bfgs.h

#pragma once
#include <pcl/pcl_macros.h>
 
#if defined __GNUC__
#pragma GCC system_header
#endif
 
#include <unsupported/Eigen/Polynomials> // for PolynomialSolver, PolynomialSolverBase
 
namespace Eigen {
template <typename _Scalar>
class PolynomialSolver<_Scalar, 2> : public PolynomialSolverBase<_Scalar, 2> {
public:
  using PS_Base = PolynomialSolverBase<_Scalar, 2>;
  EIGEN_POLYNOMIAL_SOLVER_BASE_INHERITED_TYPES(PS_Base)
 
public:
  virtual ~PolynomialSolver() {}
 
  template <typename OtherPolynomial>
  inline PolynomialSolver(const OtherPolynomial& poly, bool& hasRealRoot)
  {
    compute(poly, hasRealRoot);
  }
 
  /** Computes the complex roots of a new polynomial. */
  template <typename OtherPolynomial>
  void
  compute(const OtherPolynomial& poly, bool& hasRealRoot)
  {
    const Scalar ZERO(0);
    Scalar a2(2 * poly[2]);
    assert(ZERO != poly[poly.size() - 1]);
    Scalar discriminant((poly[1] * poly[1]) - (4 * poly[0] * poly[2]));
    if (ZERO < discriminant) {
      Scalar discriminant_root(std::sqrt(discriminant));
      m_roots[0] = (-poly[1] - discriminant_root) / (a2);
      m_roots[1] = (-poly[1] + discriminant_root) / (a2);
      hasRealRoot = true;
    }
    else {
      if (ZERO == discriminant) {
        m_roots.resize(1);
        m_roots[0] = -poly[1] / a2;
        hasRealRoot = true;
      }
      else {
        Scalar discriminant_root(std::sqrt(-discriminant));
        m_roots[0] = RootType(-poly[1] / a2, -discriminant_root / a2);
        m_roots[1] = RootType(-poly[1] / a2, discriminant_root / a2);
        hasRealRoot = false;
      }
    }
  }
 
  template <typename OtherPolynomial>
  void
  compute(const OtherPolynomial& poly)
  {
    bool hasRealRoot;
    compute(poly, hasRealRoot);
  }
 
protected:
  using PS_Base::m_roots;
};
} // namespace Eigen
 
namespace BFGSSpace {
enum Status {
  NegativeGradientEpsilon = -3,
  NotStarted = -2,
  Running = -1,
  Success = 0,
  NoProgress = 1
};
}
 
template <typename _Scalar, int NX = Eigen::Dynamic>
struct BFGSDummyFunctor {
  using Scalar = _Scalar;
  enum { InputsAtCompileTime = NX };
  using VectorType = Eigen::Matrix<Scalar, InputsAtCompileTime, 1>;
 
  const int m_inputs;
 
  BFGSDummyFunctor() : m_inputs(InputsAtCompileTime) {}
  BFGSDummyFunctor(int inputs) : m_inputs(inputs) {}
 
  virtual ~BFGSDummyFunctor() {}
  int
  inputs() const
  {
    return m_inputs;
  }
 
  virtual double
  operator()(const VectorType& x) = 0;
  virtual void
  df(const VectorType& x, VectorType& df) = 0;
  virtual void
  fdf(const VectorType& x, Scalar& f, VectorType& df) = 0;
  virtual BFGSSpace::Status
  checkGradient(const VectorType& g)
  {
    return BFGSSpace::NotStarted;
  };
};
 
/**
 * BFGS stands for Broyden–Fletcher–Goldfarb–Shanno (BFGS) method for solving
 * unconstrained nonlinear optimization problems.
 * For further details please visit: http://en.wikipedia.org/wiki/BFGS_method
 * The method provided here is almost similar to the one provided by GSL.
 * It reproduces Fletcher's original algorithm in Practical Methods of Optimization
 * algorithms : 2.6.2 and 2.6.4 and uses the same politics in GSL with cubic
 * interpolation whenever it is possible else falls to quadratic interpolation for
 * alpha parameter.
 */
template <typename FunctorType>
class BFGS {
public:
  using Scalar = typename FunctorType::Scalar;
  using FVectorType = typename FunctorType::VectorType;
 
