source: branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Algorithms.Bandits/Policies/BoltzmannExplorationPolicy.cs @ 11732

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

namespace HeuristicLab.Algorithms.Bandits {
  // also called softmax policy
  public class BoltzmannExplorationPolicy : IPolicy {
    private readonly double beta;

    public BoltzmannExplorationPolicy(double beta) {
      if (beta < 0) throw new ArgumentException();
      this.beta = beta;
    }
    public int SelectAction(Random random, IEnumerable<IPolicyActionInfo> actionInfos) {
      Debug.Assert(actionInfos.Any());

      // select best
      var myActionInfos = actionInfos.OfType<DefaultPolicyActionInfo>().ToArray(); // TODO: performance
      Debug.Assert(myActionInfos.Any(a => !a.Disabled));
      double[] w = new double[myActionInfos.Length];

      for (int a = 0; a < myActionInfos.Length; a++) {
        if (myActionInfos[a].Disabled) {
          w[a] = 0; continue;
        }
        if (myActionInfos[a].Tries == 0) return a;
        var sumReward = myActionInfos[a].SumReward;
        var tries = myActionInfos[a].Tries;
        var avgReward = sumReward / tries;
        w[a] = Math.Exp(beta * avgReward);
      }


      var bestAction = Enumerable.Range(0, w.Length).SampleProportional(random, w).First();
      Debug.Assert(bestAction >= 0);
      Debug.Assert(bestAction < w.Length);
      Debug.Assert(!myActionInfos[bestAction].Disabled);
      return bestAction;
    }

    public IPolicyActionInfo CreateActionInfo() {
      return new DefaultPolicyActionInfo();
    }

    public override string ToString() {
      return string.Format("BoltzmannExplorationPolicy({0:F2})", beta);
    }
  }
}