public void TestIrisDataset()
{
int trainingSamples = IrisDataset.input.Length;
Backprop bprop = new Backprop(1e-2, abstol: 1e-4, reltol: 1e-7, adjustThreshold: 1e-20);
bprop.UnknownCaseFaced += AddRule <GaussianRule2>;
BatchBackprop bbprop = new BatchBackprop(1e-2, abstol: 1e-4, reltol: 1e-7, adjustThreshold: 1e-20);
bbprop.UnknownCaseFaced += AddRule <GaussianRule2>;
QProp qprop = new QProp(abstol: 1e-4, reltol: 1e-7, adjustThreshold: 1e-20, InitialLearningRate: 1e-4);
qprop.UnknownCaseFaced += AddRule <GaussianRule2>;
StochasticBatch sprop = new StochasticBatch(40, 1e-2);
sprop.UnknownCaseFaced += AddRule <GaussianRule2>;
StochasticQprop sqprop = new StochasticQprop(40);
sqprop.UnknownCaseFaced += AddRule <GaussianRule2>;
double[][] x;
double[][] y;
double[][] tx;
double[][] ty;
SampleData(IrisDataset.input, IrisDataset.output, 120, out x, out y, out tx, out ty);
subtestIris(x, y, tx, ty, bprop);
subtestIris(x, y, tx, ty, bbprop);
subtestIris(x, y, tx, ty, qprop);
subtestIris(x, y, tx, ty, sprop);
subtestIris(x, y, tx, ty, sqprop);
}
public void ANFIS_OutputFromDataSet()
{
var sampleSize = RobotArmDataSet.Input.Length - 1;
//StochasticBatch sprop = new StochasticBatch(sampleSize, 1e-5);
//sprop.UnknownCaseFaced += AddRule<GaussianRule2>;
//var sprop = new Backprop(1e-1);
var sprop = new StochasticQprop(sampleSize);
var extractor = new KMEANSExtractorIO(25);
//ANFIS fis = ANFISBuilder<GaussianRule2>.Build(RobotArmDataSet.Input, RobotArmDataSet.OutputTheta1, extractor, sprop, 150);
ANFIS fis = ANFISBuilder <GaussianRule> .Build(RobotArmDataSet.Input, RobotArmDataSet.OutputTheta1, extractor, sprop, 150);
var output1 = fis.Inference(new[] { 1.10413546487088, 2.81104319371924 }).FirstOrDefault(); // 1.1
var output2 = fis.Inference(new[] { 2.31665592712393, 1.9375717475909 }).FirstOrDefault(); // 0.6
var output3 = fis.Inference(new[] { 2.88944142930409, 16.7526454098038 }).FirstOrDefault(); // 1.4
}
/// <summary>
///
/// </summary>
/// <param name="ruleNumber"></param>
/// <param name="maxIterations"></param>
/// <returns></returns>
public Task <bool> TrainANFIS(int ruleNumber, int maxIterations, bool useAnalicitalOutcomeForTraining = false)
{
return(Task.Run(() => {
if (!IsDataSetCalculated)
{
throw new ApplicationException("DataSet is not calculated or provided.");
}
var sampleSize = Positions.Count() - 1;
var dynamicObj = useAnalicitalOutcomeForTraining ? Positions.Select(x => new { Point = x, KinematicOutCome = CalculateArmJoint(x).GetAwaiter().GetResult().FirstOrDefault() }) : null;
var input = useAnalicitalOutcomeForTraining
? dynamicObj.Select(x => x.Point).ConvertToANFISParameter()
: Positions.ConvertToANFISParameter();
var theta1ANFIS = Task.Run(() => {
var sPropTheta1 = new StochasticQprop(sampleSize);
var extractorForTheta1 = new KMEANSExtractorIO(ruleNumber);
var expectedOutcome = useAnalicitalOutcomeForTraining
? dynamicObj.Select(x => new[] { x.KinematicOutCome.Theta1.ConvertRadiansToDegrees() }).ToArray()
: AnglesGrid.First().ConvertToANFISParameter();
Theta1ANFIS = ANFISBuilder <GaussianRule> .Build(input,
expectedOutcome, extractorForTheta1, sPropTheta1, maxIterations);
});
var theta2ANFIS = Task.Run(() => {
var sPropTheta2 = new StochasticQprop(sampleSize);
var extractorForTheta2 = new KMEANSExtractorIO(ruleNumber);
var expectedOutcome2 = useAnalicitalOutcomeForTraining
? dynamicObj.Select(x => new[] { x.KinematicOutCome.Theta2.ConvertRadiansToDegrees() }).ToArray()
: AnglesGrid.Last().ConvertToANFISParameter();
Theta2ANFIS = ANFISBuilder <GaussianRule> .Build(input,
expectedOutcome2, extractorForTheta2, sPropTheta2, maxIterations);
});
Task.WaitAll(theta1ANFIS, theta2ANFIS);
IsANFISTrained = true;
return true;
}));
}
public void TestOptimization1()
{
Backprop bprop = new Backprop(1e-2);
BatchBackprop bbprop = new BatchBackprop(1e-2);
QProp qprop = new QProp();
StochasticBatch sprop = new StochasticBatch(100, 1e-2);
StochasticQprop sqprop = new StochasticQprop(100);