  BFGS(FunctorType& _functor) : pnorm(0), g0norm(0), iter(-1), functor(_functor) {}
 
  using Index = Eigen::DenseIndex;
 
  struct Parameters {
    Parameters()
    : max_iters(400)
    , bracket_iters(100)
    , section_iters(100)
    , rho(0.01)
    , sigma(0.01)
    , tau1(9)
    , tau2(0.05)
    , tau3(0.5)
    , step_size(1)
    , order(3)
    {}
    Index max_iters; // maximum number of function evaluation
    Index bracket_iters;
    Index section_iters;
    Scalar rho;
    Scalar sigma;
    Scalar tau1;
    Scalar tau2;
    Scalar tau3;
    Scalar step_size;
    Index order;
  };
 
  BFGSSpace::Status
  minimize(FVectorType& x);
  BFGSSpace::Status
  minimizeInit(FVectorType& x);
  BFGSSpace::Status
  minimizeOneStep(FVectorType& x);
  BFGSSpace::Status
  testGradient();
  PCL_DEPRECATED(1, 13, "Use `testGradient()` instead")
  BFGSSpace::Status testGradient(Scalar) { return testGradient(); }
  void
  resetParameters(void)
  {
    parameters = Parameters();
  }
 
  Parameters parameters;
  Scalar f;
  FVectorType gradient;
 
private:
  BFGS&
  operator=(const BFGS&);
  BFGSSpace::Status
  lineSearch(Scalar rho,
             Scalar sigma,
             Scalar tau1,
             Scalar tau2,
             Scalar tau3,
             int order,
             Scalar alpha1,
             Scalar& alpha_new);
  Scalar
  interpolate(Scalar a,
              Scalar fa,
              Scalar fpa,
              Scalar b,
              Scalar fb,
              Scalar fpb,
              Scalar xmin,
              Scalar xmax,
              int order);
  void
  checkExtremum(const Eigen::Matrix<Scalar, 4, 1>& coefficients,
                Scalar x,
                Scalar& xmin,
                Scalar& fmin);
  void
  moveTo(Scalar alpha);
  Scalar
  slope();
  Scalar
  applyF(Scalar alpha);
  Scalar
  applyDF(Scalar alpha);
  void
  applyFDF(Scalar alpha, Scalar& f, Scalar& df);
  void
  updatePosition(Scalar alpha, FVectorType& x, Scalar& f, FVectorType& g);
  void
  changeDirection();
 
  Scalar delta_f, fp0;
  FVectorType x0, dx0, dg0, g0, dx, p;
  Scalar pnorm, g0norm;
 
  Scalar f_alpha;
  Scalar df_alpha;
  FVectorType x_alpha;
  FVectorType g_alpha;
 
  // cache "keys"
  Scalar f_cache_key;
  Scalar df_cache_key;
  Scalar x_cache_key;
  Scalar g_cache_key;
 
  Index iter;
  FunctorType& functor;
};
 
template <typename FunctorType>
void
BFGS<FunctorType>::checkExtremum(const Eigen::Matrix<Scalar, 4, 1>& coefficients,
                                 Scalar x,
                                 Scalar& xmin,
                                 Scalar& fmin)
{
  Scalar y = Eigen::poly_eval(coefficients, x);
  if (y < fmin) {
    xmin = x;
    fmin = y;
  }
}
 
template <typename FunctorType>
void
BFGS<FunctorType>::moveTo(Scalar alpha)
{
  x_alpha = x0 + alpha * p;
  x_cache_key = alpha;
}
 
template <typename FunctorType>
typename BFGS<FunctorType>::Scalar
BFGS<FunctorType>::slope()
{
  return (g_alpha.dot(p));
}
 
template <typename FunctorType>
typename BFGS<FunctorType>::Scalar
BFGS<FunctorType>::applyF(Scalar alpha)
{
  if (alpha == f_cache_key)
    return f_alpha;
  moveTo(alpha);
  f_alpha = functor(x_alpha);
  f_cache_key = alpha;
  return (f_alpha);
}
 
