source: branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/BernoulliThompsonSamplingPolicy.cs @ 12448

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

namespace HeuristicLab.Algorithms.Bandits.BanditPolicies {
  public class BernoulliThompsonSamplingPolicy : IBanditPolicy {
    // parameters of beta prior distribution
    private readonly double alpha = 1.0;
    private readonly double beta = 1.0;

    public int SelectAction(Random random, IEnumerable<IBanditPolicyActionInfo> actionInfos) {
      var myActionInfos = actionInfos.OfType<BernoulliPolicyActionInfo>();
      int bestAction = -1;
      double maxTheta = double.NegativeInfinity;
      var aIdx = -1;

      foreach (var aInfo in myActionInfos) {
        aIdx++;
        var theta = Rand.BetaRand(random, aInfo.NumSuccess + alpha, aInfo.NumFailure + beta);
        if (theta > maxTheta) {
          maxTheta = theta;
          bestAction = aIdx;
        }
      }
      Debug.Assert(bestAction > -1);
      return bestAction;
    }

    public IBanditPolicyActionInfo CreateActionInfo() {
      return new BernoulliPolicyActionInfo();
    }


    public override string ToString() {
      return "BernoulliThompsonSamplingPolicy";
    }
  }
}