Free cookie consent management tool by TermsFeed Policy Generator

source: branches/3087_Ceres_Integration/HeuristicLab.Problems.DataAnalysis.Symbolic.Regression/3.4/SingleObjective/Evaluators/SymbolicRegressionConstantOptimizationEvaluator.cs

Last change on this file was 18011, checked in by bburlacu, 3 years ago

#3087: Implement ceres-based parameter optimizer in new evaluator. Revert constant optimization evaluator to old behavior.
File size: 20.2 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 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;
25
26using HEAL.Attic;
27
28using HeuristicLab.Common;
29using HeuristicLab.Core;
30using HeuristicLab.Data;
31using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
32using HeuristicLab.Optimization;
33using HeuristicLab.Parameters;
34
35namespace HeuristicLab.Problems.DataAnalysis.Symbolic.Regression {
36  [Item("Constant Optimization Evaluator", "Calculates Pearson R² of a symbolic regression solution and optimizes the constant used.")]
37  [StorableType("24B68851-036D-4446-BD6F-3823E9028FF4")]
38  public class SymbolicRegressionConstantOptimizationEvaluator : SymbolicRegressionSingleObjectiveEvaluator {
39    private const string ConstantOptimizationIterationsParameterName = "ConstantOptimizationIterations";
40    private const string ConstantOptimizationImprovementParameterName = "ConstantOptimizationImprovement";
41    private const string ConstantOptimizationProbabilityParameterName = "ConstantOptimizationProbability";
42    private const string ConstantOptimizationRowsPercentageParameterName = "ConstantOptimizationRowsPercentage";
43    private const string UpdateConstantsInTreeParameterName = "UpdateConstantsInSymbolicExpressionTree";
44    private const string UpdateVariableWeightsParameterName = "Update Variable Weights";
45
46    private const string FunctionEvaluationsResultParameterName = "Constants Optimization Function Evaluations";
47    private const string GradientEvaluationsResultParameterName = "Constants Optimization Gradient Evaluations";
48    private const string CountEvaluationsParameterName = "Count Function and Gradient Evaluations";
49
50    public IFixedValueParameter<IntValue> ConstantOptimizationIterationsParameter {
51      get { return (IFixedValueParameter<IntValue>)Parameters[ConstantOptimizationIterationsParameterName]; }
52    }
53    public IFixedValueParameter<DoubleValue> ConstantOptimizationImprovementParameter {
54      get { return (IFixedValueParameter<DoubleValue>)Parameters[ConstantOptimizationImprovementParameterName]; }
55    }
56    public IFixedValueParameter<PercentValue> ConstantOptimizationProbabilityParameter {
57      get { return (IFixedValueParameter<PercentValue>)Parameters[ConstantOptimizationProbabilityParameterName]; }
58    }
59    public IFixedValueParameter<PercentValue> ConstantOptimizationRowsPercentageParameter {
60      get { return (IFixedValueParameter<PercentValue>)Parameters[ConstantOptimizationRowsPercentageParameterName]; }
61    }
62    public IFixedValueParameter<BoolValue> UpdateConstantsInTreeParameter {
63      get { return (IFixedValueParameter<BoolValue>)Parameters[UpdateConstantsInTreeParameterName]; }
64    }
65    public IFixedValueParameter<BoolValue> UpdateVariableWeightsParameter {
66      get { return (IFixedValueParameter<BoolValue>)Parameters[UpdateVariableWeightsParameterName]; }
67    }
68
69    public IResultParameter<IntValue> FunctionEvaluationsResultParameter {
70      get { return (IResultParameter<IntValue>)Parameters[FunctionEvaluationsResultParameterName]; }
71    }
72    public IResultParameter<IntValue> GradientEvaluationsResultParameter {
73      get { return (IResultParameter<IntValue>)Parameters[GradientEvaluationsResultParameterName]; }
74    }
75    public IFixedValueParameter<BoolValue> CountEvaluationsParameter {
76      get { return (IFixedValueParameter<BoolValue>)Parameters[CountEvaluationsParameterName]; }
77    }
78
79
80    public IntValue ConstantOptimizationIterations {
81      get { return ConstantOptimizationIterationsParameter.Value; }
82    }
83    public DoubleValue ConstantOptimizationImprovement {
