source: branches/RBFRegression/HeuristicLab.Algorithms.DataAnalysis/3.4/GBM/GradientBoostingRegressionAlgorithm.cs @ 14869

Last change on this file since 14869 was 14869, checked in by gkronber, 2 years ago

#2699: merged changesets from trunk to branch

File size: 22.6 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2016 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
4 * and the BEACON Center for the Study of Evolution in Action.
5 *
6 * This file is part of HeuristicLab.
7 *
8 * HeuristicLab is free software: you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation, either version 3 of the License, or
11 * (at your option) any later version.
12 *
13 * HeuristicLab is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
20 */
21#endregion
22
23using System;
24using System.Collections.Generic;
25using System.Linq;
26using System.Threading;
27using HeuristicLab.Analysis;
28using HeuristicLab.Common;
29using HeuristicLab.Core;
30using HeuristicLab.Data;
31using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
32using HeuristicLab.Optimization;
33using HeuristicLab.Parameters;
34using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
35using HeuristicLab.Problems.DataAnalysis;
36using HeuristicLab.Problems.DataAnalysis.Symbolic;
37using HeuristicLab.Problems.DataAnalysis.Symbolic.Regression;
38using HeuristicLab.Random;
39using HeuristicLab.Selection;
40
41namespace HeuristicLab.Algorithms.DataAnalysis.MctsSymbolicRegression {
42  [Item("Gradient Boosting Machine Regression (GBM)",
43    "Gradient boosting for any regression base learner (e.g. MCTS symbolic regression)")]
44  [StorableClass]
45  [Creatable(CreatableAttribute.Categories.DataAnalysisRegression, Priority = 350)]
46  public class GradientBoostingRegressionAlgorithm : FixedDataAnalysisAlgorithm<IRegressionProblem> {
47
48    #region ParameterNames
49
50    private const string IterationsParameterName = "Iterations";
51    private const string NuParameterName = "Nu";
52    private const string MParameterName = "M";
53    private const string RParameterName = "R";
54    private const string RegressionAlgorithmParameterName = "RegressionAlgorithm";
55    private const string SeedParameterName = "Seed";
56    private const string SetSeedRandomlyParameterName = "SetSeedRandomly";
57    private const string CreateSolutionParameterName = "CreateSolution";
58    private const string StoreRunsParameterName = "StoreRuns";
59    private const string RegressionAlgorithmSolutionResultParameterName = "RegressionAlgorithmResult";
60
61    #endregion
62
63    #region ParameterProperties
64
65    public IFixedValueParameter<IntValue> IterationsParameter {
66      get { return (IFixedValueParameter<IntValue>)Parameters[IterationsParameterName]; }
67    }
68
69    public IFixedValueParameter<DoubleValue> NuParameter {
70      get { return (IFixedValueParameter<DoubleValue>)Parameters[NuParameterName]; }
71    }
72
73    public IFixedValueParameter<DoubleValue> RParameter {
74      get { return (IFixedValueParameter<DoubleValue>)Parameters[RParameterName]; }
75    }
76
77    public IFixedValueParameter<DoubleValue> MParameter {
78      get { return (IFixedValueParameter<DoubleValue>)Parameters[MParameterName]; }
79    }
80
81    // regression algorithms are currently: DataAnalysisAlgorithms, BasicAlgorithms and engine algorithms with no common interface
82    public IConstrainedValueParameter<IAlgorithm> RegressionAlgorithmParameter {
83      get { return (IConstrainedValueParameter<IAlgorithm>)Parameters[RegressionAlgorithmParameterName]; }
84    }
85
86    public IFixedValueParameter<StringValue> RegressionAlgorithmSolutionResultParameter {
87      get { return (IFixedValueParameter<StringValue>)Parameters[RegressionAlgorithmSolutionResultParameterName]; }
88    }
89
90    public IFixedValueParameter<IntValue> SeedParameter {
