterate for each crossover point and generate it. Code copied from Mutate()
{
//layer to cp//
int i = (int)(RandHolder.NextDouble() * (parents[0].numHiddenLayers + 1)); //+1 for output
//node to cp//
int j = -1; //should cause an index out of bounds if not set by subsequent if cases
if (i == 0)
{ //if first hidden layer (same as else case)(same number of nodes as hidden layer so doesn't matter)
j = (int)(RandHolder.NextDouble() * parents[0].hiddenLayerSize);
}
else if (i == parents[0].numHiddenLayers)
{//if output layer
j = (int)(RandHolder.NextDouble() * parents[0].numOutputs);
}
else
{//if any other hidden layer
j = (int)(RandHolder.NextDouble() * parents[0].hiddenLayerSize);
}
//weight to cp//
int k = -1; //should cause an index out of bounds if not set by subsequent if cases
if (i == 0) //if first hidden layer (same as else case)(same number of nodes as hidden layer so doesn't matter)
{
k = (int)(RandHolder.NextDouble() * (parents[0].numInputs) + 1); //+1 for bias
}
else if (i == parents[0].numHiddenLayers) //if output layer same as else case
{
k = (int)(RandHolder.NextDouble() * (parents[0].hiddenLayerSize) + 1); //+1 for bias
}
else//if any other hidden layer
{
k = (int)(RandHolder.NextDouble() * (parents[0].hiddenLayerSize) + 1);//+1 for bias
}
//assign i,j,k to crossoverPoints
crossoverPoints[cpIter, 0] = i;
crossoverPoints[cpIter, 1] = j;
crossoverPoints[cpIter, 2] = k;//duplicate crossover points might exist but honestly who cares
}
//create array to hold new layers, nodes, weights
Layer[] newLayers = new Layer[parents[0].numHiddenLayers + 1];
//fill newLayers
int activeParentIndex = 0;
for (int i = 0; i < parents[0].numHiddenLayers + 1; i++) //+1 for output layer
{
if (i == 0) //first hidden layer
{
double[,] newWeights = new double[parents[0].hiddenLayerSize, parents[0].numInputs + 1]; //create empty container, +1 for bias
for (int j = 0; j < newWeights.GetLength(0); j++)
{
for (int k = 0; k < newWeights.GetLength(1); k++)//this will run for each weight and bias
{
//check if currently at a crossover point if so then swap the activeParentIndex
for (int iter = 0; iter < numCrossoverPoints; iter++)//check for each crossover point
{
if (crossoverPoints[iter, 0] == i && crossoverPoints[iter, 1] == j && crossoverPoints[iter, 2] == k)
{
activeParentIndex = (int)(RandHolder.NextDouble() * (numParents));//choose a random parent including currently active parent
}
}
newWeights[j, k] = parents[activeParentIndex].layers[i].Weights[j, k];//copy the value from the active aprent
}
}
newLayers[i] = new Layer(parents[0].numInputs, parents[0].hiddenLayerSize, newWeights); //assign new weights to newLayers array in the form of a newly instantiated Layer
}
else if (i > 0 && i < parents[0].numHiddenLayers) //all other hidden layers
{
double[,] newWeights = new double[parents[0].hiddenLayerSize, parents[0].hiddenLayerSize + 1]; //create empty container, +1 for bias
for (int j = 0; j < newWeights.GetLength(0); j++)
{
for (int k = 0; k < newWeights.GetLength(1); k++)//this will run for each weight and bias
{
//check if currently at a crossover point if so then swap the activeParentIndex
for (int iter = 0; iter < numCrossoverPoints; iter++)//check for each crossover point
{
if (crossoverPoints[iter, 0] == i && crossoverPoints[iter, 1] == j && crossoverPoints[iter, 2] == k)
{
int temp = (int)(RandHolder.NextDouble() * (numParents - 1));//the minus one is because we are avoiding the current activeParentIndex in this reassignment
if (temp < activeParentIndex)
{
activeParentIndex = temp;
}
else//if greater than or equal to activeParentIndex
{
activeParentIndex = temp + 1;
}
}
}
newWeights[j, k] = parents[activeParentIndex].layers[i].Weights[j, k];//copy the value from the active parent
}
}
newLayers[i] = new Layer(parents[0].numInputs, parents[0].hiddenLayerSize, newWeights); //assign new weights to newLayers array in the form of a newly instantiated Layer
}
else if (i == parents[0].numHiddenLayers) //output layer
{
double[,] newWeights = new double[parents[0].numOutputs, parents[0].hiddenLayerSize + 1]; //create empty container, +1 for bias
for (int j = 0; j < newWeights.GetLength(0); j++)
{
for (int k = 0; k < newWeights.GetLength(1); k++)//this will run for each weight and bias
{
//check if currently at a crossover point if so then swap the activeParentIndex
for (int iter = 0; iter < numCrossoverPoints; iter++)//check for each crossover point
{
if (crossoverPoints[iter, 0] == i && crossoverPoints[iter, 1] == j && crossoverPoints[iter, 2] == k)
{
int temp = (int)(RandHolder.NextDouble() * (numParents - 1));//the minus one is because we are avoiding the current activeParentIndex in this reassignment
if (temp < activeParentIndex)
{
activeParentIndex = temp;
}
else//if greater than or equal to activeParentIndex
{
activeParentIndex = temp + 1;
}
}
}
newWeights[j, k] = parents[activeParentIndex].layers[i].Weights[j, k];//copy the value from the active aprent
}
}
newLayers[i] = new Layer(parents[0].hiddenLayerSize, parents[0].numOutputs, newWeights);//assign new weights to newLayers array in the form of a newly instantiated Layer
}
else
{
throw new IndexOutOfRangeException();
}
}
children[childIter] = new NeuralNet(parents[0].numInputs, parents[0].numOutputs, parents[0].numHiddenLayers, parents[0].hiddenLayerSize, parents[0].TestInputSets, parents[0].TestOutputSets) //create empty with same params as parent
{
layers = newLayers //assign the data to complete the child
};
}
//convert array of type NeuralNet to type IGeneticIndividual
IGeneticIndividual[] IChildren = new IGeneticIndividual[numChildren];
for (int i = 0; i < numChildren; i++)
{
IChildren[i] = children[i] as IGeneticIndividual;
}
return(IChildren);
}