84      get { return ConstantOptimizationImprovementParameter.Value; }
85    }
86    public PercentValue ConstantOptimizationProbability {
87      get { return ConstantOptimizationProbabilityParameter.Value; }
88    }
89    public PercentValue ConstantOptimizationRowsPercentage {
90      get { return ConstantOptimizationRowsPercentageParameter.Value; }
91    }
92    public bool UpdateConstantsInTree {
93      get { return UpdateConstantsInTreeParameter.Value.Value; }
94      set { UpdateConstantsInTreeParameter.Value.Value = value; }
95    }
96
97    public bool UpdateVariableWeights {
98      get { return UpdateVariableWeightsParameter.Value.Value; }
99      set { UpdateVariableWeightsParameter.Value.Value = value; }
100    }
101
102    public bool CountEvaluations {
103      get { return CountEvaluationsParameter.Value.Value; }
104      set { CountEvaluationsParameter.Value.Value = value; }
105    }
106
107    public override bool Maximization {
108      get { return true; }
109    }
110
111    [StorableConstructor]
112    protected SymbolicRegressionConstantOptimizationEvaluator(StorableConstructorFlag _) : base(_) { }
113    protected SymbolicRegressionConstantOptimizationEvaluator(SymbolicRegressionConstantOptimizationEvaluator original, Cloner cloner)
114      : base(original, cloner) {
115    }
116    public SymbolicRegressionConstantOptimizationEvaluator()
117      : base() {
118      Parameters.Add(new FixedValueParameter<IntValue>(ConstantOptimizationIterationsParameterName, "Determines how many iterations should be calculated while optimizing the constant of a symbolic expression tree (0 indicates other or default stopping criterion).", new IntValue(10)));
119      Parameters.Add(new FixedValueParameter<DoubleValue>(ConstantOptimizationImprovementParameterName, "Determines the relative improvement which must be achieved in the constant optimization to continue with it (0 indicates other or default stopping criterion).", new DoubleValue(0)) { Hidden = true });
120      Parameters.Add(new FixedValueParameter<PercentValue>(ConstantOptimizationProbabilityParameterName, "Determines the probability that the constants are optimized", new PercentValue(1)));
121      Parameters.Add(new FixedValueParameter<PercentValue>(ConstantOptimizationRowsPercentageParameterName, "Determines the percentage of the rows which should be used for constant optimization", new PercentValue(1)));
122      Parameters.Add(new FixedValueParameter<BoolValue>(UpdateConstantsInTreeParameterName, "Determines if the constants in the tree should be overwritten by the optimized constants.", new BoolValue(true)) { Hidden = true });
123      Parameters.Add(new FixedValueParameter<BoolValue>(UpdateVariableWeightsParameterName, "Determines if the variable weights in the tree should be  optimized.", new BoolValue(true)) { Hidden = true });
124
125      Parameters.Add(new FixedValueParameter<BoolValue>(CountEvaluationsParameterName, "Determines if function and gradient evaluation should be counted.", new BoolValue(false)));
126      Parameters.Add(new ResultParameter<IntValue>(FunctionEvaluationsResultParameterName, "The number of function evaluations performed by the constants optimization evaluator", "Results", new IntValue()));
127      Parameters.Add(new ResultParameter<IntValue>(GradientEvaluationsResultParameterName, "The number of gradient evaluations performed by the constants optimization evaluator", "Results", new IntValue()));
128    }
129
130    public override IDeepCloneable Clone(Cloner cloner) {
131      return new SymbolicRegressionConstantOptimizationEvaluator(this, cloner);
132    }
133
134    [StorableHook(HookType.AfterDeserialization)]
135    private void AfterDeserialization() {
136      if (!Parameters.ContainsKey(UpdateConstantsInTreeParameterName))
137        Parameters.Add(new FixedValueParameter<BoolValue>(UpdateConstantsInTreeParameterName, "Determines if the constants in the tree should be overwritten by the optimized constants.", new BoolValue(true)));
138      if (!Parameters.ContainsKey(UpdateVariableWeightsParameterName))
