//Pushes the better networks further down the array
private static int Partition(NN[] array, int min, int max)
{
NN pivot = array[(min + (max - min) / 2)];
int i = min - 1;
int j = max + 1;
do
{
//While pivot wins to array[i]
do
{
i++;
} while (array[i].PercCorrect < pivot.PercCorrect);
//And loses to array[j]
do
{
j--;
} while (array[j].PercCorrect > pivot.PercCorrect);
if (i >= j)
{
return(j);
}
Swap(array, i, j);
} while (true);
}
public NN Mutation(NN patient, double probability)
{
var r = new Random();
if (r.NextDouble() > probability)
{
return(patient);
}
for (int i = 0; i < NumberOfMutations; i++)
{
int mutationLayer = r.Next(0, patient.Layers.Count);
int mutationPointX = r.Next(0, patient.Layers[mutationLayer].Biases.GetLength(0));
//50% to mutate weight or bias
if (r.NextDouble() < .5)
{
//"Flip the bit" so to speak to mutate
patient.Layers[mutationLayer].Biases[mutationPointX]
= 1d - patient.Layers[mutationLayer].Biases[mutationPointX];
}
else
{
int mutationPointY = r.Next(0, patient.Layers[mutationLayer].Weights.GetLength(1));
//"Flip the bit" so to speak to mutate
patient.Layers[mutationLayer].Weights[mutationPointX, mutationPointY]
= 1d - patient.Layers[mutationLayer].Weights[mutationPointX, mutationPointY];
}
}
return(patient);
}
public NN Crossover(NN parent1, NN parent2, bool side)
{
var r = new Random();
NN child = new NN();
child.Init();
for (int i = 0; i < parent1.Layers.Count; i++)
{
int crossoverPointX = r.Next(0, parent1.Layers[i].Weights.GetLength(0));
int crossoverPointY = r.Next(0, parent1.Layers[i].Weights.GetLength(1));
for (int ii = 0; ii < parent1.Layers[i].Weights.GetLength(0); ii++)
{
for (int iii = 0; iii < parent1.Layers[i].Weights.GetLength(1); iii++)
{
double gene = parent1.Layers[i].Weights[ii, iii];
if (ii > crossoverPointX && iii > crossoverPointY)
{
gene = parent2.Layers[i].Weights[ii, iii];
}
child.Layers[i].Weights[ii, iii] = gene;
}
double bgene = parent1.Layers[i].Biases[ii];
if (ii > crossoverPointX)
{
bgene = parent2.Layers[i].Biases[ii];
}
child.Layers[i].Biases[ii] = bgene;
}
}
return(child);
}
public NN[] Load()
{
var nns = new NN[PopSize];
for (int i = 0; i < PopSize; i++)
{
nns[i] = IO.Read(i);
}
return(nns);
}
void GeneratePopulation(int popsize)
{
NNs = new NN[popsize];
for (int i = 0; i < popsize; i++)
{
var nn = new NN();
nn.Init();
NNs[i] = nn;
}
Save();
}
public static void Write(NN nn, int num)
{
FileStream fs = new FileStream(BasePath + "\\" + num.ToString() + ".txt", FileMode.Create, FileAccess.Write, FileShare.None);
StreamWriter sw = new StreamWriter(fs);
sw.Write(nn.NumLayers + " ");
foreach (Layer l in nn.Layers)
{
sw.Write(l.Length + " " + l.InputLength + " ");
for (int i = 0; i < l.Length; i++)
{
for (int ii = 0; ii < l.InputLength; ii++)
{
sw.Write(l.Weights[i, ii] + " ");
}
sw.Write(l.Biases[i] + " ");
}
}
sw.Close(); fs.Close();
}
public static NN Read(int num)
{
NN nn = new NN();
if (Running)
{
throw new Exception("Already accessing file");
}
Running = true;
FileStream fs = new FileStream(BasePath + "\\" + num.ToString() + ".txt", FileMode.Open, FileAccess.Read, FileShare.None);
StreamReader sr = new StreamReader(fs);
string text = sr.ReadToEnd();
sr.Close(); fs.Close();
string[] split = text.Split(' ');
int numlayers = int.Parse(split[0]);
nn.Layers = new List <Layer>();
int iterator = 1;
for (int j = 0; j < numlayers; j++)
{
int length = int.Parse(split[iterator]); iterator++;
int inputlength = int.Parse(split[iterator]); iterator++;
nn.Layers.Add(new Layer(length, inputlength));
for (int i = 0; i < length; i++)
{
for (int ii = 0; ii < inputlength; ii++)
{
nn.Layers[j].Weights[i, ii] = double.Parse(split[iterator]);
iterator++;
}
nn.Layers[j].Biases[i] = double.Parse(split[iterator]);
iterator++;
}
}
Running = false;
return(nn);
}