source: branches/2789_MathNetNumerics-Exploration/HeuristicLab.Algorithms.DataAnalysis.Experimental/Splines.cs @ 16017

#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.Diagnostics;
using System.Linq;
using System.Threading;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.Optimization;
using HeuristicLab.Parameters;
using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
using HeuristicLab.Problems.DataAnalysis;
using HeuristicLab.Random;
using MathNet.Numerics.Interpolation;
using MathNet.Numerics.LinearAlgebra.Double;
using MathNet.Numerics.LinearAlgebra.Double.Solvers;


/*
 * Experimenting with various spline implementations.
 *
 * There are many different variants of using splines.
 *  - Spline Interpolation
 *  - Natural Spline Interpolation
 *  - Smoothing Splines (Reinsch)
 *  - P-Splines
 *  - B-Splines
 *  - Regression Splines
 *  - Penalized Regression Splines
 *  -
 *
 *
 *  Generally, a spline is a piecewise function whereby each piece connects two knots.
 *  In almost all cases splines are univariate. There are extensions to more dimensions
 *  but this is much less convenient.
 *  A linear spline is a piecewise linear function. Usually the following constaints are enforced
 *  for each knot point k and the two piecewise functions fl and fr
 *  fl(k) = fr(k), fl(k)' = fr(k)', fl(k)''=0, fr(k)''=0, additionally constraints for
 *  the boundary knots are often enforced.

 *  Spline Interpolation solves the problem of finding a smooth interpolating function
 *  which goes through all knots exactly.
 *
 *  In spline smoothing one tries to find a function minimizing the squared residuals
 *  and maximizing smoothness. For the smoothing spline g the optimization objective is
 *  || y - g(x) ||² + \lambda integrate (xmin : xmax) g(x)'' dx.
 *  The roughness penality is the integral over the second derivative of the smoothing spline
 *  over the whole domain of x. Lambda is used to balance the amount of smoothing.
 *  If lambda is set to zero, the solution is an interpolating spline. If lambda is set to
 *  infinity the solution is a maximally smooth function (= line).
 *  Interestingly, the estimation of a smoothing spline is a linear operation. As a consequence
 *  a lot of theory from linear models can be reused (e.g. cross-validation and generalized
 *  cross-validation) with minimal additional effort. GCV can also be used to tune the
 *  smoothing parameter automatically. Tuned algorithms are available to solve smoothing
 *  spline estimation efficiently O(n) or O(m²n) where m is the number of knots and n the
 *  number of data points.
 *  It is possible to calculate confidence intervals for the smoothing spline and given
 *  an error model subsequently also confidence intervals for predictions.
 *
 *  We are primarily interested in spline smoothing / regression.
 *  Alglib and Math.NET provide several different implementations for interpolation
 *  including polynomial (cubic), Hermite, Akima, Linear, Catmull-Rom, Rational, ...
 *
 *  For spline smoothing or regression there are also many differen variants.
 *
 *  Smoothing Splines by Reinsch (Reinsch, Smoothing by Spline Functions, 1967).
 *  Pseudo-code for the algorithm is given in the paper. A small change in the algorithm
 *  to improve convergence is discussed in (Reinsch, Smoothing by Spline Functions II, 1970)
 *  The algorithm uses the properties of the smoothing matrix (bandedness) to calculate the
 *  smoothing spline in O(n). Two parameters have to be set (S and rho). If the noise in the
 *  data is known then both parameter can be set directly, otherwise S should be set to n-1
 *  and rho should be found via cross-validation.
 *  It is not easy to calculate a the CV (PRESS) or GCV in the approach by Reinsch and
 *  therefore the algorithm is not ideal.
 *  Wahba and colleagues (1990) introduced CV and GCV estimation.
 *  Hutchinson and colleages improved the algorithm (Fortran implementation available as
 *  algorithm 642.f from ACM).
 *
 *  CUBGCV (Algorithm 642 Hutchingson)
 *  Several improvments over the Reinsch algorithm. In particular efficient calculation
 *  of the
 