91      get { return (IFixedValueParameter<IntValue>)Parameters[SeedParameterName]; }
92    }
93
94    public FixedValueParameter<BoolValue> SetSeedRandomlyParameter {
95      get { return (FixedValueParameter<BoolValue>)Parameters[SetSeedRandomlyParameterName]; }
96    }
97
98    public IFixedValueParameter<BoolValue> CreateSolutionParameter {
99      get { return (IFixedValueParameter<BoolValue>)Parameters[CreateSolutionParameterName]; }
100    }
101    public IFixedValueParameter<BoolValue> StoreRunsParameter {
102      get { return (IFixedValueParameter<BoolValue>)Parameters[StoreRunsParameterName]; }
103    }
104
105    #endregion
106
107    #region Properties
108
109    public int Iterations {
110      get { return IterationsParameter.Value.Value; }
111      set { IterationsParameter.Value.Value = value; }
112    }
113
114    public int Seed {
115      get { return SeedParameter.Value.Value; }
116      set { SeedParameter.Value.Value = value; }
117    }
118
119    public bool SetSeedRandomly {
120      get { return SetSeedRandomlyParameter.Value.Value; }
121      set { SetSeedRandomlyParameter.Value.Value = value; }
122    }
123
124    public double Nu {
125      get { return NuParameter.Value.Value; }
126      set { NuParameter.Value.Value = value; }
127    }
128
129    public double R {
130      get { return RParameter.Value.Value; }
131      set { RParameter.Value.Value = value; }
132    }
133
134    public double M {
135      get { return MParameter.Value.Value; }
136      set { MParameter.Value.Value = value; }
137    }
138
139    public bool CreateSolution {
140      get { return CreateSolutionParameter.Value.Value; }
141      set { CreateSolutionParameter.Value.Value = value; }
142    }
143
144    public bool StoreRuns {
145      get { return StoreRunsParameter.Value.Value; }
146      set { StoreRunsParameter.Value.Value = value; }
147    }
148
149    public IAlgorithm RegressionAlgorithm {
150      get { return RegressionAlgorithmParameter.Value; }
151    }
152
153    public string RegressionAlgorithmResult {
154      get { return RegressionAlgorithmSolutionResultParameter.Value.Value; }
155      set { RegressionAlgorithmSolutionResultParameter.Value.Value = value; }
156    }
157
158    #endregion
159
160    [StorableConstructor]
161    protected GradientBoostingRegressionAlgorithm(bool deserializing)
162      : base(deserializing) {
163    }
164
165    protected GradientBoostingRegressionAlgorithm(GradientBoostingRegressionAlgorithm original, Cloner cloner)
166      : base(original, cloner) {
167    }
168
169    public override IDeepCloneable Clone(Cloner cloner) {
170      return new GradientBoostingRegressionAlgorithm(this, cloner);
171    }
172
173    public GradientBoostingRegressionAlgorithm() {
174      Problem = new RegressionProblem(); // default problem
175      var osgp = CreateOSGP();
176      var regressionAlgs = new ItemSet<IAlgorithm>(new IAlgorithm[] {
177        new RandomForestRegression(),
178        osgp,
179      });
180      foreach (var alg in regressionAlgs) alg.Prepare();
181
182
183      Parameters.Add(new FixedValueParameter<IntValue>(IterationsParameterName,
184        "Number of iterations", new IntValue(100)));
185      Parameters.Add(new FixedValueParameter<IntValue>(SeedParameterName,
186        "The random seed used to initialize the new pseudo random number generator.", new IntValue(0)));
187      Parameters.Add(new FixedValueParameter<BoolValue>(SetSeedRandomlyParameterName,
188        "True if the random seed should be set to a random value, otherwise false.", new BoolValue(true)));
189      Parameters.Add(new FixedValueParameter<DoubleValue>(NuParameterName,
190        "The learning rate nu when updating predictions in GBM (0 < nu <= 1)", new DoubleValue(0.5)));
191      Parameters.Add(new FixedValueParameter<DoubleValue>(RParameterName,
