GSLNLSMinimizer.cxx

Go to the documentation of this file.

// @(#)root/mathmore:$Id$
// Author: L. Moneta Wed Dec 20 17:16:32 2006
 
/**********************************************************************
 *                                                                    *
 * Copyright (c) 2006  LCG ROOT Math Team, CERN/PH-SFT                *
 *                                                                    *
 *                                                                    *
 **********************************************************************/
 
// Implementation file for class GSLNLSMinimizer
 
#include "Math/GSLNLSMinimizer.h"
 
#include "Math/MinimTransformFunction.h"
#include "Math/MultiNumGradFunction.h"
#include "Math/GenAlgoOptions.h"
 
#include "Math/Error.h"
#include "GSLMultiFit.h"
#include "GSLMultiFit2.h"
#include "gsl/gsl_errno.h"
 
#include "Math/FitMethodFunction.h"
// #include "Math/Derivator.h"
 
#include <iostream>
#include <iomanip>
#include <cassert>
 
namespace ROOT {
 
namespace Math {
 
/// Internal class used by GSLNLSMinimizer to implement the transformation of the chi2
/// function used by GSL Non-linear Least-square fitting
/// The class is template on the FitMethodFunction type to support both gradient and non
/// gradient functions
template <class FMFunc>
class FitTransformFunction : public FMFunc {
 
public:
   FitTransformFunction(const FMFunc &f, std::unique_ptr<MinimTransformFunction> transFunc)
      : FMFunc(f.NDim(), f.NPoints()), fFunc(f), fTransform(std::move(transFunc)), fGrad(std::vector<double>(f.NDim()))
   {
      // constructor from a given FitMethodFunction and  Transformation object.
      // Ownership of the transformation object is passed to this class
   }
 
   virtual ~FitTransformFunction() {}
 
   // re-implement data element
   virtual double DataElement(const double *x, unsigned i, double *g = nullptr, double * = nullptr, bool = false) const
   {
      // transform from x internal to x external
      const double *xExt = fTransform->Transformation(x);
      if (g == nullptr)
         return fFunc.DataElement(xExt, i);
      // use gradient
      double val = fFunc.DataElement(xExt, i, &fGrad[0]);
      // transform gradient
      fTransform->GradientTransformation(x, &fGrad.front(), g);
      return val;
   }
 
   virtual IMultiGenFunction *Clone() const
   {
      // not supported
      return nullptr;
   }
 
   // dimension (this is number of free dimensions)
   virtual unsigned int NDim() const { return fTransform->NDim(); }
 
   unsigned int NTot() const { return fTransform->NTot(); }
 
   virtual typename FMFunc::Type_t Type() const { return fFunc.Type(); }
 
   // forward of transformation functions
   const double *Transformation(const double *x) const { return fTransform->Transformation(x); }
 
   /// inverse transformation (external -> internal)
   void InvTransformation(const double *xext, double *xint) const { fTransform->InvTransformation(xext, xint); }
 
   /// inverse transformation for steps (external -> internal) at external point x
   void InvStepTransformation(const double *x, const double *sext, double *sint) const
   {
      fTransform->InvStepTransformation(x, sext, sint);
   }
 
   /// transform gradient vector (external -> internal) at internal point x
   void GradientTransformation(const double *x, const double *gext, double *gint) const
   {
      fTransform->GradientTransformation(x, gext, gint);
   }
 
   void MatrixTransformation(const double *x, const double *cint, double *cext) const
   {
      fTransform->MatrixTransformation(x, cint, cext);
   }
 
private:
   // objects of this class are not meant for copying or assignment
   FitTransformFunction(const FitTransformFunction &rhs) = delete;
   FitTransformFunction &operator=(const FitTransformFunction &rhs) = delete;
 
   virtual double DoEval(const double *x) const { return fFunc(fTransform->Transformation(x)); }
 
   virtual double DoDerivative(const double * /* x */, unsigned int /*icoord*/) const
   {
      // not used
      throw std::runtime_error("FitTransformFunction::DoDerivative");
      return 0;
   }
 
   bool fOwnTransformation;
   const FMFunc &fFunc;                                // pointer to original fit method function
   std::unique_ptr<MinimTransformFunction> fTransform; // pointer to transformation function
   mutable std::vector<double> fGrad;                  // cached vector of gradient values
};
 