 *  Finbarr O'Sullivan BART
 *  ...
 * 
 *  B-Splines
 *  ...
 * 
 *  P-Splines (Eilers and Marx), use the B-Spline basis in combination with a penality
 *  for large variations of subsequent B-Spline coefficients. Conceptually, if the
 *  coefficients of neighbouring B-Splines are almost the same then the resulting smoothing
 *  spline is less wiggly. It has been critizized that this is rather subjective.
 *  P-Splines also allow calculation of CV and GCV for a given smoothing parameter which
 *  can be used for optimization of lambda.
 *  It is easy to use P-Splines within GAMs (no backfitting necessary).
 *  The paper contains MATLAB code also for the calculation of the B-Spline basis.
 *  The computational effort is O(?). The authors argue that it is efficient.
 *  The number of knots is not critical it is possible to use every data point as knot.
 *
 *  Alglib: penalized regression spline
 *  ...
 * 
 *  Alglib: penalized regression spline with support for weights
 *  ... 
 */

namespace HeuristicLab.Algorithms.DataAnalysis.Experimental {
  [Item("Splines", "")]
  [Creatable(CreatableAttribute.Categories.DataAnalysisRegression, Priority = 102)]
  [StorableClass]
  public sealed class Splines : FixedDataAnalysisAlgorithm<IRegressionProblem> {
    [StorableConstructor]
    private Splines(bool deserializing) : base(deserializing) { }
    [StorableHook(HookType.AfterDeserialization)]
    private void AfterDeserialization() {
    }

    private Splines(Splines original, Cloner cloner)
      : base(original, cloner) {
    }
    public override IDeepCloneable Clone(Cloner cloner) {
      return new Splines(this, cloner);
    }

    public Splines()
      : base() {
      var validTypes = new ItemSet<StringValue>(
        new[] {
          "Smoothing Spline (Reinsch)",
          "Smoothing Spline (Reinsch with automatic tolerance determination using LOOCV)",
          "B-Spline Smoothing",
          "Penalized Regression Spline (alglib)",
          "CUBGCV",
          "Finnbar O'Sullivan (sbart)"
      }.Select(s => new StringValue(s)));

      Parameters.Add(new ConstrainedValueParameter<StringValue>("Type", "The type of spline (as supported by alglib)", validTypes, validTypes.Last()));
      Parameters.Add(new ValueParameter<DoubleValue>("Lambda", "Regularization parameter for smoothing splines (0..+inf)", new DoubleValue(100)));
      Parameters.Add(new ValueParameter<DoubleValue>("StdDev (noise)", "Known error in y values. Used only be Reinsch Smoothing Splines", new DoubleValue(0.1)));
      Problem = new RegressionProblem();
    }


    protected override void Run(CancellationToken cancellationToken) {
      double[,] inputMatrix = Problem.ProblemData.Dataset.ToArray(Problem.ProblemData.AllowedInputVariables.Concat(new string[] { Problem.ProblemData.TargetVariable }), Problem.ProblemData.TrainingIndices);
      if (inputMatrix.Cast<double>().Any(x => double.IsNaN(x) || double.IsInfinity(x)))
        throw new NotSupportedException("Splines does not support NaN or infinity values in the input dataset.");

      var inputVars = Problem.ProblemData.AllowedInputVariables.ToArray();
      if (inputVars.Length > 3) throw new ArgumentException();

      var y = Problem.ProblemData.TargetVariableTrainingValues.ToArray();
      if (inputVars.Length == 1) {

