Free cookie consent management tool by TermsFeed Policy Generator

source: branches/DataAnalysis/HeuristicLab.Problems.DataAnalysis/3.3/Operators/CovariantParsimonyPressure.cs @ 4255

Last change on this file since 4255 was 4255, checked in by gkronber, 14 years ago

Added complexity reduction scheme based on validation performance for CPP. #1142
File size: 14.4 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2010 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
4 *
5 * This file is part of HeuristicLab.
6 *
7 * HeuristicLab is free software: you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation, either version 3 of the License, or
10 * (at your option) any later version.
11 *
12 * HeuristicLab is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License
18 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
19 */
20#endregion
21
22using System;
23using System.Linq;
24using alglib;
25using HeuristicLab.Core;
26using HeuristicLab.Data;
27using HeuristicLab.Operators;
28using HeuristicLab.Optimization;
29using HeuristicLab.Parameters;
30using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
31using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
32using System.Collections.Generic;
33using HeuristicLab.Problems.DataAnalysis.Evaluators;
34using HeuristicLab.Analysis;
35
36namespace HeuristicLab.Problems.DataAnalysis.Operators {
37  [Item("Covariant Parsimony Pressure", "Covariant Parsimony Pressure.")]
38  [StorableClass]
39  public class CovariantParsimonyPressure : SingleSuccessorOperator {
40    public IScopeTreeLookupParameter<SymbolicExpressionTree> SymbolicExpressionTreeParameter {
41      get { return (IScopeTreeLookupParameter<SymbolicExpressionTree>)Parameters["SymbolicExpressionTree"]; }
42    }
43    public IScopeTreeLookupParameter<DoubleValue> QualityParameter {
44      get { return (IScopeTreeLookupParameter<DoubleValue>)Parameters["Quality"]; }
45    }
46    public IScopeTreeLookupParameter<DoubleValue> AdjustedQualityParameter {
47      get { return (IScopeTreeLookupParameter<DoubleValue>)Parameters["AdjustedQuality"]; }
48    }
49
50    public ILookupParameter<BoolValue> MaximizationParameter {
51      get { return (ILookupParameter<BoolValue>)Parameters["Maximization"]; }
52    }
53    public IValueLookupParameter<DoubleValue> KParameter {
54      get { return (IValueLookupParameter<DoubleValue>)Parameters["K"]; }
55    }
56    public ILookupParameter<IntValue> GenerationsParameter {
57      get { return (ILookupParameter<IntValue>)Parameters["Generations"]; }
58    }
59    public IValueLookupParameter<IntValue> FirstGenerationParameter {
60      get { return (IValueLookupParameter<IntValue>)Parameters["FirstGenerationParameter"]; }
61    }
62    public IValueLookupParameter<BoolValue> AntiOverfitParameter {
63      get { return (IValueLookupParameter<BoolValue>)Parameters["AntiOverfit"]; }
64    }
65    public ILookupParameter<DataTable> ValidationQualityParameter {
66      get { return (ILookupParameter<DataTable>)Parameters["Validation Quality"]; }
67    }
68    public ILookupParameter<DoubleValue> CurrentBestValidationQualityParameter {
69      get { return (ILookupParameter<DoubleValue>)Parameters["Current best validation quality"]; }
70    }
71    public ILookupParameter<DoubleValue> BestValidationQualityParameter {
72      get { return (ILookupParameter<DoubleValue>)Parameters["Best solution quality (validation)"]; }
73    }
74    public ILookupParameter<DoubleValue> LengthCorrelationParameter {
75      get { return (ILookupParameter<DoubleValue>)Parameters["Correlation(Length, AdjustedFitness)"]; }
76    }
77    public ILookupParameter<DoubleValue> FitnessCorrelationParameter {
78      get { return (ILookupParameter<DoubleValue>)Parameters["Correlation(Fitness, AdjustedFitness)"]; }
