Changeset 9124 for branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Problems.Instances.DataAnalysis.GaussianProcessRegression

Timestamp:

01/08/13 11:01:56 (12 years ago)

Author:

gkronber

Message:

#1967 implemented utility app to draw random samples using a GP prior

Location:

branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Problems.Instances.DataAnalysis.GaussianProcessRegression

Files:

• GaussianProcessRegressionInstance.cs (modified) (3 diffs)
• GaussianProcessSumOfRQIso.cs (added)
• GaussianProcessSumOfSEIso.cs (added)
• Instances.DataAnalysis.GaussianProcessRegression-3.3.csproj (modified) (1 diff)
• Util.cs (modified) (3 diffs)
• VariousInstanceProvider.cs (modified) (2 diffs)

branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Problems.Instances.DataAnalysis.GaussianProcessRegression/GaussianProcessRegressionInstance.cs

@@ -30 +30 @@
 
     public override string Name {
-      get { return "Gaussian Process " + covarianceFunction.ItemName; }
+      get { return "Gaussian Process " + name; }
     }
     public override string Description {
@@ -36 +36 @@
     }
     protected override string TargetVariable { get { return "Y"; } }
-    protected override string[] VariableNames { get { return new string[] { "X1", "X2", "Y" }; } }
-    protected override string[] AllowedInputVariables { get { return new string[] { "X1", "X2" }; } }
+    protected override string[] VariableNames { get { return new string[] { "X1", "Y" }; } }
+    protected override string[] AllowedInputVariables { get { return new string[] { "X1" }; } }
     protected override int TrainingPartitionStart { get { return 0; } }
-    protected override int TrainingPartitionEnd { get { return 100; } }
-    protected override int TestPartitionStart { get { return 100; } }
-    protected override int TestPartitionEnd { get { return 200; } }
+    protected override int TrainingPartitionEnd { get { return 200; } }
+    protected override int TestPartitionStart { get { return 200; } }
+    protected override int TestPartitionEnd { get { return 400; } }
 
-    private ICovarianceFunction covarianceFunction;
+    private ParameterizedCovarianceFunction cov;
+    private string name;
 
-    public GaussianProcessRegressionInstance(ICovarianceFunction covarianceFunction) {
-      this.covarianceFunction = covarianceFunction;
+    public GaussianProcessRegressionInstance(string name, ICovarianceFunction covarianceFunction, double[] hyp) {
+      this.name = name;
+      cov = covarianceFunction.GetParameterizedCovarianceFunction(hyp, null);
     }
 
@@ -53 +55 @@
       List<List<double>> data = new List<List<double>>();
       for (int i = 0; i < AllowedInputVariables.Count(); i++) {
-        data.Add(ValueGenerator.GenerateSteps(0, 0.99, 0.01).ToList());
-        data[i].AddRange(ValueGenerator.GenerateSteps(0.005, 1, 0.01).ToList());
+        data.Add(ValueGenerator.GenerateSteps(0, 0.99, 1.0 / TrainingPartitionEnd).ToList());
+        data[i].AddRange(ValueGenerator.GenerateSteps(-0.5, 1.5, 2.0 / (TestPartitionEnd - TestPartitionStart)).ToList());
       }
       var mt = new MersenneTwister();
 
-      var noise = new CovarianceNoise();
-      noise.SetParameter(new double[] { Math.Log(Math.Sqrt(0.01)) });
-      var t = new CovarianceSum();
-      t.Terms.Add(covarianceFunction);
-      t.Terms.Add(noise);
-
-      var p = Enumerable.Range(0, t.GetNumberOfParameters(data.Count)).Select(i => mt.NextDouble() - 1).ToArray();
-
-      var cov = t.GetParameterizedCovarianceFunction(p, null);
 
       var target = Util.SampleGaussianProcess(mt, cov, data);

branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Problems.Instances.DataAnalysis.GaussianProcessRegression/Instances.DataAnalysis.GaussianProcessRegression-3.3.csproj

@@ -178 +178 @@
   </ItemGroup>
   <ItemGroup>
+    <Compile Include="GaussianProcessSumOfRQIso.cs" />
+    <Compile Include="GaussianProcessSumOfSEIso.cs" />
     <Compile Include="GaussianProcessRegressionInstance.cs" />
     <Compile Include="GaussianProcessSEIso6.cs" />

branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Problems.Instances.DataAnalysis.GaussianProcessRegression/Util.cs

@@ -31 +31 @@
 
 
+    public static List<double> SampleGaussianProcess(IRandom random, ParameterizedCovarianceFunction covFunction, List<List<double>> data) {
+      int n = data[0].Count;
 
