source: branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem.Common/3.3/KnowledgeCenter.cs @ 13797

#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 HeuristicLab.Analysis;
using HeuristicLab.Analysis.SelfOrganizingMaps;
using HeuristicLab.Collections;
using HeuristicLab.Common;
using HeuristicLab.Common.Resources;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.MainForm;
using HeuristicLab.Optimization;
using HeuristicLab.Persistence.Default.Xml;
using HeuristicLab.Problems.DataAnalysis;
using HeuristicLab.Random;
using System;
using System.Collections.Generic;
using System.Drawing;
using System.IO;
using System.Linq;
using System.Threading;
using System.Threading.Tasks;
using RunCreationClient = HeuristicLab.Clients.OKB.RunCreation.RunCreationClient;
using SingleObjectiveOKBProblem = HeuristicLab.Clients.OKB.RunCreation.SingleObjectiveOKBProblem;
using SingleObjectiveOKBSolution = HeuristicLab.Clients.OKB.RunCreation.SingleObjectiveOKBSolution;

namespace HeuristicLab.OptimizationExpertSystem.Common {
  [Item("Knowledge Center", "Currently in experimental phase, an expert system that makes algorithm suggestions based on fitness landscape analysis features and an optimization knowledge base.")]
  [Creatable(CreatableAttribute.Categories.TestingAndAnalysis, Priority = 119)]
  public sealed class KnowledgeCenter : IContent {
    private bool SuppressEvents { get; set; }

    public string Filename { get; set; }

    public static Image StaticItemImage {
      get { return VSImageLibrary.Library; }
    }

    private readonly IntValue maximumEvaluations;
    public IntValue MaximumEvaluations {
      get { return maximumEvaluations; }
    }

    private readonly DoubleValue minimumTarget;
    public DoubleValue MinimumTarget {
      get { return minimumTarget; }
    }
    
    private readonly RunCollection instanceRuns;
    public RunCollection InstanceRuns {
      get { return instanceRuns; }
    }

    private readonly RunCollection seededRuns;
    public RunCollection SeededRuns {
      get { return seededRuns; }
    }

    private readonly RunCollection knowledgeBase;
    public RunCollection KnowledgeBase {
      get { return knowledgeBase; }
    }

    private readonly SingleObjectiveOKBProblem problem;
    public SingleObjectiveOKBProblem Problem {
      get { return problem; }
    }

    private readonly ItemList<IAlgorithm> algorithmInstances;
    private readonly ReadOnlyItemList<IAlgorithm> readonlyAlgorithmInstances;
    public ReadOnlyItemList<IAlgorithm> AlgorithmInstances {
      get { return readonlyAlgorithmInstances; }
    }

    private readonly RunCollection problemInstances;
    public RunCollection ProblemInstances {
      get { return problemInstances; }
    }

    private IRecommendationModel recommendationModel;
    public IRecommendationModel RecommendationModel {
      get { return recommendationModel; }
      set {
        if (recommendationModel == value) return;
        recommendationModel = value;
        OnRecommenderModelChanged();
      }
    }

    private readonly CheckedItemList<IScope> solutionSeedingPool;
    public CheckedItemList<IScope> SolutionSeedingPool {
      get { return solutionSeedingPool; }
    }

    private readonly EnumValue<SeedingStrategyTypes> seedingStrategy;
    public EnumValue<SeedingStrategyTypes> SeedingStrategy {
      get { return seedingStrategy; }
    }
    
    private BidirectionalLookup<long, IRun> algorithmId2RunMapping;
    private BidirectionalDictionary<long, IAlgorithm> algorithmId2AlgorithmInstanceMapping;
    private BidirectionalDictionary<long, IRun> problemId2ProblemInstanceMapping;
    
    public bool Maximization {
      get { return Problem != null && Problem.ProblemId >= 0 && ((IValueParameter<BoolValue>)Problem.MaximizationParameter).Value.Value; }
    }

    public KnowledgeCenter() {
      maximumEvaluations = new IntValue(10000);
      minimumTarget = new DoubleValue(0.05);
      instanceRuns = new RunCollection();
      seededRuns = new RunCollection();
      knowledgeBase = new RunCollection();
      algorithmInstances = new ItemList<IAlgorithm>();
      readonlyAlgorithmInstances = algorithmInstances.AsReadOnly();
      problemInstances = new RunCollection();
      recommendationModel = FixedRankModel.GetEmpty();
      problem = new SingleObjectiveOKBProblem();
      algorithmId2RunMapping = new BidirectionalLookup<long, IRun>();
      algorithmId2AlgorithmInstanceMapping = new BidirectionalDictionary<long, IAlgorithm>();
      problemId2ProblemInstanceMapping = new BidirectionalDictionary<long, IRun>();
      solutionSeedingPool = new CheckedItemList<IScope>();
      seedingStrategy = new EnumValue<SeedingStrategyTypes>(SeedingStrategyTypes.NoSeeding);
      RegisterEventHandlers();
    }