139        Parameters.Add(new FixedValueParameter<BoolValue>(UpdateVariableWeightsParameterName, "Determines if the variable weights in the tree should be  optimized.", new BoolValue(true)));
140
141      if (!Parameters.ContainsKey(CountEvaluationsParameterName))
142        Parameters.Add(new FixedValueParameter<BoolValue>(CountEvaluationsParameterName, "Determines if function and gradient evaluation should be counted.", new BoolValue(false)));
143
144      if (!Parameters.ContainsKey(FunctionEvaluationsResultParameterName))
145        Parameters.Add(new ResultParameter<IntValue>(FunctionEvaluationsResultParameterName, "The number of function evaluations performed by the constants optimization evaluator", "Results", new IntValue()));
146      if (!Parameters.ContainsKey(GradientEvaluationsResultParameterName))
147        Parameters.Add(new ResultParameter<IntValue>(GradientEvaluationsResultParameterName, "The number of gradient evaluations performed by the constants optimization evaluator", "Results", new IntValue()));
148    }
149
150    private static readonly object locker = new object();
151    public override IOperation InstrumentedApply() {
152      var solution = SymbolicExpressionTreeParameter.ActualValue;
153      double quality;
154      if (RandomParameter.ActualValue.NextDouble() < ConstantOptimizationProbability.Value) {
155        IEnumerable<int> constantOptimizationRows = GenerateRowsToEvaluate(ConstantOptimizationRowsPercentage.Value);
156        var counter = new EvaluationsCounter();
157        quality = OptimizeConstants(SymbolicDataAnalysisTreeInterpreterParameter.ActualValue, solution, ProblemDataParameter.ActualValue,
158           constantOptimizationRows, ApplyLinearScalingParameter.ActualValue.Value, ConstantOptimizationIterations.Value, updateVariableWeights: UpdateVariableWeights, lowerEstimationLimit: EstimationLimitsParameter.ActualValue.Lower, upperEstimationLimit: EstimationLimitsParameter.ActualValue.Upper, updateConstantsInTree: UpdateConstantsInTree, counter: counter);
159
160        if (ConstantOptimizationRowsPercentage.Value != RelativeNumberOfEvaluatedSamplesParameter.ActualValue.Value) {
161          var evaluationRows = GenerateRowsToEvaluate();
162          quality = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(SymbolicDataAnalysisTreeInterpreterParameter.ActualValue, solution, EstimationLimitsParameter.ActualValue.Lower, EstimationLimitsParameter.ActualValue.Upper, ProblemDataParameter.ActualValue, evaluationRows, ApplyLinearScalingParameter.ActualValue.Value);
163        }
164
165        if (CountEvaluations) {
166          lock (locker) {
167            FunctionEvaluationsResultParameter.ActualValue.Value += counter.FunctionEvaluations;
168            GradientEvaluationsResultParameter.ActualValue.Value += counter.GradientEvaluations;
169          }
170        }
171
172      } else {
173        var evaluationRows = GenerateRowsToEvaluate();
174        quality = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(SymbolicDataAnalysisTreeInterpreterParameter.ActualValue, solution, EstimationLimitsParameter.ActualValue.Lower, EstimationLimitsParameter.ActualValue.Upper, ProblemDataParameter.ActualValue, evaluationRows, ApplyLinearScalingParameter.ActualValue.Value);
175      }
176      QualityParameter.ActualValue = new DoubleValue(quality);
177
178      return base.InstrumentedApply();
179    }
180
181    public override double Evaluate(IExecutionContext context, ISymbolicExpressionTree tree, IRegressionProblemData problemData, IEnumerable<int> rows) {
182      SymbolicDataAnalysisTreeInterpreterParameter.ExecutionContext = context;
183      EstimationLimitsParameter.ExecutionContext = context;
184      ApplyLinearScalingParameter.ExecutionContext = context;
185      FunctionEvaluationsResultParameter.ExecutionContext = context;
186      GradientEvaluationsResultParameter.ExecutionContext = context;
187
188      // Pearson R² evaluator is used on purpose instead of the const-opt evaluator,
189      // because Evaluate() is used to get the quality of evolved models on
190      // different partitions of the dataset (e.g., best validation model)
