//////////////////////////////////////////////////////////////////////////////// // // Copyright (c) 2011 The Regents of the University of California // // This file is part of Qbox // // Qbox is distributed under the terms of the GNU General Public License // as published by the Free Software Foundation, either version 2 of // the License, or (at your option) any later version. // See the file COPYING in the root directory of this distribution // or . // //////////////////////////////////////////////////////////////////////////////// // // CGOptimizer.C // //////////////////////////////////////////////////////////////////////////////// #include "CGOptimizer.h" #include #include #include #include "blas.h" using namespace std; //////////////////////////////////////////////////////////////////////////////// void CGOptimizer::compute_xp(const valarray& x, const double f, valarray& g, valarray& xp) { // Use the function value f and the gradient g at x to generate a new point xp // using the Fletcher-Reeves or Polak-Ribiere CG algorithm // return xp=x if the 2-norm of g is smaller than tol const double tol = 1.0e-18; const int one = 1; assert(x.size()==n_ && g.size()==n_ && xp.size()==n_); double fp; // define the descent direction if ( first_step_ ) { p_ = -g; gm_ = g; x0_ = x; f0_ = f; g0norm2_ = ddot(&n_,&g[0],&one,&g[0],&one); if ( g0norm2_ < tol ) { xp = x; return; } fp = -g0norm2_; fp0_ = fp; linmin_.reset(); alpha_ = linmin_.next_alpha(alpha_,f,fp); if ( debug_print ) cout << " CGOptimizer: first_step: alpha=" << alpha_ << " f=" << f << " fp=" << fp << endl; xp = x0_ + alpha_ * p_; first_step_ = false; } else { // fp: derivative along the current descent direction p_ // fp = df(x0+alpha*p)/dalpha at x fp = ddot(&n_,&g[0],&one,&p_[0],&one); alpha_ = linmin_.next_alpha(alpha_,f,fp); if ( debug_print ) cout << " CGOptimizer: alpha=" << alpha_ << " f=" << f << " fp=" << fp << endl; if ( linmin_.fail() ) { // line minimization failed if ( debug_print ) cout << " CGOptimizer: line minimization failed" << endl; // restart from current point p_ = -g; gm_ = g; x0_ = x; f0_ = f; g0norm2_ = ddot(&n_,&g[0],&one,&g[0],&one); if ( g0norm2_ < tol ) { xp = x; return; } fp = -g0norm2_; fp0_ = fp; linmin_.reset(); alpha_ = linmin_.next_alpha(alpha_,f,fp); if ( debug_print ) cout << " CGOptimizer: restart after fail: alpha=" << alpha_ << " f=" << f << " fp=" << fp << endl; xp = x0_ + alpha_ * p_; first_step_ = false; return; } if ( linmin_.done() ) { // wolfe1_ && wolfe2_ are true at alpha_ if ( debug_print ) cout << " CGOptimizer: done with current descent direction" << endl; // define a new descent direction p_ using the Fletcher-Reeves formula assert(g0norm2_ > 0.0); #if 0 // Fletcher-Reeves double beta = ddot(&n_,&g[0],&one,&g[0],&one) / g0norm2_; #else // Polak-Ribiere double beta = (ddot(&n_,&g[0],&one,&g[0],&one)- ddot(&n_,&gm_[0],&one,&g[0],&one)) / g0norm2_; #endif if ( beta_max_ > 0.0 && fabs(beta) > beta_max_ ) { if ( debug_print ) cout << " CGOptimizer: |beta| exceeds beta_max " << endl; beta = (beta > 0.0) ? beta_max_ : -beta_max_; } if ( debug_print ) cout << " CGOptimizer: beta = " << beta << endl; p_ = beta * p_ - g; x0_ = x; f0_ = f; // recalculate f0, fp0 // fp0 = d_e / d_alpha in direction pc_ fp = ddot(&n_,&g[0],&one,&p_[0],&one); g0norm2_ = ddot(&n_,&g[0],&one,&g[0],&one); gm_ = g; fp0_ = fp; if ( fp > 0.0 ) { // p_ is not a descent direction // restart from current point if ( debug_print ) cout << " CGOptimizer: p_ not a descent direction" << endl; p_ = -g; gm_ = g; x0_ = x; f0_ = f; g0norm2_ = ddot(&n_,&g[0],&one,&g[0],&one); fp = -g0norm2_; fp0_ = fp; } // set the starting alpha of the minimizer to be the current alpha_ if ( alpha_ < linmin_.alpha_max() ) linmin_.set_alpha_start(alpha_); else linmin_.set_alpha_start(0.5*linmin_.alpha_max()); // reset the line minimizer linmin_.reset(); alpha_ = linmin_.next_alpha(alpha_,f,fp); if ( debug_print ) cout << " CGOptimizer: restart: alpha=" << alpha_ << " f=" << f << " fp=" << fp << endl; } xp = x0_ + alpha_ * p_; } }