source: branches/2635_HeuristicLab.OSGAEvaluator/HeuristicLab.OSGAEvaluator/SymbolicRegressionSingleObjectiveOSGAEvaluator.cs @ 16135

#region License Information
/* HeuristicLab
 * Copyright (C) 2002-2016 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.Common;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
using HeuristicLab.Optimization;
using HeuristicLab.Parameters;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;

namespace HeuristicLab.Problems.DataAnalysis.Symbolic.Regression {
  [Item("SymbolicRegressionSingleObjectiveOSGAEvaluator", "An evaluator which tries to predict when a child will not be able to fullfil offspring selection criteria, to save evaluation time.")]
  [StorableClass]
  public class SymbolicRegressionSingleObjectiveOsgaEvaluator : SymbolicRegressionSingleObjectiveEvaluator {
    private const string RelativeParentChildQualityThresholdParameterName = "RelativeParentChildQualityThreshold";
    private const string RelativeFitnessEvaluationIntervalSizeParameterName = "RelativeFitnessEvaluationIntervalSize";
    private const string ResultCollectionParameterName = "Results";
    private const string AggregateStatisticsParameterName = "AggregateStatistics";
    private const string ActualSelectionPressureParameterName = "SelectionPressure";
    private const string UseAdaptiveQualityThresholdParameterName = "UseAdaptiveQualityThreshold";
    private const string UseFixedEvaluationIntervalsParameterName = "UseFixedEvaluationIntervals";
    private const string PreserveResultCompatibilityParameterName = "PreserveEvaluationResultCompatibility";

    #region parameters
    public IFixedValueParameter<BoolValue> PreserveResultCompatibilityParameter {
      get { return (IFixedValueParameter<BoolValue>)Parameters[PreserveResultCompatibilityParameterName]; }
    }
    public IFixedValueParameter<BoolValue> UseFixedEvaluationIntervalsParameter {
      get { return (IFixedValueParameter<BoolValue>)Parameters[UseFixedEvaluationIntervalsParameterName]; }
    }
    public IFixedValueParameter<BoolValue> UseAdaptiveQualityThresholdParameter {
      get { return (IFixedValueParameter<BoolValue>)Parameters[UseAdaptiveQualityThresholdParameterName]; }
    }
    public ILookupParameter<DoubleValue> ActualSelectionPressureParameter {
      get { return (ILookupParameter<DoubleValue>)Parameters[ActualSelectionPressureParameterName]; }
    }
    public ILookupParameter<ResultCollection> ResultCollectionParameter {
      get { return (ILookupParameter<ResultCollection>)Parameters[ResultCollectionParameterName]; }
    }
    public IValueParameter<BoolValue> AggregateStatisticsParameter {
      get { return (IValueParameter<BoolValue>)Parameters[AggregateStatisticsParameterName]; }
    }
    public IValueLookupParameter<DoubleValue> ComparisonFactorParameter {
      get { return (ValueLookupParameter<DoubleValue>)Parameters["ComparisonFactor"]; }
    }
    public IFixedValueParameter<PercentValue> RelativeParentChildQualityThresholdParameter {
      get { return (IFixedValueParameter<PercentValue>)Parameters[RelativeParentChildQualityThresholdParameterName]; }
    }
    public IFixedValueParameter<PercentValue> RelativeFitnessEvaluationIntervalSizeParameter {
      get { return (IFixedValueParameter<PercentValue>)Parameters[RelativeFitnessEvaluationIntervalSizeParameterName]; }
    }
    public IScopeTreeLookupParameter<DoubleValue> ParentQualitiesParameter { get { return (IScopeTreeLookupParameter<DoubleValue>)Parameters["ParentQualities"]; } }
    #endregion