    private void ProblemOnProblemChanged(object sender, EventArgs eventArgs) {
      // TODO: Potentially, knowledge base has to be re-downloaded
    }

    private void RegisterEventHandlers() {
      maximumEvaluations.ValueChanged += MaximumEvaluationsOnValueChanged;
      minimumTarget.ValueChanged += MinimumTargetOnValueChanged;
      problem.ProblemChanged += ProblemOnProblemChanged;
      problem.Solutions.ItemsAdded += ProblemSolutionsChanged;
      problem.Solutions.ItemsReplaced += ProblemSolutionsChanged;
      problem.Solutions.ItemsRemoved += ProblemSolutionsChanged;
      problem.Solutions.CollectionReset += ProblemSolutionsChanged;
      instanceRuns.CollectionReset += InformationChanged;
      instanceRuns.ItemsAdded += InformationChanged;
      instanceRuns.ItemsRemoved += InformationChanged;
      instanceRuns.Reset += InformationChanged;
      instanceRuns.UpdateOfRunsInProgressChanged += InformationChanged;
      knowledgeBase.CollectionReset += InformationChanged;
      knowledgeBase.ItemsAdded += InformationChanged;
      knowledgeBase.ItemsRemoved += InformationChanged;
    }

    private void MaximumEvaluationsOnValueChanged(object sender, EventArgs eventArgs) {

    }

    private void MinimumTargetOnValueChanged(object sender, EventArgs e) {

    }

    private void ProblemSolutionsChanged(object sender, EventArgs e) {
      foreach (var sol in Problem.Solutions.Select(x => x.Solution).OfType<IScope>()) {
        if (!SolutionSeedingPool.Contains(sol))
          SolutionSeedingPool.Add(sol, false);
      }
    }

    private void InformationChanged(object sender, EventArgs e) {
      var runCollection = sender as RunCollection;
      if (runCollection != null && runCollection.UpdateOfRunsInProgress) return;
    }
    
    public bool IsCurrentInstance(IRun run) {
      if (!problemId2ProblemInstanceMapping.ContainsSecond(run)) return false;
      return problemId2ProblemInstanceMapping.GetBySecond(run) == Problem.ProblemId;
    }

    public void UpdateInstanceProjection(string[] characteristics) {
      if (characteristics.Length == 0) return;

      var instances = GetProblemCharacteristics(characteristics);

      var key2Idx = new BidirectionalDictionary<IRun, int>();
      foreach (var kvp in instances.Select((k, i) => new { Index = i, Key = k.Key }))
        key2Idx.Add(kvp.Key, kvp.Index);

      #region MDS
      Func<double[], double[], double> euclid = (a, b) => Math.Sqrt(a.Zip(b, (x, y) => (x - y)).Sum(x => x * x));
      var num = instances.Count;
      var matrix = new DoubleMatrix(num, num);
      for (var i = 0; i < num - 1; i++) {
        for (var j = i + 1; j < num; j++) {
          matrix[i, j] = matrix[j, i] = euclid(instances[key2Idx.GetBySecond(i)], instances[key2Idx.GetBySecond(j)]);
        }
      }

      var coords = MultidimensionalScaling.KruskalShepard(matrix);
      #endregion
      #region PCA
      double[,] v = null;
      var ds = new double[instances.Count, characteristics.Length];
      if (characteristics.Length > 1) {
        foreach (var instance in instances) {
          var arr = instance.Value;
          for (var feature = 0; feature < arr.Length; feature++)
            ds[key2Idx.GetByFirst(instance.Key), feature] = arr[feature];
        }

        int info;
        double[] s2;
        alglib.pcabuildbasis(ds, ds.GetLength(0), ds.GetLength(1), out info, out s2, out v);
      }
      #endregion
      #region SOM
      var features = new DoubleMatrix(characteristics.Length, instances.Count);
      foreach (var instance in instances) {
        var arr = instance.Value;
        for (var feature = 0; feature < arr.Length; feature++)
          features[feature, key2Idx.GetByFirst(instance.Key)] = arr[feature];
      }
      var somCoords = SOM.Map(features, new MersenneTwister(42), somSize: 10, learningRadius: 20, iterations: 200, jittering: true);
      #endregion