191      double r2 = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(SymbolicDataAnalysisTreeInterpreterParameter.ActualValue, tree, EstimationLimitsParameter.ActualValue.Lower, EstimationLimitsParameter.ActualValue.Upper, problemData, rows, ApplyLinearScalingParameter.ActualValue.Value);
192
193      SymbolicDataAnalysisTreeInterpreterParameter.ExecutionContext = null;
194      EstimationLimitsParameter.ExecutionContext = null;
195      ApplyLinearScalingParameter.ExecutionContext = null;
196      FunctionEvaluationsResultParameter.ExecutionContext = null;
197      GradientEvaluationsResultParameter.ExecutionContext = null;
198
199      return r2;
200    }
201
202    public class EvaluationsCounter {
203      public int FunctionEvaluations = 0;
204      public int GradientEvaluations = 0;
205    }
206
207    public static double OptimizeConstants(ISymbolicDataAnalysisExpressionTreeInterpreter interpreter,
208      ISymbolicExpressionTree tree, IRegressionProblemData problemData, IEnumerable<int> rows, bool applyLinearScaling,
209      int maxIterations, bool updateVariableWeights = true,
210      double lowerEstimationLimit = double.MinValue, double upperEstimationLimit = double.MaxValue,
211      bool updateConstantsInTree = true, Action<double[], double, object> iterationCallback = null, EvaluationsCounter counter = null) {
212
213      // Numeric constants in the tree become variables for parameter optimization.
214      // Variables in the tree become parameters (fixed values) for parameter optimization.
215      // For each parameter (variable in the original tree) we store the
216      // variable name, variable value (for factor vars) and lag as a DataForVariable object.
217      // A dictionary is used to find parameters
218      double[] initialConstants;
219      var parameters = new List<TreeToAutoDiffTermConverter.DataForVariable>();
220
221      TreeToAutoDiffTermConverter.ParametricFunction func;
222      TreeToAutoDiffTermConverter.ParametricFunctionGradient func_grad;
223      if (!TreeToAutoDiffTermConverter.TryConvertToAutoDiff(tree, updateVariableWeights, applyLinearScaling, out parameters, out initialConstants, out func, out func_grad))
224        throw new NotSupportedException("Could not optimize constants of symbolic expression tree due to not supported symbols used in the tree.");
225      if (parameters.Count == 0) return 0.0; // constant expressions always have a R² of 0.0
226      var parameterEntries = parameters.ToArray(); // order of entries must be the same for x
227
228      // extract inital constants
229      double[] c;
230      if (applyLinearScaling) {
231        c = new double[initialConstants.Length + 2];
232        c[0] = 0.0;
233        c[1] = 1.0;
234        Array.Copy(initialConstants, 0, c, 2, initialConstants.Length);
235      } else {
236        c = (double[])initialConstants.Clone();
237      }
238
239      double originalQuality = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(interpreter, tree, lowerEstimationLimit, upperEstimationLimit, problemData, rows, applyLinearScaling);
240
241      if (counter == null) counter = new EvaluationsCounter();
242      var rowEvaluationsCounter = new EvaluationsCounter();
243
244      alglib.lsfitstate state;
245      alglib.lsfitreport rep;
246      int retVal;
247
248      IDataset ds = problemData.Dataset;
249      double[,] x = new double[rows.Count(), parameters.Count];
250      int row = 0;
251      foreach (var r in rows) {
252        int col = 0;
253        foreach (var info in parameterEntries) {
254          if (ds.VariableHasType<double>(info.variableName)) {
255            x[row, col] = ds.GetDoubleValue(info.variableName, r + info.lag);
256          } else if (ds.VariableHasType<string>(info.variableName)) {
257            x[row, col] = ds.GetStringValue(info.variableName, r) == info.variableValue ? 1 : 0;
258          } else throw new InvalidProgramException("found a variable of unknown type");
259          col++;
260        }
261        row++;
262      }
263      double[] y = ds.GetDoubleValues(problemData.TargetVariable, rows).ToArray();
264      int n = x.GetLength(0);
265      int m = x.GetLength(1);
266      int k = c.Length;
267
