internal void Run()
{
population = new List <Hunter>(PopSize);
bool successfulLoad = loadBestHunter();
for (int i = (successfulLoad? 1:0); i < PopSize; ++i)
{
Hunter temp = new Hunter().StripChromosomes();
int size = (int)Math.Ceiling(Math.Log(OptoGlobals.NumberOfClasses, 2));
for (int j = 0; j < OptoGlobals.NumberOfClasses; ++j)
{
BitArray classBits = Util.BitsFromInt(size, j);
temp.AddChromosome(new Chromosome(new System.Collections.BitArray(2, true), classBits));
int targetSize = OptoGlobals.RNG.Next(1, InitialComplexityUpperBound + 1);
while (temp.NumChromosomes < targetSize)
{
temp.AddChromosome(new Chromosome());
}
}
population.Add(temp);
}
for (generation = 0; generation < MaxGen; ++generation)
{
advanceGeneration();
Console.WriteLine("Starting Generation " + generation);
if (generation % RecordInterval == 0)
{
dumpData();
}
}
dumpData();
}
private void evaluatePopulation()
{
List <Thread> threadPool = new List <Thread>(PopSize);
foreach (Hunter x in population)
{
x.ErrorCheck();
threadPool.Add(new Thread(() => x.EvaluateSet(OptoGlobals.DaedalusTrainingSet, OptoGlobals.DaedalusTrainingY)));
}
foreach (Thread t in threadPool)
{
t.Start();
}
while (threadPool.Any(t => t.IsAlive))
{
for (int i = threadPool.Count - 1; i > 0; --i)
{
Thread t = threadPool[i];
if (!t.IsAlive)
{
threadPool.RemoveAt(i);
continue;
}
}
Thread.Sleep(5000);
}
foreach (Hunter x in population)
{
Hunter.AdjustFitnessForComplexity(x);
}
}
/// <summary>
/// First, we'll copy the elites over. Then we'll perform merger (every hunter in the breeding population has a chance of performing
/// merger. These are added to a list, and then executed without replacement.
/// </summary>
/// <param name="nextGen">Filling this</param>
/// <param name="BreedingPop">by using this</param>
private void fillListFromBreedingPop(List <Hunter> nextGen, List <Hunter> BreedingPop)
{
int elitismNum = (int)Math.Ceiling(OptoGlobals.ElitismPercent * PopSize);
Console.WriteLine("Elitism...");
for (int e = 0; e < elitismNum; ++e)
{
Console.WriteLine("Elitism: Adding hunter with fitness " + population[e].Fitness);
if (population[e].Fitness == 0)
{
nextGen.Add(new Hunter(population[e].NumChromosomes, 2));
}
nextGen.Add(population[e].EliteCopy());
}
Console.WriteLine("Merge...");
List <int> mergeList = new List <int>(15), used;
int i = -1;
Console.WriteLine("Before merge: best fitness is " + nextGen[0].Fitness);
while (++i < BreedingPop.Count)
{
if (OptoGlobals.RNG.NextDouble() < OptoGlobals.MergerPercent)
{
mergeList.Add(i);
}
}
used = new List <int>(mergeList);
foreach (int target in mergeList)
{
used.Add(target);
int k = SupportingFunctions.GetUnpickedInt(BreedingPop.Count, used);
used.Add(k);
//Console.WriteLine("Merging Hunters " + k + " and " + target);
nextGen.Add(Hunter.Merger(BreedingPop[target], BreedingPop[k]));
}
//Now, we will use Crossover for the remaining slots
while (nextGen.Count < PopSize)
{
int j = OptoGlobals.RNG.Next(0, elitismNum), k = OptoGlobals.RNG.Next(0, BreedingPop.Count);
while (k == j)
{
k = OptoGlobals.RNG.Next(0, BreedingPop.Count);
}
foreach (Hunter newGuy in Hunter.Crossover(nextGen[j], BreedingPop[k]))
{
nextGen.Add(newGuy.EliteCopy());
}
}
//Since Crossover returns 2 at a time, we will strip off any extras
while (nextGen.Count > PopSize)
{
nextGen.RemoveAt(nextGen.Count - 1);
}
}
private void generateNextGeneration()
{
Console.WriteLine("Before generation: best fitness is " + population[0].Fitness);
List <Hunter> nextGen = new List <Hunter>(PopSize), breedingPop = new List <Hunter>(PopSize / 2);
breedingPop = SupportingFunctions.StochasticUniformSample(population);
for (int i = 0; i < breedingPop.Count; ++i)
{
if (breedingPop[i].Fitness == 0)
{
breedingPop[i] = new Hunter(breedingPop[i].NumChromosomes, 1);
}
}
Console.WriteLine("Filling breeding pool");
fillListFromBreedingPop(nextGen, breedingPop);
mutatePopulation(nextGen, (int)Math.Ceiling(OptoGlobals.ElitismPercent * PopSize));
population = nextGen;
}
internal void InsertHunter(Hunter test)
{
population[PopSize - 3] = test;
}
/// <summary>
/// So we need population stats to include Complexity, Fitness, Best, average, worst, by generation
/// The confusion matrix and "DNA" of the best hunter.
