source: branches/MemPRAlgorithm/HeuristicLab.Algorithms.MemPR/3.3/MemPRContext.cs @ 14544

Last change on this file since 14544 was 14544, checked in by abeham, 3 years ago

#2701:

  • LLE: Added equality comparer
  • MemPR:
    • Added GPR to learn about heuristic performance
    • Changed Breeding to do more exhaustive search on crossover
    • Added Delinking separately to Relinking
    • Rewrote d/relinking for LLE
    • Reduce usage of local search
    • Renamed TabuWalk to AdaptiveWalk
    • Rewrote adaptive walk for binary problems
    • Renamed LLE namespace to Grouping to avoid namespace clashes
File size: 35.6 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2016 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.Collections.Generic;
24using System.Linq;
25using System.Runtime.CompilerServices;
26using System.Threading;
27using HeuristicLab.Algorithms.DataAnalysis;
28using HeuristicLab.Algorithms.MemPR.Interfaces;
29using HeuristicLab.Common;
30using HeuristicLab.Core;
31using HeuristicLab.Data;
32using HeuristicLab.Optimization;
33using HeuristicLab.Parameters;
34using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
35using HeuristicLab.Problems.DataAnalysis;
36using HeuristicLab.Random;
37using ExecutionContext = HeuristicLab.Core.ExecutionContext;
38
39namespace HeuristicLab.Algorithms.MemPR {
40  [Item("MemPRContext", "Abstract base class for MemPR contexts.")]
41  [StorableClass]
42  public abstract class MemPRPopulationContext<TProblem, TSolution, TPopulationContext, TSolutionContext> : ParameterizedNamedItem,
43    IPopulationBasedHeuristicAlgorithmContext<TProblem, TSolution>, ISolutionModelContext<TSolution>
44      where TProblem : class, IItem, ISingleObjectiveProblemDefinition
45      where TSolution : class, IItem
46      where TPopulationContext : MemPRPopulationContext<TProblem, TSolution, TPopulationContext, TSolutionContext>
47      where TSolutionContext : MemPRSolutionContext<TProblem, TSolution, TPopulationContext, TSolutionContext> {
48
49    private IExecutionContext parent;
50    public IExecutionContext Parent {
51      get { return parent; }
52      set { parent = value; }
53    }
54
55    [Storable]
56    private IScope scope;
57    public IScope Scope {
58      get { return scope; }
59      private set { scope = value; }
60    }
61
62    IKeyedItemCollection<string, IParameter> IExecutionContext.Parameters {
63      get { return Parameters; }
64    }
65
66    [Storable]
67    private IValueParameter<TProblem> problem;
68    public TProblem Problem {
69      get { return problem.Value; }
70      set { problem.Value = value; }
71    }
72
73    [Storable]
74    private IValueParameter<BoolValue> initialized;
75    public bool Initialized {
76      get { return initialized.Value.Value; }
77      set { initialized.Value.Value = value; }
78    }
79
80    [Storable]
81    private IValueParameter<IntValue> iterations;
82    public int Iterations {
83      get { return iterations.Value.Value; }
84      set { iterations.Value.Value = value; }
85    }
86
87    [Storable]
88    private IValueParameter<IntValue> evaluatedSolutions;
89    public int EvaluatedSolutions {
90      get { return evaluatedSolutions.Value.Value; }
91      set { evaluatedSolutions.Value.Value = value; }
92    }
93
94    [Storable]
95    private IValueParameter<DoubleValue> bestQuality;
96    public double BestQuality {
97      get { return bestQuality.Value.Value; }
98      set { bestQuality.Value.Value = value; }
99    }
100
101    [Storable]
102    private IValueParameter<TSolution> bestSolution;
103    public TSolution BestSolution {
104      get { return bestSolution.Value; }
105      set { bestSolution.Value = value; }
106    }
107
108    [Storable]
109    private IValueParameter<IntValue> localSearchEvaluations;
110    public int LocalSearchEvaluations {
111      get { return localSearchEvaluations.Value.Value; }
112      set { localSearchEvaluations.Value.Value = value; }
113    }
114
115    [Storable]
116    private IValueParameter<IntValue> byBreeding;
117    public int ByBreeding {
118      get { return byBreeding.Value.Value; }
119      set { byBreeding.Value.Value = value; }
120    }
121
122    [Storable]
123    private IValueParameter<IntValue> byRelinking;
124    public int ByRelinking {
125      get { return byRelinking.Value.Value; }
126      set { byRelinking.Value.Value = value; }
127    }
128
129    [Storable]
130    private IValueParameter<IntValue> byDelinking;
131    public int ByDelinking {
132      get { return byDelinking.Value.Value; }
133      set { byDelinking.Value.Value = value; }
134    }
135
136    [Storable]
137    private IValueParameter<IntValue> bySampling;
138    public int BySampling {