    #region parameter properties
    public bool AggregateStatistics {
      get { return AggregateStatisticsParameter.Value.Value; }
      set { AggregateStatisticsParameter.Value.Value = value; }
    }
    public bool PreserveResultCompatibility {
      get { return PreserveResultCompatibilityParameter.Value.Value; }
      set { PreserveResultCompatibilityParameter.Value.Value = value; }
    }
    public bool UseFixedEvaluationIntervals {
      get { return UseFixedEvaluationIntervalsParameter.Value.Value; }
      set { UseFixedEvaluationIntervalsParameter.Value.Value = value; }
    }
    public bool UseAdaptiveQualityThreshold {
      get { return UseAdaptiveQualityThresholdParameter.Value.Value; }
      set { UseAdaptiveQualityThresholdParameter.Value.Value = value; }
    }
    public double RelativeParentChildQualityThreshold {
      get { return RelativeParentChildQualityThresholdParameter.Value.Value; }
      set { RelativeParentChildQualityThresholdParameter.Value.Value = value; }
    }
    public double RelativeFitnessEvaluationIntervalSize {
      get { return RelativeFitnessEvaluationIntervalSizeParameter.Value.Value; }
      set { RelativeFitnessEvaluationIntervalSizeParameter.Value.Value = value; }
    }

    #endregion

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

    // keep track of statistics
    public double AdjustedEvaluatedSolutions { get; set; }
    public IntMatrix RejectedStats { get; set; }
    public IntMatrix TotalStats { get; set; }

    public SymbolicRegressionSingleObjectiveOsgaEvaluator() {
      Parameters.Add(new ValueLookupParameter<DoubleValue>("ComparisonFactor", "Determines if the quality should be compared to the better parent (1.0), to the worse (0.0) or to any linearly interpolated value between them."));
      Parameters.Add(new FixedValueParameter<PercentValue>(RelativeParentChildQualityThresholdParameterName, new PercentValue(0.9)));
      Parameters.Add(new FixedValueParameter<PercentValue>(RelativeFitnessEvaluationIntervalSizeParameterName, new PercentValue(0.1)));
      Parameters.Add(new LookupParameter<ResultCollection>(ResultCollectionParameterName));
      Parameters.Add(new ScopeTreeLookupParameter<DoubleValue>("ParentQualities") { ActualName = "Quality" });
      Parameters.Add(new ValueParameter<BoolValue>(AggregateStatisticsParameterName, new BoolValue(false)));
      Parameters.Add(new LookupParameter<DoubleValue>(ActualSelectionPressureParameterName));
      Parameters.Add(new FixedValueParameter<BoolValue>(UseAdaptiveQualityThresholdParameterName, new BoolValue(false)));
      Parameters.Add(new FixedValueParameter<BoolValue>(UseFixedEvaluationIntervalsParameterName, new BoolValue(false)));
      Parameters.Add(new FixedValueParameter<BoolValue>(PreserveResultCompatibilityParameterName, new BoolValue(false)));

      RejectedStats = new IntMatrix();
      TotalStats = new IntMatrix();
    }

    [StorableHook(HookType.AfterDeserialization)]
    private void AfterDeserialization() {
      if (!Parameters.ContainsKey(ActualSelectionPressureParameterName))
        Parameters.Add(new LookupParameter<DoubleValue>(ActualSelectionPressureParameterName));

      if (!Parameters.ContainsKey(UseAdaptiveQualityThresholdParameterName))
        Parameters.Add(new FixedValueParameter<BoolValue>(UseAdaptiveQualityThresholdParameterName, new BoolValue(false)));

      if (!Parameters.ContainsKey(UseFixedEvaluationIntervalsParameterName))
        Parameters.Add(new FixedValueParameter<BoolValue>(UseFixedEvaluationIntervalsParameterName, new BoolValue(false)));

      if (!Parameters.ContainsKey(PreserveResultCompatibilityParameterName))
        Parameters.Add(new FixedValueParameter<BoolValue>(PreserveResultCompatibilityParameterName, new BoolValue(false)));
    }