268      alglib.ndimensional_pfunc function_cx_1_func = CreatePFunc(func);
269      alglib.ndimensional_pgrad function_cx_1_grad = CreatePGrad(func_grad);
270      alglib.ndimensional_rep xrep = (p, f, obj) => iterationCallback(p, f, obj);
271
272      try {
273        alglib.lsfitcreatefg(x, y, c, n, m, k, false, out state);
274        alglib.lsfitsetcond(state, 0.0, maxIterations);
275        alglib.lsfitsetxrep(state, iterationCallback != null);
276        alglib.lsfitfit(state, function_cx_1_func, function_cx_1_grad, xrep, rowEvaluationsCounter);
277        alglib.lsfitresults(state, out retVal, out c, out rep);
278      } catch (ArithmeticException) {
279        return originalQuality;
280      } catch (alglib.alglibexception) {
281        return originalQuality;
282      }
283
284      counter.FunctionEvaluations += rowEvaluationsCounter.FunctionEvaluations / n;
285      counter.GradientEvaluations += rowEvaluationsCounter.GradientEvaluations / n;
286
287      //retVal == -7  => constant optimization failed due to wrong gradient
288      //          -8  => optimizer detected  NAN / INF  in  the target
289      //                 function and/ or gradient
290      if (retVal != -7 && retVal != -8) {
291        if (applyLinearScaling) {
292          var tmp = new double[c.Length - 2];
293          Array.Copy(c, 2, tmp, 0, tmp.Length);
294          UpdateConstants(tree, tmp, updateVariableWeights);
295        } else UpdateConstants(tree, c, updateVariableWeights);
296      }
297      var quality = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(interpreter, tree, lowerEstimationLimit, upperEstimationLimit, problemData, rows, applyLinearScaling);
298
299      if (!updateConstantsInTree) UpdateConstants(tree, initialConstants, updateVariableWeights);
300
301      if (originalQuality - quality > 0.001 || double.IsNaN(quality)) {
302        UpdateConstants(tree, initialConstants, updateVariableWeights);
303        return originalQuality;
304      }
305      return quality;
306    }
307
308    private static void UpdateConstants(ISymbolicExpressionTree tree, double[] constants, bool updateVariableWeights) {
309      int i = 0;
310      foreach (var node in tree.Root.IterateNodesPrefix().OfType<SymbolicExpressionTreeTerminalNode>()) {
311        ConstantTreeNode constantTreeNode = node as ConstantTreeNode;
312        VariableTreeNodeBase variableTreeNodeBase = node as VariableTreeNodeBase;
313        FactorVariableTreeNode factorVarTreeNode = node as FactorVariableTreeNode;
314        if (constantTreeNode != null) {
315          if (constantTreeNode.Parent.Symbol is Power
316              && constantTreeNode.Parent.GetSubtree(1) == constantTreeNode) continue; // exponents in powers are not optimizated (see TreeToAutoDiffTermConverter)
317          constantTreeNode.Value = constants[i++];
318        } else if (updateVariableWeights && variableTreeNodeBase != null)
319          variableTreeNodeBase.Weight = constants[i++];
320        else if (factorVarTreeNode != null) {
321          for (int j = 0; j < factorVarTreeNode.Weights.Length; j++)
322            factorVarTreeNode.Weights[j] = constants[i++];
323        }
324      }
325    }
326
327    private static alglib.ndimensional_pfunc CreatePFunc(TreeToAutoDiffTermConverter.ParametricFunction func) {
328      return (double[] c, double[] x, ref double fx, object o) => {
329        fx = func(c, x);
330        var counter = (EvaluationsCounter)o;
331        counter.FunctionEvaluations++;
332      };
333    }
334
335    private static alglib.ndimensional_pgrad CreatePGrad(TreeToAutoDiffTermConverter.ParametricFunctionGradient func_grad) {
336      return (double[] c, double[] x, ref double fx, double[] grad, object o) => {
337        var tuple = func_grad(c, x);
338        fx = tuple.Item2;
339        Array.Copy(tuple.Item1, grad, grad.Length);
340        var counter = (EvaluationsCounter)o;
341        counter.GradientEvaluations++;
342      };
343    }
344    public static bool CanOptimizeConstants(ISymbolicExpressionTree tree) {
345      return TreeToAutoDiffTermConverter.IsCompatible(tree);
346    }
347  }
348}
Note: See TracBrowser for help on using the repository browser.