source: branches/3040_VectorBasedGP/HeuristicLab.Problems.DataAnalysis.Symbolic/3.4/Mutators/NestedOptimizerSubVectorImprovementManipulator.cs @ 18229

using System;
using System.Collections.Generic;
using System.Linq;
using HEAL.Attic;
using HeuristicLab.Algorithms.EvolutionStrategy;
using HeuristicLab.Algorithms.GeneticAlgorithm;
using HeuristicLab.Algorithms.OffspringSelectionGeneticAlgorithm;
using HeuristicLab.Algorithms.RandomSearch;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.Encodings.IntegerVectorEncoding;
using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
using HeuristicLab.Optimization;
using HeuristicLab.Parameters;
using HeuristicLab.Random;
using HeuristicLab.Selection;

namespace HeuristicLab.Problems.DataAnalysis.Symbolic {

  [Item("NestedOptimizerSubVectorImprovementManipulator", "Mutator that optimizes the ranges for a subvector symbol by utilizing a nested optimizer.")]
  [StorableType("32E58EEE-97B4-4396-98A8-B98AB897E3F0")]
  public class NestedOptimizerSubVectorImprovementManipulator<T> : SymbolicDataAnalysisExpressionManipulator<T> where T : class, IDataAnalysisProblemData {
    private const string BestSolutionParameterName = "BestSolution";

    [Item("SubVectorOptimizationProblem", "")]
    [StorableType("EA3D3221-B274-4F2F-8B58-23CB2D091FD7")]
    private class SubVectorOptimizationProblem : SingleObjectiveBasicProblem<IntegerVectorEncoding> {
      #region Fixed Problem Parameters
      [Storable]
      private ISymbolicDataAnalysisSingleObjectiveEvaluator<T> evaluator;
      [Storable]
      private T problemData;
      [Storable]
      private List<int> rows;
      [Storable]
      private IExecutionContext executionContext;
      #endregion

      #region Instance Parameters
      [Storable]
      private ISymbolicExpressionTree tree;
      [Storable]
      private IList<int> selectedSubVectorNodes;
      #endregion

      public override bool Maximization { get { return false; } }

      public SubVectorOptimizationProblem() {
        Encoding = new IntegerVectorEncoding("bounds");
        Parameters.Add(new ResultParameter<IntegerVector>(BestSolutionParameterName, ""));
      }
      private SubVectorOptimizationProblem(SubVectorOptimizationProblem original, Cloner cloner)
        : base(original, cloner) { }
      public override IDeepCloneable Clone(Cloner cloner) {
        return new SubVectorOptimizationProblem(this, cloner);
      }
      [StorableConstructor]
      private SubVectorOptimizationProblem(StorableConstructorFlag _) : base(_) { }

      public override double Evaluate(Individual individual, IRandom random) {
        var solution = individual.IntegerVector(Encoding.Name);

        var updatedTree = (ISymbolicExpressionTree)tree.Clone();
        UpdateFromVector(updatedTree, selectedSubVectorNodes, solution, Encoding.Bounds[0, 1]);
        
        var quality = evaluator.Evaluate(executionContext, updatedTree, problemData, rows);
        if (evaluator.Maximization)
          quality = -quality;
        return quality;
      }

      public override void Analyze(Individual[] individuals, double[] qualities, ResultCollection results, IRandom random) {
        var best = GetBestIndividual(individuals, qualities);
        var vector = best.Item1.IntegerVector(Encoding.Name);

        results.AddOrUpdateResult(BestSolutionParameterName, vector);
      }

      public void SetProblemData(ISymbolicDataAnalysisSingleObjectiveEvaluator<T> evaluator, T problemData, List<int> rows, IExecutionContext executionContext) {
        this.evaluator = evaluator;
        this.problemData = problemData;
        this.rows = rows;
        this.executionContext = executionContext;
      }
      public void SetInstanceData(ISymbolicExpressionTree tree, List<int> selectedSubVectorNodes, int vectorLength) {
        this.tree = tree;
        this.selectedSubVectorNodes = selectedSubVectorNodes;
        Encoding.Length = selectedSubVectorNodes.Count * 2;
        Encoding.Bounds = new IntMatrix(new int[,] { { 0, vectorLength } });
      }
    }


    #region Parameter Properties
    public IConstrainedValueParameter<IAlgorithm> NestedOptimizerParameter {
      get { return (IConstrainedValueParameter<IAlgorithm>)Parameters["NestedOptimizer"]; }
    }

    public IFixedValueParameter<PercentValue> PercentOptimizedSubVectorNodesParameter {
      get { return (IFixedValueParameter<PercentValue>)Parameters["PercentOptimizedSubVectorNodes"]; }
    }
    #endregion

    #region Properties
    public IOptimizer NestedOptimizer {
      get { return NestedOptimizerParameter.Value; }
    }

    public PercentValue PercentOptimizedSubVectorNodes {
      get { return PercentOptimizedSubVectorNodesParameter.Value; }
    }
    #endregion

    public NestedOptimizerSubVectorImprovementManipulator() : base() {
      var problem = new SubVectorOptimizationProblem();