        var x = Problem.ProblemData.Dataset.GetDoubleValues(inputVars.First(), Problem.ProblemData.TrainingIndices).ToArray();
        alglib.spline1dinterpolant c;
        var type = ((StringValue)Parameters["Type"].ActualValue).Value;
        switch (type) {
          case "Smoothing Spline (Reinsch)": {
              CalculateSmoothingSplineReinsch(x, y, inputVars);
              break;
            }
          case "Smoothing Spline (Reinsch with automatic tolerance determination using LOOCV)": {
              double optTol, looRMSE;
              var model = CalculateSmoothingSplineReinsch(x, y, inputVars, Problem.ProblemData.TargetVariable, out optTol, out looRMSE);
              Results.Add(new Result("Solution", model.CreateRegressionSolution((IRegressionProblemData)Problem.ProblemData.Clone())));
              Results.Add(new Result("Optimal tolerance", new DoubleValue(optTol)));
              Results.Add(new Result("RMSE (LOO-CV)", new DoubleValue(looRMSE)));
              break;
            }
          case "CUBGCV": {
              CubicSplineGCV.CubGcvReport report;
              var model =
                CubicSplineGCV.CalculateCubicSpline(x, y,
                Problem.ProblemData.TargetVariable, inputVars, out report);
              var targetVar = Problem.ProblemData.TargetVariable;
              var problemData = (IRegressionProblemData)Problem.ProblemData.Clone();
              Results.Add(new Result("Solution", model.CreateRegressionSolution(problemData)));
              Results.Add(new Result("GCV", new DoubleValue(report.generalizedCrossValidation)));
              Results.Add(new Result("Estimated error variance", new DoubleValue(report.estimatedErrorVariance)));
              Results.Add(new Result("Estimated RSS degrees of freedom", new DoubleValue(report.estimatedRSSDegreesOfFreedom)));
              Results.Add(new Result("Estimated true mean squared error at data points", new DoubleValue(report.estimatedTrueMeanSquaredErrorAtDataPoints)));
              Results.Add(new Result("Mean square of DF", new DoubleValue(report.meanSquareOfDf)));
              Results.Add(new Result("Mean square residual", new DoubleValue(report.meanSquareResiudal)));
              Results.Add(new Result("Smoothing parameter (optimized for GCV)", new DoubleValue(report.smoothingParameter)));
              break;
            }
          case "Finnbar O'Sullivan (sbart)": {
              SBART.SBART_Report report;
           
              var model =
                SBART.CalculateSBART(x, y,
                Problem.ProblemData.TargetVariable, inputVars, out report);
              var targetVar = Problem.ProblemData.TargetVariable;
              var problemData = (IRegressionProblemData)Problem.ProblemData.Clone();
              Results.Add(new Result("Solution", model.CreateRegressionSolution(problemData)));
              Results.Add(new Result("GCV", new DoubleValue(report.gcv)));
              Results.Add(new Result("Smoothing parameter (optimized for GCV)", new DoubleValue(report.smoothingParameter)));
              break;

            }
          case "Penalized Regression Spline (alglib)": {
              var lambda = ((IValueParameter<DoubleValue>)Parameters["Lambda"]).Value.Value; // controls extent of smoothing
              var model = CalculatePenalizedRegressionSpline(x, y, lambda, x.Length, Problem.ProblemData.TargetVariable, inputVars);
              var targetVar = Problem.ProblemData.TargetVariable;
              var problemData = (IRegressionProblemData)Problem.ProblemData.Clone();
              Results.Add(new Result("Solution", model.CreateRegressionSolution(problemData)));

              break;
            }
          default: throw new NotSupportedException();
        }

      }
    }


    public static IRegressionModel CalculatePenalizedRegressionSpline(double[] x, double[] y, double lambda, int numKnots, string targetVar, string[] inputVars) {
      int info;
      alglib.spline1dinterpolant interpolant;
      alglib.spline1dfitreport rep;

      alglib.spline1dfitpenalized(x, y, numKnots, lambda, out info, out interpolant, out rep);
      return new Spline1dModel(interpolant, targetVar, inputVars);
    }

    public static IRegressionEnsembleModel CalculatePenalizedRegressionSpline(double[] x, double[] y, double lambda, string targetVar, string[] inputVars,
      out double avgTrainRMSE,
      out double cvRMSE, out double[] predictions) {
      int info;
      alglib.spline1dinterpolant interpolant;
      alglib.spline1dfitreport rep;
      int n = x.Length;
      n = (int)Math.Max(50, 3 * Math.Sqrt(n));