192        "The fraction of rows that are sampled randomly for the base learner in each iteration (0 < r <= 1)",
193        new DoubleValue(1)));
194      Parameters.Add(new FixedValueParameter<DoubleValue>(MParameterName,
195        "The fraction of variables that are sampled randomly for the base learner in each iteration (0 < m <= 1)",
196        new DoubleValue(0.5)));
197      Parameters.Add(new ConstrainedValueParameter<IAlgorithm>(RegressionAlgorithmParameterName,
198        "The regression algorithm to use as a base learner", regressionAlgs, osgp));
199      Parameters.Add(new FixedValueParameter<StringValue>(RegressionAlgorithmSolutionResultParameterName,
200        "The name of the solution produced by the regression algorithm", new StringValue("Solution")));
201      Parameters[RegressionAlgorithmSolutionResultParameterName].Hidden = true;
202      Parameters.Add(new FixedValueParameter<BoolValue>(CreateSolutionParameterName,
203        "Flag that indicates if a solution should be produced at the end of the run", new BoolValue(true)));
204      Parameters[CreateSolutionParameterName].Hidden = true;
205      Parameters.Add(new FixedValueParameter<BoolValue>(StoreRunsParameterName,
206        "Flag that indicates if the results of the individual runs should be stored for detailed analysis", new BoolValue(false)));
207      Parameters[StoreRunsParameterName].Hidden = true;
208    }
209
210    protected override void Run(CancellationToken cancellationToken) {
211      // Set up the algorithm
212      if (SetSeedRandomly) Seed = new System.Random().Next();
213      var rand = new MersenneTwister((uint)Seed);
214
215      // Set up the results display
216      var iterations = new IntValue(0);
217      Results.Add(new Result("Iterations", iterations));
218
219      var table = new DataTable("Qualities");
220      table.Rows.Add(new DataRow("R² (train)"));
221      table.Rows.Add(new DataRow("R² (test)"));
222      Results.Add(new Result("Qualities", table));
223      var curLoss = new DoubleValue();
224      var curTestLoss = new DoubleValue();
225      Results.Add(new Result("R² (train)", curLoss));
226      Results.Add(new Result("R² (test)", curTestLoss));
227      var runCollection = new RunCollection();
228      if (StoreRuns)
229        Results.Add(new Result("Runs", runCollection));
230
231      // init
232      var problemData = Problem.ProblemData;
233      var targetVarName = problemData.TargetVariable;
234      var activeVariables = problemData.AllowedInputVariables.Concat(new string[] { problemData.TargetVariable });
235      var modifiableDataset = new ModifiableDataset(
236        activeVariables,
237        activeVariables.Select(v => problemData.Dataset.GetDoubleValues(v).ToList()));
238
239      var trainingRows = problemData.TrainingIndices;
240      var testRows = problemData.TestIndices;
241      var yPred = new double[trainingRows.Count()];
242      var yPredTest = new double[testRows.Count()];
243      var y = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, problemData.TrainingIndices).ToArray();
244      var curY = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, problemData.TrainingIndices).ToArray();
245
246      var yTest = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, problemData.TestIndices).ToArray();
247      var curYTest = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, problemData.TestIndices).ToArray();
248      var nu = Nu;
249      var mVars = (int)Math.Ceiling(M * problemData.AllowedInputVariables.Count());
250      var rRows = (int)Math.Ceiling(R * problemData.TrainingIndices.Count());
251      var alg = RegressionAlgorithm;
252      List<IRegressionModel> models = new List<IRegressionModel>();
253      try {
254
255        // Loop until iteration limit reached or canceled.
