source: branches/HeuristicLab.Algorithms.IteratedSentenceConstruction/HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction/3.3/QualityFunctions/TabularStateValueFunctionBase.cs @ 12955

using System;
using System.Collections.Generic;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Parameters;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;

namespace HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction {
  [StorableClass]
  public abstract class TabularStateValueFunctionBase : ParameterizedNamedItem, ITabularStateValueFunction {
    [Storable]
    private readonly Dictionary<object, double> q = new Dictionary<object, double>();
    [Storable]
    private readonly Dictionary<object, double> qVariance = new Dictionary<object, double>();
    [Storable]
    private readonly Dictionary<object, int> tries = new Dictionary<object, int>();

    public IStateFunction StateFunction {
      get {
        return ((IValueParameter<IStateFunction>)Parameters["State function"]).Value;
      }
      set { ((IValueParameter<IStateFunction>)Parameters["State function"]).Value = value; }
    }

    protected TabularStateValueFunctionBase()
      : base() {
      Parameters.Add(new ValueParameter<IStateFunction>("State function", "The function that is used to map partial trees to states", new DefaultStateFunction()));
    }

    public int Tries(object state) {
      var t = 0;
      if (!tries.TryGetValue(state, out t)) return 0;
      return t;
    }

    public double Value(object state) {
      // an action that has never been tried has q == infinity
      double quality;
      if (!q.TryGetValue(state, out quality)) return double.PositiveInfinity;
      return quality;
    }

    public double ValueVariance(object state) {
      // an action that has never been tried has qVariance == infinity
      double var;
      if (!qVariance.TryGetValue(state, out var)) return double.PositiveInfinity;
      return var / Tries(state);
    }

    public virtual void Update(object state, double observedQuality) {
      int t;
      if (!tries.TryGetValue(state, out t)) {
        t = 0;
        tries.Add(state, t + 1);
        q.Add(state, observedQuality);
        qVariance.Add(state, 0); // naive initialization
      } else {
        tries[state] = t + 1;
        var delta = observedQuality - q[state];
        var curMean = CalculateNewQ(state, observedQuality);
        q[state] = curMean;
        qVariance[state] = qVariance[state] + delta * (observedQuality - curMean); // iterative calculation of mean
      }
    }

    protected abstract double CalculateNewQ(object state, double observedQuality);


    #region item
    [StorableConstructor]
    protected TabularStateValueFunctionBase(bool deserializing) : base(deserializing) { }
    protected TabularStateValueFunctionBase(TabularStateValueFunctionBase original, Cloner cloner)
      : base(original, cloner) {
      // TODO: these become really large when using this class only from BasicAlgorithms it would not be necessary to clone and reset everything, (pause is not allowed)

      this.q = new Dictionary<object, double>(original.q);
      this.tries = new Dictionary<object, int>(original.tries);
      this.qVariance = new Dictionary<object, double>(original.qVariance);
    }
    #endregion

    //public void InitializeState() {
    //  ClearState();
    //}
    //
    //public void ClearState() {
    //  q.Clear();
    //  tries.Clear();
    //  qVariance.Clear();
    //}
  }
}