139      get { return bySampling.Value.Value; }
140      set { bySampling.Value.Value = value; }
141    }
142
143    [Storable]
144    private IValueParameter<IntValue> byHillclimbing;
145    public int ByHillclimbing {
146      get { return byHillclimbing.Value.Value; }
147      set { byHillclimbing.Value.Value = value; }
148    }
149
150    [Storable]
151    private IValueParameter<IntValue> byAdaptivewalking;
152    public int ByAdaptivewalking {
153      get { return byAdaptivewalking.Value.Value; }
154      set { byAdaptivewalking.Value.Value = value; }
155    }
156
157    [Storable]
158    private IValueParameter<IRandom> random;
159    public IRandom Random {
160      get { return random.Value; }
161      set { random.Value = value; }
162    }
163   
164    public IEnumerable<ISingleObjectiveSolutionScope<TSolution>> Population {
165      get { return scope.SubScopes.OfType<ISingleObjectiveSolutionScope<TSolution>>(); }
166    }
167    public void AddToPopulation(ISingleObjectiveSolutionScope<TSolution> solScope) {
168      scope.SubScopes.Add(solScope);
169    }
170    public void ReplaceAtPopulation(int index, ISingleObjectiveSolutionScope<TSolution> solScope) {
171      scope.SubScopes[index] = solScope;
172    }
173    public ISingleObjectiveSolutionScope<TSolution> AtPopulation(int index) {
174      return scope.SubScopes[index] as ISingleObjectiveSolutionScope<TSolution>;
175    }
176    public void SortPopulation() {
177      scope.SubScopes.Replace(scope.SubScopes.OfType<ISingleObjectiveSolutionScope<TSolution>>().OrderBy(x => Problem.Maximization ? -x.Fitness : x.Fitness).ToList());
178    }
179    public int PopulationCount {
180      get { return scope.SubScopes.Count; }
181    }
182
183    [Storable]
184    private IConfidenceRegressionModel breedingPerformanceModel;
185    public IConfidenceRegressionModel BreedingPerformanceModel {
186      get { return breedingPerformanceModel; }
187    }
188    [Storable]
189    private List<Tuple<double, double, double>> breedingStat;
190    public List<Tuple<double, double, double>> BreedingStat {
191      get { return breedingStat; }
192    }
193    [Storable]
194    private IConfidenceRegressionModel relinkingPerformanceModel;
195    public IConfidenceRegressionModel RelinkingPerformanceModel {
196      get { return relinkingPerformanceModel; }
197    }
198    [Storable]
199    private List<Tuple<double, double, double>> relinkingStat;
200    public List<Tuple<double, double, double>> RelinkingStat {
201      get { return relinkingStat; }
202    }
203    [Storable]
204    private IConfidenceRegressionModel delinkingPerformanceModel;
205    public IConfidenceRegressionModel DelinkingPerformanceModel {
206      get { return delinkingPerformanceModel; }
207    }
208    [Storable]
209    private List<Tuple<double, double, double>> delinkingStat;
210    public List<Tuple<double, double, double>> DelinkingStat {
211      get { return delinkingStat; }
212    }
213    [Storable]
214    private IConfidenceRegressionModel samplingPerformanceModel;
215    public IConfidenceRegressionModel SamplingPerformanceModel {
216      get { return samplingPerformanceModel; }
217    }
218    [Storable]
219    private List<Tuple<double, double>> samplingStat;
220    public List<Tuple<double, double>> SamplingStat {
221      get { return samplingStat; }
222    }
223    [Storable]
224    private IConfidenceRegressionModel hillclimbingPerformanceModel;
225    public IConfidenceRegressionModel HillclimbingPerformanceModel {
226      get { return hillclimbingPerformanceModel; }
227    }
228    [Storable]
229    private List<Tuple<double, double>> hillclimbingStat;
230    public List<Tuple<double, double>> HillclimbingStat {
231      get { return hillclimbingStat; }
232    }
233    [Storable]
234    private IConfidenceRegressionModel adaptiveWalkPerformanceModel;
235    public IConfidenceRegressionModel AdaptiveWalkPerformanceModel {
236      get { return adaptiveWalkPerformanceModel; }
237    }
238    [Storable]
239    private List<Tuple<double, double>> adaptivewalkingStat;
240    public List<Tuple<double, double>> AdaptivewalkingStat {
241      get { return adaptivewalkingStat; }
242    }
243
244    [Storable]
245    public ISolutionModel<TSolution> Model { get; set; }
246
247    [StorableConstructor]
248    protected MemPRPopulationContext(bool deserializing) : base(deserializing) { }
249    protected MemPRPopulationContext(MemPRPopulationContext<TProblem, TSolution, TPopulationContext, TSolutionContext> original, Cloner cloner)
250      : base(original, cloner) {
251      scope = cloner.Clone(original.scope);
252      problem = cloner.Clone(original.problem);
253      initialized = cloner.Clone(original.initialized);