256        for (int i = 0; i < Iterations; i++) {
257          cancellationToken.ThrowIfCancellationRequested();
258
259          modifiableDataset.RemoveVariable(targetVarName);
260          modifiableDataset.AddVariable(targetVarName, curY.Concat(curYTest));
261
262          SampleTrainingData(rand, modifiableDataset, rRows, problemData.Dataset, curY, problemData.TargetVariable, problemData.TrainingIndices); // all training indices from the original problem data are allowed
263          var modifiableProblemData = new RegressionProblemData(modifiableDataset,
264            problemData.AllowedInputVariables.SampleRandomWithoutRepetition(rand, mVars),
265            problemData.TargetVariable);
266          modifiableProblemData.TrainingPartition.Start = 0;
267          modifiableProblemData.TrainingPartition.End = rRows;
268          modifiableProblemData.TestPartition.Start = problemData.TestPartition.Start;
269          modifiableProblemData.TestPartition.End = problemData.TestPartition.End;
270
271          if (!TrySetProblemData(alg, modifiableProblemData))
272            throw new NotSupportedException("The algorithm cannot be used with GBM.");
273
274          IRegressionModel model;
275          IRun run;
276
277          // try to find a model. The algorithm might fail to produce a model. In this case we just retry until the iterations are exhausted
278          if (TryExecute(alg, rand.Next(), RegressionAlgorithmResult, out model, out run)) {
279            int row = 0;
280            // update predictions for training and test
281            // update new targets (in the case of squared error loss we simply use negative residuals)
282            foreach (var pred in model.GetEstimatedValues(problemData.Dataset, trainingRows)) {
283              yPred[row] = yPred[row] + nu * pred;
284              curY[row] = y[row] - yPred[row];
285              row++;
286            }
287            row = 0;
288            foreach (var pred in model.GetEstimatedValues(problemData.Dataset, testRows)) {
289              yPredTest[row] = yPredTest[row] + nu * pred;
290              curYTest[row] = yTest[row] - yPredTest[row];
291              row++;
292            }
293            // determine quality
294            OnlineCalculatorError error;
295            var trainR = OnlinePearsonsRCalculator.Calculate(yPred, y, out error);
296            var testR = OnlinePearsonsRCalculator.Calculate(yPredTest, yTest, out error);
297
298            // iteration results
299            curLoss.Value = error == OnlineCalculatorError.None ? trainR * trainR : 0.0;
300            curTestLoss.Value = error == OnlineCalculatorError.None ? testR * testR : 0.0;
301
302            models.Add(model);
303
304
305          }
306
307          if (StoreRuns)
308            runCollection.Add(run);
309          table.Rows["R² (train)"].Values.Add(curLoss.Value);
310          table.Rows["R² (test)"].Values.Add(curTestLoss.Value);
311          iterations.Value = i + 1;
312        }
313
314        // produce solution
315        if (CreateSolution) {
316          // when all our models are symbolic models we can easily combine them to a single model
317          if (models.All(m => m is ISymbolicRegressionModel)) {
318            Results.Add(new Result("Solution", CreateSymbolicSolution(models, Nu, (IRegressionProblemData)problemData.Clone())));
319          }
320          // just produce an ensemble solution for now (TODO: correct scaling or linear regression for ensemble model weights)
321
322          var ensembleSolution = CreateEnsembleSolution(models, (IRegressionProblemData)problemData.Clone());
323          Results.Add(new Result("EnsembleSolution", ensembleSolution));
324        }
325      }
326      finally {
327        // reset everything
328        alg.Prepare(true);
329      }
330    }
331
332    private static IRegressionEnsembleSolution CreateEnsembleSolution(List<IRegressionModel> models,
333      IRegressionProblemData problemData) {
334      var rows = problemData.TrainingPartition.Size;
335      var features = models.Count;
336      double[,] inputMatrix = new double[rows, features + 1];