      int folds = 10;
      int foldSize = x.Length <= 30 ? 1 : x.Length / folds;
      if (foldSize == 1) folds = x.Length;

      double[] x_loo = new double[(folds - 1) * foldSize];
      double[] y_loo = new double[(folds - 1) * foldSize];
      double[] w_loo = Enumerable.Repeat(1.0, x_loo.Length).ToArray();

      var models = new List<IRegressionModel>();
      var sumTrainSE = 0.0;
      int nTrain = 0;
      var sumTestSE = 0.0;
      int nTest = 0;

      predictions = new double[y.Length];

      for (int i = 0; i < folds; i++) {
        if (i < folds - 1) {
          //Console.WriteLine("Copy from {0} to {1} n = {2}", (i + 1) * foldSize, i * foldSize, (folds - i - 1) * foldSize);
          Array.Copy(x, (i + 1) * foldSize, x_loo, i * foldSize, (folds - i - 1) * foldSize);
          Array.Copy(y, (i + 1) * foldSize, y_loo, i * foldSize, (folds - i - 1) * foldSize);
        }
        alglib.spline1dfitpenalizedw(x_loo, y_loo, w_loo, n, lambda, out info, out interpolant, out rep);
        //Debug.Assert(y_loo.All(yi => yi != 0.0));
        //Debug.Assert(x_loo.All(xi => xi != 0.0));

        // training error
        for (int j = 0; j < x_loo.Length; j++) {
          var y_pred = alglib.spline1dcalc(interpolant, x[j]);
          double res = y[j] - y_pred;
          sumTrainSE += res * res;
          nTrain++;
        }
        // test
        for (int j = i * foldSize; j < (i + 1) * foldSize; j++) {
          predictions[j] = alglib.spline1dcalc(interpolant, x[j]);
          double res = y[j] - predictions[j];
          sumTestSE += res * res;
          nTest++;
        }

        models.Add(new Spline1dModel(interpolant, targetVar, inputVars));

        if (i < folds - 1) {
          //Console.WriteLine("Copy from {0} to {1} n = {2}", i * foldSize, i * foldSize, foldSize);
          // add current fold for next iteration
          Array.Copy(x, i * foldSize, x_loo, i * foldSize, foldSize);
          Array.Copy(y, i * foldSize, y_loo, i * foldSize, foldSize);
        }
      }

      cvRMSE = Math.Sqrt(sumTestSE / nTest);
      avgTrainRMSE = Math.Sqrt(sumTrainSE / nTrain);

      return new RegressionEnsembleModel(models);
    }

    private struct SplineData {
      public double a, b, c, d, x, y;
    };


    /*
     * The procedure Quincunx, takes as arguments
       the vectors u, v and w which are respectively the diagonal, the first supradiagonal
       and the second supradiagonal of the banded matrix. The vector on
       the LHS of the equation (80) is placed in q which contains the solution on the
       completion of the procedure.
    */
    private void Quincunx(int n, MyArray<double> u, MyArray<double> v, MyArray<double> w, MyArray<double> q) {
      // { factorisation}
      u[-1] = 0;
      u[0] = 0;
      for (int j = 1; j <= n - 1; j++) {
        u[j] = u[j] - u[j - 2] * Math.Pow(w[j - 2], 2) - u[j - 1] * Math.Pow(v[j - 1], 2);
        v[j] = (v[j] - u[j - 1] * v[j - 1] * w[j - 1]) / u[j];
        w[j] = w[j] / u[j];
      }
      // { forward substitution}
      for (int j = 1; j <= n - 1; j++) {
        q[j] = q[j] - v[j - 1] * q[j - 1] - w[j - 2] * q[j - 2];
      }
      for (int j = 1; j < n - 1; j++) {
        q[j] = q[j] / u[j];
      }
      // { back substitution}
      q[n + 1] = 0;
      q[n] = 0;
      for (int j = n - 1; j >= 1; j--) {
        q[j] = q[j] - v[j] * q[j + 1] - w[j] * q[j + 2];
      }
    }

    public void CalculateSmoothingSplineReinsch(double[] x, double[] y, string[] inputVars) {
      double s = x.Length + 1;
      double stdTol = ((IValueParameter<DoubleValue>)Parameters["StdDev (noise)"]).Value.Value; // controls extent of smoothing

      var model = CalculateSmoothingSplineReinsch(x, y, inputVars, stdTol, s, Problem.ProblemData.TargetVariable);

