Changeset 12876 for branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits

Timestamp:

08/17/15 19:13:19 (9 years ago)

Author:

gkronber

Message:

#2283: implemented first crude version of extreme hunter algorithm in branch

Location:

branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits

Files:

• GaussianBandit.cs (modified) (3 diffs)
• HeuristicLab.Problems.Bandits.csproj (modified) (1 diff)
• IBandit.cs (modified) (1 diff)
• ParetoBandit.cs (added)

branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits/GaussianBandit.cs

@@ -12 +12 @@
     public int OptimalExpectedRewardArm { get; private set; }
     public int OptimalMaximalRewardArm { get; private set; }
-
+    public double MaxReward { get; private set; }
+    public double MinReward { get; private set; }
     private readonly Random random;
     private readonly double[] exp;
     private readonly double[] stdDev;
-    public GaussianBandit(Random random, int nArms) {
+    public GaussianBandit(Random random, int nArms, double minReward = double.NegativeInfinity, double maxReward = double.PositiveInfinity) {
+      this.MaxReward = maxReward;
+      this.MinReward = minReward;
       this.random = random;
       this.NumArms = nArms;
@@ -31 +34 @@
           OptimalExpectedRewardArm = i;
         }
-        var q = alglib.invnormaldistribution(0.99) * stdDev[i] + exp[i];
+        var q = alglib.invnormaldistribution(0.999) * stdDev[i] + exp[i];
         if (q > bestQ) {
           bestQ = q;
@@ -39 +42 @@
     }
 
-    // pulling an arm results in a truncated normally distributed reward 
-    // with mean expReward[i] and std.dev 0.1
+    // pulling an arm results in a normally distributed reward 
+    // with mean expReward[i] and std.dev 
     public double Pull(int arm) {
-      var z = Rand.RandNormal(random);
-      var x = z * stdDev[arm] + exp[arm];
+      double x;
+      do {
+        var z = Rand.RandNormal(random);
+        x = z * stdDev[arm] + exp[arm];
+      } while (x <= MinReward || x > MaxReward);
       return x;
     }

branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits/HeuristicLab.Problems.Bandits.csproj

@@ -41 +41 @@
     <Compile Include="BanditHelper.cs" />
     <Compile Include="BernoulliBandit.cs" />
+    <Compile Include="ParetoBandit.cs" />
     <Compile Include="GaussianBandit.cs" />
     <Compile Include="GaussianMixtureBandit.cs" />

branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits/IBandit.cs

@@ -12 +12 @@
     int OptimalMaximalRewardArm { get; } // arm which is optimal for optimization of maximal reward
 
-    double Pull(int arm); // pulling an arm returns a regret
+    double Pull(int arm); // pulling an arm returns a reward
   }
 }