337      //add model estimates
338      for (int m = 0; m < models.Count; m++) {
339        var model = models[m];
340        var estimates = model.GetEstimatedValues(problemData.Dataset, problemData.TrainingIndices);
341        int estimatesCounter = 0;
342        foreach (var estimate in estimates) {
343          inputMatrix[estimatesCounter, m] = estimate;
344          estimatesCounter++;
345        }
346      }
347
348      //add target
349      var targets = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, problemData.TrainingIndices);
350      int targetCounter = 0;
351      foreach (var target in targets) {
352        inputMatrix[targetCounter, models.Count] = target;
353        targetCounter++;
354      }
355
356      alglib.linearmodel lm = new alglib.linearmodel();
357      alglib.lrreport ar = new alglib.lrreport();
358      double[] coefficients;
359      int retVal = 1;
360      alglib.lrbuildz(inputMatrix, rows, features, out retVal, out lm, out ar);
361      if (retVal != 1) throw new ArgumentException("Error in calculation of linear regression solution");
362
363      alglib.lrunpack(lm, out coefficients, out features);
364
365      var ensembleModel = new RegressionEnsembleModel(models, coefficients.Take(models.Count)) { AverageModelEstimates = false };
366      var ensembleSolution = (IRegressionEnsembleSolution)ensembleModel.CreateRegressionSolution(problemData);
367      return ensembleSolution;
368    }
369
370
371    private IAlgorithm CreateOSGP() {
372      // configure strict osgp
373      var alg = new OffspringSelectionGeneticAlgorithm.OffspringSelectionGeneticAlgorithm();
374      var prob = new SymbolicRegressionSingleObjectiveProblem();
375      prob.MaximumSymbolicExpressionTreeDepth.Value = 7;
376      prob.MaximumSymbolicExpressionTreeLength.Value = 15;
377      alg.Problem = prob;
378      alg.SuccessRatio.Value = 1.0;
379      alg.ComparisonFactorLowerBound.Value = 1.0;
380      alg.ComparisonFactorUpperBound.Value = 1.0;
381      alg.MutationProbability.Value = 0.15;
382      alg.PopulationSize.Value = 200;
383      alg.MaximumSelectionPressure.Value = 100;
384      alg.MaximumEvaluatedSolutions.Value = 20000;
385      alg.SelectorParameter.Value = alg.SelectorParameter.ValidValues.OfType<GenderSpecificSelector>().First();
386      alg.MutatorParameter.Value = alg.MutatorParameter.ValidValues.OfType<MultiSymbolicExpressionTreeManipulator>().First();
387      alg.StoreAlgorithmInEachRun = false;
388      return alg;
389    }
390
391    private void SampleTrainingData(MersenneTwister rand, ModifiableDataset ds, int rRows,
392      IDataset sourceDs, double[] curTarget, string targetVarName, IEnumerable<int> trainingIndices) {
393      var selectedRows = trainingIndices.SampleRandomWithoutRepetition(rand, rRows).ToArray();
394      int t = 0;
395      object[] srcRow = new object[ds.Columns];
396      var varNames = ds.DoubleVariables.ToArray();
397      foreach (var r in selectedRows) {
398        // take all values from the original dataset
399        for (int c = 0; c < srcRow.Length; c++) {
400          var col = sourceDs.GetReadOnlyDoubleValues(varNames[c]);
401          srcRow[c] = col[r];
402        }
403        ds.ReplaceRow(t, srcRow);
404        // but use the updated target values
405        ds.SetVariableValue(curTarget[r], targetVarName, t);
406        t++;
407      }
408    }
409
410    private static ISymbolicRegressionSolution CreateSymbolicSolution(List<IRegressionModel> models, double nu, IRegressionProblemData problemData) {
411      var symbModels = models.OfType<ISymbolicRegressionModel>();
412      var lowerLimit = symbModels.Min(m => m.LowerEstimationLimit);
413      var upperLimit = symbModels.Max(m => m.UpperEstimationLimit);
414      var interpreter = new SymbolicDataAnalysisExpressionTreeLinearInterpreter();
415      var progRootNode = new ProgramRootSymbol().CreateTreeNode();
416      var startNode = new StartSymbol().CreateTreeNode();
417
418      var addNode = new Addition().CreateTreeNode();
419      var mulNode = new Multiplication().CreateTreeNode();
