source: branches/Sliding Window GP/HeuristicLab.Problems.DataAnalysis.Symbolic.Regression/3.4/SymbolicDataAnalysisRegressionSolutionPruningOptimizer.cs @ 10396

#region License Information
/* HeuristicLab
 * Copyright (C) 2002-2013 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.Linq;
using HeuristicLab.Core;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
using HeuristicLab.Problems.DataAnalysis.Symbolic.Regression;

namespace HeuristicLab.Problems.DataAnalysis.Symbolic {
  [Item("SymbolicDataAnalysisSolutionPruningOptimizer", "An operator which automatically removes nodes that have a negative impact from the tree model, optimizing the remaining constants.")]
  [StorableClass]
  public class SymbolicDataAnalysisRegressionSolutionPruningOptimizer : SymbolicDataAnalysisSolutionPruningOptimizer {
    public override ISymbolicDataAnalysisSolution PruneAndOptimizeSolution(ISymbolicDataAnalysisSolution solution) {
      var regressionSolution = (ISymbolicRegressionSolution)solution;
      return PruneAndOptimizeRegressionSolution(regressionSolution);
    }
    /// <summary>
    /// This method will walk all the levels of the symbolic regression solution model root starting from the deepest level and:
    /// - it calculates the impact value of every originalNode on that level
    /// - it prunes (replaces with a constant) the originalNode with the lowest negative impact value (0 or positive impacts are left unchanged)
    /// - when no more nodes can be pruned, it moves on the upper level in the tree
    /// - if the pruned and optimized solution is worse than the original solution (it can happen sometimes), then the original solution is returned
    /// </summary>
    /// <param name="solution"></param>
    /// <returns></returns>
    private ISymbolicRegressionSolution PruneAndOptimizeRegressionSolution(ISymbolicRegressionSolution solution) {
      var calculator = new SymbolicRegressionSolutionImpactValuesCalculator();
      var model = (ISymbolicRegressionModel)solution.Model;
      var problemData = solution.ProblemData;
      var rows = problemData.TrainingIndices.ToList();
      // get tree levels and iterate each level from the bottom up
      var root = model.SymbolicExpressionTree.Root.GetSubtree(0).GetSubtree(0);
      var levels = root.IterateNodesBreadth().GroupBy(root.GetBranchLevel).OrderByDescending(g => g.Key);

      OptimizeConstants(solution);

      foreach (var level in levels) {
        var nodes = level.ToArray();
        double minImpact;
        do {
          minImpact = 0.0;
          int minImpactIndex = -1;
          for (int i = 0; i < nodes.Length; ++i) {
            // parent will be null if the node has already been replaced with a constant node in the tree (therefore it can be skipped)
            if (nodes[i] is ConstantTreeNode || nodes[i].Parent == null) continue;
            var impact = calculator.CalculateImpactValue(model, nodes[i], problemData, rows);
            if (!(impact < minImpact)) continue;
            minImpact = impact;
            minImpactIndex = i;
          }
          if (minImpact >= 0) continue;
          var node = nodes[minImpactIndex];
          var replacementValue = calculator.CalculateReplacementValue(model, node, problemData, rows);
          var constantNode = MakeConstantTreeNode(replacementValue);
          ReplaceWithConstantNode(node, constantNode);
          nodes[minImpactIndex] = constantNode;
          OptimizeConstants(solution);
        } while (minImpact < 0);
      }

      var newSolution = (ISymbolicRegressionSolution)model.CreateRegressionSolution(problemData);
      return newSolution.TrainingRSquared > solution.TrainingRSquared ? newSolution : solution;
    }

    private static void OptimizeConstants(ISymbolicRegressionSolution solution) {
      const int MAX_ITERATIONS = 50;
      var model = solution.Model;
      var interpreter = model.Interpreter;
      var tree = model.SymbolicExpressionTree;
      var problemData = solution.ProblemData;
      var rows = problemData.TrainingIndices;

      SymbolicRegressionConstantOptimizationEvaluator.OptimizeConstants(interpreter, tree, problemData, rows, true, MAX_ITERATIONS);
    }
  }
}