source: ThirdParty/mpqc_open/src/lib/math/optimize/opt.cc@ 23612c

//
// opt.cc
//
// Copyright (C) 1996 Limit Point Systems, Inc.
//
// Author: Curtis Janssen <cljanss@limitpt.com>
// Maintainer: LPS
//
// This file is part of the SC Toolkit.
//
// The SC Toolkit is free software; you can redistribute it and/or modify
// it under the terms of the GNU Library General Public License as published by
// the Free Software Foundation; either version 2, or (at your option)
// any later version.
//
// The SC Toolkit 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 Library General Public License for more details.
//
// You should have received a copy of the GNU Library General Public License
// along with the SC Toolkit; see the file COPYING.LIB.  If not, write to
// the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
//
// The U.S. Government is granted a limited license as per AL 91-7.
//

#ifdef __GNUC__
#pragma implementation
#endif

#include <math.h>
#include <deque>

#include <math/optimize/opt.h>
#include <util/keyval/keyval.h>
#include <util/misc/formio.h>
#include <util/misc/timer.h>
#include <util/state/stateio.h>
#include <util/state/state_bin.h>

using namespace std;
using namespace sc;

/////////////////////////////////////////////////////////////////////////
// Optimize

static ClassDesc Optimize_cd(
  typeid(Optimize),"Optimize",2,"virtual public SavableState",
  0, 0, 0);

Optimize::Optimize() :
  ckpt_(0), ckpt_file(0)
{
}

Optimize::Optimize(StateIn&s):
  SavableState(s)
{
  s.get(ckpt_,"checkpoint");
  s.getstring(ckpt_file);
  s.get(max_iterations_,"max_iterations");
  s.get(max_stepsize_,"max_stepsize");
  if (s.version(::class_desc<Optimize>()) > 1) {
      s.get(print_timings_,"print_timings");
    }
  n_iterations_ = 0;
  conv_ << SavableState::restore_state(s);
  function_ << SavableState::key_restore_state(s,"function");
}

Optimize::Optimize(const Ref<KeyVal>&keyval)
{
  print_timings_ = keyval->booleanvalue("print_timings");
  if (keyval->error() != KeyVal::OK) print_timings_ = 0;
  ckpt_ = keyval->booleanvalue("checkpoint");
  if (keyval->error() != KeyVal::OK) ckpt_ = 0;
  ckpt_file = keyval->pcharvalue("checkpoint_file");
  if (keyval->error() != KeyVal::OK) {
    ckpt_file = new char[13];
    strcpy(ckpt_file,"opt_ckpt.dat");
  }

  max_iterations_ = keyval->intvalue("max_iterations");
  if (keyval->error() != KeyVal::OK) max_iterations_ = 10;
  n_iterations_ = 0;

  max_stepsize_ = keyval->doublevalue("max_stepsize");
  if (keyval->error() != KeyVal::OK) max_stepsize_ = 0.6;

  function_ << keyval->describedclassvalue("function");
//  if (function_.null()) {
//      ExEnv::err0() << "Optimize requires a function keyword" << endl;
//      ExEnv::err0() << "which is an object of type Function" << endl;
//      abort();
//    }
// can't assume lineopt's have a function keyword

  conv_ << keyval->describedclassvalue("convergence");
  if (conv_.null()) {
      double convergence = keyval->doublevalue("convergence");
      if (keyval->error() == KeyVal::OK) {
          conv_ = new Convergence(convergence);
        }
    }
  if (conv_.null()) conv_ = new Convergence();
}

Optimize::~Optimize()
{
  if (ckpt_file) delete[] ckpt_file;
  ckpt_file=0;
}

void
Optimize::save_data_state(StateOut&s)
{
  s.put(ckpt_);
  s.putstring(ckpt_file);
  s.put(max_iterations_);
  s.put(max_stepsize_);
  s.put(print_timings_);
  SavableState::save_state(conv_.pointer(),s);
  SavableState::save_state(function_.pointer(),s);
}

void
Optimize::init()
{
  n_iterations_ = 0;
}

void
Optimize::set_checkpoint()
{
  ckpt_=1;
}

void
Optimize::set_max_iterations(int mi)
{
  max_iterations_ = mi;
}

void
Optimize::set_checkpoint_file(const char *path)
{
  if (ckpt_file) delete[] ckpt_file;
  if (path) {
    ckpt_file = new char[strlen(path)+1];
    strcpy(ckpt_file,path);
  } else
    ckpt_file=0;
}
  
void
Optimize::set_function(const Ref<Function>& f)
{
  function_ = f;
}

#ifndef OPTSTATEOUT
#define OPTSTATEOUT StateOutBin
#endif

int
Optimize::optimize()
{
  int result=0;
  while((n_iterations_ < max_iterations_) && (!(result = update()))) {
      ++n_iterations_;
      if (ckpt_) {
        OPTSTATEOUT so(ckpt_file);
        this->save_state(so);
      }
      if (print_timings_) {
          tim_print(0);
        }
    }
  return result;
}

void
Optimize::apply_transform(const Ref<NonlinearTransform> &t)
{
}

/////////////////////////////////////////////////////////////////////////
// LineOpt

static ClassDesc LineOpt_cd(
  typeid(LineOpt),"LineOpt",1,"public Optimize",
  0, 0, 0);