//________________________________________________________________________________
/**
    LSResidualFunc class description.
    Internal class used for accessing the residuals of the Least Square function
    and their derivatives which are estimated numerically using GSL numerical derivation.
    The class contains a pointer to the fit method function and an index specifying
    the i-th residual and wraps it in a multi-dim gradient function interface
    ROOT::Math::IGradientFunctionMultiDim.
    The class is used by ROOT::Math::GSLNLSMinimizer (GSL non linear least square fitter)
 
    @ingroup MultiMin
*/
template <class Func>
class LSResidualFunc : public IMultiGradFunction {
public:
   // default ctor (required by CINT)
   LSResidualFunc() : fIndex(0), fChi2(0) {}
 
   LSResidualFunc(const Func &func, unsigned int i) : fIndex(i), fChi2(&func) {}
 
   // copy ctor
   LSResidualFunc(const LSResidualFunc<Func> &rhs) : IMultiGenFunction(), IMultiGradFunction() { operator=(rhs); }
 
   // assignment
   LSResidualFunc<Func> &operator=(const LSResidualFunc<Func> &rhs)
   {
      fIndex = rhs.fIndex;
      fChi2 = rhs.fChi2;
      return *this;
   }
 
   IMultiGenFunction *Clone() const override { return new LSResidualFunc<Func>(*fChi2, fIndex); }
 
   unsigned int NDim() const override { return fChi2->NDim(); }
 
   void Gradient(const double *x, double *g) const override
   {
      double f0 = 0;
      FdF(x, f0, g);
   }
 
   bool IsLSType() const { return fChi2->Type() == fChi2->kLeastSquare; }
 
   void FdF(const double *x, double &f, double *g) const override { f = fChi2->DataElement(x, fIndex, g); }
 
private:
   double DoEval(const double *x) const override { return fChi2->DataElement(x, fIndex, nullptr); }
 
   double DoDerivative(const double * /* x */, unsigned int /* icoord */) const override
   {
      // this function should not be called by GSL
      throw std::runtime_error("LSRESidualFunc::DoDerivative");
      return 0;
   }
 
   unsigned int fIndex;
   const Func *fChi2;
};
 
int GetTypeFromName(const char *name)
{
   std::string tName(name);
   if (tName.empty())
      return 0;
   if (tName == "lms_old")
      return 1;
   if (tName == "lm_old")
      return 2;
   if (tName == "trust")
      return 3;
   if (tName == "trust_lm")
      return 4;
   if (tName == "trust_lmaccel")
      return 5;
   if (tName == "trust_dogleg")
      return 6;
   if (tName == "trust_ddogleg")
      return 7;
   if (tName == "trust_subspace2D" || tName == "trust_2D")
      return 8;
   return 0;
}
 
// GSLNLSMinimizer implementation
GSLNLSMinimizer::GSLNLSMinimizer(const char *name) : GSLNLSMinimizer(GetTypeFromName(name)) {}
 
GSLNLSMinimizer::GSLNLSMinimizer(int type)
{
   // Constructor implementation : create GSLMultiFit wrapper object
   const gsl_multifit_fdfsolver_type *gsl_old_type = nullptr; // use default type defined in GSLMultiFit
   if (type == 1)
      gsl_old_type = gsl_multifit_fdfsolver_lmsder; // scaled lmder version
   if (type == 2)
      gsl_old_type = gsl_multifit_fdfsolver_lmder; // unscaled version
 
   // const gsl_multifit_nlinear_type * gsl_new_type = nullptr; //
   // if (type == 3) gsl_new_type =   gsl_multifit_nlinear_trust; // trust region default
 
   if (gsl_old_type)
      fGSLMultiFit = new GSLMultiFit(gsl_old_type);
   else
      fGSLMultiFit2 = new GSLMultiFit2(type - 3);
 
   fEdm = -1;
 
