source: branches/3.2/sources/HeuristicLab.GP.StructureIdentification/3.3/TournamentPruning.cs @ 9015

#region License Information
/* HeuristicLab
 * Copyright (C) 2002-2008 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
 *
 * This file is part of HeuristicLab.
 *
 * HeuristicLab is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * HeuristicLab is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
 */
#endregion

using System.Collections.Generic;
using System.Linq;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.GP.Interfaces;
using System;
using HeuristicLab.DataAnalysis;
using HeuristicLab.Modeling;

namespace HeuristicLab.GP.StructureIdentification {
  public class TournamentPruning : OperatorBase {
    public TournamentPruning()
      : base() {
      AddVariableInfo(new VariableInfo("Random", "", typeof(IRandom), VariableKind.In));
      AddVariableInfo(new VariableInfo("FunctionTree", "The tree to analyse", typeof(IGeneticProgrammingModel), VariableKind.In));
      AddVariableInfo(new VariableInfo("Dataset", "Dataset", typeof(Dataset), VariableKind.In));
      AddVariableInfo(new VariableInfo("TargetVariable", "", typeof(StringData), VariableKind.In));
      AddVariableInfo(new VariableInfo("TrainingSamplesStart", "Samples start", typeof(IntData), VariableKind.In));
      AddVariableInfo(new VariableInfo("TrainingSamplesEnd", "Samples end", typeof(IntData), VariableKind.In));
      AddVariableInfo(new VariableInfo("TreeEvaluator", "", typeof(ITreeEvaluator), VariableKind.In));
      AddVariableInfo(new VariableInfo("MaxPruningRatio", "Maximale relative size of the pruned branch", typeof(DoubleData), VariableKind.In));
      AddVariableInfo(new VariableInfo("TournamentSize", "Number of branches to compare for pruning", typeof(IntData), VariableKind.
In));
      AddVariableInfo(new VariableInfo("PopulationPercentileStart", "", typeof(DoubleData), VariableKind.In));
      AddVariableInfo(new VariableInfo("PopulationPercentileEnd", "", typeof(DoubleData), VariableKind.In));
      AddVariableInfo(new VariableInfo("QualityGainWeight", "", typeof(DoubleData), VariableKind.In));
    }

    public override IOperation Apply(IScope scope) {
      IRandom random = scope.GetVariableValue<IRandom>("Random", true);
      double percentileStart = scope.GetVariableValue<DoubleData>("PopulationPercentileStart", true).Data;
      double percentileEnd = scope.GetVariableValue<DoubleData>("PopulationPercentileEnd", true).Data;
      int tournamentSize = scope.GetVariableValue<IntData>("TournamentSize", true).Data;
      Dataset dataset = scope.GetVariableValue<Dataset>("Dataset", true);
      string targetVariable = scope.GetVariableValue<StringData>("TargetVariable", true).Data;
      int samplesStart = scope.GetVariableValue<IntData>("TrainingSamplesStart", true).Data;
      int samplesEnd = scope.GetVariableValue<IntData>("TrainingSamplesEnd", true).Data;
      ITreeEvaluator evaluator = scope.GetVariableValue<ITreeEvaluator>("TreeEvaluator", true);
      double maxPruningRatio = scope.GetVariableValue<DoubleData>("MaxPruningRatio", true).Data;
      double qualityGainWeight = scope.GetVariableValue<DoubleData>("QualityGainWeight", true).Data;
      int n = scope.SubScopes.Count;
      // for each tree in the given percentile
      var trees = (from subScope in scope.SubScopes
                   select subScope.GetVariableValue<IGeneticProgrammingModel>("FunctionTree", false))
                  .Skip((int)(n * percentileStart))
                  .Take((int)(n * (percentileEnd - percentileStart)));
      foreach (var tree in trees) {
        tree.FunctionTree = Prune(random, tree.FunctionTree, tournamentSize, dataset, targetVariable, samplesStart, samplesEnd, evaluator, maxPruningRatio, qualityGainWeight);
      }
      return null;
    }

    public static IFunctionTree Prune(IRandom random, IFunctionTree tree, int tournamentSize,
      Dataset dataset, string targetVariable, int samplesStart, int samplesEnd, ITreeEvaluator evaluator,
      double maxPruningRatio, double qualityGainWeight) {
      var evaluatedRows = Enumerable.Range(samplesStart, samplesEnd - samplesStart);
      var estimatedValues = evaluator.Evaluate(dataset, tree, evaluatedRows).ToArray();
      var targetValues = dataset.GetVariableValues(targetVariable, samplesStart, samplesEnd);
      int originalSize = tree.GetSize();
      double originalMse = SimpleMSEEvaluator.Calculate(Matrix<double>.Create(targetValues, estimatedValues));

      int maxPrunedBranchSize = (int)(tree.GetSize() * maxPruningRatio);


      IFunctionTree bestTree = tree;
      double bestGain = double.PositiveInfinity;

      for (int i = 0; i < tournamentSize; i++) {
        var clonedTree = (IFunctionTree)tree.Clone();
        var prunePoints = (from node in FunctionTreeIterator.IteratePrefix(clonedTree)
                           from subTree in node.SubTrees
                           where subTree.GetSize() <= maxPrunedBranchSize
                           select new { Parent = node, Branch = subTree, SubTreeIndex = node.SubTrees.IndexOf(subTree) })
               .ToList();

        var selectedPrunePoint = prunePoints[random.Next(prunePoints.Count)];
        var branchValues = evaluator.Evaluate(dataset, selectedPrunePoint.Branch, evaluatedRows);
        var branchMean = branchValues.Average();

        selectedPrunePoint.Parent.RemoveSubTree(selectedPrunePoint.SubTreeIndex);
        var constNode = CreateConstant(branchMean);
        selectedPrunePoint.Parent.InsertSubTree(selectedPrunePoint.SubTreeIndex, constNode);

        estimatedValues = evaluator.Evaluate(dataset, clonedTree, evaluatedRows).ToArray();
        double prunedMse = SimpleMSEEvaluator.Calculate(Matrix<double>.Create(targetValues, estimatedValues));
        double prunedSize = clonedTree.GetSize();
        // MSE of the pruned tree is larger than the original tree in most cases
        // size of the pruned tree is always smaller than the size of the original tree
        // same change in quality => prefer pruning operation that removes a larger tree
        double gain = ((prunedMse / originalMse) * qualityGainWeight) /
                       (originalSize / prunedSize);
        if (gain < bestGain) {
          bestGain = gain;
          bestTree = clonedTree;
        }
      }

      return bestTree;
    }

    private static FunctionTree CreateConstant(double constantValue) {
      var node = (ConstantFunctionTree)(new Constant()).GetTreeNode();
      node.Value = constantValue;
      return node;
    }
  }
}