source: branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/UCB1Policy.cs @ 11728

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

namespace HeuristicLab.Algorithms.Bandits {
  public class UCB1Policy : BanditPolicy {
    private readonly int[] tries;
    private readonly double[] sumReward;
    private int totalTries = 0;
    public UCB1Policy(int numActions)
      : base(numActions) {
      this.tries = new int[numActions];
      this.sumReward = new double[numActions];
    }

    public override int SelectAction() {
      int bestAction = -1;
      double bestQ = double.NegativeInfinity;
      foreach (var a in Actions) {
        if (tries[a] == 0) return a;
        var q = sumReward[a] / tries[a] + Math.Sqrt((2 * Math.Log(totalTries)) / tries[a]);
        if (q > bestQ) {
          bestQ = q;
          bestAction = a;
        }
      }
      return bestAction;
    }
    public override void UpdateReward(int action, double reward) {
      Debug.Assert(Actions.Contains(action));
      totalTries++;
      tries[action]++;
      sumReward[action] += reward;
    }

    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);
    }
  }
}