      ProblemInstances.UpdateOfRunsInProgress = true;
      try {
        foreach (var instance in ProblemInstances) {
          IItem item;
          if (v != null) {
            double x = 0, y = 0;
            for (var feature = 0; feature < ds.GetLength(1); feature++) {
              x += ds[key2Idx.GetByFirst(instance), feature] * v[feature, 0];
              y += ds[key2Idx.GetByFirst(instance), feature] * v[feature, 1];
            }

            if (instance.Results.TryGetValue("Projection.PCA.X", out item)) {
              ((DoubleValue)item).Value = x;
            } else instance.Results.Add("Projection.PCA.X", new DoubleValue(x));
            if (instance.Results.TryGetValue("Projection.PCA.Y", out item)) {
              ((DoubleValue)item).Value = y;
            } else instance.Results.Add("Projection.PCA.Y", new DoubleValue(y));
          } else {
            instance.Results.Remove("Projection.PCA.X");
            instance.Results.Remove("Projection.PCA.Y");
          }

          if (instance.Results.TryGetValue("Projection.MDS.X", out item)) {
            ((DoubleValue)item).Value = coords[key2Idx.GetByFirst(instance), 0];
          } else instance.Results.Add("Projection.MDS.X", new DoubleValue(coords[key2Idx.GetByFirst(instance), 0]));
          if (instance.Results.TryGetValue("Projection.MDS.Y", out item)) {
            ((DoubleValue)item).Value = coords[key2Idx.GetByFirst(instance), 1];
          } else instance.Results.Add("Projection.MDS.Y", new DoubleValue(coords[key2Idx.GetByFirst(instance), 1]));

          if (instance.Results.TryGetValue("Projection.SOM.X", out item)) {
            ((DoubleValue)item).Value = somCoords[key2Idx.GetByFirst(instance), 0];
          } else instance.Results.Add("Projection.SOM.X", new DoubleValue(somCoords[key2Idx.GetByFirst(instance), 0]));
          if (instance.Results.TryGetValue("Projection.SOM.Y", out item)) {
            ((DoubleValue)item).Value = somCoords[key2Idx.GetByFirst(instance), 1];
          } else instance.Results.Add("Projection.SOM.Y", new DoubleValue(somCoords[key2Idx.GetByFirst(instance), 1]));
        }
      } finally { ProblemInstances.UpdateOfRunsInProgress = false; }
    }

    private static readonly HashSet<string> InterestingValueNames = new HashSet<string>() {
      "QualityPerEvaluations", "Problem Name", "Problem Type", "Algorithm Name", "Algorithm Type", "Maximization", "BestKnownQuality"
    };

    public Task<ResultCollection> StartAlgorithmAsync(int index) {
      return StartAlgorithmAsync(index, CancellationToken.None);
    }

    public Task<ResultCollection> StartAlgorithmAsync(int index, CancellationToken cancellation) {
      var selectedInstance = algorithmInstances[index];
      var algorithmClone = (IAlgorithm)selectedInstance.Clone();
      var problemClone = Problem.CloneProblem() as ISingleObjectiveHeuristicOptimizationProblem;
      if (problemClone == null) throw new InvalidOperationException("Problem is not of type " + typeof(ISingleObjectiveHeuristicOptimizationProblem).FullName);
      // TODO: It is assumed the problem instance by default is configured using no preexisting solution creator
      var seedingStrategyLocal = SeedingStrategy.Value;
      if (seedingStrategyLocal != SeedingStrategyTypes.NoSeeding) {
        if (!SolutionSeedingPool.CheckedItems.Any()) throw new InvalidOperationException("There are no solutions selected for seeding.");
        // TODO: It would be necessary to specify the solution creator somewhere (property and GUI)
        var seedingCreator = problemClone.Operators.OfType<IPreexistingSolutionCreator>().FirstOrDefault();
        if (seedingCreator == null) throw new InvalidOperationException("The problem does not contain a solution creator that allows seeding.");
        seedingCreator.PreexistingSolutionsParameter.Value.Replace(SolutionSeedingPool.CheckedItems.Select(x => x.Value));
        seedingCreator.SampleFromPreexistingParameter.Value.Value = seedingStrategyLocal == SeedingStrategyTypes.SeedBySampling;
        // TODO: WHY!? WHY??!?
        ((dynamic)problemClone.SolutionCreatorParameter).Value = (dynamic)seedingCreator;
      }
      algorithmClone.Problem = problemClone;
      algorithmClone.Prepare(true);
      IParameter stopParam;
      var monitorStop = true;
      if (algorithmClone.Parameters.TryGetValue("MaximumEvaluations", out stopParam)) {
        var maxEvalParam = stopParam as IValueParameter<Data.IntValue>;
        if (maxEvalParam != null) {
          maxEvalParam.Value.Value = MaximumEvaluations.Value;
          monitorStop = false;
        }
      }