79    }
80    public IValueLookupParameter<IntValue> GenerationSpanParameter {
81      get { return (IValueLookupParameter<IntValue>)Parameters["GenerationSpan"]; }
82    }
83    public IValueLookupParameter<PercentValue> OverfittingLimitParameter {
84      get { return (IValueLookupParameter<PercentValue>)Parameters["OverfittingLimit"]; }
85    }
86    public IValueLookupParameter<PercentValue> ComplexityAdaptionParameter {
87      get { return (IValueLookupParameter<PercentValue>)Parameters["ComplexityAdaption"]; }
88    }
89    public ILookupParameter<DataTable> QualitiesParameter {
90      get { return (ILookupParameter<DataTable>)Parameters["Qualities"]; }
91    }
92
93    public CovariantParsimonyPressure(bool deserializing) : base(deserializing) { }
94    public CovariantParsimonyPressure()
95      : base() {
96      Parameters.Add(new ScopeTreeLookupParameter<SymbolicExpressionTree>("SymbolicExpressionTree"));
97      Parameters.Add(new ScopeTreeLookupParameter<DoubleValue>("Quality"));
98      Parameters.Add(new ScopeTreeLookupParameter<DoubleValue>("AdjustedQuality"));
99      Parameters.Add(new LookupParameter<BoolValue>("Maximization"));
100      Parameters.Add(new ValueLookupParameter<DoubleValue>("K", new DoubleValue(1.0)));
101      Parameters.Add(new LookupParameter<IntValue>("Generations"));
102      Parameters.Add(new ValueLookupParameter<IntValue>("FirstGenerationParameter", new IntValue(5)));
103      Parameters.Add(new ValueLookupParameter<BoolValue>("AntiOverfit", new BoolValue(false)));
104      //Parameters.Add(new LookupParameter<DoubleValue>("Current best validation quality"));
105      //Parameters.Add(new LookupParameter<DoubleValue>("Best solution quality (validation)"));
106      Parameters.Add(new LookupParameter<DataTable>("Validation Quality"));
107      Parameters.Add(new LookupParameter<DataTable>("Qualities"));
108      Parameters.Add(new ValueLookupParameter<IntValue>("GenerationSpan", new IntValue(5)));
109      Parameters.Add(new ValueLookupParameter<PercentValue>("OverfittingLimit", new PercentValue(5)));
110      Parameters.Add(new ValueLookupParameter<PercentValue>("ComplexityAdaption", new PercentValue(-5)));
111      Parameters.Add(new LookupParameter<DoubleValue>("Correlation(Length, AdjustedFitness)"));
112      Parameters.Add(new LookupParameter<DoubleValue>("Correlation(Fitness, AdjustedFitness)"));
113    }
114
115    [StorableHook(Persistence.Default.CompositeSerializers.Storable.HookType.AfterDeserialization)]
116    private void AfterDeserialization() {
117      if (!Parameters.ContainsKey("Maximization"))
118        Parameters.Add(new LookupParameter<BoolValue>("Maximization"));
119      if (!Parameters.ContainsKey("K"))
120        Parameters.Add(new ValueLookupParameter<DoubleValue>("K", new DoubleValue(1.0)));
121      if (!Parameters.ContainsKey("AdjustedQuality")) {
122        Parameters.Add(new ScopeTreeLookupParameter<DoubleValue>("AdjustedQuality"));
123      }
124      if (!Parameters.ContainsKey("Generations")) {
125        Parameters.Add(new LookupParameter<IntValue>("Generations"));
126      }
127      if (!Parameters.ContainsKey("FirstGenerationParameter")) {
128        Parameters.Add(new ValueLookupParameter<IntValue>("FirstGenerationParameter", new IntValue(5)));
129      }
130      if (!Parameters.ContainsKey("AntiOverfit")) {
131        Parameters.Add(new ValueLookupParameter<BoolValue>("AntiOverfit", new BoolValue(false)));
132      }
133      //if (!Parameters.ContainsKey("Current best validation quality")) {
134      //  Parameters.Add(new LookupParameter<DoubleValue>("Current best validation quality"));
135      //}
136      //if (!Parameters.ContainsKey("Best solution quality (validation)")) {
137      //  Parameters.Add(new LookupParameter<DoubleValue>("Best solution quality (validation)"));