    [StorableConstructor]
    protected SymbolicRegressionSingleObjectiveOsgaEvaluator(bool deserializing) : base(deserializing) {
      TotalStats = new IntMatrix();
      RejectedStats = new IntMatrix();
    }

    protected SymbolicRegressionSingleObjectiveOsgaEvaluator(SymbolicRegressionSingleObjectiveOsgaEvaluator original,
      Cloner cloner) : base(original, cloner) {
      if (original.TotalStats != null)
        TotalStats = cloner.Clone(original.TotalStats);

      if (original.RejectedStats != null)
        RejectedStats = cloner.Clone(original.RejectedStats);
    }

    public override IDeepCloneable Clone(Cloner cloner) {
      return new SymbolicRegressionSingleObjectiveOsgaEvaluator(this, cloner);
    }

    public override void ClearState() {
      base.ClearState();
      RejectedStats = new IntMatrix();
      TotalStats = new IntMatrix();
      AdjustedEvaluatedSolutions = 0;
    }

    public override IOperation InstrumentedApply() {
      var solution = SymbolicExpressionTreeParameter.ActualValue;
      IEnumerable<int> rows = GenerateRowsToEvaluate();

      var interpreter = SymbolicDataAnalysisTreeInterpreterParameter.ActualValue;
      var estimationLimits = EstimationLimitsParameter.ActualValue;
      var problemData = ProblemDataParameter.ActualValue;
      var applyLinearScaling = ApplyLinearScalingParameter.ActualValue.Value;

      double quality;
      var parentQualities = ParentQualitiesParameter.ActualValue;

      // parent subscopes are not present during evaluation of the initial population
      if (parentQualities.Length > 0) {
        quality = Calculate(interpreter, solution, estimationLimits, problemData, rows);
      } else {
        quality = Calculate(interpreter, solution, estimationLimits.Lower, estimationLimits.Upper, problemData, rows, applyLinearScaling);
      }
      QualityParameter.ActualValue = new DoubleValue(quality);

      return base.InstrumentedApply();
    }

    public static double Calculate(ISymbolicDataAnalysisExpressionTreeInterpreter interpreter, ISymbolicExpressionTree solution, double lowerEstimationLimit, double upperEstimationLimit, IRegressionProblemData problemData, IEnumerable<int> rows, bool applyLinearScaling) {
      IEnumerable<double> estimatedValues = interpreter.GetSymbolicExpressionTreeValues(solution, problemData.Dataset, rows);
      IEnumerable<double> targetValues = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows);
      OnlineCalculatorError errorState;

      double r;
      if (applyLinearScaling) {
        var rCalculator = new OnlinePearsonsRCalculator();
        CalculateWithScaling(targetValues, estimatedValues, lowerEstimationLimit, upperEstimationLimit, rCalculator, problemData.Dataset.Rows);
        errorState = rCalculator.ErrorState;
        r = rCalculator.R;
      } else {
        IEnumerable<double> boundedEstimatedValues = estimatedValues.LimitToRange(lowerEstimationLimit, upperEstimationLimit);
        r = OnlinePearsonsRCalculator.Calculate(targetValues, boundedEstimatedValues, out errorState);
      }
      if (errorState != OnlineCalculatorError.None) return double.NaN;
      return r * r;
    }

    private double Calculate(ISymbolicDataAnalysisExpressionTreeInterpreter interpreter, ISymbolicExpressionTree solution, DoubleLimit estimationLimits, IRegressionProblemData problemData, IEnumerable<int> rows) {
      var lowerEstimationLimit = EstimationLimitsParameter.ActualValue.Lower;
      var upperEstimationLimit = EstimationLimitsParameter.ActualValue.Upper;
      var estimatedValues = interpreter.GetSymbolicExpressionTreeValues(solution, problemData.Dataset, rows).LimitToRange(lowerEstimationLimit, upperEstimationLimit);
      var targetValues = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows).ToList();
      var parentQualities = ParentQualitiesParameter.ActualValue.Select(x => x.Value);
      var minQuality = double.MaxValue;
      var maxQuality = double.MinValue;

      foreach (var quality in parentQualities) {
        if (minQuality > quality) minQuality = quality;
        if (maxQuality < quality) maxQuality = quality;
      }

      var comparisonFactor = ComparisonFactorParameter.ActualValue.Value;
      var parentQuality = minQuality + (maxQuality - minQuality) * comparisonFactor;