      // TODO: The following can be simplified when we have async implementation patterns for our algorithms:
      // TODO: The closures can be removed and replaced with private member methods
      var waitHandle = new AutoResetEvent(false);

      #region EventHandler closures
      EventHandler exeStateChanged = (sender, e) => {
        if (algorithmClone.ExecutionState == ExecutionState.Stopped) {
          lock (Problem.Solutions) {
            foreach (var solution in algorithmClone.Results.Where(x => x.Name.ToLower().Contains("solution")).Select(x => x.Value).OfType<IScope>()) {
              Problem.Solutions.Add(new SingleObjectiveOKBSolution(Problem.ProblemId) {
                Quality = solution.Variables.ContainsKey(Problem.Problem.Evaluator.QualityParameter.ActualName) ? ((DoubleValue)solution.Variables[Problem.Problem.Evaluator.QualityParameter.ActualName].Value).Value : double.NaN,
                Solution = (IItem)solution.Clone()
              });
            }
          }
          if (seedingStrategyLocal == SeedingStrategyTypes.NoSeeding) {
            lock (InstanceRuns) {
              InstanceRuns.Add(algorithmClone.Runs.Last());
            }
          } else {
            lock (SeededRuns) {
              SeededRuns.Add(algorithmClone.Runs.Last());
            }
          }
          waitHandle.Set();
        }
      };

      EventHandler<EventArgs<Exception>> exceptionOccurred = (sender, e) => {
        waitHandle.Set();
      };

      EventHandler timeChanged = (sender, e) => {
        IResult evalSolResult;
        if (!algorithmClone.Results.TryGetValue("EvaluatedSolutions", out evalSolResult) || !(evalSolResult.Value is Data.IntValue)) return;
        var evalSols = ((Data.IntValue)evalSolResult.Value).Value;
        if (evalSols >= MaximumEvaluations.Value && algorithmClone.ExecutionState == ExecutionState.Started)
          algorithmClone.Stop();
      };
      #endregion

      algorithmClone.ExecutionStateChanged += exeStateChanged;
      algorithmClone.ExceptionOccurred += exceptionOccurred;
      if (monitorStop) algorithmClone.ExecutionTimeChanged += timeChanged;

      return Task.Factory.StartNew(() => {
        algorithmClone.Start();
        OnAlgorithmInstanceStarted(algorithmClone);
        var cancelRequested = false;
        while (!waitHandle.WaitOne(200)) {
          if (cancellation.IsCancellationRequested) {
            cancelRequested = true;
            break;
          }
        }
        if (cancelRequested) {
          try { algorithmClone.Stop(); } catch { } // ignore race condition if it is stopped in the meantime
          waitHandle.WaitOne();
        }
        waitHandle.Dispose();
        return algorithmClone.Results;
      }, TaskCreationOptions.LongRunning);
    }

    public ResultCollection StartAlgorithm(int index, CancellationToken cancellation) {
      var task = StartAlgorithmAsync(index, cancellation);
      task.Wait(cancellation);
      return task.Result;
    }

    public Task UpdateKnowledgeBaseAsync(IProgress progress = null) {
      if (progress == null) progress = new Progress(string.Empty);
      progress.Start("Updating Knowledge Base from OKB");
      OnDownloadStarted(progress);
      return Task.Factory.StartNew(() => { DoUpdateKnowledgeBase(progress); }, TaskCreationOptions.LongRunning);
    }

    public void UpdateKnowledgeBase(IProgress progress = null) {
      UpdateKnowledgeBaseAsync(progress).Wait();
    }

    private void DoUpdateKnowledgeBase(IProgress progress) {
      var queryClient = Clients.OKB.Query.QueryClient.Instance;
      var adminClient = Clients.OKB.Administration.AdministrationClient.Instance;
      try {
        progress.Status = "Connecting to OKB...";
        progress.ProgressValue = 0;
        // FIXME: How to tell if refresh is necessary?
        var refreshTasks = new[] {
          Task.Factory.StartNew(() => queryClient.Refresh()),
          Task.Factory.StartNew(() => adminClient.Refresh())
        };
        Task.WaitAll(refreshTasks);

        var probInstance = adminClient.Problems.SingleOrDefault(x => x.Id == Problem.ProblemId);
        if (probInstance == null) throw new InvalidOperationException("The chosen problem instance cannot be found in the OKB.");
        var probClassId = probInstance.ProblemClassId;

        var problemClassFilter = (Clients.OKB.Query.StringComparisonAvailableValuesFilter)queryClient.Filters.Single(x => x.Label == "Problem Class Name");
        problemClassFilter.Value = adminClient.ProblemClasses.Single(x => x.Id == probClassId).Name;