template <typename FunctorType>
typename BFGS<FunctorType>::Scalar
BFGS<FunctorType>::applyDF(Scalar alpha)
{
  if (alpha == df_cache_key)
    return df_alpha;
  moveTo(alpha);
  if (alpha != g_cache_key) {
    functor.df(x_alpha, g_alpha);
    g_cache_key = alpha;
  }
  df_alpha = slope();
  df_cache_key = alpha;
  return (df_alpha);
}
 
template <typename FunctorType>
void
BFGS<FunctorType>::applyFDF(Scalar alpha, Scalar& f, Scalar& df)
{
  if (alpha == f_cache_key && alpha == df_cache_key) {
    f = f_alpha;
    df = df_alpha;
    return;
  }
 
  if (alpha == f_cache_key || alpha == df_cache_key) {
    f = applyF(alpha);
    df = applyDF(alpha);
    return;
  }
 
  moveTo(alpha);
  functor.fdf(x_alpha, f_alpha, g_alpha);
  f_cache_key = alpha;
  g_cache_key = alpha;
  df_alpha = slope();
  df_cache_key = alpha;
  f = f_alpha;
  df = df_alpha;
}
 
template <typename FunctorType>
void
BFGS<FunctorType>::updatePosition(Scalar alpha,
                                  FVectorType& x,
                                  Scalar& f,
                                  FVectorType& g)
{
  {
    Scalar f_alpha, df_alpha;
    applyFDF(alpha, f_alpha, df_alpha);
  };
 
  f = f_alpha;
  x = x_alpha;
  g = g_alpha;
}
 
template <typename FunctorType>
void
BFGS<FunctorType>::changeDirection()
{
  x_alpha = x0;
  x_cache_key = 0.0;
  f_cache_key = 0.0;
  g_alpha = g0;
  g_cache_key = 0.0;
  df_alpha = slope();
  df_cache_key = 0.0;
}
 
template <typename FunctorType>
BFGSSpace::Status
BFGS<FunctorType>::minimize(FVectorType& x)
{
  BFGSSpace::Status status = minimizeInit(x);
  do {
    status = minimizeOneStep(x);
    iter++;
  } while (status == BFGSSpace::Success && iter < parameters.max_iters);
  return status;
}
 
template <typename FunctorType>
BFGSSpace::Status
BFGS<FunctorType>::minimizeInit(FVectorType& x)
{
  iter = 0;
  delta_f = 0;
  dx.setZero();
  functor.fdf(x, f, gradient);
  x0 = x;
  g0 = gradient;
  g0norm = g0.norm();
  p = gradient * -1 / g0norm;
  pnorm = p.norm();
  fp0 = -g0norm;
 
  {
    x_alpha = x0;
    x_cache_key = 0;
 
    f_alpha = f;
    f_cache_key = 0;
 
    g_alpha = g0;
    g_cache_key = 0;
 
    df_alpha = slope();
    df_cache_key = 0;
  }
 
  return BFGSSpace::NotStarted;
}
 
template <typename FunctorType>
BFGSSpace::Status
BFGS<FunctorType>::minimizeOneStep(FVectorType& x)
{
  Scalar alpha = 0.0, alpha1;
  Scalar f0 = f;
  if (pnorm == 0.0 || g0norm == 0.0 || fp0 == 0) {
    dx.setZero();
    return BFGSSpace::NoProgress;
  }
 
  if (delta_f < 0) {
    Scalar del =
        std::max(-delta_f, 10 * std::numeric_limits<Scalar>::epsilon() * std::abs(f0));
    alpha1 = std::min(1.0, 2.0 * del / (-fp0));
  }
  else
    alpha1 = std::abs(parameters.step_size);
 
  BFGSSpace::Status status = lineSearch(parameters.rho,
                                        parameters.sigma,
                                        parameters.tau1,
                                        parameters.tau2,
                                        parameters.tau3,
                                        parameters.order,
                                        alpha1,
                                        alpha);
 
  if (status != BFGSSpace::Success)
    return status;
 
  updatePosition(alpha, x, f, gradient);
 
  delta_f = f - f0;
 
  /* Choose a new direction for the next step */
  {
    /* This is the BFGS update: */
    /* p' = g1 - A dx - B dg */
    /* A = - (1+ dg.dg/dx.dg) B + dg.g/dx.dg */
    /* B = dx.g/dx.dg */
 
    Scalar dxg, dgg, dxdg, dgnorm, A, B;
 
