source: branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/BanditPolicies/UCB1Policy.cs @ 11742

using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace HeuristicLab.Algorithms.Bandits.BanditPolicies {
  // policy for k-armed bandit (see Auer et al. 2002)
  public class UCB1Policy : IBanditPolicy {
    public int SelectAction(Random random, IEnumerable<IBanditPolicyActionInfo> actionInfos) {
      var myActionInfos = actionInfos.OfType<DefaultPolicyActionInfo>().ToArray(); // TODO: performance
      int bestAction = -1;
      double bestQ = double.NegativeInfinity;
      int totalTries = myActionInfos.Where(a => !a.Disabled).Sum(a => a.Tries);

      for (int a = 0; a < myActionInfos.Length; a++) {
        if (myActionInfos[a].Disabled) continue;
        if (myActionInfos[a].Tries == 0) return a;
        var q = myActionInfos[a].SumReward / myActionInfos[a].Tries + Math.Sqrt((2 * Math.Log(totalTries)) / myActionInfos[a].Tries);
        if (q > bestQ) {
          bestQ = q;
          bestAction = a;
        }
      }
      Debug.Assert(bestAction > -1);
      return bestAction;
    }

    public IBanditPolicyActionInfo CreateActionInfo() {
      return new DefaultPolicyActionInfo();
    }
    public override string ToString() {
      return "UCB1Policy";
    }
  }
}