      Results.Add(new Result("Solution", model.CreateRegressionSolution((IRegressionProblemData)Problem.ProblemData.Clone())));
    }


    public static IRegressionEnsembleModel CalculateSmoothingSplineReinsch(double[] x, double[] y, double tol, string targetVar, string[] inputVars,
      out double avgTrainRMSE,
      out double cvRMSE) {
      int info;

      int folds = 10;
      int foldSize = x.Length <= 30 ? 1 : x.Length / folds;
      if (foldSize == 1) folds = x.Length;

      double[] x_loo = new double[(folds - 1) * foldSize];
      double[] y_loo = new double[(folds - 1) * foldSize];

      var models = new List<IRegressionModel>();
      var sumTrainSE = 0.0;
      int nTrain = 0;
      var sumTestSE = 0.0;
      int nTest = 0;


      for (int i = 0; i < folds; i++) {
        if (i < folds - 1) {
          Array.Copy(x, (i + 1) * foldSize, x_loo, i * foldSize, x.Length - ((i + 1) * foldSize) - 1);
          Array.Copy(y, (i + 1) * foldSize, y_loo, i * foldSize, y.Length - ((i + 1) * foldSize) - 1);
        }
        var model = CalculateSmoothingSplineReinsch(x_loo, y_loo, inputVars, tol, ((folds - 1) * foldSize) - 1, targetVar);

        // training error
        for(int j = 0;j<x_loo.Length;j++) {
          var y_pred = model.GetEstimatedValue(x[j]);
          double res = y[j] - y_pred;
          sumTrainSE += res * res;
          nTrain++;
        }
        // test
        for (int j = i * foldSize; j < (i + 1) * foldSize; j++) {
          var y_pred = model.GetEstimatedValue(x[j]);
          double res = y[j] - y_pred;
          sumTestSE += res * res;
          nTest++;
        }

        models.Add(model);

        if (i < folds - 1) {
          // add current fold for next iteration
          Array.Copy(x, i * foldSize, x_loo, i * foldSize, foldSize);
          Array.Copy(y, i * foldSize, y_loo, i * foldSize, foldSize);
        }
      }

      cvRMSE = Math.Sqrt(sumTestSE / nTest);
      avgTrainRMSE = Math.Sqrt(sumTrainSE / nTrain);

      return new RegressionEnsembleModel(models);
    }


    // automatically determines tolerance using cv, SLOW!
    public static IRegressionModel CalculateSmoothingSplineReinsch(double[] x, double[] y, string[] inputVars, string targetVar,
      out double optTolerance, out double cvRMSE) {
      var maxTolerance = y.StandardDeviation();
      var minTolerance = maxTolerance * 1E-6;
      optTolerance = maxTolerance;
      var tol = maxTolerance;
      cvRMSE = double.PositiveInfinity;
      IRegressionModel bestModel = new ConstantModel(y.Average(), targetVar);
      while (tol > minTolerance) {
        double curCVRMSE;
        double avgTrainRmse;
        var model = Splines.CalculateSmoothingSplineReinsch(
          x, y, tol, targetVar, inputVars, out avgTrainRmse, out curCVRMSE);

        if (curCVRMSE < cvRMSE) {
          cvRMSE = curCVRMSE;
          optTolerance = tol;
          bestModel = model;
        }
        tol *= 0.85;
      }
      return bestModel;
    }

    public static ReinschSmoothingSplineModel CalculateSmoothingSplineReinsch(double[] xOrig, double[] yOrig, string[] inputVars, double stdTol, double s, string targetVar) {
      var minX = xOrig.Min();
      var maxX = xOrig.Max();
      var range = maxX - minX;

      double[] w = Enumerable.Repeat(stdTol, xOrig.Length).ToArray();
      SortAndBin((double[])xOrig.Clone(), (double[])yOrig.Clone(), w, out xOrig, out yOrig, out w, scaling: false);