    /* dx0 = x - x0 */
    dx0 = x - x0;
    dx = dx0; /* keep a copy */
 
    /* dg0 = g - g0 */
    dg0 = gradient - g0;
    dxg = dx0.dot(gradient);
    dgg = dg0.dot(gradient);
    dxdg = dx0.dot(dg0);
    dgnorm = dg0.norm();
 
    if (dxdg != 0) {
      B = dxg / dxdg;
      A = -(1.0 + dgnorm * dgnorm / dxdg) * B + dgg / dxdg;
    }
    else {
      B = 0;
      A = 0;
    }
 
    p = -A * dx0;
    p += gradient;
    p += -B * dg0;
  }
 
  g0 = gradient;
  x0 = x;
  g0norm = g0.norm();
  pnorm = p.norm();
 
  Scalar dir = ((p.dot(gradient)) > 0) ? -1.0 : 1.0;
  p *= dir / pnorm;
  pnorm = p.norm();
  fp0 = p.dot(g0);
 
  changeDirection();
  return BFGSSpace::Success;
}
 
template <typename FunctorType>
typename BFGSSpace::Status
BFGS<FunctorType>::testGradient()
{
  return functor.checkGradient(gradient);
}
 
template <typename FunctorType>
typename BFGS<FunctorType>::Scalar
BFGS<FunctorType>::interpolate(Scalar a,
                               Scalar fa,
                               Scalar fpa,
                               Scalar b,
                               Scalar fb,
                               Scalar fpb,
                               Scalar xmin,
                               Scalar xmax,
                               int order)
{
  /* Map [a,b] to [0,1] */
  Scalar y, alpha, ymin, ymax, fmin;
 
  ymin = (xmin - a) / (b - a);
  ymax = (xmax - a) / (b - a);
 
  // Ensure ymin <= ymax
  if (ymin > ymax) {
    Scalar tmp = ymin;
    ymin = ymax;
    ymax = tmp;
  };
 
  if (order > 2 && !(fpb != fpa) && fpb != std::numeric_limits<Scalar>::infinity()) {
    fpa = fpa * (b - a);
    fpb = fpb * (b - a);
 
    Scalar eta = 3 * (fb - fa) - 2 * fpa - fpb;
    Scalar xi = fpa + fpb - 2 * (fb - fa);
    Scalar c0 = fa, c1 = fpa, c2 = eta, c3 = xi;
    Scalar y0, y1;
    Eigen::Matrix<Scalar, 4, 1> coefficients;
    coefficients << c0, c1, c2, c3;
 
    y = ymin;
    // Evaluate the cubic polyinomial at ymin;
    fmin = Eigen::poly_eval(coefficients, ymin);
    checkExtremum(coefficients, ymax, y, fmin);
    {
      // Solve quadratic polynomial for the derivate
      Eigen::Matrix<Scalar, 3, 1> coefficients2;
      coefficients2 << c1, 2 * c2, 3 * c3;
      bool real_roots;
      Eigen::PolynomialSolver<Scalar, 2> solver(coefficients2, real_roots);
      if (real_roots) {
        if ((solver.roots()).size() == 2) /* found 2 roots */
        {
          y0 = std::real(solver.roots()[0]);
          y1 = std::real(solver.roots()[1]);
          if (y0 > y1) {
            Scalar tmp(y0);
            y0 = y1;
            y1 = tmp;
          }
          if (y0 > ymin && y0 < ymax)
            checkExtremum(coefficients, y0, y, fmin);
          if (y1 > ymin && y1 < ymax)
            checkExtremum(coefficients, y1, y, fmin);
        }
        else if ((solver.roots()).size() == 1) /* found 1 root */
        {
          y0 = std::real(solver.roots()[0]);
          if (y0 > ymin && y0 < ymax)
            checkExtremum(coefficients, y0, y, fmin);
        }
      }
    }
  }
  else {
    fpa = fpa * (b - a);
    Scalar fl = fa + ymin * (fpa + ymin * (fb - fa - fpa));
    Scalar fh = fa + ymax * (fpa + ymax * (fb - fa - fpa));
    Scalar c = 2 * (fb - fa - fpa); /* curvature */
    y = ymin;
    fmin = fl;
 