        problemId2ProblemInstanceMapping.Clear();
        progress.Status = "Downloading problem instances...";
        progress.ProgressValue = 0;
        int[] p = { 0 };
        ProblemInstances.UpdateOfRunsInProgress = true;
        ProblemInstances.Clear();
        var characteristics = new HashSet<string>();
        var totalProblems = adminClient.Problems.Count(x => x.ProblemClassId == probClassId);
        Parallel.ForEach(adminClient.Problems.Where(x => x.ProblemClassId == probClassId), new ParallelOptions { MaxDegreeOfParallelism = 8 }, (pInst) => { 
          var charas = new List<string>();
          IRun probRun = null;
          var data = Clients.OKB.Administration.AdministrationClient.GetProblemData(pInst.Id);
          if (data != null) {
            using (var stream = new MemoryStream(data)) {
              try {
                var prob = (IProblem)XmlParser.Deserialize<IContent>(stream);
                probRun = new Run() { Name = prob.Name };
                prob.CollectParameterValues(probRun.Parameters);
                probRun.Parameters["Problem Name"] = new StringValue(prob.Name);
                probRun.Parameters["Problem Type"] = new StringValue(prob.GetType().Name);
                foreach (var v in RunCreationClient.Instance.GetCharacteristicValues(pInst.Id)) {
                  probRun.Results.Add("Characteristic." + v.Name, RunCreationClient.Instance.ConvertToItem(v));
                  charas.Add("Characteristic." + v.Name);
                }
              } catch { }
              stream.Close();
            }
            if (probRun != null) {
              lock (characteristics) {
                problemId2ProblemInstanceMapping.Add(pInst.Id, probRun);
                ProblemInstances.Add(probRun);
                progress.Status = string.Format("Downloaded problem {0} (okb-id: {1})....", pInst.Name, pInst.Id);
                p[0]++;
                progress.ProgressValue = p[0] / (double)totalProblems;
                foreach (var c in charas) characteristics.Add(c);
              }
            }
          }
        });

        algorithmId2AlgorithmInstanceMapping.Clear();
        progress.Status = "Downloading algorithm instances...";
        progress.ProgressValue = 0;
        p[0] = 0;
        Parallel.ForEach(adminClient.Algorithms, new ParallelOptions { MaxDegreeOfParallelism = 8 }, (algInst) => {
          IAlgorithm alg = null;
          var data = Clients.OKB.Administration.AdministrationClient.GetAlgorithmData(algInst.Id);
          if (data != null) {
            using (var stream = new MemoryStream(data)) {
              try {
                alg = (IAlgorithm)XmlParser.Deserialize<IContent>(stream);
              } catch { }
              stream.Close();
            }
            if (alg != null) {
              lock (algorithmId2AlgorithmInstanceMapping) {
                algorithmInstances.Add(alg);
                algorithmId2AlgorithmInstanceMapping.Add(algInst.Id, alg);
                progress.Status = string.Format("Downloaded algorithm {0} (okb-id: {1})...", algInst.Name, algInst.Id);
                p[0]++;
                progress.ProgressValue = p[0] / (double)adminClient.Algorithms.Count;
              }
            }
          }
        });

        var interestingValues = queryClient.ValueNames.Where(x => InterestingValueNames.Contains(x.Name)).ToList();

        progress.Status = "Downloading runs...";
        progress.ProgressValue = 0;
        p[0] = 0;
        var count = queryClient.GetNumberOfRuns(problemClassFilter);
        if (count == 0) return;
        
        var runList = new List<IRun>();
        var runIds = queryClient.GetRunIds(problemClassFilter).ToList();
        var batches = runIds.Select((v, i) => new { Idx = i, Val = v }).GroupBy(x => x.Idx / 500, x => x.Val);
        Parallel.ForEach(batches.Select(x => x.ToList()), new ParallelOptions { MaxDegreeOfParallelism = 4 }, (batch) => {
          var okbRuns = queryClient.GetRunsWithValues(batch, true, interestingValues);
          var hlRuns = okbRuns.AsParallel().Select(x => new { AlgorithmId = x.Algorithm.Id, Run = queryClient.ConvertToOptimizationRun(x) }).ToList();
          lock (runList) {
            foreach (var r in hlRuns) {
              algorithmId2RunMapping.Add(r.AlgorithmId, r.Run);
              runList.Add(r.Run);
            }
            progress.Status = string.Format("Downloaded runs {0} to {1} of {2}...", p[0], p[0] + batch.Count, count);
            p[0] += batch.Count;
            progress.ProgressValue = p[0] / (double)count;
          }
        });

