source: branches/gp-crossover/HeuristicLab.Problems.DataAnalysis.Symbolic/3.4/Crossovers/SymbolicDataAnalysisExpressionProbabilisticFunctionalCrossover.cs @ 7303

#region License Information
/* HeuristicLab
 * Copyright (C) 2002-2011 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;
using System.Collections.Generic;
using System.Linq;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;

namespace HeuristicLab.Problems.DataAnalysis.Symbolic {

  [Item("ProbabilisticFunctionalCrossover", "An operator which performs subtree swapping based on the behavioral similarity between subtrees.")]
  public sealed class SymbolicDataAnalysisExpressionProbabilisticFunctionalCrossover<T> : SymbolicDataAnalysisExpressionCrossover<T> where T : class, IDataAnalysisProblemData {
    [StorableConstructor]
    private SymbolicDataAnalysisExpressionProbabilisticFunctionalCrossover(bool deserializing) : base(deserializing) { }
    private SymbolicDataAnalysisExpressionProbabilisticFunctionalCrossover(SymbolicDataAnalysisExpressionCrossover<T> original, Cloner cloner)
      : base(original, cloner) { }
    public SymbolicDataAnalysisExpressionProbabilisticFunctionalCrossover()
      : base() {
      Name = "ProbabilisticFunctionalCrossover";
    }
    public override IDeepCloneable Clone(Cloner cloner) { return new SymbolicDataAnalysisExpressionProbabilisticFunctionalCrossover<T>(this, cloner); }

    protected override ISymbolicExpressionTree Cross(IRandom random, ISymbolicExpressionTree parent0, ISymbolicExpressionTree parent1) {
      ISymbolicDataAnalysisExpressionTreeInterpreter interpreter = SymbolicDataAnalysisTreeInterpreterParameter.ActualValue;
      List<int> rows = GenerateRowsToEvaluate().ToList();
      T problemData = ProblemDataParameter.ActualValue;
      return Cross(random, parent0, parent1, interpreter, problemData,
                   rows, MaximumSymbolicExpressionTreeDepth.Value, MaximumSymbolicExpressionTreeLength.Value);
    }

    public override ISymbolicExpressionTree Crossover(IRandom random, ISymbolicExpressionTree parent0, ISymbolicExpressionTree parent1) {
      return Cross(random, parent0, parent1);
    }

    /// <summary>
    /// Takes two parent individuals P0 and P1. 
    /// Randomly choose a node i from the first parent, then for each matching node j from the second parent, calculate the behavioral distance based on the range:
    /// d(i,j) = 0.5 * ( abs(max(i) - max(j)) + abs(min(i) - min(j)) ).
    /// Next, assign probabilities for the selection of the second cross point based on the inversed and normalized behavioral distance and
    /// choose the second crosspoint via a random weighted selection procedure.
    /// </summary>
    public static ISymbolicExpressionTree Cross(IRandom random, ISymbolicExpressionTree parent0, ISymbolicExpressionTree parent1,
                                                ISymbolicDataAnalysisExpressionTreeInterpreter interpreter, T problemData, IList<int> rows, int maxDepth, int maxLength) {
      var crossoverPoints0 = new List<CutPoint>();
      parent0.Root.ForEachNodePostfix((n) => {
        if (n.Subtrees.Any() && n != parent0.Root)
          foreach (var child in n.Subtrees)
            crossoverPoints0.Add(new CutPoint(n, child));
      });
      var crossoverPoint0 = crossoverPoints0.SelectRandom(random);
      int level = parent0.Root.GetBranchLevel(crossoverPoint0.Child);
      int length = parent0.Root.GetLength() - crossoverPoint0.Child.GetLength();

      var allowedBranches = new List<ISymbolicExpressionTreeNode>();
      parent1.Root.ForEachNodePostfix((n) => {
        if (n.Subtrees.Any() && n != parent1.Root)
          allowedBranches.AddRange(n.Subtrees.Where(s => crossoverPoint0.IsMatchingPointType(s) && s.GetDepth() + level <= maxDepth && s.GetLength() + length <= maxLength));
      });

      if (allowedBranches.Count == 0)
        return parent0;

      var dataset = problemData.Dataset;