    if (fh < fmin) {
      y = ymax;
      fmin = fh;
    }
 
    if (c > a) /* positive curvature required for a minimum */
    {
      Scalar z = -fpa / c; /* location of minimum */
      if (z > ymin && z < ymax) {
        Scalar f = fa + z * (fpa + z * (fb - fa - fpa));
        if (f < fmin) {
          y = z;
          fmin = f;
        };
      }
    }
  }
 
  alpha = a + y * (b - a);
  return alpha;
}
 
template <typename FunctorType>
BFGSSpace::Status
BFGS<FunctorType>::lineSearch(Scalar rho,
                              Scalar sigma,
                              Scalar tau1,
                              Scalar tau2,
                              Scalar tau3,
                              int order,
                              Scalar alpha1,
                              Scalar& alpha_new)
{
  Scalar f0, fp0, falpha, falpha_prev, fpalpha, fpalpha_prev, delta, alpha_next;
  Scalar alpha = alpha1, alpha_prev = 0.0;
  Scalar a, b, fa, fb, fpa, fpb;
  Index i = 0;
 
  applyFDF(0.0, f0, fp0);
 
  falpha_prev = f0;
  fpalpha_prev = fp0;
 
  /* Avoid uninitialized variables morning */
  a = 0.0;
  b = alpha;
  fa = f0;
  fb = 0.0;
  fpa = fp0;
  fpb = 0.0;
 
  /* Begin bracketing */
 
  while (i++ < parameters.bracket_iters) {
    falpha = applyF(alpha);
 
    if (falpha > f0 + alpha * rho * fp0 || falpha >= falpha_prev) {
      a = alpha_prev;
      fa = falpha_prev;
      fpa = fpalpha_prev;
      b = alpha;
      fb = falpha;
      fpb = std::numeric_limits<Scalar>::quiet_NaN();
      break;
    }
 
    fpalpha = applyDF(alpha);
 
    /* Fletcher's sigma test */
    if (std::abs(fpalpha) <= -sigma * fp0) {
      alpha_new = alpha;
      return BFGSSpace::Success;
    }
 
    if (fpalpha >= 0) {
      a = alpha;
      fa = falpha;
      fpa = fpalpha;
      b = alpha_prev;
      fb = falpha_prev;
      fpb = fpalpha_prev;
      break; /* goto sectioning */
    }
 
    delta = alpha - alpha_prev;
 
    {
      Scalar lower = alpha + delta;
      Scalar upper = alpha + tau1 * delta;
 
      alpha_next = interpolate(alpha_prev,
                               falpha_prev,
                               fpalpha_prev,
                               alpha,
                               falpha,
                               fpalpha,
                               lower,
                               upper,
                               order);
    }
 
    alpha_prev = alpha;
    falpha_prev = falpha;
    fpalpha_prev = fpalpha;
    alpha = alpha_next;
  }
  /*  Sectioning of bracket [a,b] */
  while (i++ < parameters.section_iters) {
    delta = b - a;
 
    {
      Scalar lower = a + tau2 * delta;
      Scalar upper = b - tau3 * delta;
 
      alpha = interpolate(a, fa, fpa, b, fb, fpb, lower, upper, order);
    }
    falpha = applyF(alpha);
    if ((a - alpha) * fpa <= std::numeric_limits<Scalar>::epsilon()) {
      /* roundoff prevents progress */
      return BFGSSpace::NoProgress;
    };
 
    if (falpha > f0 + rho * alpha * fp0 || falpha >= fa) {
      /*  a_next = a; */
      b = alpha;
      fb = falpha;
      fpb = std::numeric_limits<Scalar>::quiet_NaN();
    }
    else {
      fpalpha = applyDF(alpha);
 
      if (std::abs(fpalpha) <= -sigma * fp0) {
        alpha_new = alpha;
        return BFGSSpace::Success; /* terminate */
      }
 
      if (((b - a) >= 0 && fpalpha >= 0) || ((b - a) <= 0 && fpalpha <= 0)) {
        b = a;
        fb = fa;
        fpb = fpa;
        a = alpha;
        fa = falpha;
        fpa = fpalpha;
      }
      else {
        a = alpha;
        fa = falpha;
        fpa = fpalpha;
      }
    }
  }
  return BFGSSpace::Success;
}