      #region Create nested Algorithms
      var rs = new RandomSearchAlgorithm() {
        Problem = problem,
        BatchSize = 10,
        MaximumEvaluatedSolutions = 100
      };

      var es = new EvolutionStrategy() {
        Problem = problem,
        PlusSelection = new BoolValue(true),
        PopulationSize = new IntValue(1),
        Children = new IntValue(10),
        MaximumGenerations = new IntValue(100)
      };
      es.Mutator = es.MutatorParameter.ValidValues.OfType<UniformSomePositionsManipulator>().Single();

      var ga = new GeneticAlgorithm() {
        Problem = problem,
        PopulationSize = new IntValue(10),
        MutationProbability = new PercentValue(0.1),
        MaximumGenerations = new IntValue(100)
      };
      ga.Selector = ga.SelectorParameter.ValidValues.OfType<TournamentSelector>().Single();
      ga.Crossover = ga.CrossoverParameter.ValidValues.OfType<RoundedBlendAlphaBetaCrossover>().Single();
      ga.Mutator = ga.MutatorParameter.ValidValues.OfType<UniformOnePositionManipulator>().Single();

      var osga = new OffspringSelectionGeneticAlgorithm() {
        Problem = problem,
        PopulationSize = new IntValue(10),
        ComparisonFactorLowerBound = new DoubleValue(1.0),
        ComparisonFactorUpperBound = new DoubleValue(1.0),
        MutationProbability = new PercentValue(0.1),
        MaximumGenerations = new IntValue(100),
        MaximumEvaluatedSolutions = new IntValue(1000)
      };
      osga.Selector = osga.SelectorParameter.ValidValues.OfType<TournamentSelector>().Single();
      osga.Crossover = osga.CrossoverParameter.ValidValues.OfType<RoundedBlendAlphaBetaCrossover>().Single();
      osga.Mutator = osga.MutatorParameter.ValidValues.OfType<UniformOnePositionManipulator>().Single();
      #endregion

      var optimizers = new ItemSet<IAlgorithm>() { rs, es, ga, osga };

      Parameters.Add(new ConstrainedValueParameter<IAlgorithm>("NestedOptimizer", optimizers, rs));
      Parameters.Add(new FixedValueParameter<PercentValue>("PercentOptimizedSubVectorNodes", new PercentValue(1.0)));
    }

    private NestedOptimizerSubVectorImprovementManipulator(NestedOptimizerSubVectorImprovementManipulator<T> original, Cloner cloner) : base(original, cloner) { }

    public override IDeepCloneable Clone(Cloner cloner) {
      return new NestedOptimizerSubVectorImprovementManipulator<T>(this, cloner);
    }

    [StorableConstructor]
    private NestedOptimizerSubVectorImprovementManipulator(StorableConstructorFlag _) : base(_) { }

    public override void Manipulate(IRandom random, ISymbolicExpressionTree symbolicExpressionTree) {
      int vectorLengths = GetVectorLengths(ProblemDataParameter.ActualValue);
      
      var selectedSubVectorNodes = GetSelectedSubVectorNodes(symbolicExpressionTree, random);
      if (selectedSubVectorNodes.Count == 0) 
        return;

      var algorithm = (IAlgorithm)NestedOptimizer.Clone();
      PrepareAlgorithm(algorithm, symbolicExpressionTree, selectedSubVectorNodes, vectorLengths);

      algorithm.Start(CancellationToken);

      if (algorithm.ExecutionState != ExecutionState.Stopped)
        throw new InvalidOperationException("Nested Algorithm did not finish.");

      var solution = (IntegerVector)algorithm.Results[BestSolutionParameterName].Value;
      UpdateFromVector(symbolicExpressionTree, selectedSubVectorNodes, solution, vectorLengths);
    }

    private void PrepareAlgorithm(IAlgorithm algorithm, ISymbolicExpressionTree symbolicExpressionTree, List<int> selectedSubVectorNodes, int vectorLengths) {
      var problem = (SubVectorOptimizationProblem)algorithm.Problem;
      problem.SetProblemData(EvaluatorParameter.ActualValue, ProblemDataParameter.ActualValue, GenerateRowsToEvaluate().ToList(), ExecutionContext);
      problem.SetInstanceData(symbolicExpressionTree, selectedSubVectorNodes, vectorLengths);
    }

    private List<int> GetSelectedSubVectorNodes(ISymbolicExpressionTree symbolicExpressionTree, IRandom random) {
      var subVectorNodes = GetSubVectorNodes(symbolicExpressionTree).ToList();

      int numSelect = (int)Math.Round(subVectorNodes.Count * PercentOptimizedSubVectorNodes.Value);
      var selectedSubVectorNodes = Enumerable.Range(0, subVectorNodes.Count).SampleRandomWithoutRepetition(random, numSelect).ToList();
      return selectedSubVectorNodes;
    }

    private static int GetVectorLengths(T problemData) {  // ToDo evaluate a tree to get vector length per node
      var vectorLengths = problemData.Dataset.DoubleVectorVariables
        .Select(v => problemData.Dataset.GetDoubleVectorValue(v, row: 0).Count)
        .Distinct();
      return vectorLengths.Single();
    }

    private static void UpdateFromVector(ISymbolicExpressionTree tree, IList<int> selectedNodes, IntegerVector solution, int vectorLength) {
      var nodes = GetSubVectorNodes(tree).ToList();

      int i = 0;
      foreach (var nodeIdx in selectedNodes) {
        var node = nodes[nodeIdx];
        node.Offset = (double)solution[i++] / (vectorLength - 1);
        node.Length = (double)solution[i++] / (vectorLength - 1);
      }
    }

    private static IEnumerable<WindowedSymbolTreeNode> GetSubVectorNodes(ISymbolicExpressionTree tree) {
      return ActualRoot(tree)
        .IterateNodesBreadth()
        .OfType<WindowedSymbolTreeNode>()
        .Where(n => n.HasLocalParameters);
    }
    private static ISymbolicExpressionTreeNode ActualRoot(ISymbolicExpressionTree tree) {
      return tree.Root.GetSubtree(0).GetSubtree(0);
    }

  }
}