source: branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits/ParetoBandit.cs @ 13805

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

namespace HeuristicLab.Algorithms.Bandits {
  // for test case 1 in Extreme Bandits paper (Carpentier, NIPS 2014)
  public class ParetoBandit : IBandit {
    private double[] alpha;
    private double[] pZero;
    public int NumArms { get { return alpha.Length; } }
    public int OptimalExpectedRewardArm { get; private set; }
    public int OptimalMaximalRewardArm { get; private set; }
    private readonly Random random;

    public ParetoBandit(Random random, IEnumerable<double> alpha) {
      this.alpha = alpha.ToArray();
      this.pZero = new double[this.alpha.Length];
      this.random = random;
      OptimalExpectedRewardArm = Array.IndexOf(this.alpha, alpha.Min());
      OptimalMaximalRewardArm = OptimalExpectedRewardArm;
    }
    public ParetoBandit(Random random, IEnumerable<double> alpha, IEnumerable<double> pZero, int bestExpRewardArm, int bestMaxRewardArm) { // probability of a zero reward
      this.alpha = alpha.ToArray();
      this.pZero = pZero.ToArray();
      this.random = random;
      OptimalExpectedRewardArm = bestExpRewardArm;
      OptimalMaximalRewardArm = bestMaxRewardArm;
    }

    public double Pull(int arm) {
      if (random.NextDouble() < pZero[arm]) return 0.0;
      var u = random.NextDouble();
      return Math.Pow(1.0 - u, (-1 / alpha[arm]));
    }
  }
}