254      iterations = cloner.Clone(original.iterations);
255      evaluatedSolutions = cloner.Clone(original.evaluatedSolutions);
256      bestQuality = cloner.Clone(original.bestQuality);
257      bestSolution = cloner.Clone(original.bestSolution);
258      localSearchEvaluations = cloner.Clone(original.localSearchEvaluations);
259      byBreeding = cloner.Clone(original.byBreeding);
260      byRelinking = cloner.Clone(original.byRelinking);
261      byDelinking = cloner.Clone(original.byDelinking);
262      bySampling = cloner.Clone(original.bySampling);
263      byHillclimbing = cloner.Clone(original.byHillclimbing);
264      byAdaptivewalking = cloner.Clone(original.byAdaptivewalking);
265      random = cloner.Clone(original.random);
266      breedingPerformanceModel = cloner.Clone(original.breedingPerformanceModel);
267      breedingStat = original.breedingStat.Select(x => Tuple.Create(x.Item1, x.Item2, x.Item3)).ToList();
268      relinkingPerformanceModel = cloner.Clone(original.relinkingPerformanceModel);
269      relinkingStat = original.relinkingStat.Select(x => Tuple.Create(x.Item1, x.Item2, x.Item3)).ToList();
270      delinkingPerformanceModel = cloner.Clone(original.delinkingPerformanceModel);
271      delinkingStat = original.delinkingStat.Select(x => Tuple.Create(x.Item1, x.Item2, x.Item3)).ToList();
272      samplingPerformanceModel = cloner.Clone(original.samplingPerformanceModel);
273      samplingStat = original.samplingStat.Select(x => Tuple.Create(x.Item1, x.Item2)).ToList();
274      hillclimbingPerformanceModel = cloner.Clone(original.hillclimbingPerformanceModel);
275      hillclimbingStat = original.hillclimbingStat.Select(x => Tuple.Create(x.Item1, x.Item2)).ToList();
276      adaptiveWalkPerformanceModel = cloner.Clone(original.adaptiveWalkPerformanceModel);
277      adaptivewalkingStat = original.adaptivewalkingStat.Select(x => Tuple.Create(x.Item1, x.Item2)).ToList();
278     
279      Model = cloner.Clone(original.Model);
280    }
281    public MemPRPopulationContext() : this("MemPRContext") { }
282    public MemPRPopulationContext(string name) : base(name) {
283      scope = new Scope("Global");
284
285      Parameters.Add(problem = new ValueParameter<TProblem>("Problem"));
286      Parameters.Add(initialized = new ValueParameter<BoolValue>("Initialized", new BoolValue(false)));
287      Parameters.Add(iterations = new ValueParameter<IntValue>("Iterations", new IntValue(0)));
288      Parameters.Add(evaluatedSolutions = new ValueParameter<IntValue>("EvaluatedSolutions", new IntValue(0)));
289      Parameters.Add(bestQuality = new ValueParameter<DoubleValue>("BestQuality", new DoubleValue(double.NaN)));
290      Parameters.Add(bestSolution = new ValueParameter<TSolution>("BestSolution"));
291      Parameters.Add(localSearchEvaluations = new ValueParameter<IntValue>("LocalSearchEvaluations", new IntValue(0)));
292      Parameters.Add(byBreeding = new ValueParameter<IntValue>("ByBreeding", new IntValue(0)));
293      Parameters.Add(byRelinking = new ValueParameter<IntValue>("ByRelinking", new IntValue(0)));
294      Parameters.Add(byDelinking = new ValueParameter<IntValue>("ByDelinking", new IntValue(0)));
295      Parameters.Add(bySampling = new ValueParameter<IntValue>("BySampling", new IntValue(0)));
296      Parameters.Add(byHillclimbing = new ValueParameter<IntValue>("ByHillclimbing", new IntValue(0)));
297      Parameters.Add(byAdaptivewalking = new ValueParameter<IntValue>("ByAdaptivewalking", new IntValue(0)));
298      Parameters.Add(random = new ValueParameter<IRandom>("Random", new MersenneTwister()));
299
300      breedingStat = new List<Tuple<double, double, double>>();
301      relinkingStat = new List<Tuple<double, double, double>>();
302      delinkingStat = new List<Tuple<double, double, double>>();
303      samplingStat = new List<Tuple<double, double>>();
304      hillclimbingStat = new List<Tuple<double, double>>();
305      adaptivewalkingStat = new List<Tuple<double, double>>();
306    }
307
308    public abstract TSolutionContext CreateSingleSolutionContext(ISingleObjectiveSolutionScope<TSolution> solution);
309
310    public void IncrementEvaluatedSolutions(int byEvaluations) {
311      if (byEvaluations < 0) throw new ArgumentException("Can only increment and not decrement evaluated solutions.");
312      EvaluatedSolutions += byEvaluations;
313    }
314
315    public void RelearnBreedingPerformanceModel() {
316      breedingPerformanceModel = RunRegression(PrepareRegression(BreedingStat), breedingPerformanceModel).Model;
317    }
318    public bool BreedingSuited(ISingleObjectiveSolutionScope<TSolution> p1, ISingleObjectiveSolutionScope<TSolution> p2) {
319      if (breedingPerformanceModel == null) return true;
