public GeneticAlgorithmFloat(List <DataFP> dataList, int numberOfRules, int totalGenerations,
float crossoverRate, int tournamentSize, float mutationRange, float mutationRate, int?populationSize = null)
{
// Condition length inferred from supplied data.
this.boundryLength = dataList[0].cond.Length;
// Population defaults to the size of the supplied dataList.
this.populationSize = populationSize ?? dataList.Count;
this.numberOfRules = numberOfRules;
this.totalGenerations = totalGenerations;
this.crossoverRate = crossoverRate;
this.mutationRate = mutationRate;
this.mutationRange = mutationRange;
this.mutationRateOriginal = mutationRate;
this.mutationRangeOriginal = mutationRange;
this.tournamentSize = tournamentSize;
this.tournamentSizeOriginal = tournamentSize;
// Split up our data into training and evaluation sets.
IndividualFP.SetTrainingAndEvaluationData(dataList.Take(1000).ToList(),
dataList.Skip(1000).Take(1000).ToList());
generationCounter = 1;
population = new List <IndividualFP>();
fitnessLookup = new Dictionary <IndividualFP, int>();
}
/// <summary>
/// We perform a bitwise crossover on the supplied population.
/// </summary>
/// <returns></returns>
public List <IndividualFP> CrossoverSinglePoint(List <IndividualFP> population)
{
Random random = new Random();
var crossoverOutput = new List <IndividualFP>(population);
for (int i = 0; i < crossoverOutput.Count; i += 2)
{
var tempParent1 = new IndividualFP(crossoverOutput[i]);
var tempParent2 = new IndividualFP(crossoverOutput[i + 1]);
int crossoverPoint = random.Next(numberOfRules);
// Check for crossover.
if (random.NextDouble() < crossoverRate)
{
for (int j = crossoverPoint; j < numberOfRules; j++)
{
tempParent1.Rulebase[j] = crossoverOutput[i + 1].Rulebase[j];
tempParent2.Rulebase[j] = crossoverOutput[i].Rulebase[j];
}
}
crossoverOutput[i] = tempParent1;
crossoverOutput[i + 1] = tempParent2;
}
return(crossoverOutput);
}
/// <summary>
/// We generate offspring vis tournament selection.
/// </summary>
/// <param name="tournamentSize"></param>
/// <returns></returns>
public List <IndividualFP> TournamentSelection(List <IndividualFP> population)
{
Random random = new Random();
var offspring = new List <IndividualFP>();
for (int i = 0; i < population.Count; i++)
{
//Select the parents from the population
var parents = new List <IndividualFP>();
for (int j = 0; j < tournamentSize; j++)
{
//Pick random parents from the population
parents.Add(population[random.Next(populationSize)]);
}
int fitnessMax = 0;
var bestParent = new IndividualFP(population[random.Next(populationSize)]);
foreach (var parent in parents)
{
if (fitnessLookup[parent] > fitnessMax)
{
fitnessMax = fitnessLookup[parent];
bestParent = parent;
}
}
//Select the best parent to be the offspring
offspring.Add(new IndividualFP(bestParent));
}
return(offspring);
}
/// <summary>
/// Used for cloning purposes.
/// </summary>
/// <param name="clonedIndividual"></param>
public IndividualFP(IndividualFP clonedIndividual)
{
Rulebase = new List <RuleFP>();
foreach (var rule in clonedIndividual.Rulebase)
{
Rulebase.Add(new RuleFP(rule));
}
}
/// <summary>
/// We perform a uniform crossover on the supplied population.
/// </summary>
/// <param name="population"></param>
/// <returns></returns>
public List <IndividualFP> CrossoverUniform(List <IndividualFP> population)
{
Random random = new Random();
var crossoverOutput = new List <IndividualFP>(population);
for (int i = 0; i < crossoverOutput.Count; i += 2)
{
var tempParent1 = new IndividualFP(crossoverOutput[i]);
var tempParent2 = new IndividualFP(crossoverOutput[i + 1]);
// Check for crossover.
if (random.NextDouble() < crossoverRate)
{
for (int j = 0; j < numberOfRules; j++)
{
for (int k = 0; k < boundryLength; k++)
{
if (random.Next(2) > 0)
{
if (random.Next(2) > 0)
{
tempParent1.Rulebase[j].condBoundry[k].high =
crossoverOutput[i + 1].Rulebase[j].condBoundry[k].high;
tempParent2.Rulebase[j].condBoundry[k].high =
crossoverOutput[i].Rulebase[j].condBoundry[k].high;
}
else
{
tempParent1.Rulebase[j].condBoundry[k].low =
crossoverOutput[i + 1].Rulebase[j].condBoundry[k].low;
tempParent2.Rulebase[j].condBoundry[k].low =
crossoverOutput[i].Rulebase[j].condBoundry[k].low;
}
}
}
if (random.Next(2) > 0)
{
tempParent1.Rulebase[j].output = crossoverOutput[i + 1].Rulebase[j].output;
tempParent2.Rulebase[j].output = crossoverOutput[i].Rulebase[j].output;
}
}
}
crossoverOutput[i] = tempParent1;
crossoverOutput[i + 1] = tempParent2;
}
return(crossoverOutput);
}
/// <summary>
/// We initiate and evolve our GA here until it meets the termination conditions.
