source: branches/EfficientGlobalOptimization/HeuristicLab.Algorithms.EGO/Operators/ModelBuilder.cs @ 15338

#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.Linq;
using System.Threading;
using HeuristicLab.Algorithms.DataAnalysis;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.Operators;
using HeuristicLab.Optimization;
using HeuristicLab.Parameters;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
using HeuristicLab.Problems.DataAnalysis;
using HeuristicLab.Problems.SurrogateProblem;

namespace HeuristicLab.Algorithms.EGO {
  /// <summary>
  /// A base class for operators that manipulate real-valued vectors.
  /// </summary>
  [Item("ModelBuilder", "Builds a model from a dataset and a given RegressionAlgorithm")]
  [StorableClass]
  public class ModelBuilder : InstrumentedOperator, IStochasticOperator, ICancellableOperator {
    public override bool CanChangeName => true;
    public CancellationToken Cancellation { get; set; }

    #region Parameter properties
    public ILookupParameter<IDataAnalysisAlgorithm<IRegressionProblem>> RegressionAlgorithmParameter => (ILookupParameter<IDataAnalysisAlgorithm<IRegressionProblem>>)Parameters["RegressionAlgorithm"];
    public ILookupParameter<IRegressionSolution> ModelParameter => (ILookupParameter<IRegressionSolution>)Parameters["Model"];
    public ILookupParameter<ModifiableDataset> DatasetParameter => (ILookupParameter<ModifiableDataset>)Parameters["Dataset"];
    public ILookupParameter<IRandom> RandomParameter => (ILookupParameter<IRandom>)Parameters["Random"];
    public ILookupParameter<IntValue> MaxModelSizeParameter => (ILookupParameter<IntValue>)Parameters["Maximal Model Size"];
    public ILookupParameter<DoubleMatrix> InfillBoundsParameter => (ILookupParameter<DoubleMatrix>)Parameters["InfillBounds"];
    #endregion

    [StorableConstructor]
    protected ModelBuilder(bool deserializing) : base(deserializing) { }
    protected ModelBuilder(ModelBuilder original, Cloner cloner) : base(original, cloner) { }
    public ModelBuilder() {
      Parameters.Add(new LookupParameter<IDataAnalysisAlgorithm<IRegressionProblem>>("RegressionAlgorithm", "The algorithm used to build a model") { Hidden = true });
      Parameters.Add(new LookupParameter<IRegressionSolution>("Model", "The resulting model") { Hidden = true });
      Parameters.Add(new LookupParameter<ModifiableDataset>("Dataset", "The Dataset from which the model is created") { Hidden = true });
      Parameters.Add(new LookupParameter<IRandom>("Random", "A random number generator") { Hidden = true });
      Parameters.Add(new LookupParameter<IntValue>("Maximal Model Size", "The maximum number of sample points used to build the model (Set -1 for infinite size") { Hidden = true });
      Parameters.Add(new LookupParameter<DoubleMatrix>("InfillBounds", "The bounds applied for infill solving") { Hidden = true });
    }

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

    public override IOperation InstrumentedApply() {
      var regressionAlg = RegressionAlgorithmParameter.ActualValue;
      IDataset data = DatasetParameter.ActualValue;
      var random = RandomParameter.ActualValue;
      var oldModel = ModelParameter.ActualValue;
      var max = MaxModelSizeParameter.ActualValue.Value;
      if (data.Rows > max && max > 0) {
        data = SelectBestSamples(data, max);
        InfillBoundsParameter.ActualValue = GetBounds(data);
      }
      ModelParameter.ActualValue = BuildModel(random, regressionAlg, data, oldModel);
      return base.InstrumentedApply();
    }

    private DoubleMatrix GetBounds(IDataset data) {
      var res = new DoubleMatrix(data.Columns - 1, 2);
      var names = data.DoubleVariables.ToArray();
      for (var i = 0; i < names.Length - 1; i++) {
        res[i, 0] = data.GetDoubleValues(names[i]).Min();
        res[i, 1] = data.GetDoubleValues(names[i]).Max();
      }
      return res;
    }