320      double minI1 = double.MaxValue, minI2 = double.MaxValue, maxI1 = double.MinValue, maxI2 = double.MinValue;
321      foreach (var d in BreedingStat) {
322        if (d.Item1 < minI1) minI1 = d.Item1;
323        if (d.Item1 > maxI1) maxI1 = d.Item1;
324        if (d.Item2 < minI2) minI2 = d.Item2;
325        if (d.Item2 > maxI2) maxI2 = d.Item2;
326      }
327      if (IsBetter(p1, p2)) {
328        if (p1.Fitness < minI1 || p1.Fitness > maxI1 || p2.Fitness < minI2 || p2.Fitness > maxI2)
329          return true;
330        return Random.NextDouble() < ProbabilityAccept3dModel(p1.Fitness, p2.Fitness, breedingPerformanceModel);
331      }
332      if (p1.Fitness < minI2 || p1.Fitness > maxI2 || p2.Fitness < minI1 || p2.Fitness > maxI1)
333        return true;
334      return Random.NextDouble() < ProbabilityAccept3dModel(p2.Fitness, p1.Fitness, breedingPerformanceModel);
335    }
336
337    public void RelearnRelinkingPerformanceModel() {
338      relinkingPerformanceModel = RunRegression(PrepareRegression(RelinkingStat), relinkingPerformanceModel).Model;
339    }
340    public bool RelinkSuited(ISingleObjectiveSolutionScope<TSolution> p1, ISingleObjectiveSolutionScope<TSolution> p2) {
341      if (relinkingPerformanceModel == null) return true;
342      double minI1 = double.MaxValue, minI2 = double.MaxValue, maxI1 = double.MinValue, maxI2 = double.MinValue;
343      foreach (var d in RelinkingStat) {
344        if (d.Item1 < minI1) minI1 = d.Item1;
345        if (d.Item1 > maxI1) maxI1 = d.Item1;
346        if (d.Item2 < minI2) minI2 = d.Item2;
347        if (d.Item2 > maxI2) maxI2 = d.Item2;
348      }
349      if (IsBetter(p1, p2)) {
350        if (p1.Fitness < minI1 || p1.Fitness > maxI1 || p2.Fitness < minI2 || p2.Fitness > maxI2)
351          return true;
352        return Random.NextDouble() < ProbabilityAccept3dModel(p1.Fitness, p2.Fitness, relinkingPerformanceModel);
353      }
354      if (p1.Fitness < minI2 || p1.Fitness > maxI2 || p2.Fitness < minI1 || p2.Fitness > maxI1)
355        return true;
356      return Random.NextDouble() < ProbabilityAccept3dModel(p2.Fitness, p1.Fitness, relinkingPerformanceModel);
357    }
358
359    public void RelearnDelinkingPerformanceModel() {
360      delinkingPerformanceModel = RunRegression(PrepareRegression(DelinkingStat), delinkingPerformanceModel).Model;
361    }
362    public bool DelinkSuited(ISingleObjectiveSolutionScope<TSolution> p1, ISingleObjectiveSolutionScope<TSolution> p2) {
363      if (delinkingPerformanceModel == null) return true;
364      double minI1 = double.MaxValue, minI2 = double.MaxValue, maxI1 = double.MinValue, maxI2 = double.MinValue;
365      foreach (var d in DelinkingStat) {
366        if (d.Item1 < minI1) minI1 = d.Item1;
367        if (d.Item1 > maxI1) maxI1 = d.Item1;
368        if (d.Item2 < minI2) minI2 = d.Item2;
369        if (d.Item2 > maxI2) maxI2 = d.Item2;
370      }
371      if (IsBetter(p1, p2)) {
372        if (p1.Fitness < minI1 || p1.Fitness > maxI1 || p2.Fitness < minI2 || p2.Fitness > maxI2)
373          return true;
374        return Random.NextDouble() < ProbabilityAccept3dModel(p1.Fitness, p2.Fitness, delinkingPerformanceModel);
375      }
376      if (p1.Fitness < minI2 || p1.Fitness > maxI2 || p2.Fitness < minI1 || p2.Fitness > maxI1)
377        return true;
378      return Random.NextDouble() < ProbabilityAccept3dModel(p2.Fitness, p1.Fitness, delinkingPerformanceModel);
379    }
380
381    public void RelearnSamplingPerformanceModel() {
382      samplingPerformanceModel = RunRegression(PrepareRegression(SamplingStat), samplingPerformanceModel).Model;
383    }
384    public bool SamplingSuited() {
385      if (samplingPerformanceModel == null) return true;
386      return Random.NextDouble() < ProbabilityAccept2dModel(Population.Average(x => x.Fitness), samplingPerformanceModel);
387    }
388
389    public void RelearnHillclimbingPerformanceModel() {
390      hillclimbingPerformanceModel = RunRegression(PrepareRegression(HillclimbingStat), hillclimbingPerformanceModel).Model;
391    }
392    public bool HillclimbingSuited(ISingleObjectiveSolutionScope<TSolution> scope) {
393      if (hillclimbingPerformanceModel == null) return true;
394      if (scope.Fitness < HillclimbingStat.Min(x => x.Item1) || scope.Fitness > HillclimbingStat.Max(x => x.Item1))
395        return true;
396      return Random.NextDouble() < ProbabilityAccept2dModel(scope.Fitness, hillclimbingPerformanceModel);
397    }
398    public bool HillclimbingSuited(double startingFitness) {
399      if (hillclimbingPerformanceModel == null) return true;
400      if (startingFitness < HillclimbingStat.Min(x => x.Item1) || startingFitness > HillclimbingStat.Max(x => x.Item1))