      #region fixed intervals
      if (UseFixedEvaluationIntervals) {
        double threshold = parentQuality * RelativeParentChildQualityThreshold;

        if (UseAdaptiveQualityThreshold) {
          var actualSelectionPressure = ActualSelectionPressureParameter.ActualValue;
          if (actualSelectionPressure != null)
            threshold = parentQuality * (1 - actualSelectionPressure.Value / 100.0);
        }

        var estimatedEnumerator = estimatedValues.GetEnumerator();
        var targetEnumerator = targetValues.GetEnumerator();

        var rcalc = new OnlinePearsonsRCalculator();
        var trainingPartitionSize = problemData.TrainingPartition.Size;
        var interval = (int)Math.Floor(trainingPartitionSize * RelativeFitnessEvaluationIntervalSize);

        var calculatedRows = 0;
        #region aggregate statistics
        if (AggregateStatistics) {
          var trainingEnd = problemData.TrainingPartition.End;

          double quality = 0;
          int intervalCount = 0, rejectionInterval = 0;
          var predictedRejected = false;

          while (estimatedEnumerator.MoveNext() & targetEnumerator.MoveNext()) {
            var estimated = estimatedEnumerator.Current;
            var target = targetEnumerator.Current;
            rcalc.Add(target, estimated);
            ++calculatedRows;
            if (calculatedRows % interval == 0 || calculatedRows == trainingPartitionSize) {
              intervalCount++;
              if (predictedRejected) continue;
              var r = rcalc.ErrorState == OnlineCalculatorError.None ? rcalc.R : 0d;
              quality = r * r;
              if (!(quality > threshold)) {
                rejectionInterval = intervalCount - 1;
                predictedRejected = true;
              }
            }
          }
          var actualQuality = rcalc.ErrorState == OnlineCalculatorError.None ? rcalc.R : 0d;
          actualQuality *= actualQuality;

          if (!predictedRejected) quality = actualQuality;

          var actuallyRejected = !(actualQuality > parentQuality);

          if (RejectedStats.Rows == 0 || TotalStats.Rows == 0) {
            RejectedStats = new IntMatrix(2, intervalCount + 1);
            RejectedStats.RowNames = new[] { "Predicted", "Actual" };
            RejectedStats.ColumnNames = Enumerable.Range(1, RejectedStats.Columns).Select(x => string.Format("0-{0}", Math.Min(trainingEnd, x * interval)));
            TotalStats = new IntMatrix(2, 1);
            TotalStats.RowNames = new[] { "Predicted", "Actual" };
            TotalStats.ColumnNames = new[] { "Rejected" };
          }

          if (actuallyRejected) {
            TotalStats[0, 0]++; // prediction true
            TotalStats[1, 0]++;
            RejectedStats[0, rejectionInterval]++;
            RejectedStats[1, rejectionInterval]++;
          } else {
            if (predictedRejected) {
              RejectedStats[0, rejectionInterval]++;
              TotalStats[0, 0]++;
            }
          }
          return quality;
        }
        #endregion
        else {
          while (estimatedEnumerator.MoveNext() & targetEnumerator.MoveNext()) {
            rcalc.Add(targetEnumerator.Current, estimatedEnumerator.Current);
            ++calculatedRows;
            if (calculatedRows % interval == 0 || calculatedRows == trainingPartitionSize) {
              var q = rcalc.ErrorState != OnlineCalculatorError.None ? double.NaN : rcalc.R;
              var quality = q * q;
              if (!(quality > threshold)) {
                AdjustedEvaluatedSolutions += (double)calculatedRows / problemData.TrainingPartition.Size;
                return quality;
              }
            }
          }
          var r = rcalc.ErrorState != OnlineCalculatorError.None ? double.NaN : rcalc.R;
          var actualQuality = r * r;
          AdjustedEvaluatedSolutions += 1d;
          return actualQuality;
        }
        #endregion
      } else {
        var lsc = new OnlineLinearScalingParameterCalculator();
        var rcalc = new OnlinePearsonsRCalculator();
        var interval = (int)Math.Round(RelativeFitnessEvaluationIntervalSize * problemData.TrainingPartition.Size);
        var quality = 0d;
        var calculatedRows = 0;

