source: branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/EpsGreedyPolicy.cs @ 11711

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

namespace HeuristicLab.Algorithms.Bandits {
  public class EpsGreedyPolicy : BanditPolicy {
    private readonly Random random;
    private readonly double eps;
    private readonly int[] tries;
    private readonly double[] sumReward;
    public EpsGreedyPolicy(Random random, int numActions, double eps)
      : base(numActions) {
      this.random = random;
      this.eps = eps;
      this.tries = new int[NumActions];
      this.sumReward = new double[NumActions];
    }

    public override int SelectAction() {
      if (random.NextDouble() > eps) {
        // select best
        var maxReward = double.NegativeInfinity;
        int bestAction = -1;
        for (int i = 0; i < NumActions; i++) {
          if (tries[i] == 0) return i;
          var avgReward = sumReward[i] / tries[i];
          if (maxReward < avgReward) {
            maxReward = avgReward;
            bestAction = i;
          }
        }
        return bestAction;
      } else {
        // select random
        return random.Next(NumActions);
      }
    }
    public override void UpdateReward(int action, double reward) {
      tries[action]++;
      sumReward[action] += reward;
    }
    public override void Reset() {
      Array.Clear(tries, 0, tries.Length);
      Array.Clear(sumReward, 0, sumReward.Length);
    }
  }
}