/// </summary>
private void dumpData()
{
string directory = "./" + OptoGlobals.EnvironmentTag + OptoGlobals.DataSetName + "Daedalus/";
if (!Directory.Exists(directory))
{
Directory.CreateDirectory(directory);
}
using (StreamWriter fout = new StreamWriter(new FileStream(directory + popfileName, FileMode.Create)))
{
StringBuilder x = new StringBuilder("Generation:,");
for (int i = 0; i < generation; ++i)
{
x.Append(i + ",");
}
x.Remove(x.Length - 1, 1);//trim the last comma
fout.WriteLine(x.ToString());
fout.WriteLine(makeLineFromData(_sBestFitness, "Best Fitness:,").ToString());
fout.WriteLine(makeLineFromData(_sComplexityOfBestHunter, "Complexity Of Best Hunter:,").ToString());
fout.WriteLine(makeLineFromData(_sAverageFitness, "Average Fitness:,").ToString());
fout.WriteLine(makeLineFromData(_sWorstFitness, "Worst Fitness,").ToString());
fout.WriteLine(makeLineFromData(_sMaxComplexity, "Highest Complexity:,").ToString());
fout.WriteLine(makeLineFromData(_sMinComplexity, "Lowest Complexity:,").ToString());
fout.WriteLine(makeLineFromData(_sAverageComplexity, "Average Complexity:,").ToString());
fout.WriteLine(makeLineFromData(_sFitnessStdDev, "Fitness Standard Deviations:,").ToString());
fout.WriteLine(makeLineFromData(_sComplexityStdDev, "Complexity Standard Deviations:,").ToString());
}
using (StreamWriter fout = new StreamWriter(new FileStream(directory + hunterFileName, FileMode.Create)))
{
Hunter best = population[0];
fout.WriteLine(best.Serialize());
fout.WriteLine();
fout.WriteLine("Best Fitness = " + best.Fitness);
fout.WriteLine(best.HumanReadableHunter);//Dump the human readable portion
int[,] cm = best.ConfusionMatrix;
StringBuilder x = new StringBuilder();
x.Append("Predicted/Actual,");
for (int i = 0; i < OptoGlobals.NumberOfClasses; ++i)
{
x.Append(OptoGlobals.ClassList[i] + ",");
}
x.DeleteLastChar();
fout.WriteLine(x.ToString());
x = new StringBuilder();
if (cm == null)
{
best.EvaluateSet(OptoGlobals.DaedalusTrainingSet, OptoGlobals.DaedalusTrainingY, false);
cm = best.ConfusionMatrix;
}
fout.WriteLine("Training Confusion Matrix");
for (int i = 0; i < OptoGlobals.NumberOfClasses; ++i)
{
x.Append(OptoGlobals.ClassList[i] + ",");
for (int j = 0; j < OptoGlobals.NumberOfClasses; ++j)
{
x.Append(cm[i, j]);
x.Append(",");
}
x.DeleteLastChar();
fout.WriteLine(x.ToString());
x = new StringBuilder();
}
cm = best.ValidationMatrix;
if (cm == null)
{
best.EvaluateSet(OptoGlobals.DaedalusValidationSet, OptoGlobals.DaedalusValidationY, true);
cm = best.ValidationMatrix;
}
fout.WriteLine("Validation Confusion Matrix");
x.Append("Predicted/Actual,");
for (int i = 0; i < OptoGlobals.NumberOfClasses; ++i)
{
x.Append(OptoGlobals.ClassList[i] + ",");
}
x.AppendLine();
for (int i = 0; i < OptoGlobals.NumberOfClasses; ++i)
{
x.Append(OptoGlobals.ClassList[i] + ",");
for (int j = 0; j < OptoGlobals.NumberOfClasses; ++j)
{
x.Append(cm[i, j]);
x.Append(",");
}
x.DeleteLastChar();
fout.WriteLine(x.ToString());
x = new StringBuilder();
}
}
}