Changeset 12893 for branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Algorithms.Bandits/ActionInfos/ExtremeHunterActionInfo.cs

Timestamp:

08/24/15 13:56:27 (9 years ago)

Author:

gkronber

Message:

#2283: experiments on grammatical optimization algorithms (maxreward instead of avg reward, ...)

File:

• branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Algorithms.Bandits/ActionInfos/ExtremeHunterActionInfo.cs (modified) (5 diffs)

branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Algorithms.Bandits/ActionInfos/ExtremeHunterActionInfo.cs

@@ -22 +22 @@
           if (minHeap.Count <= 1) return double.PositiveInfinity;
           double xk = minHeap.GetMin();
-          if (xk.IsAlmost(0.0)) return double.NegativeInfinity;
+          if (xk.IsAlmost(0.0)) return double.PositiveInfinity;
           var alpha = 1.0 / (minHeap.Count - 1) * minHeap.Skip(1).Sum(x => Math.Log(x) - Math.Log(xk));
           Debug.Assert(alpha > 0);
@@ -55 +55 @@
 
         Debug.Assert(minHeap.Count == ((int)Math.Floor(n * R)));
-        Debug.Assert(maxHeap.Count == 0 || minHeap.Count == 0 || maxHeap.GetMin() < minHeap.GetMin());
+        Debug.Assert(maxHeap.Count == 0 || minHeap.Count == 0 || maxHeap.GetMin() <= minHeap.GetMin());
       }
     }
@@ -64 +64 @@
     private OnlineHillEstimator hillEstimator;
     private List<double> rewards;
-
+    public double MaxReward { get; private set; }
     public double Value {
       get {
@@ -76 +76 @@
     public void UpdateReward(double reward) {
       if (reward < 0.0) throw new ArgumentException("reward");
+      MaxReward = Math.Max(MaxReward, reward);
       Tries++;
+      reward = (1 / (1 - reward)); // transformation from [0..1]
       rewards.Add(reward);
       hillEstimator.Update(reward);
@@ -82 +84 @@
 
     public void Reset() {
+      MaxReward = double.NegativeInfinity;
+
       this.hillEstimator = new OnlineHillEstimator();
       this.rewards = new List<double>();