source: branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits/MixtureBandit.cs @ 12893

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

namespace HeuristicLab.Algorithms.Bandits {
  // custom testcase for extreme hunter policy
  public class MixtureBandit : IBandit {
    private double[] mu;
    private double[] sigma;
    private double[] pZero;
    public int NumArms { get { return mu.Length; } }
    public double OptimalExpectedReward { get; private set; } // reward of the best arm, for calculating regret
    public int OptimalExpectedRewardArm { get; private set; }
    public int OptimalMaximalRewardArm { get; private set; }
    public double MaxReward { get; private set; }
    public double MinReward { get; private set; }
    private readonly Random random;
    public MixtureBandit(Random random, IEnumerable<double> mu) : this(random, mu, mu.Select(_ => 1.0), mu.Select(_ => 0.0), double.PositiveInfinity) { }
    public MixtureBandit(Random random, IEnumerable<double> mu, IEnumerable<double> sigma, IEnumerable<double> pZero, double maxReward) { // probability of a zero reward
      this.mu = mu.ToArray();
      this.sigma = sigma.ToArray();
      this.pZero = pZero.ToArray();
      this.MaxReward = maxReward;
      this.random = random;

      double optimalExpectedReward;
      int bestArmForMaxReward, bestArmForExpReward;

      BanditHelper.SampleArms(random, this, 100000, out optimalExpectedReward, out bestArmForExpReward, out bestArmForMaxReward);
      OptimalExpectedReward = optimalExpectedReward;
      OptimalExpectedRewardArm = bestArmForExpReward;
      OptimalMaximalRewardArm = bestArmForMaxReward;
    }

    public double Pull(int arm) {
      if (random.NextDouble() < pZero[arm]) return 0.0;
      var z = Rand.RandNormal(random);
      var x = z * sigma[arm] + mu[arm];
      return Math.Max(0, Math.Min(MaxReward, x));
    }
  }
}