/// </summary>
/// <param name="requiredFitness"></param>
/// <returns></returns>
public Tuple <List <RuleFP>, int, int, int> RunGA(int requiredFitness)
{
// For tracking purposes.
int[] maximumFitness = new int[totalGenerations + 1];
double[] averageFitness = new double[totalGenerations + 1];
List <double> meanToMaxRatioOverTime = new List <double>();
// These bools are an expansion on the adaptive mutation idea.
bool stagnation = false;
bool superStagnation = false;
// We fill our population with randomly generated chromosomes.
InitiatePopulation();
// Each chromosome's fitness is calculated and appended to a lookup table.
fitnessToLookup();
// We keep the best individual to ensure it survives the selection process.
IndividualFP best = new IndividualFP(FindBestIndividual());
maximumFitness[0] = maxFitness;
averageFitness[0] = meanFitness;
// We enter the main generationasl loop of the GA.
for (int k = 0; k < totalGenerations; k++)
{
Random random = new Random();
// Tournament selection of offspring.
var offspring = TournamentSelection(population);
// We switch our crossover strategy to Uniform if gene stagnation is detected.
if (stagnation && random.Next(100) < 10)
{
offspring = CrossoverUniform(offspring);
}
else
{
offspring = CrossoverSinglePoint(offspring);
}
// We mutate our offspring.
offspring = Mutate(offspring);
// Offspring added to population.
population.Clear();
population.AddRange(offspring);
// Best fitness individual added back into population.
population[random.Next(populationSize)] = best;
fitnessToLookup();
best = new IndividualFP(FindBestIndividual());
// Check for termination condition.
if (best.fitness >= requiredFitness || k == totalGenerations - 1)
{
return(new Tuple <List <RuleFP>, int, int, int>(best.Rulebase.OrderByDescending(x => x.fitness).ToList(), generationCounter, best.fitness, best.evaluationFitness));
}
offspring.Clear();
generationCounter++;
// Getting stats.
totalFitness = 0;
maxFitness = 0;
meanFitness = 0;
foreach (var ind in fitnessLookup)
{
if (ind.Value > maxFitness)
{
maxFitness = ind.Value;
}
totalFitness += ind.Value;
}
meanFitness = (float)totalFitness / (float)populationSize;
maximumFitness[k + 1] = maxFitness;
averageFitness[k + 1] = meanFitness;
double meanToMaxRatio = (double)meanFitness / (double)maxFitness;
if (generationCounter % 10 == 0)
{
Console.WriteLine($"Gen = {generationCounter}, mean fitness = {(int)meanFitness}, max fitness = {(int)maxFitness} MR/MRan = {mutationRate:N3}/{mutationRange:N3} S:{stagnation} MToMRatio:{meanToMaxRatio:N2}");
}
// Here we set super stagnation to true if the max fitness of the population hasnt increased in 200 generations.
if (k > 200)
{
if (maxFitness == maximumFitness[k - 200])
{
superStagnation = true;
}
else
{
superStagnation = false;
}
}
// Here we keep a running average of the change in the mean to max ration over time.
meanToMaxRatioOverTime.Add(meanToMaxRatio);
// We wait until we have 60 generations worth of data.
if (meanToMaxRatioOverTime.Count > 60)
{
var recentChange = meanToMaxRatioOverTime.Skip(meanToMaxRatioOverTime.Count - 30).Take(30).Average();
// Has the mean to max ratio not changed much? if so set stagnation to true.
// Super stagnation allows for a creater gap between maxfitness and meanfitness for exploration.
if (meanToMaxRatio > (recentChange - 0.07) && meanToMaxRatio < (recentChange + 0.07) && meanToMaxRatio > (superStagnation?0.90:0.96))
{
stagnation = true;
}
else
{
stagnation = false;
}
}
// Adaptive mutation.
if (stagnation)
{
tournamentSize = 2;
mutationRate += 0.0016f;
if (random.NextDouble() < 0.95 || !superStagnation)
{
mutationRange += (mutationRange / 10);
}
else
{
double u1 = 1.0 - random.NextDouble(); //uniform(0,1] random doubles
double u2 = 1.0 - random.NextDouble();
double randStdNormal = Math.Sqrt(-2.0 * Math.Log(u1)) *
Math.Sin(2.0 * Math.PI * u2); //random normal(0,1)
double randNormal = 0.3 * randStdNormal;
if (randNormal < 0.0)
{
randNormal = -randNormal;
}
// We will randomly shift the mutation range to either very low or a random value to encourage exploration.
if (random.Next(2) > 0)
{
mutationRange = mutationRangeOriginal / 50;
}
else
{
mutationRate = mutationRateOriginal / 2;
mutationRange = (float)(randNormal);
}
}
if (mutationRate > 1.0f)
{
mutationRate = 1.0f;
}
if (mutationRange > 1.0f)
{
mutationRange = 1.0f;
}
}
else
{
// If superstagnation is active, we use an agressive adaptive mutation.
if (superStagnation)
{
tournamentSize = 2;
mutationRange -= (mutationRange / 10);
mutationRate -= 0.0016f;
if (mutationRate < mutationRateOriginal)
{
mutationRate = mutationRateOriginal / 50;
}
if (mutationRange < mutationRangeOriginal)
{
mutationRange = mutationRangeOriginal / 50;
}
}
// Otherwise its business as usual.
else
{
tournamentSize = tournamentSizeOriginal;
mutationRate = mutationRateOriginal;
mutationRange = mutationRangeOriginal;
}
}
}
return(null);
}