   // default tolerance and max iterations
   int niter = ROOT::Math::MinimizerOptions::DefaultMaxIterations();
   if (niter <= 0)
      niter = 100;
   SetMaxIterations(niter);
 
   fLSTolerance = ROOT::Math::MinimizerOptions::DefaultTolerance();
   if (fLSTolerance <= 0)
      fLSTolerance = 0.0001; // default internal value
 
   SetPrintLevel(ROOT::Math::MinimizerOptions::DefaultPrintLevel());
 
   // set the default options
   if (fGSLMultiFit2) {
      fOptions.SetExtraOptions(fGSLMultiFit2->GetDefaultOptions());
      if (type == 0 || type == 3)
         fOptions.SetMinimizerAlgorithm("trust_lm");
 
      fOptions.ExtraOptions()->SetValue("scale", "marquardt");
   }
}
 
GSLNLSMinimizer::~GSLNLSMinimizer()
{
   if (fGSLMultiFit)
      delete fGSLMultiFit;
   if (fGSLMultiFit2)
      delete fGSLMultiFit2;
}
 
void GSLNLSMinimizer::SetFunction(const ROOT::Math::IMultiGenFunction &func)
{
   // set the function to minimizer
   // need to create vector of functions to be passed to GSL multifit
 
   // call base class method. It will clone the function and set number of dimensions
   BasicMinimizer::SetFunction(func);
   fNFree = NDim();
   fUseGradFunction = func.HasGradient();
}
 
bool GSLNLSMinimizer::Minimize()
{
 
   if (ObjFunction() == nullptr) {
      MATH_ERROR_MSG("GSLNLSMinimizer::Minimize", "Function has not been  set");
      return false;
   }
   // check type of function (if it provides gradient)
   auto fitFunc = (!fUseGradFunction) ? dynamic_cast<const ROOT::Math::FitMethodFunction *>(ObjFunction()) : nullptr;
   auto fitGradFunc =
      (fUseGradFunction) ? dynamic_cast<const ROOT::Math::FitMethodGradFunction *>(ObjFunction()) : nullptr;
   if (fitFunc == nullptr && fitGradFunc == nullptr) {
      if (PrintLevel() > 0)
         std::cout << "GSLNLSMinimizer: Invalid function set - only FitMethodFunction types are supported" << std::endl;
      return false;
   }
 
   if (fGSLMultiFit) {
 
      if (PrintLevel() > 0)
         std::cout << "GLSNLSMinimizer::Minimize - Using old GSLMultiFit with method " << fOptions.MinimizerAlgorithm()
                << std::endl;
 
      if (fitGradFunc)
         return DoMinimize<ROOT::Math::FitMethodGradFunction, GSLMultiFit>(*fitGradFunc, fGSLMultiFit);
      else
         return DoMinimize<ROOT::Math::FitMethodFunction, GSLMultiFit>(*fitFunc, fGSLMultiFit);
   }
   if (fGSLMultiFit2) {
 
      // set specific minimizer parameters
      fGSLMultiFit2->SetParameters(fOptions);
 
      if (PrintLevel() > 0)
         std::cout << "GLSNLSMinimizer::Minimize - Using new GSLMultiFit with trs method " << fOptions.MinimizerAlgorithm()
                << std::endl;
 
      if (fitGradFunc)
         return DoMinimize<ROOT::Math::FitMethodGradFunction, GSLMultiFit2>(*fitGradFunc, fGSLMultiFit2);
      else
         return DoMinimize<ROOT::Math::FitMethodFunction, GSLMultiFit2>(*fitFunc, fGSLMultiFit2);
   }
   return false;
}
 
template <class Func, class FitterType>
bool GSLNLSMinimizer::DoMinimize(const Func &fitFunc, FitterType *fitter)
{
 
   unsigned int size = fitFunc.NPoints();
   fNCalls = 0; // reset number of function calls
 
   std::vector<LSResidualFunc<Func>> residualFuncs;
   residualFuncs.reserve(size);
 