401        return true;
402      return Random.NextDouble() < ProbabilityAccept2dModel(startingFitness, hillclimbingPerformanceModel);
403    }
404
405    public void RelearnAdaptiveWalkPerformanceModel() {
406      adaptiveWalkPerformanceModel = RunRegression(PrepareRegression(AdaptivewalkingStat), adaptiveWalkPerformanceModel).Model;
407    }
408    public bool AdaptivewalkingSuited(ISingleObjectiveSolutionScope<TSolution> scope) {
409      if (adaptiveWalkPerformanceModel == null) return true;
410      if (scope.Fitness < AdaptivewalkingStat.Min(x => x.Item1) || scope.Fitness > AdaptivewalkingStat.Max(x => x.Item1))
411        return true;
412      return Random.NextDouble() < ProbabilityAccept2dModel(scope.Fitness, adaptiveWalkPerformanceModel);
413    }
414    public bool AdaptivewalkingSuited(double startingFitness) {
415      if (adaptiveWalkPerformanceModel == null) return true;
416      if (startingFitness < AdaptivewalkingStat.Min(x => x.Item1) || startingFitness > AdaptivewalkingStat.Max(x => x.Item1))
417        return true;
418      return Random.NextDouble() < ProbabilityAccept2dModel(startingFitness, adaptiveWalkPerformanceModel);
419    }
420
421    public IConfidenceRegressionSolution GetSolution(IConfidenceRegressionModel model, List<Tuple<double, double>> data) {
422      return new ConfidenceRegressionSolution(model, PrepareRegression(data));
423    }
424    public IConfidenceRegressionSolution GetSolution(IConfidenceRegressionModel model, List<Tuple<double, double, double>> data) {
425      return new ConfidenceRegressionSolution(model, PrepareRegression(data));
426    }
427
428    protected RegressionProblemData PrepareRegression(List<Tuple<double, double>> sample) {
429      var inCol = new List<double>();
430      var outCol = new List<double>();
431      foreach (var next in sample.Shuffle(Random)) {
432        inCol.Add(next.Item1);
433        outCol.Add(next.Item2);
434      }
435      var ds = new Dataset(new[] { "in", "out" }, new[] { inCol, outCol });
436      var regPrb = new RegressionProblemData(ds, new[] { "in" }, "out") {
437        TrainingPartition = { Start = 0, End = Math.Min(50, sample.Count) },
438        TestPartition = { Start = Math.Min(50, sample.Count), End = sample.Count }
439      };
440      return regPrb;
441    }
442
443    protected RegressionProblemData PrepareRegression(List<Tuple<double, double, double>> sample) {
444      var in1Col = new List<double>();
445      var in2Col = new List<double>();
446      var outCol = new List<double>();
447      foreach (var next in sample.Shuffle(Random)) {
448        in1Col.Add(next.Item1);
449        in2Col.Add(next.Item2);
450        outCol.Add(next.Item3);
451      }
452      var ds = new Dataset(new[] { "in1", "in2", "out" }, new[] { in1Col, in2Col, outCol });
453      var regPrb = new RegressionProblemData(ds, new[] { "in1", "in2" }, "out") {
454        TrainingPartition = { Start = 0, End = Math.Min(50, sample.Count) },
455        TestPartition = { Start = Math.Min(50, sample.Count), End = sample.Count }
456      };
457      return regPrb;
458    }
459
460    protected static IConfidenceRegressionSolution RunRegression(RegressionProblemData trainingData, IConfidenceRegressionModel baseLineModel = null) {
461      var baseline = baseLineModel != null ? new ConfidenceRegressionSolution(baseLineModel, trainingData) : null;
462      var gpr = new GaussianProcessRegression { Problem = { ProblemData = trainingData } };
463      if (trainingData.InputVariables.CheckedItems.Any(x => alglib.pearsoncorr2(trainingData.Dataset.GetDoubleValues(x.Value.Value).ToArray(), trainingData.TargetVariableValues.ToArray()) > 0.8)) {
464        gpr.MeanFunction = new MeanZero();
465        var cov1 = new CovarianceSum();
466        cov1.Terms.Add(new CovarianceLinearArd());
467        cov1.Terms.Add(new CovarianceConst());
468        gpr.CovarianceFunction = cov1;
469      }
470      IConfidenceRegressionSolution solution = null;
471      var cnt = 0;
472      do {
473        ExecuteAlgorithm(gpr);
474        solution = (IConfidenceRegressionSolution)gpr.Results["Solution"].Value;
475        cnt++;
476      } while (cnt < 10 && (solution == null || solution.TrainingRSquared.IsAlmost(0)));
477      if (baseline == null) return solution;
478      if (trainingData.Dataset.Rows < 60)
479        return solution.TrainingMeanAbsoluteError < baseline.TrainingMeanAbsoluteError ? solution : baseline;
480      return solution.TestMeanAbsoluteError < baseline.TestMeanAbsoluteError ? solution : baseline;
481    }
482
483    protected static void ExecuteAlgorithm(IAlgorithm algorithm) {
484      using (var evt = new AutoResetEvent(false)) {