int trainingSamples = 1000;
double[][] x = new double[trainingSamples][];
double[][] y = new double[trainingSamples][];
double[][] tx = new double[trainingSamples][];
double[][] ty = new double[trainingSamples][];
Random rnd = new Random();
for (int i = 0; i < trainingSamples; i++)
{
double valx = 0.5 - rnd.NextDouble();
double valy = 0.5 - rnd.NextDouble();
x[i] = new double[] { valx, valy };
y[i] = new double[] { 1 };
valx = 0.5 - rnd.NextDouble();
valy = 0.5 - rnd.NextDouble();
tx[i] = new double[] { valx, valy };
ty[i] = new double[] { 1 };
}
subTestOptimization1(bprop, x, y, tx, ty);
subTestOptimization1(bbprop, x, y, tx, ty);
subTestOptimization1(qprop, x, y, tx, ty);
subTestOptimization1(sprop, x, y, tx, ty);
subTestOptimization1(sqprop, x, y, tx, ty);
}
public void TestOptimization2()
{
Backprop bprop = new Backprop(1e-2);
BatchBackprop bbprop = new BatchBackprop(1e-2);
QProp qprop = new QProp();
StochasticBatch sprop = new StochasticBatch(100, 1e-2);
StochasticQprop sqprop = new StochasticQprop(100);
int trainingSamples = 100;
double[][] x = new double[trainingSamples][];
double[][] y = new double[trainingSamples][];
double[][] tx = new double[trainingSamples][];
double[][] ty = new double[trainingSamples][];
Random rnd = new Random();
for (int i = 0; i < trainingSamples; i++)
{
bool isRigth = i % 2 == 0;
double valx = (isRigth ? 1 : -1) + (0.5 - rnd.NextDouble());
x[i] = new double[] { valx };
y[i] = new double[] { isRigth ? 1 : 0, isRigth ? 0 : 1 };
valx = (isRigth ? 1 : -1) + (0.5 - rnd.NextDouble());
tx[i] = new double[] { valx };
ty[i] = new double[] { isRigth ? 1 : 0, isRigth ? 0 : 1 };
}
subTestOptimization2(bprop, x, y, tx, ty);
subTestOptimization2(bbprop, x, y, tx, ty);
subTestOptimization2(qprop, x, y, tx, ty);
subTestOptimization2(sprop, x, y, tx, ty);
subTestOptimization2(sqprop, x, y, tx, ty);
}
public void TestLogisticMap()
{
int trainingSamples = 2000;
double[][] x = new double[trainingSamples][];
double[][] y = new double[trainingSamples][];
double[][] tx = new double[trainingSamples][];
double[][] ty = new double[trainingSamples][];
double px = 0.1;
double r = 3.8;//3.56995;
double lx = r * px * (1 - px);
for (int i = 0; i < trainingSamples; i++)
{
x[i] = new double[] { px, lx };
px = lx;
lx = r * lx * (1 - lx);
y[i] = new double[] { lx };
}
for (int i = 0; i < trainingSamples; i++)
{
tx[i] = new double[] { px, lx };
px = lx;
lx = r * lx * (1 - lx);
ty[i] = new double[] { lx };
}
Backprop bprop = new Backprop(1e-2);
bprop.UnknownCaseFaced += AddRule <GaussianRule2>;
BatchBackprop bbprop = new BatchBackprop(1e-2);
bbprop.UnknownCaseFaced += AddRule <GaussianRule2>;
QProp qprop = new QProp();
qprop.UnknownCaseFaced += AddRule <GaussianRule2>;
StochasticBatch sprop = new StochasticBatch(500, 1e-2);
sprop.UnknownCaseFaced += AddRule <GaussianRule2>;
StochasticQprop sqprop = new StochasticQprop(500);
sqprop.UnknownCaseFaced += AddRule <GaussianRule2>;
subtestLogisticsMap <LinearRule>(x, y, tx, ty, bprop);
subtestLogisticsMap <LinearRule>(x, y, tx, ty, bbprop);
subtestLogisticsMap <LinearRule>(x, y, tx, ty, qprop);
subtestLogisticsMap <LinearRule>(x, y, tx, ty, sprop);
subtestLogisticsMap <LinearRule>(x, y, tx, ty, sqprop);
bprop = new Backprop(1e-2);
bprop.UnknownCaseFaced += AddRule <GaussianRule2>;
bbprop = new BatchBackprop(1e-2);
bbprop.UnknownCaseFaced += AddRule <GaussianRule2>;
qprop = new QProp();
qprop.UnknownCaseFaced += AddRule <GaussianRule2>;
sprop = new StochasticBatch(500, 1e-2);
sprop.UnknownCaseFaced += AddRule <GaussianRule2>;
sqprop = new StochasticQprop(500);
sqprop.UnknownCaseFaced += AddRule <GaussianRule2>;
subtestLogisticsMap <GaussianRule>(x, y, tx, ty, bprop);
subtestLogisticsMap <GaussianRule>(x, y, tx, ty, bbprop);
subtestLogisticsMap <GaussianRule>(x, y, tx, ty, qprop);
subtestLogisticsMap <GaussianRule>(x, y, tx, ty, sprop);
subtestLogisticsMap <GaussianRule>(x, y, tx, ty, sqprop);
}