      // See Smoothing by Spline Functions, Reinsch, 1967
      // move x and y into an array indexed with 1..n to match indexing in Reinsch paper
      int n = xOrig.Length;
      var x = new MyArray<double>(1, xOrig);
      var y = new MyArray<double>(1, yOrig);
      var inv_dy = new MyArray<double>(1, w);

      int n1 = 1;
      int n2 = n;

      // results
      var a = new MyArray<double>(n1, n);
      var b = new MyArray<double>(n1, n);
      var c = new MyArray<double>(n1, n);
      var d = new MyArray<double>(n1, n);

      // smooth
      {
        int i, m1, m2; double e, f, f2, g = 0.0, h, p;
        MyArray<double>
          r = new MyArray<double>(n1 - 1, n + 2),
          r1 = new MyArray<double>(n1 - 1, n + 2),
          r2 = new MyArray<double>(n1 - 1, n + 2),
          t = new MyArray<double>(n1 - 1, n + 2),
          t1 = new MyArray<double>(n1 - 1, n + 2),
          u = new MyArray<double>(n1 - 1, n + 2),
          v = new MyArray<double>(n1 - 1, n + 2);
        m1 = n1 - 1;
        m2 = n2 + 1;
        r[m1] = r[n1] = r1[n2] = r2[n2] = r2[m2] =
          u[m1] = u[n1] = u[n2] = u[m2] = p = 0.0;
        m1 = n1 + 1;
        m2 = n2 - 1;
        h = x[m1] - x[n1]; f = (y[m1] - y[n1]) / h;
        for (i = m1; i <= m2; i++) {
          g = h; h = x[i + 1] - x[i];
          e = f; f = (y[i + 1] - y[i]) / h;
          a[i] = f - e; t[i] = 2 * (g + h) / 3; t1[i] = h / 3;
          r2[i] = inv_dy[i - 1] / g; r[i] = inv_dy[i + 1] / h;
          r1[i] = -inv_dy[i] / g - inv_dy[i] / h;
        }
        for (i = m1; i <= m2; i++) {
          b[i] = r[i] * r[i] + r1[i] * r1[i] + r2[i] * r2[i];
          c[i] = r[i] * r1[i + 1] + r1[i] * r2[i + 1];
          d[i] = r[i] * r2[i + 2];
        }
        f2 = -s;
        next:
        for (i = m1; i <= m2; i++) {
          r1[i - 1] = f * r[i - 1]; r2[i - 2] = g * r[i - 2];
          r[i] = 1 / (p * b[i] + t[i] - f * r1[i - 1] - g * r2[i - 2]);
          u[i] = a[i] - r1[i - 1] * u[i - 1] - r2[i - 2] * u[i - 2];
          f = p * c[i] + t1[i] - h * r1[i - 1]; g = h; h = d[i] * p;
        }
        for (i = m2; i >= m1; i--) {
          u[i] = r[i] * u[i] - r1[i] * u[i + 1] - r2[i] * u[i + 2];
        }
        e = h = 0;
        for (i = n1; i <= m2; i++) {
          g = h; h = (u[i + 1] - u[i]) / (x[i + 1] - x[i]);
          v[i] = (h - g) * inv_dy[i] * inv_dy[i]; e = e + v[i] * (h - g);
        }
        g = v[n2] = -h * inv_dy[n2] * inv_dy[n2]; e = e - g * h;
        g = f2; f2 = e * p * p;
        if (f2 >= s || f2 <= g) goto fin;
        f = 0; h = (v[m1] - v[n1]) / (x[m1] - x[n1]);
        for (i = m1; i <= m2; i++) {
          g = h; h = (v[i + 1] - v[i]) / (x[i + 1] - x[i]);
          g = h - g - r1[i - 1] * r[i - 1] - r2[i - 2] * r[i - 2];
          f = f + g * r[i] * g; r[i] = g;
        }
        h = e - p * f; if (h <= 0) goto fin;
        p = p + (s - f2) / ((Math.Sqrt(s / e) + p) * h); goto next;