485        EventHandler exeStateChanged = (o, args) => {
486          if (algorithm.ExecutionState == ExecutionState.Paused || algorithm.ExecutionState == ExecutionState.Stopped)
487            evt.Set();
488        };
489        algorithm.ExecutionStateChanged += exeStateChanged;
490        algorithm.Prepare(true);
491        algorithm.Start();
492        evt.WaitOne();
493        algorithm.ExecutionStateChanged -= exeStateChanged;
494      }
495    }
496
497    protected double ProbabilityAccept(ISingleObjectiveSolutionScope<TSolution> scope, IList<Tuple<double, double>> data) {
498      if (double.IsNaN(scope.Fitness)) throw new ArgumentException("solution not evaluated or quality unknown", "scope");
499      return ProbabilityAccept2d(scope.Fitness, data);
500    }
501
502    private double ProbabilityAccept2dModel(double a, IConfidenceRegressionModel model) {
503      var ds = new Dataset(new[] { "in", "out" }, new[] { new List<double> { a }, new List<double> { double.NaN } });
504      var mean = model.GetEstimatedValues(ds, new[] { 0 }).Single();
505      var sdev = Math.Sqrt(model.GetEstimatedVariances(ds, new[] { 0 }).Single());
506
507      var goal = Problem.Maximization ? Population.Min(x => x.Fitness) : Population.Max(x => x.Fitness);
508      var z = (goal - mean) / sdev;
509      return Problem.Maximization ? 1.0 - Phi(z) /* P(X >= z) */ : Phi(z); // P(X <= z)
510    }
511    protected double ProbabilityAccept2d(double startingFitness, IList<Tuple<double, double>> data) {
512      if (data.Count < 10) return 1.0;
513      var samples = 0;
514      double meanStart = 0, meanStartOld = 0, meanEnd = 0, meanEndOld = 0;
515      double varStart = 0, varStartOld = 0, varEnd = 0, varEndOld = 0;
516      for (var i = 0; i < data.Count; i++) {
517        samples++;
518        var x = data[i].Item1;
519        var y = data[i].Item2;
520
521        if (samples == 1) {
522          meanStartOld = x;
523          meanEndOld = y;
524        } else {
525          meanStart = meanStartOld + (x - meanStartOld) / samples;
526          meanEnd = meanEndOld + (y - meanEndOld) / samples;
527          varStart = varStartOld + (x - meanStartOld) * (x - meanStart) / (samples - 1);
528          varEnd = varEndOld + (y - meanEndOld) * (y - meanEnd) / (samples - 1);
529
530          meanStartOld = meanStart;
531          meanEndOld = meanEnd;
532          varStartOld = varStart;
533          varEndOld = varEnd;
534        }
535      }
536      var cov = data.Select((v, i) => new { Index = i, Value = v }).Select(x => x.Value).Sum(x => (x.Item1 - meanStart) * (x.Item2 - meanEnd)) / data.Count;
537
538      var biasedMean = meanEnd + cov / varStart * (startingFitness - meanStart);
539      var biasedStdev = Math.Sqrt(varEnd - (cov * cov) / varStart);
540
541      if (Problem.Maximization) {
542        var goal = Population.Min(x => x.Fitness);
543        var z = (goal - biasedMean) / biasedStdev;
544        return 1.0 - Phi(z); // P(X >= z)
545      } else {
546        var goal = Population.Max(x => x.Fitness);
547        var z = (goal - biasedMean) / biasedStdev;
548        return Phi(z); // P(X <= z)
549      }
550    }
551
552    private double ProbabilityAccept3dModel(double a, double b, IConfidenceRegressionModel model) {
553      var ds = new Dataset(new[] { "in1", "in2", "out" }, new[] { new List<double> { a }, new List<double> { b }, new List<double> { double.NaN } });
554      var mean = model.GetEstimatedValues(ds, new[] { 0 }).Single();
555      var sdev = Math.Sqrt(model.GetEstimatedVariances(ds, new[] { 0 }).Single());
556
557      var goal = Problem.Maximization ? Population.Min(x => x.Fitness) : Population.Max(x => x.Fitness);
558      var z = (goal - mean) / sdev;
559      return Problem.Maximization ? 1.0 - Phi(z) /* P(X >= z) */ : Phi(z); // P(X <= z)
560    }
561    protected double ProbabilityAccept3d(double startingFitness1, double startingFitness2, IList<Tuple<double, double, double>> data) {
562      if (data.Count < 20) return 1.0;
563      var samples = 0;
564      double meanX1 = 0, meanX1Old = 0, meanX2 = 0, meanX2Old = 0, meanY = 0, meanYOld = 0;
565      double varX1 = 0, varX1Old = 0, varX2 = 0, varX2Old = 0, varY = 0, varYOld = 0;
566      for (var i = 0; i < data.Count; i++) {
567        samples++;
568        var x1 = data[i].Item1;
569        var x2 = data[i].Item2;
570        var y = data[i].Item3;
571
572        if (samples == 1) {
573          meanX1Old = x1;
574          meanX2Old = x2;
575          meanYOld = y;
576        } else {
577          meanX1 = meanX1Old + (x1 - meanX1Old) / samples;
578          meanX2 = meanX2Old + (x2 - meanX2Old) / samples;
579          meanY = meanYOld + (y - meanYOld) / samples;
