Changeset 11747 for branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Models/GaussianMixtureModel.cs

Timestamp:

01/12/15 21:23:01 (9 years ago)

Author:

gkronber

Message:

#2283: implemented test problems for MCTS

File:

• branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Models/GaussianMixtureModel.cs (modified) (4 diffs)

branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Models/GaussianMixtureModel.cs

@@ -9 +9 @@
 namespace HeuristicLab.Algorithms.Bandits.Models {
   public class GaussianMixtureModel : IModel {
-    private readonly double[] componentMeans;
-    private readonly double[] componentVars;
-    private readonly double[] componentProbs;
+    private double[] componentMeans;
+    private double[] componentVars;
+    private double[] componentProbs;
+    private readonly List<double> allRewards = new List<double>();
 
     private int numComponents;
@@ -17 +18 @@
     public GaussianMixtureModel(int nComponents = 5) {
       this.numComponents = nComponents;
-      this.componentProbs = new double[nComponents];
-      this.componentMeans = new double[nComponents];
-      this.componentVars = new double[nComponents];
+
+      Reset();
     }
 
@@ -29 +29 @@
 
     public void Update(double reward) {
-      // see http://www.cs.toronto.edu/~mackay/itprnn/ps/302.320.pdf Algorithm 22.2 soft k-means
-      throw new NotImplementedException();
+      allRewards.Add(reward);
+      throw new NotSupportedException("this does not yet work");
+      if (allRewards.Count < 1000 && allRewards.Count % 10 == 0) {
+        // see http://www.cs.toronto.edu/~mackay/itprnn/ps/302.320.pdf Algorithm 22.2 soft k-means
+        Reset();
+        for (int i = 0; i < 20; i++) {
+          var responsibilities = allRewards.Select(r => CalcResponsibility(r)).ToArray();
+
+
+          var sumWeightedRewards = new double[numComponents];
+          var sumResponsibilities = new double[numComponents];
+          foreach (var p in allRewards.Zip(responsibilities, Tuple.Create)) {
+            for (int k = 0; k < numComponents; k++) {
+              sumWeightedRewards[k] += p.Item2[k] * p.Item1;
+              sumResponsibilities[k] += p.Item2[k];
+            }
+          }
+          for (int k = 0; k < numComponents; k++) {
+            componentMeans[k] = sumWeightedRewards[k] / sumResponsibilities[k];
+          }
+
+          sumWeightedRewards = new double[numComponents];
+          foreach (var p in allRewards.Zip(responsibilities, Tuple.Create)) {
+            for (int k = 0; k < numComponents; k++) {
+              sumWeightedRewards[k] += p.Item2[k] * Math.Pow(p.Item1 - componentMeans[k], 2);
+            }
+          }
+          for (int k = 0; k < numComponents; k++) {
+            componentVars[k] = sumWeightedRewards[k] / sumResponsibilities[k];
+            componentProbs[k] = sumResponsibilities[k] / sumResponsibilities.Sum();
+          }
+        }
+      }
+    }
+
+    private double[] CalcResponsibility(double r) {
+      var res = new double[numComponents];
+      for (int k = 0; k < numComponents; k++) {
+        componentVars[k] = Math.Max(componentVars[k], 0.001);
+        res[k] = componentProbs[k] * alglib.normaldistribution((r - componentMeans[k]) / Math.Sqrt(componentVars[k]));
+        res[k] = Math.Max(res[k], 0.0001);
+      }
+      var sum = res.Sum();
+      for (int k = 0; k < numComponents; k++) {
+        res[k] /= sum;
+      }
+      return res;
     }
 
@@ -44 +89 @@
 
     public void Reset() {
-      Array.Clear(componentMeans, 0, numComponents);
-      Array.Clear(componentVars, 0, numComponents);
-      Array.Clear(componentProbs, 0, numComponents);
+      var rand = new Random();
+      this.componentProbs = Enumerable.Range(0, numComponents).Select((_) => rand.NextDouble()).ToArray();
+      var sum = componentProbs.Sum();
+      for (int i = 0; i < componentProbs.Length; i++) componentProbs[i] /= sum;
+      this.componentMeans = Enumerable.Range(0, numComponents).Select((_) => Rand.RandNormal(rand)).ToArray();
+      this.componentVars = Enumerable.Range(0, numComponents).Select((_) => 0.01).ToArray();
     }