        fin:
        for (i = n1; i <= n2; i++) {
          a[i] = y[i] - p * v[i];
          c[i] = u[i];
        }
        for (i = n1; i <= m2; i++) {
          h = x[i + 1] - x[i];
          d[i] = (c[i + 1] - c[i]) / (3 * h);
          b[i] = (a[i + 1] - a[i]) / h - (h * d[i] + c[i]) * h;
        }
      }

      // extrapolate for xx > x[n2]
      b[b.Length] = b[b.Length - 1];
      d[1] = 0;

      return new ReinschSmoothingSplineModel(a, b, c, d, x, targetVar, inputVars);
    }

    public static void SortAndBin(double[] x, double[] y, double[] w, out double[] x2, out double[] y2, out double[] w2, bool scaling = false) {
      var sortedIdx = Enumerable.Range(0, x.Length).ToArray();
      // sort by x
      Array.Sort(x, sortedIdx);

      var xl = new List<double>();
      var yl = new List<double>();
      var wl = new List<double>();

      int n = x.Length;
      var range = x[n - 1] - x[0];
      var binSize = range / n / 10;
      {
        // binning
        int i = 0;
        while (i < n) {
          int k = 0;
          int j = i;
          double sumX = 0.0;
          double sumY = 0.0;
          double sumW = 0.0;
          while (j < n && x[j] - x[i] <= binSize) {
            k++;
            sumX += x[j];
            sumY += y[sortedIdx[j]];
            sumW += w[sortedIdx[j]];
            j++;
          }
          var avgX = sumX / k;
          if (scaling) avgX = (avgX - x[0]) / range;
          xl.Add(avgX);
          yl.Add(sumY / k);
          wl.Add(sumW);
          i += k;
        }
      }

      x2 = xl.ToArray();
      y2 = yl.ToArray();
      w2 = wl.ToArray();
    }
  }


  // array with non-zero lower index
  internal class MyArray<T> {
    private T[] arr;
    private int lowerBound;

    public int Length { get { return arr.Length; } }

    public T this[int key] {
      get {
        return arr[key - lowerBound];
      }
      set {
        arr[key - lowerBound] = value;
      }
    }

    public MyArray(int lowerBound, int numElements) {
      this.lowerBound = lowerBound;
      arr = new T[numElements];
    }
    public MyArray(int lowerBound, T[] source) : this(lowerBound, source.Length) {
      Array.Copy(source, arr, source.Length);
    }

    public T[] ToArray() {
      var res = new T[arr.Length];
      Array.Copy(arr, res, res.Length);
      return res;
    }
  }


  // UNFINISHED
  public class ReinschSmoothingSplineModel : NamedItem, IRegressionModel {
    private MyArray<double> a;
    private MyArray<double> b;
    private MyArray<double> c;
    private MyArray<double> d;
    private MyArray<double> x;
    private double offset;
    private double scale;

    public string TargetVariable { get; set; }

    public IEnumerable<string> VariablesUsedForPrediction { get; private set; }

    public event EventHandler TargetVariableChanged;

    public ReinschSmoothingSplineModel(ReinschSmoothingSplineModel orig, Cloner cloner) : base(orig, cloner) {
      this.TargetVariable = orig.TargetVariable;
      this.VariablesUsedForPrediction = orig.VariablesUsedForPrediction.ToArray();
      this.a = orig.a;
      this.b = orig.b;
      this.c = orig.c;
      this.d = orig.d;
      this.x = orig.x;
      this.scale = orig.scale;
      this.offset = orig.offset;
    }
    internal ReinschSmoothingSplineModel(
      MyArray<double> a,
      MyArray<double> b,
      MyArray<double> c,
      MyArray<double> d,
      MyArray<double> x,
      string targetVar, string[] inputs, double offset = 0, double scale = 1) : base("SplineModel", "SplineModel") {
      this.a = a;
      this.b = b;
      this.c = c;
      this.d = d;
      this.x = x;
      this.TargetVariable = targetVar;
      this.VariablesUsedForPrediction = inputs;
      this.scale = scale;
      this.offset = offset;
    }