   // set initial parameters of the minimizer
   int debugLevel = PrintLevel();
 
   unsigned int npar = NPar();
   unsigned int ndim = NDim();
   if (npar == 0 || npar < ndim) {
      MATH_ERROR_MSGVAL("GSLNLSMinimizer::Minimize", "Wrong number of parameters", npar);
      return false;
   }
 
   // set residual functions and check if a transformation is needed
   std::vector<double> startValues;
 
   // transformation need a grad function.
   std::unique_ptr<MultiNumGradFunction> gradFunction;
   std::unique_ptr<MinimTransformFunction> trFuncRaw;
   if (!fUseGradFunction) {
      gradFunction = std::make_unique<MultiNumGradFunction>(fitFunc);
      trFuncRaw.reset(CreateTransformation(startValues, gradFunction.get()));
   } else {
      // use pointer stored in BasicMinimizer
      trFuncRaw.reset(CreateTransformation(startValues));
   }
   // need to transform in a FitTransformFunction which is set in the residual functions
   std::unique_ptr<FitTransformFunction<Func>> trFunc;
   if (trFuncRaw) {
      // pass ownership of trFuncRaw to FitTransformFunction
      trFunc = std::make_unique<FitTransformFunction<Func>>(fitFunc, std::move(trFuncRaw));
      assert(npar == trFunc->NTot());
      for (unsigned int ires = 0; ires < size; ++ires) {
         residualFuncs.emplace_back(LSResidualFunc<Func>(*trFunc, ires));
      }
   } else {
      for (unsigned int ires = 0; ires < size; ++ires) {
         residualFuncs.emplace_back(LSResidualFunc<Func>(fitFunc, ires));
      }
   }
 
   if (debugLevel >= 1)
      std::cout << "Minimize using GSLNLSMinimizer " << std::endl;
 
   int iret = fitter->Set(residualFuncs, &startValues.front());
 
   if (iret) {
      MATH_ERROR_MSGVAL("GSLNLSMinimizer::Minimize", "Error setting the residual functions ", iret);
      return false;
   }
 
   int status = 0;
   bool minFound = false;
   unsigned int iter = 0;
   if (fGSLMultiFit) {
      // case of using old solver
 
      if (debugLevel >= 1)
         std::cout << "GSLNLSMinimizer: " << fGSLMultiFit->Name() << " - start iterating......... " << std::endl;
 
      // start iteration
      do {
         status = fitter->Iterate();
 
         if (debugLevel >= 1) {
            std::cout << "----------> Iteration " << iter << " / " << MaxIterations() << " status "
                      << gsl_strerror(status) << std::endl;
            const double *x = fitter->X();
            if (trFunc)
               x = trFunc->Transformation(x);
            int pr = std::cout.precision(18);
            std::cout << "            FVAL = " << (fitFunc)(x) << std::endl;
            std::cout.precision(pr);
            std::cout << "            X Values : ";
            for (unsigned int i = 0; i < NDim(); ++i)
               std::cout << " " << VariableName(i) << " = " << x[i];
            std::cout << std::endl;
         }
 
         if (status)
            break;
 
         // check also the delta in X()
         status = fitter->TestDelta(Tolerance(), Tolerance());
         if (status == GSL_SUCCESS) {
            minFound = true;
         }
 
         // double-check with the gradient
         int status2 = fitter->TestGradient(Tolerance());
         if (minFound && status2 != GSL_SUCCESS) {
            // check now edm
            fEdm = fitter->Edm();
            if (fEdm > Tolerance()) {
               // continue the iteration
               status = status2;
               minFound = false;
            }
         }
 
         if (debugLevel >= 1) {
            std::cout << "          after Gradient and Delta tests:  " << gsl_strerror(status);
            if (fEdm > 0)
               std::cout << ", edm is:  " << fEdm;
            std::cout << std::endl;
         }
 
         iter++;
 