580          varX1 = varX1Old + (x1 - meanX1Old) * (x1 - meanX1) / (samples - 1);
581          varX2 = varX2Old + (x2 - meanX2Old) * (x2 - meanX2) / (samples - 1);
582          varY = varYOld + (y - meanYOld) * (y - meanY) / (samples - 1);
583
584          meanX1Old = meanX1;
585          meanX2Old = meanX2;
586          meanYOld = meanY;
587          varX1Old = varX1;
588          varX2Old = varX2;
589          varYOld = varY;
590        }
591      }
592      double covX1X2 = 0, covX1Y = 0, covX2Y = 0;
593      for (var i = 0; i < data.Count; i++) {
594        covX1X2 += (data[i].Item1 - meanX1) * (data[i].Item2 - meanX2) / data.Count;
595        covX1Y += (data[i].Item1 - meanX1) * (data[i].Item3 - meanY) / data.Count;
596        covX2Y += (data[i].Item2 - meanX2) * (data[i].Item3 - meanY) / data.Count;
597      }
598
599      var biasedMean = meanY + ((covX1Y * varX2 - covX2Y * covX1X2) * (startingFitness1 - meanX1) - (covX1Y * covX1X2 - covX2Y * varX1) * (startingFitness2 - meanX2)) / (varX1 * varX2 - covX1X2 * covX1X2);
600      var biasedStdev = Math.Sqrt(varY - (covX1Y * covX1Y * varX2 - 2 * covX1Y * covX2Y * covX1X2 + covX2Y * covX2Y * varX1) / (varX1 * varX2 - covX1X2 * covX1X2));
601      if (Problem.Maximization) {
602        var goal = Population.Min(x => x.Fitness);
603        var z = (goal - biasedMean) / biasedStdev;
604        return 1.0 - Phi(z); // P(X >= z)
605      } else {
606        var goal = Population.Max(x => x.Fitness);
607        var z = (goal - biasedMean) / biasedStdev;
608        return Phi(z); // P(X <= z)
609      }
610    }
611
612    [MethodImpl(MethodImplOptions.AggressiveInlining)]
613    public bool IsBetter(ISingleObjectiveSolutionScope<TSolution> a, ISingleObjectiveSolutionScope<TSolution> b) {
614      return IsBetter(a.Fitness, b.Fitness);
615    }
616    [MethodImpl(MethodImplOptions.AggressiveInlining)]
617    public bool IsBetter(double a, double b) {
618      return double.IsNaN(b) && !double.IsNaN(a)
619        || Problem.Maximization && a > b
620        || !Problem.Maximization && a < b;
621    }
622
623    public void AddBreedingResult(ISingleObjectiveSolutionScope<TSolution> a, ISingleObjectiveSolutionScope<TSolution> b, ISingleObjectiveSolutionScope<TSolution> child) {
624      if (IsBetter(a, b))
625        BreedingStat.Add(Tuple.Create(a.Fitness, b.Fitness, child.Fitness));
626      else BreedingStat.Add(Tuple.Create(b.Fitness, a.Fitness, child.Fitness));
627    }
628
629    public void AddRelinkingResult(ISingleObjectiveSolutionScope<TSolution> a, ISingleObjectiveSolutionScope<TSolution> b, ISingleObjectiveSolutionScope<TSolution> child) {
630      if (IsBetter(a, b))
631        RelinkingStat.Add(Tuple.Create(a.Fitness, b.Fitness, child.Fitness));
632      else RelinkingStat.Add(Tuple.Create(b.Fitness, a.Fitness, child.Fitness));
633    }
634
635    public void AddDelinkingResult(ISingleObjectiveSolutionScope<TSolution> a, ISingleObjectiveSolutionScope<TSolution> b, ISingleObjectiveSolutionScope<TSolution> child) {
636      if (IsBetter(a, b))
637        DelinkingStat.Add(Tuple.Create(a.Fitness, b.Fitness, child.Fitness));
638      else DelinkingStat.Add(Tuple.Create(b.Fitness, a.Fitness, child.Fitness));
639    }
640
641    public void AddSamplingResult(ISingleObjectiveSolutionScope<TSolution> sample) {
642      SamplingStat.Add(Tuple.Create(Population.Average(x => x.Fitness), sample.Fitness));
643    }
644
645    public void AddHillclimbingResult(ISingleObjectiveSolutionScope<TSolution> input, ISingleObjectiveSolutionScope<TSolution> outcome) {
646      HillclimbingStat.Add(Tuple.Create(input.Fitness, outcome.Fitness));
647    }
648
649    public void AddTabuwalkingResult(ISingleObjectiveSolutionScope<TSolution> input, ISingleObjectiveSolutionScope<TSolution> outcome) {
650      AdaptivewalkingStat.Add(Tuple.Create(input.Fitness, outcome.Fitness));
651    }
652
653    #region IExecutionContext members
654    public IAtomicOperation CreateOperation(IOperator op) {
655      return new ExecutionContext(this, op, Scope);
656    }
657
658    public IAtomicOperation CreateOperation(IOperator op, IScope s) {
659      return new ExecutionContext(this, op, s);
660    }
661
662    public IAtomicOperation CreateChildOperation(IOperator op) {
663      return new ExecutionContext(this, op, Scope);
664    }
665
666    public IAtomicOperation CreateChildOperation(IOperator op, IScope s) {
667      return new ExecutionContext(this, op, s);
668    }
669    #endregion
670
671    #region Math Helper
672    // normal distribution CDF (left of x) for N(0;1) standard normal distribution
673    // from http://www.johndcook.com/blog/csharp_phi/
674    // license: "This code is in the public domain. Do whatever you want with it, no strings attached."