      // create symbols in order to improvize an ad-hoc tree so that the child can be evaluated
      var rootSymbol = new ProgramRootSymbol();
      var startSymbol = new StartSymbol();
      var tree0 = CreateTreeFromNode(random, crossoverPoint0.Child, rootSymbol, startSymbol); // this will change crossoverPoint0.Child.Parent
      List<double> estimatedValues0 = interpreter.GetSymbolicExpressionTreeValues(tree0, dataset, rows).ToList();
      double min0 = estimatedValues0.Min();
      double max0 = estimatedValues0.Max();
      crossoverPoint0.Child.Parent = crossoverPoint0.Parent; // restore correct parent 

      var weights = new List<double>();
      foreach (var node in allowedBranches) {
        var parent = node.Parent;
        var tree1 = CreateTreeFromNode(random, node, rootSymbol, startSymbol);
        List<double> estimatedValues1 = interpreter.GetSymbolicExpressionTreeValues(tree1, dataset, rows).ToList();
        double min1 = estimatedValues1.Min();
        double max1 = estimatedValues1.Max();
        double behavioralDistance = (Math.Abs(min0 - min1) + Math.Abs(max0 - max1)) / 2; // this can be NaN of Infinity because some trees are crazy like exp(exp(exp(...))), we correct that below
        weights.Add(behavioralDistance);
        node.Parent = parent; // restore correct node parent
      }

      // remove branches with an infinite or NaN behavioral distance
      int count = weights.Count, idx = 0;
      while (idx < count) {
        if (Double.IsNaN(weights[idx]) || Double.IsInfinity(weights[idx])) {
          weights.RemoveAt(idx);
          allowedBranches.RemoveAt(idx);
          --count;
        } else {
          ++idx;
        }
      }

      // check if there are any allowed branches left
      if (allowedBranches.Count == 0)
        return parent0;

      ISymbolicExpressionTreeNode selectedBranch;
      double sum = weights.Sum();
      if (sum == 0.0)
        selectedBranch = allowedBranches[0]; // just return the first, since we don't care (all weights are zero)
      else {
        // transform similarity distances into probabilities by normalizing and inverting the values
        for (int i = 0; i != weights.Count; ++i)
          weights[i] = (1 - weights[i] / sum);

        //selectedBranch = allowedBranches.SelectRandom(weights, random);
        selectedBranch = SelectRandomBranch(random, allowedBranches, weights);
      }
      swap(crossoverPoint0, selectedBranch);
      return parent0;
    }

    private static void swap(CutPoint crossoverPoint, ISymbolicExpressionTreeNode selectedBranch) {
      if (crossoverPoint.Child != null) {
        // manipulate the tree of parent0 in place
        // replace the branch in tree0 with the selected branch from tree1
        crossoverPoint.Parent.RemoveSubtree(crossoverPoint.ChildIndex);
        if (selectedBranch != null) {
          crossoverPoint.Parent.InsertSubtree(crossoverPoint.ChildIndex, selectedBranch);
        }
      } else {
        // child is null (additional child should be added under the parent)
        if (selectedBranch != null) {
          crossoverPoint.Parent.AddSubtree(selectedBranch);
        }
      }
    }

    private static ISymbolicExpressionTreeNode SelectRandomBranch(IRandom random, IList<ISymbolicExpressionTreeNode> nodes, IList<double> weights) {
      double r = weights.Sum() * random.NextDouble();
      for (int i = 0; i != nodes.Count; ++i) {
        if (r < weights[i])
          return nodes[i];
        r -= weights[i];
      }
      return nodes.Last();
    }
  }
}