Changeset 11732 for branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/ChernoffIntervalEstimationPolicy.cs

Timestamp:

01/07/15 09:21:46 (9 years ago)

Author:

gkronber

Message:

#2283: refactoring and bug fixes

File:

• branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/ChernoffIntervalEstimationPolicy.cs (modified) (2 diffs)

branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/ChernoffIntervalEstimationPolicy.cs

@@ -10 +10 @@
 International Conference, CP 2006, Nantes, France, September 25-29, 2006. pp 560-574 */
 
-  public class ChernoffIntervalEstimationPolicy : BanditPolicy {
-    private readonly int[] tries;
-    private readonly double[] sumReward;
-    private int totalTries = 0;
+  public class ChernoffIntervalEstimationPolicy : IPolicy {
     private readonly double delta;
 
-    public ChernoffIntervalEstimationPolicy(int numActions, double delta = 0.01)
-      : base(numActions) {
+    public ChernoffIntervalEstimationPolicy(double delta = 0.01) {
       this.delta = delta;
-      this.tries = new int[numActions];
-      this.sumReward = new double[numActions];
     }
-
-    public override int SelectAction() {
+    public int SelectAction(Random random, IEnumerable<IPolicyActionInfo> actionInfos) {
+      Debug.Assert(actionInfos.Any());
+      // select best
+      var myActionInfos = actionInfos.OfType<DefaultPolicyActionInfo>().ToArray(); // TODO: performance
+      int k = myActionInfos.Length;
+      int totalTries = myActionInfos.Where(a => !a.Disabled).Sum(a => a.Tries);
       int bestAction = -1;
       double bestQ = double.NegativeInfinity;
-      double k = Actions.Count();
-      Debug.Assert(k > 0);
-      foreach (var a in Actions) {
-        if (tries[a] == 0) return a;
+      for (int a = 0; a < myActionInfos.Length; a++) {
+        if (myActionInfos[a].Disabled) continue;
+        if (myActionInfos[a].Tries == 0) return a;
+
+        var sumReward = myActionInfos[a].SumReward;
+        var tries = myActionInfos[a].Tries;
+
+        var avgReward = sumReward / tries;
+
         // page 5 of "A simple distribution-free appraoch to the max k-armed bandit problem"
         // var alpha = Math.Log(2 * totalTries * k / delta);
         double alpha = Math.Log(2) + Math.Log(totalTries) + Math.Log(k) - Math.Log(delta); // total tries is max tries in the original paper
-        double mu = sumReward[a] / tries[a];
-        var q = mu + (alpha + Math.Sqrt(2 * tries[a] * mu * alpha + alpha * alpha)) / tries[a];
+        var q = avgReward + (alpha + Math.Sqrt(2 * tries * avgReward * alpha + alpha * alpha)) / tries;
         if (q > bestQ) {
           bestQ = q;
@@ -40 +42 @@
         }
       }
+      Debug.Assert(bestAction >= 0);
       return bestAction;
     }
-    public override void UpdateReward(int action, double reward) {
-      Debug.Assert(Actions.Contains(action));
-      totalTries++;
-      tries[action]++;
-      sumReward[action] += reward;
+
+    public IPolicyActionInfo CreateActionInfo() {
+      return new DefaultPolicyActionInfo();
     }
 
-    public override void DisableAction(int action) {
-      base.DisableAction(action);
-      totalTries -= tries[action];
-      tries[action] = -1;
-      sumReward[action] = 0;
-    }
-
-    public override void Reset() {
-      base.Reset();
-      totalTries = 0;
-      Array.Clear(tries, 0, tries.Length);
-      Array.Clear(sumReward, 0, sumReward.Length);
-    }
-
-    public override void PrintStats() {
-      for (int i = 0; i < sumReward.Length; i++) {
-        if (tries[i] >= 0) {
-          Console.Write("{0,5:F2}", sumReward[i] / tries[i]);
-        } else {
-          Console.Write("{0,5}", "");
-        }
-      }
-      Console.WriteLine();
-    }
     public override string ToString() {
       return string.Format("ChernoffIntervalEstimationPolicy({0:F2})", delta);