        progress.Status = "Finishing...";
        
        try {
          KnowledgeBase.UpdateOfRunsInProgress = true;
          KnowledgeBase.Clear();
          KnowledgeBase.AddRange(runList);
        } finally { KnowledgeBase.UpdateOfRunsInProgress = false; }

        var algInstRunDict = runList.Where(x => x.Parameters.ContainsKey("Problem Name") && x.Parameters["Problem Name"] is StringValue)
                                          .GroupBy(x => ((StringValue)x.Parameters["Problem Name"]).Value)
                                          .ToDictionary(x => x.Key, x => x.GroupBy(y => ((StringValue)y.Parameters["Algorithm Name"]).Value)
                                                                                  .ToDictionary(y => y.Key, y => y.ToList()));

        // set best-known quality to best-found in case it is not known
        foreach (var kvp in algInstRunDict) {
          var prob = ProblemInstances.SingleOrDefault(x => ((StringValue)x.Parameters["Problem Name"]).Value == kvp.Key);
          if (prob == null) continue;
          var maximization = ((BoolValue)prob.Parameters["Maximization"]).Value;

          IItem bkParam;
          if (!prob.Parameters.TryGetValue("BestKnownQuality", out bkParam) || !(bkParam is DoubleValue)) {
            var list = kvp.Value.SelectMany(x => x.Value)
              .Where(x => x.Results.ContainsKey("QualityPerEvaluations"))
              .Select(x => ((IndexedDataTable<double>)x.Results["QualityPerEvaluations"]).Rows.First().Values.Last().Item2);
            if (!list.Any()) continue;
            bkParam = new DoubleValue(maximization ? list.Max() : list.Min());
            prob.Parameters["BestKnownQuality"] = bkParam;
          }
        }

        // add algorithm instance ranks as features to the problem instances for a range of targets
        foreach (var target in new[] {0, 0.01, 0.05, 0.1, 0.2, 0.5}) {
          var cls = GetPerformanceClasses(target, 5);
          foreach (var kvp in cls) {
            var prob = kvp.Key;
            foreach (var kvp2 in kvp.Value) {
              var resultName = "Rank." + algorithmId2AlgorithmInstanceMapping.GetByFirst(kvp2.Key) + "@" + (target * 100) + "%";
              prob.Results[resultName] = new IntValue(kvp2.Value);
            }
          }
        }
      } finally { progress.Finish(); ProblemInstances.UpdateOfRunsInProgress = false; }
      UpdateInstanceProjection(ProblemInstances.ResultNames.Where(x => x.StartsWith("Characteristic.")).ToArray());
    }

    public static double[][] GetFeatures(IRun[] problemInstances, string[] characteristics, double[] medianValues = null) {
      var instances = new double[problemInstances.Length][];
      for (var p = 0; p < problemInstances.Length; p++) {
        instances[p] = new double[characteristics.Length];
        for (var f = 0; f < characteristics.Length; f++) {
          IItem item;
          if (problemInstances[p].Results.TryGetValue(characteristics[f], out item)) {
            double val = 0;
            var dItem = item as DoubleValue;
            if (dItem != null) {
              val = dItem.Value;
            } else {
              var iItem = item as IntValue;
              if (iItem != null) val = iItem.Value;
              else val = double.NaN;
            }
            if (double.IsNaN(val) && medianValues != null)
              instances[p][f] = medianValues[f];
            else instances[p][f] = val;
          } else instances[p][f] = medianValues != null ? medianValues[f] : double.NaN;
        }
      }
      return instances;
    }

    public static double[] GetMedianValues(IRun[] problemInstances, string[] characteristics) {
      var values = new List<double>[characteristics.Length];
      foreach (var problemInstance in problemInstances) {
        for (var f = 0; f < characteristics.Length; f++) {
          if (values[f] == null) values[f] = new List<double>(problemInstances.Length);
          IItem item;
          if (problemInstance.Results.TryGetValue(characteristics[f], out item)) {
            var dItem = item as DoubleValue;
            if (dItem != null) values[f].Add(dItem.Value);
            else {
              var iItem = item as IntValue;
              if (iItem != null) values[f].Add(iItem.Value);
            }
          }
        }
      }
      return values.Select(x => x.Count == 0 ? 0.0 : x.Median()).ToArray();
    }

    public Dictionary<IRun, double[]> GetProblemCharacteristics(string[] characteristics) {
      var map = ProblemInstances.Select((v, i) => new { Index = i, ProblemInstance = v }).ToDictionary(x => x.Index, x => x.ProblemInstance);
      var instances = GetFeatures(ProblemInstances.ToArray(), characteristics);
      var median = GetMedianValues(ProblemInstances.ToArray(), characteristics);

      var allValues = instances.Select(x => x.Select((f, i) => new { Idx = i, Val = double.IsNaN(f) ? median[i] : f }).ToList())
        .SelectMany(x => x)
        .GroupBy(x => x.Idx, x => x.Val)
        .OrderBy(x => x.Key).ToList();
      var avg = allValues.Select(x => x.Average()).ToList();
      var stdev = allValues.Select(x => x.StandardDeviation()).ToList();

