[12876] | 1 | using System;
|
---|
| 2 | using System.Collections.Generic;
|
---|
| 3 | using System.Diagnostics;
|
---|
| 4 | using System.Linq;
|
---|
| 5 | using System.Text;
|
---|
| 6 | using System.Threading.Tasks;
|
---|
| 7 | using HeuristicLab.Common;
|
---|
| 8 |
|
---|
| 9 | namespace HeuristicLab.Algorithms.Bandits.BanditPolicies {
|
---|
| 10 | // Extreme Bandits, NIPS2014
|
---|
| 11 | public class ExtremeHunterPolicy : IBanditPolicy {
|
---|
| 12 |
|
---|
| 13 |
|
---|
| 14 | public double E { get; set; }
|
---|
| 15 | public double D { get; set; }
|
---|
| 16 | public double delta { get; set; }
|
---|
| 17 | public double b { get; set; }
|
---|
[12893] | 18 | public double n { get; set; }
|
---|
| 19 | public int minPulls { get; set; }
|
---|
[12876] | 20 |
|
---|
[12893] | 21 | public ExtremeHunterPolicy(double E = 1.0E-3, double D = 1.0E-2, double b = 1.0, double n = 1.0E4, int minPulls = 100) {
|
---|
[12876] | 22 | this.E = E; // parameter TODO
|
---|
| 23 | this.D = D; // parameter TODO
|
---|
[12893] | 24 | this.b = b; // parameter: set to 1 in original paper "to consider a wide class of distributions"
|
---|
[12876] | 25 | // private communication with Alexandra Carpentier:
|
---|
| 26 | // For instance, on our synthetic experiments, we calibrated the constants by
|
---|
| 27 | // cross validation, using exact Pareto distributions and b=1, and we found
|
---|
| 28 | // out that taking E = 1e-3 is acceptable. For all the datasets
|
---|
| 29 | // (exact Pareto, approximate Pareto, and network data), we kept this same constant
|
---|
[12893] | 30 |
|
---|
| 31 | // minPulls seems to be set to 100 in the experiments in extreme bandit paper
|
---|
| 32 | this.minPulls = minPulls; // parameter: TODO (there are conditions for N given in the paper)
|
---|
| 33 | this.n = n;
|
---|
[12876] | 34 | }
|
---|
| 35 |
|
---|
| 36 | public int SelectAction(Random random, IEnumerable<IBanditPolicyActionInfo> actionInfos) {
|
---|
| 37 | var myActionInfos = actionInfos.OfType<ExtremeHunterActionInfo>();
|
---|
| 38 | double bestQ = double.NegativeInfinity;
|
---|
[12893] | 39 | // int totalTries = myActionInfos.Sum(a => a.Tries);
|
---|
[12876] | 40 | int K = myActionInfos.Count();
|
---|
| 41 |
|
---|
| 42 | this.delta = Math.Exp(-Math.Log(Math.Log(n))) / (2.0 * n * K); // TODO
|
---|
| 43 |
|
---|
| 44 | var bestActions = new List<int>();
|
---|
| 45 | int aIdx = -1;
|
---|
| 46 | foreach (var aInfo in myActionInfos) {
|
---|
| 47 | aIdx++;
|
---|
| 48 | double q;
|
---|
| 49 | if (aInfo.Tries <= minPulls) {
|
---|
| 50 | q = double.PositiveInfinity;
|
---|
| 51 | } else {
|
---|
| 52 | double t = aInfo.Tries;
|
---|
| 53 | double h = aInfo.Value;
|
---|
[12893] | 54 | if (double.IsInfinity(h)) q = 0;
|
---|
| 55 | else {
|
---|
| 56 | var thres = Math.Pow(t, h / (2 * b + 1));
|
---|
| 57 | double c = Math.Pow(t, 1.0 / (2 * b + 1)) * ((1.0 / t) * aInfo.Rewards.Count(r => r >= thres));
|
---|
| 58 | q = Math.Pow((c + B2(t)) * n, h + B1(t)) * Gamma(h, B1(t)); // eqn (5)
|
---|
| 59 | Debug.Assert(q > 0);
|
---|
| 60 | }
|
---|
[12876] | 61 | }
|
---|
| 62 | if (q > bestQ) {
|
---|
| 63 | bestQ = q;
|
---|
| 64 | bestActions.Clear();
|
---|
| 65 | bestActions.Add(aIdx);
|
---|
| 66 | } else if (q.IsAlmost(bestQ)) {
|
---|
| 67 | bestActions.Add(aIdx);
|
---|
| 68 | }
|
---|
| 69 | }
|
---|
| 70 | Debug.Assert(bestActions.Any());
|
---|
| 71 | return bestActions.SelectRandom(random);
|
---|
| 72 | }
|
---|
| 73 |
|
---|
| 74 | public double Gamma(double x, double y) {
|
---|
| 75 | if (1.0 - x - y <= 0) return double.PositiveInfinity;
|
---|
| 76 | else return alglib.gammafunction(1.0 - x - y); // comment on eqn 5
|
---|
| 77 | }
|
---|
| 78 |
|
---|
| 79 | // eqn 2
|
---|
| 80 | public double B1(double t) {
|
---|
| 81 | return D * Math.Sqrt(Math.Log(1.0 / delta)) * Math.Pow(t, -b / (2 * b + 1));
|
---|
| 82 | }
|
---|
| 83 |
|
---|
| 84 | // eqn 4
|
---|
| 85 | public double B2(double t) {
|
---|
| 86 | return E * Math.Sqrt(Math.Log(t / delta)) * Math.Log(t) * Math.Pow(t, -b / (2 * b + 1));
|
---|
| 87 | }
|
---|
| 88 |
|
---|
| 89 | public IBanditPolicyActionInfo CreateActionInfo() {
|
---|
| 90 | return new ExtremeHunterActionInfo();
|
---|
| 91 | }
|
---|
| 92 | public override string ToString() {
|
---|
[12893] | 93 | return string.Format("ExtremeHunter(E={0:F2},D={1:F2},b={2:F2},n={3:F0},minPulls={4:F0}", E, D, b, n, minPulls);
|
---|
[12876] | 94 | }
|
---|
| 95 | }
|
---|
| 96 | }
|
---|