675    // added: 2016-11-19 21:46 CET
676    protected static double Phi(double x) {
677      // constants
678      double a1 = 0.254829592;
679      double a2 = -0.284496736;
680      double a3 = 1.421413741;
681      double a4 = -1.453152027;
682      double a5 = 1.061405429;
683      double p = 0.3275911;
684
685      // Save the sign of x
686      int sign = 1;
687      if (x < 0)
688        sign = -1;
689      x = Math.Abs(x) / Math.Sqrt(2.0);
690
691      // A&S formula 7.1.26
692      double t = 1.0 / (1.0 + p * x);
693      double y = 1.0 - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * Math.Exp(-x * x);
694
695      return 0.5 * (1.0 + sign * y);
696    }
697    #endregion
698  }
699
700  [Item("SingleSolutionMemPRContext", "Abstract base class for single solution MemPR contexts.")]
701  [StorableClass]
702  public abstract class MemPRSolutionContext<TProblem, TSolution, TContext, TSolutionContext> : ParameterizedNamedItem,
703    ISingleSolutionHeuristicAlgorithmContext<TProblem, TSolution>
704      where TProblem : class, IItem, ISingleObjectiveProblemDefinition
705      where TSolution : class, IItem
706      where TContext : MemPRPopulationContext<TProblem, TSolution, TContext, TSolutionContext>
707      where TSolutionContext : MemPRSolutionContext<TProblem, TSolution, TContext, TSolutionContext> {
708
709    private TContext parent;
710    public IExecutionContext Parent {
711      get { return parent; }
712      set { throw new InvalidOperationException("Cannot set the parent of a single-solution context."); }
713    }
714
715    [Storable]
716    private ISingleObjectiveSolutionScope<TSolution> scope;
717    public IScope Scope {
718      get { return scope; }
719    }
720
721    IKeyedItemCollection<string, IParameter> IExecutionContext.Parameters {
722      get { return Parameters; }
723    }
724
725    public TProblem Problem {
726      get { return parent.Problem; }
727    }
728
729    public double BestQuality {
730      get { return parent.BestQuality; }
731      set { parent.BestQuality = value; }
732    }
733
734    public TSolution BestSolution {
735      get { return parent.BestSolution; }
736      set { parent.BestSolution = value; }
737    }
738
739    public IRandom Random {
740      get { return parent.Random; }
741    }
742
743    [Storable]
744    private IValueParameter<IntValue> evaluatedSolutions;
745    public int EvaluatedSolutions {
746      get { return evaluatedSolutions.Value.Value; }
747      set { evaluatedSolutions.Value.Value = value; }
748    }
749
750    [Storable]
751    private IValueParameter<IntValue> iterations;
752    public int Iterations {
753      get { return iterations.Value.Value; }
754      set { iterations.Value.Value = value; }
755    }
756
757    ISingleObjectiveSolutionScope<TSolution> ISingleSolutionHeuristicAlgorithmContext<TProblem, TSolution>.Solution {
758      get { return scope; }
759    }
760
761    [StorableConstructor]
762    protected MemPRSolutionContext(bool deserializing) : base(deserializing) { }
763    protected MemPRSolutionContext(MemPRSolutionContext<TProblem, TSolution, TContext, TSolutionContext> original, Cloner cloner)
764      : base(original, cloner) {
765      scope = cloner.Clone(original.scope);
766      evaluatedSolutions = cloner.Clone(original.evaluatedSolutions);
767      iterations = cloner.Clone(original.iterations);
768    }
769    public MemPRSolutionContext(TContext baseContext, ISingleObjectiveSolutionScope<TSolution> solution) {
770      parent = baseContext;
771      scope = solution;
772     
773      Parameters.Add(evaluatedSolutions = new ValueParameter<IntValue>("EvaluatedSolutions", new IntValue(0)));
774      Parameters.Add(iterations = new ValueParameter<IntValue>("Iterations", new IntValue(0)));
775    }
776
777    public void IncrementEvaluatedSolutions(int byEvaluations) {
778      if (byEvaluations < 0) throw new ArgumentException("Can only increment and not decrement evaluated solutions.");
779      EvaluatedSolutions += byEvaluations;
780    }
781
782    #region IExecutionContext members
783    public IAtomicOperation CreateOperation(IOperator op) {
784      return new ExecutionContext(this, op, Scope);
785    }
786
787    public IAtomicOperation CreateOperation(IOperator op, IScope s) {
788      return new ExecutionContext(this, op, s);
789    }
790
791    public IAtomicOperation CreateChildOperation(IOperator op) {
792      return new ExecutionContext(this, op, Scope);
793    }
794
795    public IAtomicOperation CreateChildOperation(IOperator op, IScope s) {
796      return new ExecutionContext(this, op, s);
797    }
798    #endregion
799  }
800}
Note: See TracBrowser for help on using the repository browser.