-    public static List<double> SampleGaussianProcess(IRandom random, ParameterizedCovarianceFunction covFunction, List<List<double>> data) {
+      var normalRand = new NormalDistributedRandom(random, 0, 1);
+      var alpha = (from i in Enumerable.Range(0, n)
+                   select normalRand.NextDouble()).ToArray();
+      return SampleGaussianProcess(random, covFunction, data, alpha);
+    }
+
+    public static List<double> SampleGaussianProcess(IRandom random, ParameterizedCovarianceFunction covFunction, List<List<double>> data, double[] alpha) {
+      if (alpha.Length != data[0].Count) throw new ArgumentException();
 
       double[,] x = new double[data[0].Count, data.Count];
@@ -43 +52 @@
           K[i, j] = covFunction.Covariance(x, i, j);
 
-      var normalRand = new NormalDistributedRandom(random, 0, 1);
-
       if (!alglib.spdmatrixcholesky(ref K, K.GetLength(0), true)) throw new ArgumentException();
-      var r = (from i in Enumerable.Range(0, K.GetLength(0))
-               select normalRand.NextDouble()).ToArray();
 
       List<double> target = new List<double>(K.GetLength(0));
@@ -54 +59 @@
         for (int j = K.GetLength(0) - 1; j >= 0; j--) {
 
-          s += K[j, i] * r[j];
+          s += K[j, i] * alpha[j];
         }
 

branches/HeuristicLab.Problems.GaussianProcessTuning/HeuristicLab.Problems.Instances.DataAnalysis.GaussianProcessRegression/VariousInstanceProvider.cs

@@ -22 +22 @@
 using System;
 using System.Collections.Generic;
+using System.Linq;
 using HeuristicLab.Algorithms.DataAnalysis;
 
@@ -51 +52 @@
       descriptorList.Add(new GaussianProcessSEIsoDependentNoise());
 
-      var covs = new ICovarianceFunction[] {
-        new CovarianceSquaredExponentialIso(), 
-        new CovarianceSquaredExponentialArd(), 
-        new CovarianceLinear(), 
-        new CovarianceLinearArd(), 
-        new CovarianceMaternIso(), 
-        new CovariancePeriodic(), 
-        new CovarianceRationalQuadraticArd(), 
-        new CovarianceRationalQuadraticIso()
-      };
-      foreach (var cov in covs) {
-        descriptorList.Add(new GaussianProcessRegressionInstance(cov));
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovarianceSquaredExponentialIso());
+        cov.Terms.Add(new CovarianceSquaredExponentialIso());
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { -2.8, -0.1, 0.5, 0.3, -1.5 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("SE+SE", cov, hyp));
       }
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovarianceRationalQuadraticIso());
+        cov.Terms.Add(new CovarianceRationalQuadraticIso());
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { -3, 0, 0, -1.5, 0, 2.5, -1.5 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("RQ+RQ", cov, hyp));
+      }
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovariancePeriodic());
+        cov.Terms.Add(new CovariancePeriodic());
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { 0, -1.8, -1.5, 0, -0.5, -1, -2.1 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("Periodic+Periodic", cov, hyp));
+      }
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovarianceMaternIso());
+        cov.Terms.Add(new CovarianceMaternIso());
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { 0, 0, -1, 1, -4 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("Matern1+Matern1", cov, hyp));
+      }
+      {
+        var cov = new CovarianceSum();
+        var m1 = new CovarianceMaternIso();
+        m1.DParameter.Value = m1.DParameter.ValidValues.First(v => v.Value == 3);
+        var m2 = new CovarianceMaternIso();
+        m2.DParameter.Value = m2.DParameter.ValidValues.First(v => v.Value == 3);
+        cov.Terms.Add(m1);
+        cov.Terms.Add(m2);
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { -2.7, 0, -1, 1, -1.5 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("Matern3+Matern3", cov, hyp));
+      }
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovarianceRationalQuadraticIso());
+        cov.Terms.Add(new CovarianceSquaredExponentialIso());
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { -1.5, -0.5, -3, -1, -1, -3 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("RQ+SE", cov, hyp));
+      }
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovarianceSquaredExponentialIso());
+        var prod = new CovarianceProduct();
+        prod.Factors.Add(new CovarianceLinear());
+        prod.Factors.Add(new CovarianceNoise());
+        cov.Terms.Add(prod);
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { -3, 0, 0, -1.5 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("SE+Linear*Noise", cov, hyp));
+      }
+      {
+        var cov = new CovarianceSum();
+        cov.Terms.Add(new CovarianceSquaredExponentialIso());
+        cov.Terms.Add(new CovariancePeriodic());
+        cov.Terms.Add(new CovarianceNoise());
+        var hyp = new double[] { -1, 0, 0, -1.5, 0, -2 };
+        descriptorList.Add(new GaussianProcessRegressionInstance("SE+Periodic", cov, hyp));
+      }
+
 
       return descriptorList;