source: branches/HeuristicLab.Algorithms.IteratedSentenceConstruction/HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction/3.3/Policies/UcbTunedSymbolicExpressionConstructionPolicy.cs @ 13398

using System;
using System.Collections.Generic;
using System.Linq;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
using HeuristicLab.Parameters;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
using HeuristicLab.PluginInfrastructure;
using HeuristicLab.Random;

namespace HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction {
  [StorableClass]
  [Item("UcbTunedSymbolicExpressionConstructionPolicy", "Also uses an estimate of the variance")]
  public class UcbTunedSymbolicExpressionConstructionPolicy : SymbolicExpressionConstructionPolicyBase {
    public double R {
      get { return ((IFixedValueParameter<DoubleValue>)Parameters["R"]).Value.Value; }
      set { ((IFixedValueParameter<DoubleValue>)Parameters["R"]).Value.Value = value; }
    }

    public ITabularStateValueFunction StateValueFunction {
      get {
        return ((IValueParameter<ITabularStateValueFunction>)Parameters["Quality function"]).Value;
      }
      set { ((IValueParameter<ITabularStateValueFunction>)Parameters["Quality function"]).Value = value; }
    }


    protected UcbTunedSymbolicExpressionConstructionPolicy(UcbTunedSymbolicExpressionConstructionPolicy original, Cloner cloner)
      : base(original, cloner) {
    }

    [StorableConstructor]
    protected UcbTunedSymbolicExpressionConstructionPolicy(bool deserializing) : base(deserializing) { }


    public UcbTunedSymbolicExpressionConstructionPolicy()
      : base() {
      Parameters.Add(new FixedValueParameter<DoubleValue>("R", "The weighting factor for the confidence bound (should be scaled based on the range or the fitness values)", new DoubleValue(1.0)));
      Parameters.Add(new ValueParameter<ITabularStateValueFunction>("Quality function", "The quality function to use", new TabularAvgStateValueFunction()));
    }

    protected sealed override int Select(IReadOnlyList<object> followStates, IRandom random) {
      var bestFollowStates = new List<int>();
      var bestQuality = double.NegativeInfinity;
      int totalTries = followStates.Sum(s => StateValueFunction.Tries(s));
      for (int idx = 0; idx < followStates.Count; idx++) {
        var s = followStates[idx];
        double quality;
        if (StateValueFunction.Tries(s) == 0) {
          quality = double.PositiveInfinity;
        } else {
          double v = StateValueFunction.ValueVariance(s) + Math.Sqrt(2 * Math.Log(totalTries) / StateValueFunction.Tries(s));
          quality = StateValueFunction.Value(s) + R * Math.Sqrt(Math.Log(totalTries) / StateValueFunction.Tries(s) * v);
        }
        if (quality >= bestQuality) {
          if (quality > bestQuality) {
            bestFollowStates.Clear();
            bestQuality = quality;
          }
          bestFollowStates.Add(idx);
        }
      }
      return bestFollowStates.SampleRandom(random);
    }

    public sealed override void Update(IEnumerable<object> stateSequence, double quality) {
      foreach (var state in stateSequence) {
        StateValueFunction.Update(state, quality);
      }
    }

    protected override object CreateState(ISymbolicExpressionTreeNode root, List<ISymbol> actionSequence, ISymbolicExpressionTreeNode parent, int childIdx) {
      return StateValueFunction.StateFunction.CreateState(root, actionSequence, parent, childIdx);
    }

    public override IDeepCloneable Clone(HeuristicLab.Common.Cloner cloner) {
      return new UcbTunedSymbolicExpressionConstructionPolicy(this, cloner);
    }
  }
}