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

    public IRegressionSolution CreateRegressionSolution(IRegressionProblemData problemData) {
      return new RegressionSolution(this, (IRegressionProblemData)problemData.Clone());
    }

    public double GetEstimatedValue(double xx) {
      xx = (xx - offset) * scale;
      int n = a.Length;
      if (xx <= x[1]) {
        double h = xx - x[1];
        return a[1] + h * (b[1] + h * (c[1] + h * d[1]));
      } else if (xx >= x[n]) {
        double h = xx - x[n];
        return a[n] + h * (b[n] + h * (c[n] + h * d[n]));
      } else {
        // binary search
        int lower = 1;
        int upper = n;
        while (true) {
          if (upper < lower) throw new InvalidProgramException();
          int i = lower + (upper - lower) / 2;
          if (x[i] <= xx && xx < x[i + 1]) {
            double h = xx - x[i];
            return a[i] + h * (b[i] + h * (c[i] + h * d[i]));
          } else if (xx < x[i]) {
            upper = i - 1;
          } else {
            lower = i + 1;
          }
        }
      }
    }

    public IEnumerable<double> GetEstimatedValues(IDataset dataset, IEnumerable<int> rows) {
      int n = x.Length;
      var product = dataset.GetDoubleValues(VariablesUsedForPrediction.First(), rows).ToArray();
      foreach (var factor in VariablesUsedForPrediction.Skip(1)) {
        product = product.Zip(dataset.GetDoubleValues(factor, rows), (pi, fi) => pi * fi).ToArray();
      }
      foreach (var xx in product) {
        yield return GetEstimatedValue(xx);
      }
    }
  }

  // UNFINISHED
  public class Spline1dModel : NamedItem, IRegressionModel {
    alglib.spline1dinterpolant interpolant;

    public string TargetVariable { get; set; }

    public IEnumerable<string> VariablesUsedForPrediction { get; private set; }

    public event EventHandler TargetVariableChanged;

    public Spline1dModel(Spline1dModel orig, Cloner cloner) : base(orig, cloner) {
      this.TargetVariable = orig.TargetVariable;
      this.VariablesUsedForPrediction = orig.VariablesUsedForPrediction.ToArray();

      this.interpolant = (alglib.spline1dinterpolant)orig.interpolant.make_copy();
    }
    public Spline1dModel(
      alglib.spline1dinterpolant interpolant,
      string targetVar, string[] inputs, double offset = 0, double scale = 1) : base("SplineModel", "SplineModel") {
      this.interpolant = (alglib.spline1dinterpolant)interpolant.make_copy();
      this.TargetVariable = targetVar;
      this.VariablesUsedForPrediction = inputs;
    }

    public override IDeepCloneable Clone(Cloner cloner) {
      return new Spline1dModel(this, cloner);
    }

    public IRegressionSolution CreateRegressionSolution(IRegressionProblemData problemData) {
      return new RegressionSolution(this, (IRegressionProblemData)problemData.Clone());
    }

    public double GetEstimatedValue(double xx) {
      return alglib.spline1dcalc(interpolant, xx);
    }

    public IEnumerable<double> GetEstimatedValues(IDataset dataset, IEnumerable<int> rows) {
      var product = dataset.GetDoubleValues(VariablesUsedForPrediction.First(), rows).ToArray();
      foreach (var factor in VariablesUsedForPrediction.Skip(1)) {
        product = product.Zip(dataset.GetDoubleValues(factor, rows), (pi, fi) => pi * fi).ToArray();
      }
      foreach (var xx in product) {
        yield return GetEstimatedValue(xx);
      }
    }
  }
}