        var cache = PreserveResultCompatibility ? new List<double>(problemData.TrainingPartition.Size) : null;
        foreach (var target in estimatedValues.Zip(targetValues, (e, t) => new { EstimatedValue = e, ActualValue = t })) {
          if (cache != null)
            cache.Add(target.EstimatedValue);

          lsc.Add(target.EstimatedValue, target.ActualValue);
          rcalc.Add(target.EstimatedValue, target.ActualValue);

          calculatedRows++;

          if (calculatedRows % interval != 0) continue;

          var alpha = lsc.Alpha;
          var beta = lsc.Beta;
          if (lsc.ErrorState != OnlineCalculatorError.None) {
            alpha = 0;
            beta = 1;
          }

          var calc = (OnlinePearsonsRCalculator)rcalc.Clone();
          foreach (var t in targetValues.Skip(calculatedRows)) {
            var s = (t - alpha) / beta; // scaled target
            calc.Add(s, t); // add pair (scaled, target) to the calculator
          }
          var r = calc.ErrorState == OnlineCalculatorError.None ? calc.R : 0d;
          quality = r * r;

          if (!(quality > parentQuality)) {
            AdjustedEvaluatedSolutions += (double)calculatedRows / problemData.TrainingPartition.Size;
            return quality;
          }
        }
        if (PreserveResultCompatibility) {
          // get quality for all the rows. to ensure reproducibility of results between this evaluator
          // and the standard one, we calculate the quality in an identical way (otherwise the returned 
          // quality could be slightly off due to rounding errors (in the range 1e-15 to 1e-16)
          var applyLinearScaling = ApplyLinearScalingParameter.ActualValue.Value;
          double r;
          OnlineCalculatorError calculatorError;

          if (applyLinearScaling) {
            var alpha = lsc.Alpha;
            var beta = lsc.Beta;
            if (lsc.ErrorState != OnlineCalculatorError.None) {
              alpha = 0;
              beta = 1;
            }
            var boundedEstimatedValues = cache.Select(x => x * beta + alpha).LimitToRange(estimationLimits.Lower, estimationLimits.Upper);
            r = OnlinePearsonsRCalculator.Calculate(boundedEstimatedValues, targetValues, out calculatorError);
          } else {
            var boundedEstimatedValues = cache.LimitToRange(estimationLimits.Lower, estimationLimits.Upper);
            r = OnlinePearsonsRCalculator.Calculate(boundedEstimatedValues, targetValues, out calculatorError);
          }
          quality = calculatorError == OnlineCalculatorError.None ? r * r : 0d;
        }
        AdjustedEvaluatedSolutions += 1d;
        return quality;
      }
    }

    public override double Evaluate(IExecutionContext context, ISymbolicExpressionTree tree, IRegressionProblemData problemData, IEnumerable<int> rows) {
      SymbolicDataAnalysisTreeInterpreterParameter.ExecutionContext = context;
      EstimationLimitsParameter.ExecutionContext = context;
      ApplyLinearScalingParameter.ExecutionContext = context;

      var interpreter = SymbolicDataAnalysisTreeInterpreterParameter.ActualValue;
      var estimationLimits = EstimationLimitsParameter.ActualValue;
      var applyLinearScaling = ApplyLinearScalingParameter.ActualValue.Value;

      double r2 = Calculate(interpreter, tree, estimationLimits.Lower, estimationLimits.Upper, problemData, rows, applyLinearScaling);

      SymbolicDataAnalysisTreeInterpreterParameter.ExecutionContext = null;
      EstimationLimitsParameter.ExecutionContext = null;
      ApplyLinearScalingParameter.ExecutionContext = null;

      return r2;
    }
  }
}