    private static Dataset SelectBestSamples(IDataset data, int max) {
      var bestSampleIndices = data.GetDoubleValues("output").Select((d, i) => Tuple.Create(d, i)).OrderBy(x => x.Item1).Take(max).Select(x => x.Item2).ToArray();
      return new Dataset(data.VariableNames, data.VariableNames.Select(v => data.GetDoubleValues(v, bestSampleIndices).ToList()));
    }

    private IRegressionSolution BuildModel(IRandom random, IDataAnalysisAlgorithm<IRegressionProblem> regressionAlgorithm, IDataset dataset, IRegressionSolution oldSolution) {
      //var dataset = EgoUtilities.GetDataSet(dataSamples, RemoveDuplicates);
      var problemdata = new RegressionProblemData(dataset, dataset.VariableNames.Where(x => !x.Equals("output")), "output");
      problemdata.TrainingPartition.Start = 0;
      problemdata.TrainingPartition.End = dataset.Rows;
      problemdata.TestPartition.Start = dataset.Rows;
      problemdata.TestPartition.End = dataset.Rows;

      //train 
      var problem = (RegressionProblem)regressionAlgorithm.Problem;
      problem.ProblemDataParameter.Value = problemdata;
      var i = 0;
      IRegressionSolution solution = null;

      while (solution == null && i++ < 100) {
        var results = EgoUtilities.SyncRunSubAlgorithm(regressionAlgorithm, random.Next(int.MaxValue), Cancellation);
        solution = results.Select(x => x.Value).OfType<IRegressionSolution>().SingleOrDefault();
      }

      if (regressionAlgorithm is GaussianProcessRegression && oldSolution != null)
        solution = SanitizeGaussianProcess(oldSolution as GaussianProcessRegressionSolution, solution as GaussianProcessRegressionSolution, Cancellation);

      if (regressionAlgorithm is M5RegressionTree && oldSolution != null)
        solution = SanitizeM5Regression(oldSolution.Model as M5Model, solution, random, Cancellation);


      regressionAlgorithm.Runs.Clear();
      return solution;

    }

    private static IRegressionSolution SanitizeM5Regression(M5Model oldmodel, IRegressionSolution newSolution, IRandom random, CancellationToken cancellation) {
      var problemdata = newSolution.ProblemData;
      oldmodel.UpdateLeafModels(problemdata, problemdata.AllIndices, random, cancellation);
      var oldSolution = oldmodel.CreateRegressionSolution(problemdata);
      var magicDecision = newSolution.TrainingRSquared < oldSolution.TrainingRSquared - 0.05;
      return magicDecision ? newSolution : oldmodel.CreateRegressionSolution(problemdata);
    }

    //try creating a model with old hyperparameters and new dataset;
    private static IRegressionSolution SanitizeGaussianProcess(GaussianProcessRegressionSolution oldmodel, GaussianProcessRegressionSolution newSolution, CancellationToken cancellation) {
      var problemdata = newSolution.ProblemData;
      var mean = (IMeanFunction)oldmodel.Model.MeanFunction.Clone();
      var cov = (ICovarianceFunction)oldmodel.Model.CovarianceFunction.Clone();
      try {
        var model = new GaussianProcessModel(problemdata.Dataset, problemdata.TargetVariable, problemdata.AllowedInputVariables, problemdata.TrainingIndices, new[] { 0.0 }, mean, cov);
        cancellation.ThrowIfCancellationRequested();
        model.FixParameters();
        var sol = new GaussianProcessRegressionSolution(model, problemdata);
        if (newSolution.TrainingMeanSquaredError > sol.TrainingMeanSquaredError) {
          newSolution = sol;
        }
      } catch (ArgumentException) { }
      return newSolution;
    }

  }
}