Free cookie consent management tool by TermsFeed Policy Generator

Changeset 18178 for branches


Ignore:
Timestamp:
01/04/22 10:43:14 (3 years ago)
Author:
gkronber
Message:

#3136 copied code with minor modifications from ParameterOptimizationEvaluator into the NMSEConstraintsEvaluator because the code in ParameterOptimizationEvaluator uses R² internally and is incompatible to the NMSEEvaluator.

Location:
branches/3136_Structural_GP/HeuristicLab.Problems.DataAnalysis.Symbolic.Regression/3.4/SingleObjective
Files:
2 edited

Legend:

Unmodified
Added
Removed
  • branches/3136_Structural_GP/HeuristicLab.Problems.DataAnalysis.Symbolic.Regression/3.4/SingleObjective/Evaluators/NMSESingleObjectiveConstraintsEvaluator.cs

    r18146 r18178  
    247247
    248248      if (OptimizeParameters)
    249         SymbolicRegressionParameterOptimizationEvaluator.OptimizeParameters(
     249        Optimize(
    250250          interpreter, tree,
    251251          problemData, rows,
    252           applyLinearScaling: false, // Tree already contains scaling terms
    253252          ParameterOptimizationIterations,
    254253          updateVariableWeights: true,
     
    268267        PenalityFactor);
    269268    }
     269
     270    public static double Optimize(ISymbolicDataAnalysisExpressionTreeInterpreter interpreter,
     271          ISymbolicExpressionTree tree, IRegressionProblemData problemData, IEnumerable<int> rows,
     272          int maxIterations, bool updateVariableWeights = true,
     273          double lowerEstimationLimit = double.MinValue, double upperEstimationLimit = double.MaxValue,
     274          bool updateParametersInTree = true, Action<double[], double, object> iterationCallback = null, EvaluationsCounter counter = null) {
     275
     276      // Numeric parameters in the tree become variables for parameter optimization.
     277      // Variables in the tree become parameters (fixed values) for parameter optimization.
     278      // For each parameter (variable in the original tree) we store the
     279      // variable name, variable value (for factor vars) and lag as a DataForVariable object.
     280      // A dictionary is used to find parameters
     281      double[] initialParameters;
     282      var parameters = new List<TreeToAutoDiffTermConverter.DataForVariable>();
     283
     284      TreeToAutoDiffTermConverter.ParametricFunction func;
     285      TreeToAutoDiffTermConverter.ParametricFunctionGradient func_grad;
     286      if (!TreeToAutoDiffTermConverter.TryConvertToAutoDiff(tree, updateVariableWeights, addLinearScalingTerms: false, out parameters, out initialParameters, out func, out func_grad))
     287        throw new NotSupportedException("Could not optimize parameters of symbolic expression tree due to not supported symbols used in the tree.");
     288      var parameterEntries = parameters.ToArray(); // order of entries must be the same for x
     289
     290      // extract inital parameters
     291      double[] c = (double[])initialParameters.Clone();
     292
     293      double originalQuality = SymbolicRegressionSingleObjectiveMeanSquaredErrorEvaluator.Calculate(
     294        tree, problemData, rows,
     295        interpreter, applyLinearScaling: false,
     296        lowerEstimationLimit,
     297        upperEstimationLimit);
     298
     299      if (counter == null) counter = new EvaluationsCounter();
     300      var rowEvaluationsCounter = new EvaluationsCounter();
     301
     302      alglib.lsfitstate state;
     303      alglib.lsfitreport rep;
     304      int retVal;
     305
     306      IDataset ds = problemData.Dataset;
     307      double[,] x = new double[rows.Count(), parameters.Count];
     308      int row = 0;
     309      foreach (var r in rows) {
     310        int col = 0;
     311        foreach (var info in parameterEntries) {
     312          if (ds.VariableHasType<double>(info.variableName)) {
     313            x[row, col] = ds.GetDoubleValue(info.variableName, r + info.lag);
     314          } else if (ds.VariableHasType<string>(info.variableName)) {
     315            x[row, col] = ds.GetStringValue(info.variableName, r) == info.variableValue ? 1 : 0;
     316          } else throw new InvalidProgramException("found a variable of unknown type");
     317          col++;
     318        }
     319        row++;
     320      }
     321      double[] y = ds.GetDoubleValues(problemData.TargetVariable, rows).ToArray();
     322      int n = x.GetLength(0);
     323      int m = x.GetLength(1);
     324      int k = c.Length;
     325
     326      alglib.ndimensional_pfunc function_cx_1_func = CreatePFunc(func);
     327      alglib.ndimensional_pgrad function_cx_1_grad = CreatePGrad(func_grad);
     328      alglib.ndimensional_rep xrep = (p, f, obj) => iterationCallback(p, f, obj);
     329
     330      try {
     331        alglib.lsfitcreatefg(x, y, c, n, m, k, false, out state);
     332        alglib.lsfitsetcond(state, 0.0, maxIterations);