420      var scaleNode = (ConstantTreeNode)new Constant().CreateTreeNode(); // all models are scaled using the same nu
421      scaleNode.Value = nu;
422
423      foreach (var m in symbModels) {
424        var relevantPart = m.SymbolicExpressionTree.Root.GetSubtree(0).GetSubtree(0); // skip root and start
425        addNode.AddSubtree((ISymbolicExpressionTreeNode)relevantPart.Clone());
426      }
427
428      mulNode.AddSubtree(addNode);
429      mulNode.AddSubtree(scaleNode);
430      startNode.AddSubtree(mulNode);
431      progRootNode.AddSubtree(startNode);
432      var t = new SymbolicExpressionTree(progRootNode);
433      var combinedModel = new SymbolicRegressionModel(problemData.TargetVariable, t, interpreter, lowerLimit, upperLimit);
434      var sol = new SymbolicRegressionSolution(combinedModel, problemData);
435      return sol;
436    }
437
438    private static bool TrySetProblemData(IAlgorithm alg, IRegressionProblemData problemData) {
439      var prob = alg.Problem as IRegressionProblem;
440      // there is already a problem and it is compatible -> just set problem data
441      if (prob != null) {
442        prob.ProblemDataParameter.Value = problemData;
443        return true;
444      } else return false;
445    }
446
447    private static bool TryExecute(IAlgorithm alg, int seed, string regressionAlgorithmResultName, out IRegressionModel model, out IRun run) {
448      model = null;
449      SetSeed(alg, seed);
450      using (var wh = new AutoResetEvent(false)) {
451        Exception ex = null;
452        EventHandler<EventArgs<Exception>> handler = (sender, args) => {
453          ex = args.Value;
454          wh.Set();
455        };
456        EventHandler handler2 = (sender, args) => wh.Set();
457        alg.ExceptionOccurred += handler;
458        alg.Stopped += handler2;
459        try {
460          alg.Prepare();
461          alg.Start();
462          wh.WaitOne();
463
464          if (ex != null) throw new AggregateException(ex);
465          run = alg.Runs.Last();
466          alg.Runs.Clear();
467          var sols = alg.Results.Select(r => r.Value).OfType<IRegressionSolution>();
468          if (!sols.Any()) return false;
469          var sol = sols.First();
470          if (sols.Skip(1).Any()) {
471            // more than one solution => use regressionAlgorithmResult
472            if (alg.Results.ContainsKey(regressionAlgorithmResultName)) {
473              sol = (IRegressionSolution)alg.Results[regressionAlgorithmResultName].Value;
474            }
475          }
476          var symbRegSol = sol as SymbolicRegressionSolution;
477          // only accept symb reg solutions that do not hit the estimation limits
478          // NaN evaluations would not be critical but are problematic if we want to combine all symbolic models into a single symbolic model
479          if (symbRegSol == null ||
480            (symbRegSol.TrainingLowerEstimationLimitHits == 0 && symbRegSol.TrainingUpperEstimationLimitHits == 0 &&
481             symbRegSol.TestLowerEstimationLimitHits == 0 && symbRegSol.TestUpperEstimationLimitHits == 0) &&
482            symbRegSol.TrainingNaNEvaluations == 0 && symbRegSol.TestNaNEvaluations == 0) {
483            model = sol.Model;
484          }
485        }
486        finally {
487          alg.ExceptionOccurred -= handler;
488          alg.Stopped -= handler2;
489        }
490      }
491      return model != null;
492    }
493
494    private static void SetSeed(IAlgorithm alg, int seed) {
495      // no common interface for algs that use a PRNG -> use naming convention to set seed
496      var paramItem = alg as IParameterizedItem;
497
498      if (paramItem.Parameters.ContainsKey("SetSeedRandomly")) {
499        ((BoolValue)paramItem.Parameters["SetSeedRandomly"].ActualValue).Value = false;
500        ((IntValue)paramItem.Parameters["Seed"].ActualValue).Value = seed;
501      } else {
502        throw new ArgumentException("Base learner does not have a seed parameter (algorithm {0})", alg.Name);
503      }
504
505    }
506  }
507}
Note: See TracBrowser for help on using the repository browser.