static public double[][] SubstractiveClustering(double[][] x, double arad, double brad)
{
double[] P = new double[x.Length];
arad = (arad / 2.0) * (arad / 2.0);
brad = (brad / 2.0) * (brad / 2.0);
int progress = 0;
Parallel.For(0, x.Length, row =>
{
for (int col = row + 1; col < x.Length; col++)
{
double dist = math.EuclidianDistance2(x[row], x[col]);
double add = Math.Exp(-dist / arad);
if (add != 0)
{
lock (o)
{
P[row] += add;
P[col] += add;
}
}
}
if (progress++ % 100 == 0)
{
InMemoryLogger.PrintMessage($"Potential {progress}/{x.Length}");
}
});
List <double[]> c = new List <double[]>();
double D = P.Max();
double cap = 0.01 * D;
double[] wn = x[P.FindIndex(z => z == D)];
c.Add(wn);
while (D > cap)
{
InMemoryLogger.PrintMessage($"Iteration {c.Count} [{D} cap {cap}]");
progress = 0;
Parallel.For(0, x.Length, row =>
{
double dist = math.EuclidianDistance2(x[row], wn);
double add = -D * Math.Exp(-dist / brad);
P[row] += add;
//if (progress++ % 100 == 0)
//{
// Console.Write("\rPotential {0}/{1}", progress, x.Length);
//}
});
D = P.Max();
wn = x[P.FindIndex(z => z == D)];
c.Add(wn);
}
return(c.ToArray());
}
public static double[][] clustering(double[][] x, int k, kmeansType initType, out double error)
{
System.Diagnostics.Stopwatch sw = new System.Diagnostics.Stopwatch();
double[][] m = null;
//Console.WriteLine("Initialization...");
switch (initType)
{
case kmeansType.Forgy: m = ForgyInit(x, k); break;
case kmeansType.kmeanspp: m = kmeansppinit(x, k); break;
case kmeansType.RandomPartition: m = RandomPartion(x, k); break;
}
int confirmations = 0;
int[] a = new int[x.Length];
int reassigned = 0;
int iters = 0;
bool[] ntu = new bool[k];
DateTime start = DateTime.Now;
do
{
for (int i = 0; i < k; i++)
{
ntu[i] = true;
}
do
{
sw.Restart();
reassigned = Assignment(x, m, ref a, ref ntu);
if (reassigned != 0)
{
m = Update(x, m, a, ntu, k);
}
int unmoved = ntu.Count(z => !z);
InMemoryLogger.PrintMessage($"[{(DateTime.Now - start).TotalMilliseconds} ms] Iteration {iters++}, reassigned {reassigned} unmoved {unmoved} Time {sw.ElapsedMilliseconds} ms ");
if (reassigned != 0)
{
confirmations = 0;
}
}while (reassigned != 0);
confirmations++;
InMemoryLogger.PrintMessage($"CONFIRMATION {confirmations}");
} while (confirmations < 2);
error = Error(x, m, a);
return(m);
}
public static double[][] clustering(double[][] x, int k, int restarts, kmeansType initType)
{
double[][][] pool = new double[restarts][][];
double[] error = new double[restarts];
for (int r = 0; r < restarts; r++)
{
InMemoryLogger.PrintMessage($"Restart {r}");
double er = 0.0;
pool[r] = clustering(x, k, initType, out er);
error[r] = er;
InMemoryLogger.PrintMessage($"Error {er}");
}
double min = error.Min();
int opt = error.FindIndex(z => z == min);
return(pool[opt]);
}
private void Solve(double[][] x, double[][] y, double[][] tx, double[][] ty, ITraining bprop)
{
KMEANSExtractorI extractor = new KMEANSExtractorI(int.Parse(txtbxRulesCount.Text));
var timer = Stopwatch.StartNew();
var fis = ANFISBuilder <GaussianRule2> .Build(x, y, extractor, bprop, int.Parse(txtbxMaxIterCount.Text));
timer.Stop();
double err = bprop.Error(tx, ty, fis.RuleBase);
string line = "";
double correctClass = 0;
for (int i = 0; i < tx.Length; i++)
{
double[] o = fis.Inference(tx[i]);
if (tx[i].Length == 4 && o.Length == 3)
{
line = $"input: [{tx[i][0]}, {tx[i][1]}, {tx[i][2]}, {tx[i][3]}] output:[{o[0].ToString("F2")}, {o[1].ToString("F2")}, {o[2].ToString("F2")}] expected output: [{ty[i][0]}, {ty[i][1]}, {ty[i][2]}]";
}
for (int j = 0; j < ty[i].Length; j++)
{
if (ty[i][j] == 1.0 && o[j] == o.Max())
{
correctClass++;
line += " OK";
}
}
if (tx[i].Length == 4 && o.Length == 3)
{
InMemoryLogger.PrintMessage(line);
}
}
InMemoryLogger.PrintMessage(string.Format("Correct answers {5}\tClassification Error {4}\tElapsed {2}\tRuleBase {3}", err, bprop.GetType().Name, timer.Elapsed, fis.RuleBase.Length, 1.0 - correctClass / ty.Length, correctClass));
}
public double Iteration(double[][] x, double[][] y, IList <IRule> ruleBase)
{
Restart:
_isStop = false;
if (x.Length != y.Length)
{
throw new Exception("Input and desired output lengths not match");
}
if (ruleBase == null || ruleBase.Count == 0)
{
throw new Exception("Incorrect rulebase");
}
int outputDim = ruleBase[0].Z.Length;
int numOfRules = ruleBase.Count;
double globalError = 0.0;
double[] firings = new double[numOfRules];
for (int sample = 0; sample < x.Length; sample++)
{
double[] o = new double[outputDim];
double firingSum = 0.0;
for (int i = 0; i < numOfRules; i++)
{
firings[i] = ruleBase[i].Membership(x[sample]);
firingSum += firings[i];
}
if (UnknownCaseFaced != null && firingSum < _adjustThreshold)
{
int neig = math.NearestNeighbourhood(ruleBase.Select(z => z.Centroid).ToArray(), x[sample]);
UnknownCaseFaced(ruleBase, x[sample], y[sample], ruleBase[neig].Centroid);
InMemoryLogger.PrintMessage($"Adjusting rule base. Now {ruleBase.Count} are in base.");
goto Restart;
}
for (int i = 0; i < numOfRules; i++)
{
for (int C = 0; C < outputDim; C++)
{
o[C] += firings[i] / firingSum * ruleBase[i].Z[C];
}
}
for (int rule = 0; rule < ruleBase.Count; rule++)
{
double[] parm = ruleBase[rule].Parameters;
double[] grad = ruleBase[rule].GetGradient(x[sample]);
for (int p = 0; p < parm.Length; p++)
{
double g = dEdP(y[sample], o, ruleBase, firings, grad, firingSum, rule, outputDim, numOfRules, p);
parm[p] -= _learningRate * g;
}
}
for (int i = 0; i < numOfRules; i++)
{
for (int C = 0; C < outputDim; C++)
{
ruleBase[i].Z[C] -= _learningRate * (o[C] - y[sample][C]) * firings[i] / firingSum;
}
}
for (int C = 0; C < outputDim; C++)
{
globalError += Math.Abs(o[C] - y[sample][C]);
}
}
checkStop(globalError);
return(globalError / x.Length);
}