      } while (status == GSL_CONTINUE && iter < MaxIterations());
 
      // check edm
      fEdm = fitter->Edm();
      if (fEdm < Tolerance()) {
         minFound = true;
      }
 
   } else if (fGSLMultiFit2) {
      // case using new solver and given driver
      status = fGSLMultiFit2->Solve();
      if (status == GSL_SUCCESS)
         minFound = true;
      iter = fGSLMultiFit2->NIter();
   }
 
   // save state with values and function value
   const double *x = fitter->X();
   if (x == nullptr)
      return false;
   // apply transformation outside SetFinalValues(..)
   // because trFunc is not a MinimTransformFunction but a FitTransFormFunction
   if (trFunc)
      x = trFunc->Transformation(x);
   SetFinalValues(x);
 
   SetMinValue((fitFunc)(x));
   fStatus = status;
   fNCalls = fitFunc.NCalls();
   fErrors.resize(NDim());
 
   // get errors from cov matrix
   const double *cov = fitter->CovarMatrix();
   if (cov) {
 
      fCovMatrix.resize(ndim * ndim);
 
      if (trFunc) {
         trFunc->MatrixTransformation(x, fitter->CovarMatrix(), fCovMatrix.data());
      } else {
         std::copy(cov, cov + fCovMatrix.size(), fCovMatrix.begin());
      }
 
      for (unsigned int i = 0; i < ndim; ++i)
         fErrors[i] = std::sqrt(fCovMatrix[i * ndim + i]);
   }
 
   if (minFound) {
 
      if (debugLevel >= 1) {
         std::cout << "GSLNLSMinimizer: Minimum Found" << std::endl;
         int pr = std::cout.precision(18);
         std::cout << "FVAL         = " << MinValue() << std::endl;
         std::cout << "Edm          = " << fEdm << std::endl;
         std::cout.precision(pr);
         std::cout << "NIterations  = " << iter << std::endl;
         std::cout << "NFuncCalls   = " << fitFunc.NCalls() << std::endl;
         for (unsigned int i = 0; i < NDim(); ++i)
            std::cout << std::setw(12) << VariableName(i) << " = " << std::setw(12) << X()[i] << "   +/-   "
                      << std::setw(12) << fErrors[i] << std::endl;
      }
 
      return true;
   } else {
      if (debugLevel >= 0) {
         std::cout << "GSLNLSMinimizer: Minimization did not converge: " << std::endl;
         if (status == GSL_ENOPROG) // case status 27
            std::cout << "\t iteration is not making progress towards solution" << std::endl;
         else
            std::cout << "\t failed with status " << status << std::endl;
      }
      if (debugLevel >= 1) {
         std::cout << "FVAL         = " << MinValue() << std::endl;
         std::cout << "Edm   = " << fitter->Edm() << std::endl;
         std::cout << "Niterations  = " << iter << std::endl;
      }
      return false;
   }
   return false;
}
 
const double *GSLNLSMinimizer::MinGradient() const
{
   // return gradient (internal values) - only when using old fitter
   return (fGSLMultiFit) ? fGSLMultiFit->Gradient() : nullptr;
}
 
double GSLNLSMinimizer::CovMatrix(unsigned int i, unsigned int j) const
{
   // return covariance matrix element
   unsigned int ndim = NDim();
   if (fCovMatrix.empty())
      return 0;
   if (i > ndim || j > ndim)
      return 0;
   return fCovMatrix[i * ndim + j];
}
 
int GSLNLSMinimizer::CovMatrixStatus() const
{
   // return covariance  matrix status = 0 not computed,
   // 1 computed but is approximate because minimum is not valid, 3 is fine
   if (fCovMatrix.empty())
      return 0;
   // case minimization did not finished correctly
   if (fStatus != GSL_SUCCESS)
      return 1;
   return 3;
}
 
} // end namespace Math
 
} // end namespace ROOT