138      //}
139      if (!Parameters.ContainsKey("Correlation(Length, AdjustedFitness)")) {
140        Parameters.Add(new LookupParameter<DoubleValue>("Correlation(Length, AdjustedFitness)"));
141      }
142      if (!Parameters.ContainsKey("Correlation(Fitness, AdjustedFitness)")) {
143        Parameters.Add(new LookupParameter<DoubleValue>("Correlation(Fitness, AdjustedFitness)"));
144      }
145      if (!Parameters.ContainsKey("Validation Quality")) {
146        Parameters.Add(new LookupParameter<DataTable>("Validation Quality"));
147      }
148      if (!Parameters.ContainsKey("Qualities")) {
149        Parameters.Add(new LookupParameter<DataTable>("Qualities"));
150      }
151      if (!Parameters.ContainsKey("GenerationSpan")) {
152        Parameters.Add(new ValueLookupParameter<IntValue>("GenerationSpan", new IntValue(5)));
153      }
154      if (!Parameters.ContainsKey("OverfittingLimit")) {
155        Parameters.Add(new ValueLookupParameter<PercentValue>("OverfittingLimit", new PercentValue(5)));
156      }
157      if (!Parameters.ContainsKey("ComplexityAdaption")) {
158        Parameters.Add(new ValueLookupParameter<PercentValue>("ComplexityAdaption", new PercentValue(-5)));
159      }
160    }
161
162    public override IOperation Apply() {
163      ItemArray<SymbolicExpressionTree> trees = SymbolicExpressionTreeParameter.ActualValue;
164      ItemArray<DoubleValue> qualities = QualityParameter.ActualValue;
165      // always apply Parsimony pressure if anti-overfit is false
166      // otherwise appliy PP only when we are currently overfitting
167      if (GenerationsParameter.ActualValue != null && GenerationsParameter.ActualValue.Value >= FirstGenerationParameter.ActualValue.Value &&
168         (AntiOverfitParameter.ActualValue.Value == false || IsOverfitting())) {
169        var lengths = from tree in trees
170                      select tree.Size;
171        double k = KParameter.ActualValue.Value;
172
173        // calculate cov(f, l) and cov(l, l^k)
174        OnlineCovarianceEvaluator lengthFitnessCovEvaluator = new OnlineCovarianceEvaluator();
175        OnlineCovarianceEvaluator lengthAdjLengthCovEvaluator = new OnlineCovarianceEvaluator();
176        OnlineMeanAndVarianceCalculator lengthMeanCalculator = new OnlineMeanAndVarianceCalculator();
177        OnlineMeanAndVarianceCalculator fitnessMeanCalculator = new OnlineMeanAndVarianceCalculator();
178        OnlineMeanAndVarianceCalculator adjLengthMeanCalculator = new OnlineMeanAndVarianceCalculator();
179        var lengthEnumerator = lengths.GetEnumerator();
180        var qualityEnumerator = qualities.GetEnumerator();
181        while (lengthEnumerator.MoveNext() & qualityEnumerator.MoveNext()) {
182          double fitness = qualityEnumerator.Current.Value;
183          if (!MaximizationParameter.ActualValue.Value) {
184            // use f = 1 / (1 + quality) for minimization problems
185            fitness = 1.0 / (1.0 + fitness);
186          }
187          lengthFitnessCovEvaluator.Add(lengthEnumerator.Current, fitness);
188          lengthAdjLengthCovEvaluator.Add(lengthEnumerator.Current, Math.Pow(lengthEnumerator.Current, k));
189          lengthMeanCalculator.Add(lengthEnumerator.Current);
190          fitnessMeanCalculator.Add(fitness);
191          adjLengthMeanCalculator.Add(Math.Pow(lengthEnumerator.Current, k));
192        }
193
194        double sizeAdaption = lengthMeanCalculator.Mean * ComplexityAdaptionParameter.ActualValue.Value;
195        if (sizeAdaption < 0) sizeAdaption = Math.Floor(sizeAdaption);
196        else sizeAdaption = Math.Ceiling(sizeAdaption);
197        double g = lengthMeanCalculator.Mean + sizeAdaption;
198
199        //            cov(l, f) - (g(t+1) - mu(t)) avgF