      // normalize characteristic values by transforming them to their z-score
      foreach (var features in instances) {
        for (var i = 0; i < features.Length; i++) {
          if (double.IsNaN(features[i])) features[i] = median[i];
          if (stdev[i] > 0) features[i] = (features[i] - avg[i]) / stdev[i];
        }
      }
      return instances.Select((v, i) => new { ProblemInstance = map[i], Features = v }).ToDictionary(x => x.ProblemInstance, x => x.Features);
    }

    public Dictionary<IAlgorithm, double> GetAlgorithmPerformance(IRun problemInstance) {
      if (!problemInstance.Parameters.ContainsKey("BestKnownQuality")) return new Dictionary<IAlgorithm, double>();
      var target = GetTarget(((DoubleValue)problemInstance.Parameters["BestKnownQuality"]).Value, MinimumTarget.Value, Maximization);
      return knowledgeBase.Where(x => ((StringValue)x.Parameters["Problem Name"]).Value == ((StringValue)problemInstance.Parameters["Problem Name"]).Value)
                          .GroupBy(x => algorithmId2AlgorithmInstanceMapping.GetByFirst(algorithmId2RunMapping.GetBySecond(x).Single()))
                          .ToDictionary(x => x.Key, x => ExpectedRuntimeHelper.CalculateErt(x.ToList(), "QualityPerEvaluations", target, Maximization).ExpectedRuntime);
    }

    public Dictionary<IAlgorithm, List<IRun>> GetAlgorithmRuns(IRun problemInstance) {
      return knowledgeBase.Where(x => ((StringValue)x.Parameters["Problem Name"]).Value == ((StringValue)problemInstance.Parameters["Problem Name"]).Value)
                          .GroupBy(x => algorithmId2AlgorithmInstanceMapping.GetByFirst(algorithmId2RunMapping.GetBySecond(x).Single()))
                          .ToDictionary(x => x.Key, x => x.ToList());
    } 

    public Dictionary<IAlgorithm, List<IRun>> GetKnowledgeBaseByAlgorithm() {
      return KnowledgeBase.GroupBy(x => algorithmId2AlgorithmInstanceMapping.GetByFirst(algorithmId2RunMapping.GetBySecond(x).Single()))
                          .ToDictionary(x => x.Key, x => x.ToList());
    }

    public IEnumerable<IRegressionProblem> GetRegressionProblemPerAlgorithmInstance(double target, string[] characteristics) {
      if (Problem == null) yield break;
      var features = GetProblemCharacteristics(characteristics);
      // TODO: knowledgebase only stores problem name as a string
      // this doesn't work if there are two equally named problem instances
      var problemMap = ProblemInstances.Select(x => new { Key = ((StringValue)x.Parameters["Problem Name"]).Value, Value = x })
                                       .ToDictionary(x => x.Key, x => x.Value);
      foreach (var relevantRuns in knowledgeBase.GroupBy(x => algorithmId2RunMapping.GetBySecond(x).Single())) {
        var problemRuns = relevantRuns.GroupBy(x => ((StringValue)x.Parameters["Problem Name"]).Value).ToList();
        var ds = new ModifiableDataset();
        ds.AddVariable("Problem Name", new List<string>());
        foreach (var pc in characteristics)
          ds.AddVariable(pc, new List<double>());
        ds.AddVariable("ERT", new List<double>());
        foreach (var pr in problemRuns) {
          var prob = problemMap[pr.Key];
          var f = features[prob];
          var max = ((BoolValue)prob.Parameters["Maximization"]).Value;
          var bkq = ((DoubleValue)prob.Parameters["BestKnownQuality"]).Value;
          var ert = ExpectedRuntimeHelper.CalculateErt(pr.ToList(), "QualityPerEvaluations", GetTarget(bkq, target, max), max).ExpectedRuntime;
          if (double.IsNaN(ert)) ert = int.MaxValue;
          ds.AddRow(new object[] { pr.Key }.Concat(f.Cast<object>()).Concat(new object[] { ert }));
        }
        var datAnalysisData = new RegressionProblemData(ds, characteristics, "ERT");
        var result = new RegressionProblem() {
          Name = algorithmId2AlgorithmInstanceMapping.GetByFirst(relevantRuns.Key).Name
        };
        result.ProblemDataParameter.Value = datAnalysisData;
        yield return result;
      }
    }

