Changeset 12922 for branches/HeuristicLab.Algorithms.IteratedSentenceConstruction/HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction/3.3/QualityFunctions/TabularAvgQualityFunction.cs

Timestamp:

08/27/15 16:55:07 (9 years ago)

Author:

gkronber

Message:

#2471

• added a simple implementation of Koza-style symbolic regression
• added ucb tuned and estimation of variance to tabular quality functions

File:

• branches/HeuristicLab.Algorithms.IteratedSentenceConstruction/HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction/3.3/QualityFunctions/TabularAvgQualityFunction.cs (modified) (3 diffs)

branches/HeuristicLab.Algorithms.IteratedSentenceConstruction/HeuristicLab.Algorithms.IteratedSymbolicExpressionConstruction/3.3/QualityFunctions/TabularAvgQualityFunction.cs

@@ -9 +9 @@
   [StorableClass]
   [Item("TabularAvgQualityFunction", "")]
-  internal class TabularAvgQualityFunction : ParameterizedNamedItem, ITabularQualityFunction {
-    [Storable]
-    private readonly Dictionary<object, Dictionary<int, double>> q = new Dictionary<object, Dictionary<int, double>>();
-    [Storable]
-    private readonly Dictionary<object, Dictionary<int, int>> tries = new Dictionary<object, Dictionary<int, int>>();
+  internal class TabularAvgQualityFunction : TabularQualityFunctionBase {
 
-    public IStateFunction StateFunction {
-      get {
-        return ((IValueParameter<IStateFunction>)Parameters["State function"]).Value;
-      }
-      set { ((IValueParameter<IStateFunction>)Parameters["State function"]).Value = value; }
-    }
-
-    public TabularAvgQualityFunction() {
-      Parameters.Add(new ValueParameter<IStateFunction>("State function", "The function that is used to map partial trees to states", new DefaultStateFunction()));
-    }
-
-    public double Q(object state, int action) {
-      // an action that has never been tried has q == infinity
-      Dictionary<int, double> stateActionQualities;
-      if (!q.TryGetValue(state, out stateActionQualities)) return double.PositiveInfinity;
-      double quality;
-      if (!stateActionQualities.TryGetValue(action, out quality)) return double.PositiveInfinity;
-      return quality;
-    }
-
-    public void Update(object state, int action, double observedQuality) {
-      Dictionary<int, int> stateActionTries;
-      if (!tries.TryGetValue(state, out stateActionTries)) {
-        stateActionTries = new Dictionary<int, int>();
-        tries.Add(state, stateActionTries);
-      }
-      Dictionary<int, double> stateActionQualities = null;
-      if (!q.TryGetValue(state, out stateActionQualities)) {
-        stateActionQualities = new Dictionary<int, double>();
-        q.Add(state, stateActionQualities);
-      }
-      int t;
-      if (!stateActionTries.TryGetValue(action, out t)) {
-        t = 0;
-        stateActionTries.Add(action, t + 1);
-        stateActionQualities.Add(action, observedQuality);
-      } else {
-        stateActionTries[action] = t + 1;
-        var delta = observedQuality - stateActionQualities[action];
-        stateActionQualities[action] = stateActionQualities[action] + (1.0 / (t + 1)) * delta; // incremental calculation of average
-      }
-    }
-
-    public int Tries(object state, int action) {
-      Dictionary<int, int> stateTries;
-      if (!tries.TryGetValue(state, out stateTries)) return 0;
-      int stateActionTries;
-      if (!stateTries.TryGetValue(action, out stateActionTries)) return 0;
-      return stateActionTries;
+    public TabularAvgQualityFunction()
+      : base() {
     }
 
@@ -71 +20 @@
     protected TabularAvgQualityFunction(TabularAvgQualityFunction original, Cloner cloner)
       : base(original, cloner) {
-      this.q = new Dictionary<object, Dictionary<int, double>>(original.q);
-      this.tries = new Dictionary<object, Dictionary<int, int>>(original.tries);
     }
     public override IDeepCloneable Clone(Cloner cloner) {
@@ -79 +26 @@
     #endregion
 
-    public void InitializeState() {
-      ClearState();
-    }
-
-    public void ClearState() {
-      q.Clear();
-      tries.Clear();
+    protected override double CalculateNewQ(object state, int action, double observedQuality) {
+      var delta = observedQuality - Q(state, action);
+      return Q(state, action) + (1.0 / Tries(state, action)) * delta; // iterative calculation of mean
     }
   }