200        // c(t) =  --------------------------------------------
201        //           cov(l, l^k) - (g(t+1) - mu(t)) E[l^k]
202        double c = lengthFitnessCovEvaluator.Covariance - (g - lengthMeanCalculator.Mean) * fitnessMeanCalculator.Mean;
203        c /= lengthAdjLengthCovEvaluator.Covariance - (g - lengthMeanCalculator.Mean) * adjLengthMeanCalculator.Mean;
204
205        // adjust fitness
206        bool maximization = MaximizationParameter.ActualValue.Value;
207
208        lengthEnumerator = lengths.GetEnumerator();
209        qualityEnumerator = qualities.GetEnumerator();
210        int i = 0;
211        ItemArray<DoubleValue> adjQualities = new ItemArray<DoubleValue>(qualities.Length);
212
213        while (lengthEnumerator.MoveNext() & qualityEnumerator.MoveNext()) {
214          adjQualities[i++] = new DoubleValue(qualityEnumerator.Current.Value - c * Math.Pow(lengthEnumerator.Current, k));
215        }
216        AdjustedQualityParameter.ActualValue = adjQualities;
217        double[] lengthArr = lengths.Select(x => (double)x).ToArray<double>();
218
219        double[] adjFitess = (from f in AdjustedQualityParameter.ActualValue
220                              select f.Value).ToArray<double>();
221        double[] fitnessArr = (from f in QualityParameter.ActualValue
222                               let normFit = maximization ? f.Value : 1.0 / (1.0 + f.Value)
223                               select normFit).ToArray<double>();
224
225        LengthCorrelationParameter.ActualValue = new DoubleValue(alglib.correlation.spearmanrankcorrelation(lengthArr, adjFitess, lengthArr.Length));
226        FitnessCorrelationParameter.ActualValue = new DoubleValue(alglib.correlation.spearmanrankcorrelation(fitnessArr, adjFitess, lengthArr.Length));
227
228      } else {
229        // adjusted fitness is equal to fitness
230        AdjustedQualityParameter.ActualValue = (ItemArray<DoubleValue>)QualityParameter.ActualValue.Clone();
231        FitnessCorrelationParameter.ActualValue = new DoubleValue(1.0);
232
233        double[] lengths = (from tree in trees
234                            select (double)tree.Size).ToArray<double>();
235
236        double[] fitess = (from f in AdjustedQualityParameter.ActualValue
237                           select f.Value).ToArray<double>();
238
239        LengthCorrelationParameter.ActualValue = new DoubleValue(alglib.correlation.spearmanrankcorrelation(lengths, fitess, lengths.Length));
240      }
241      return base.Apply();
242    }
243
244    private bool IsOverfitting() {
245      bool maximization = MaximizationParameter.ActualValue.Value;
246      DataTable trainingQualities = QualitiesParameter.ActualValue;
247      DataTable validationQualities = ValidationQualityParameter.ActualValue;
248      int genSpan = GenerationSpanParameter.ActualValue.Value;
249      if (validationQualities == null || trainingQualities == null) return false;
250      if (validationQualities.Rows["Best solution quality (validation)"].Values.Count < genSpan) return false;
251
252      IEnumerable<double> bestTrainingQualities = trainingQualities.Rows["CurrentBestQuality"].Values;
253      IEnumerable<double> bestValidationQualities = validationQualities.Rows["Current best validation quality"].Values;
254
255      double trainingAvg = bestTrainingQualities.Reverse().Take(genSpan).Average();
256      double validationAvg = bestValidationQualities.Reverse().Take(genSpan).Average();
257
258      double maxPercentDiff = OverfittingLimitParameter.ActualValue.Value;
259
260      double percentDiff = maximization ? trainingAvg / validationAvg - 1 : validationAvg / trainingAvg - 1;
261      return percentDiff > maxPercentDiff;
262    }
263  }
264}
Note: See TracBrowser for help on using the repository browser.