Free cookie consent management tool by TermsFeed Policy Generator

# source:branches/HeuristicLab.Problems.GrammaticalOptimization-gkr/HeuristicLab.Problems.Bandits/GaussianBandit.cs@12876

Last change on this file since 12876 was 12876, checked in by gkronber, 8 years ago

#2283: implemented first crude version of extreme hunter algorithm in branch

File size: 2.1 KB
Line
1using System;
2using System.Collections.Generic;
3using System.Linq;
4using System.Text;
6using HeuristicLab.Common;
7
8namespace HeuristicLab.Algorithms.Bandits {
9  public class GaussianBandit : IBandit {
10    public int NumArms { get; private set; }
11    public double OptimalExpectedReward { get; private set; } // reward of the best arm, for calculating regret
12    public int OptimalExpectedRewardArm { get; private set; }
13    public int OptimalMaximalRewardArm { get; private set; }
14    public double MaxReward { get; private set; }
15    public double MinReward { get; private set; }
19    public GaussianBandit(Random random, int nArms, double minReward = double.NegativeInfinity, double maxReward = double.PositiveInfinity) {
20      this.MaxReward = maxReward;
21      this.MinReward = minReward;
22      this.random = random;
23      this.NumArms = nArms;
24      // expected reward of arms is iid and uniformly distributed
25      exp = new double[nArms];
26      stdDev = new double[nArms];
27      OptimalExpectedReward = double.NegativeInfinity;
28      var bestQ = double.NegativeInfinity;
29      for (int i = 0; i < nArms; i++) {
30        exp[i] = Rand.RandNormal(random);  // exp values for arms is N(0,1) distributed
31        stdDev[i] = 1.0 / Rand.GammaRand(random, 1); // variance is inv-gamma distributed
32        if (exp[i] > OptimalExpectedReward) {
33          OptimalExpectedReward = exp[i];
34          OptimalExpectedRewardArm = i;
35        }
36        var q = alglib.invnormaldistribution(0.999) * stdDev[i] + exp[i];
37        if (q > bestQ) {
38          bestQ = q;
39          OptimalMaximalRewardArm = i;
40        }
41      }
42    }
43
44    // pulling an arm results in a normally distributed reward
45    // with mean expReward[i] and std.dev
46    public double Pull(int arm) {
47      double x;
48      do {
49        var z = Rand.RandNormal(random);
50        x = z * stdDev[arm] + exp[arm];
51      } while (x <= MinReward || x > MaxReward);
52      return x;
53    }
54  }
55}
Note: See TracBrowser for help on using the repository browser.