     333        alglib.lsfitsetxrep(state, iterationCallback != null);
     334        alglib.lsfitfit(state, function_cx_1_func, function_cx_1_grad, xrep, rowEvaluationsCounter);
     335        alglib.lsfitresults(state, out retVal, out c, out rep);
     336      } catch (ArithmeticException) {
     337        return originalQuality;
     338      } catch (alglib.alglibexception) {
     339        return originalQuality;
     340      }
     341
     342      counter.FunctionEvaluations += rowEvaluationsCounter.FunctionEvaluations / n;
     343      counter.GradientEvaluations += rowEvaluationsCounter.GradientEvaluations / n;
     344
     345      //retVal == -7  => parameter optimization failed due to wrong gradient
     346      //          -8  => optimizer detected  NAN / INF  in  the target
     347      //                 function and/ or gradient
     348      if (retVal != -7 && retVal != -8) {
     349        UpdateParameters(tree, c, updateVariableWeights);
     350      }
     351      var quality = SymbolicRegressionSingleObjectiveMeanSquaredErrorEvaluator.Calculate(
     352        tree, problemData, rows,
     353        interpreter, applyLinearScaling: false,
     354        lowerEstimationLimit, upperEstimationLimit);
     355
     356      if (!updateParametersInTree) UpdateParameters(tree, initialParameters, updateVariableWeights);
     357
     358      if (originalQuality < quality || double.IsNaN(quality)) {
     359        UpdateParameters(tree, initialParameters, updateVariableWeights);
     360        return originalQuality;
     361      }
     362      return quality;
     363    }
     364
     365    private static void UpdateParameters(ISymbolicExpressionTree tree, double[] parameters, bool updateVariableWeights) {
     366      int i = 0;
     367      foreach (var node in tree.Root.IterateNodesPrefix().OfType<SymbolicExpressionTreeTerminalNode>()) {
     368        NumberTreeNode numberTreeNode = node as NumberTreeNode;
     369        VariableTreeNodeBase variableTreeNodeBase = node as VariableTreeNodeBase;
     370        FactorVariableTreeNode factorVarTreeNode = node as FactorVariableTreeNode;
     371        if (numberTreeNode != null) {
     372          if (numberTreeNode.Parent.Symbol is Power
     373              && numberTreeNode.Parent.GetSubtree(1) == numberTreeNode) continue; // exponents in powers are not optimizated (see TreeToAutoDiffTermConverter)
     374          numberTreeNode.Value = parameters[i++];
     375        } else if (updateVariableWeights && variableTreeNodeBase != null)
     376          variableTreeNodeBase.Weight = parameters[i++];
     377        else if (factorVarTreeNode != null) {
     378          for (int j = 0; j < factorVarTreeNode.Weights.Length; j++)
     379            factorVarTreeNode.Weights[j] = parameters[i++];
     380        }
     381      }
     382    }
     383
     384    private static alglib.ndimensional_pfunc CreatePFunc(TreeToAutoDiffTermConverter.ParametricFunction func) {
     385      return (double[] c, double[] x, ref double fx, object o) => {
     386        fx = func(c, x);
     387        var counter = (EvaluationsCounter)o;
     388        counter.FunctionEvaluations++;
     389      };
     390    }
     391
     392    private static alglib.ndimensional_pgrad CreatePGrad(TreeToAutoDiffTermConverter.ParametricFunctionGradient func_grad) {
     393      return (double[] c, double[] x, ref double fx, double[] grad, object o) => {
     394        var tuple = func_grad(c, x);
     395        fx = tuple.Item2;
     396        Array.Copy(tuple.Item1, grad, grad.Length);
     397        var counter = (EvaluationsCounter)o;
     398        counter.GradientEvaluations++;
     399      };
     400    }
     401
     402    public class EvaluationsCounter {
     403      public int FunctionEvaluations = 0;
     404      public int GradientEvaluations = 0;
     405    }
    270406  }
    271407}
  • branches/3136_Structural_GP/HeuristicLab.Problems.DataAnalysis.Symbolic.Regression/3.4/SingleObjective/StructuredSymbolicRegressionSingleObjectiveProblem.cs

    r18177 r18178  
    5757    }
    5858
     59    public SymbolicRegressionSingleObjectiveEvaluator TreeEvaluator => TreeEvaluatorParameter.Value;
     60
    5961    public StructureTemplate StructureTemplate => StructureTemplateParameter.Value;
    6062
     
    199201      var tree = BuildTree(individual);
    200202
    201       if (StructureTemplate.ApplyLinearScaling) {
     203      // NMSEConstraintsEvaluator sets linear scaling terms itself
     204      if (StructureTemplate.ApplyLinearScaling && !(TreeEvaluator is NMSESingleObjectiveConstraintsEvaluator)) {
    202205        AdjustLinearScalingParams(ProblemData, tree, Interpreter);
    203206      }
     
    205208      individual[SymbolicExpressionTreeName] = tree;
    206209
    207       return TreeEvaluatorParameter.Value.Evaluate(
     210      return TreeEvaluator.Evaluate(
    208211        tree, ProblemData,
    209212        ProblemData.TrainingIndices,
Note: See TracChangeset for help on using the changeset viewer.