Changeset 11742 for branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Models/GaussianModel.cs

Timestamp:

01/09/15 14:57:28 (9 years ago)

Author:

gkronber

Message:

#2283 refactoring

File:

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

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

@@ -7 +7 @@
   // 2) unknown mean and unknown variance
   public class GaussianModel : IModel {
+
     private OnlineMeanAndVarianceEstimator estimator = new OnlineMeanAndVarianceEstimator();
 
@@ -82 +83 @@
       }
 
+
       // sample from the posterior marginal for mu (expected value) equ. 91
       // p(µ|D) = T2αn (µ| µn, βn/(αnκn)) 
 
-      // sample from Tk distribution : http://stats.stackexchange.com/a/70270
       var t2alpha = alglib.invstudenttdistribution((int)(2 * posteriorAlpha), random.NextDouble());
 
@@ -91 +92 @@
       return theta;
 
-      //return alglib.invnormaldistribution(random.NextDouble()) * + theta;
-      //return alglib.invstudenttdistribution((int)(2 * posteriorAlpha), 0.99) * (posteriorBeta*posteriorK + posteriorBeta) / (posteriorAlpha*posteriorK) + posteriorMean;
+
+      /*
+       * value function : 0.99-quantile
+      // sample posterior mean and posterior variance independently
+      var sampledPrec = Rand.GammaRand(random, posteriorAlpha) * posteriorBeta;
+      var t2alpha = alglib.invstudenttdistribution((int)(2 * posteriorAlpha), random.NextDouble());
+      var sampledMean = t2alpha * posteriorBeta / (posteriorAlpha * posteriorK) + posteriorMean;
+
+      return alglib.invnormaldistribution(0.99) / Math.Sqrt(sampledPrec) + sampledMean;
+       */
     }
 
@@ -114 +123 @@
         return new GaussianModel(meanPriorMu, meanPriorVariance, precisionPriorAlpha, precisionPriorBeta);
     }
+
+    public override string ToString() {
+      if (knownVariance) {
+        var posteriorMeanVariance = 1.0 / (estimator.N / rewardVariance + 1.0 / meanPriorVariance);
+        var posteriorMeanMean = posteriorMeanVariance * (meanPriorMu / meanPriorVariance + estimator.Sum / rewardVariance);
+        return string.Format("Gaussian(mu, var=0.1), mu ~ Gaussian(mu'={0:F3}, var'={1:F3})", posteriorMeanMean, posteriorMeanVariance);
+      } else {
+        var posteriorMean = (priorK * meanPriorMu + estimator.Sum) / (priorK + estimator.N);
+        var posteriorK = priorK + estimator.N;
+        var posteriorAlpha = precisionPriorAlpha + estimator.N / 2.0;
+        double posteriorBeta;
+        if (estimator.N > 0) {
+          posteriorBeta = precisionPriorBeta + 0.5 * estimator.N * estimator.Variance + priorK * estimator.N * Math.Pow(estimator.Avg - meanPriorMu, 2) / (2.0 * (priorK + estimator.N));
+        } else {
+          posteriorBeta = precisionPriorBeta;
+        }
+        var nu = (int)(2 * posteriorAlpha);
+        var meanVariance = posteriorBeta / (posteriorAlpha * posteriorK) * (nu / (double)(nu - 2));
+        return string.Format("Gaussian(mu, var), mu ~ T{0}(mu'={1:F3}, var'={2:F3}), 1.0/var ~ Gamma(mu={3:F3}, var={4:F3})",
+          nu, posteriorMean, meanVariance,
+          posteriorAlpha / posteriorBeta, posteriorAlpha / (posteriorBeta * posteriorBeta));
+      }
+    }
   }
 }