    public IEnumerable<IClassificationProblem> GetClassificationProblemPerAlgorithmInstance(double target, string[] characteristics) {
      if (Problem == null) yield break;

      var classes = GetPerformanceClasses(target, 5);
      var features = GetProblemCharacteristics(characteristics);

      foreach (var alg in AlgorithmInstances) {
        var ds = new ModifiableDataset();
        ds.AddVariable("Problem Name", new List<string>());
        foreach (var pc in characteristics)
          ds.AddVariable(pc, new List<double>());
        ds.AddVariable("Class", new List<double>());

        foreach (var c in classes) {
          int cls;
          if (c.Value.TryGetValue(algorithmId2AlgorithmInstanceMapping.GetBySecond(alg), out cls)) {
            ds.AddRow(new object[] { ((StringValue)c.Key.Parameters["Problem Name"]).Value }
              .Concat(features[c.Key].Cast<object>()).Concat(new object[] { cls }));
          }
        }
        var datAnalysisData = new ClassificationProblemData(ds, characteristics, "Class");
        var result = new ClassificationProblem() {
          Name = alg.Name
        };
        result.ProblemDataParameter.Value = datAnalysisData;
        yield return result;
      }
    }

    public Dictionary<IRun, double> GetProblemDistances(string[] characteristics) {
      var result = new Dictionary<IRun, double>();
      var currentInstance = problemId2ProblemInstanceMapping.GetByFirst(Problem.ProblemId);
      var features = GetProblemCharacteristics(characteristics);
      var cF = features[currentInstance];
      foreach (var b in ProblemInstances) {
        if (b == currentInstance) continue;
        var sum = features[b].Select((t, f) => (cF[f] - t) * (cF[f] - t)).Sum();
        result[b] = Math.Sqrt(sum);
      }
      return result;
    }

    public Dictionary<IRun, Dictionary<long, int>> GetPerformanceClasses(double target, int nClasses) {
      var result = new Dictionary<IRun, Dictionary<long, int>>();
      var problemMap = ProblemInstances.Select(x => new { Key = ((StringValue)x.Parameters["Problem Name"]).Value, Value = x })
                                       .ToDictionary(x => x.Key, x => x.Value);
      foreach (var pr in KnowledgeBase.GroupBy(x => ((StringValue)x.Parameters["Problem Name"]).Value).ToList()) {
        var bkq = ((DoubleValue)problemMap[pr.Key].Parameters["BestKnownQuality"]).Value;
        var max = ((BoolValue)problemMap[pr.Key].Parameters["Maximization"]).Value;

        result[problemMap[pr.Key]] = new Dictionary<long, int>();

        var values = pr.GroupBy(x => algorithmId2RunMapping.GetBySecond(x).Single())
                       .ToDictionary(x => x.Key, x => ExpectedRuntimeHelper.CalculateErt(x.ToList(), "QualityPerEvaluations", GetTarget(bkq, target, max), max).ExpectedRuntime);
        var ranks = ClusteringHelper<long>.Cluster(nClasses, values, x => double.IsNaN(x.Value))
          .GetByCluster().ToList();
        foreach (var c in ranks) {
          foreach (var a in c.Value)
            result[problemMap[pr.Key]][a.Key] = c.Key;
        }
      }
      return result;
    }

    public double GetTarget(double bestKnownQuality, double target, bool maximization) {
      return bestKnownQuality * (maximization ? (1 - target) : (1 + target));
    }

    public event EventHandler<EventArgs<IProgress>> DownloadStarted;
    private void OnDownloadStarted(IProgress progress) {
      var handler = DownloadStarted;
      if (handler != null) handler(this, new EventArgs<IProgress>(progress));
    }

    public event EventHandler<EventArgs<IAlgorithm>> AlgorithmInstanceStarted;
    private void OnAlgorithmInstanceStarted(IAlgorithm instance) {
      var handler = AlgorithmInstanceStarted;
      if (handler != null) handler(this, new EventArgs<IAlgorithm>(instance));
    }

    public event EventHandler RecommendationModelChanged;
    private void OnRecommenderModelChanged() {
      var handler = RecommendationModelChanged;
      if (handler != null) handler(this, EventArgs.Empty);
    }

    public IEnumerable<KeyValuePair<IAlgorithm, double>> GetAlgorithmInstanceRanking() {
      return RecommendationModel.GetRanking(ProblemInstances.Single(IsCurrentInstance));
    }
  }
}