LineOpt::LineOpt(StateIn&s):
  Optimize(s), SavableState(s)
{
  search_direction_ = matrixkit()->vector(dimension());
  search_direction_.restore(s);
}

LineOpt::LineOpt(const Ref<KeyVal>&keyval)
{
  decrease_factor_ = keyval->doublevalue("decrease_factor");    
  if (keyval->error() != KeyVal::OK) decrease_factor_ = 0.1;
}

LineOpt::~LineOpt()
{
}

void
LineOpt::save_data_state(StateOut&s)
{
  search_direction_.save(s);
}

void
LineOpt::init(RefSCVector& direction)
{
  if (function().null()) {
      ExEnv::err0() << "LineOpt requires a function object through" << endl;
      ExEnv::err0() << "constructor or init method" << endl;
      abort();
  }
  search_direction_ = direction.copy();
  initial_x_ = function()->get_x();
  initial_value_ = function()->value();
  initial_grad_ = function()->gradient();
  Optimize::init();
}

void
LineOpt::init(RefSCVector& direction, Ref<Function> function )
{
  set_function(function);
  init(direction);
}

int
LineOpt::sufficient_decrease(RefSCVector& step) {

  double ftarget = initial_value_ + decrease_factor_ *
    initial_grad_.scalar_product(step);
  
  RefSCVector xnext = initial_x_ + step;
  function()->set_x(xnext);
  Ref<NonlinearTransform> t = function()->change_coordinates();
  apply_transform(t);

  return function()->value() <= ftarget;
}

void
LineOpt::apply_transform(const Ref<NonlinearTransform> &t)
{
  if (t.null()) return;
  apply_transform(t);
  t->transform_gradient(search_direction_);
}

/////////////////////////////////////////////////////////////////////////
// Backtrack

static ClassDesc Backtrack_cd(
  typeid(Backtrack),"Backtrack",1,"public LineOpt",
  0, create<Backtrack>, 0);

Backtrack::Backtrack(const Ref<KeyVal>& keyval) 
  : LineOpt(keyval)
{ 
  backtrack_factor_ = keyval->doublevalue("backtrack_factor");
  if (keyval->error() != KeyVal::OK) backtrack_factor_ = 0.1;
}

int
Backtrack::update() {
  
  deque<double> values;
  int acceptable=0;
  int descent=1;
  int took_step=0;
  int using_step;

  RefSCVector backtrack = -1.0 * backtrack_factor_ * search_direction_;
  RefSCVector step = search_direction_.copy();
  
  // check if line search is needed
  if( sufficient_decrease(step) ) {
    ExEnv::out0() << endl << indent << 
      "Unscaled initial step yields sufficient decrease." << endl;
    return 1;
  }

  ExEnv::out0() << endl << indent 
    << "Unscaled initial step does not yield a sufficient decrease."
    << endl << indent
    << "Initiating backtracking line search." << endl; 

  // perform a simple backtrack
  values.push_back( function()->value() );  
  for(int i=0; i<max_iterations_ && !acceptable && descent; ++i) {

    step = step + backtrack;

    if ( sqrt(step.scalar_product(step)) >= 0.1 * 
         sqrt(search_direction_.scalar_product(search_direction_)) ) {

      ++took_step;    
      if( sufficient_decrease(step) ) {
        ExEnv::out0() << endl << indent << "Backtrack " << i+1 
                      << " yields a sufficient decrease." << endl;
        acceptable = 1; 
        using_step = i+1;
      }
      
      else if ( values.back() < function()->value() ) {
        ExEnv::out0() << endl << indent << "Backtrack " << i+1 
                      << " increases value; terminating search." << endl;
        acceptable = 1;
        using_step = i;
      }
      
      else {
        ExEnv::out0() << endl << indent << "Backtrack " << i+1 
                      << " does not yield a sufficient decrease." << endl;
        using_step = i+1;
      }

      values.push_back( function()->value() );
    }

    else { 
      ExEnv::out0() << indent << 
        "Search direction does not appear to be a descent direction;" <<
        " terminating search." << endl;
      descent = 0;
    }
  }
             

  if ( !acceptable && descent ) {
    ExEnv::out0() << indent << 
      "Maximum number of backtrack iterations has been exceeded." << endl;
    acceptable = 1;
  }
 
  for(int i=0; i <= took_step; ++i) {
    if(i==0) ExEnv::out0() << indent << "initial step    " << " value: ";
    else ExEnv::out0() << indent << "backtrack step " << i << " value: ";
    ExEnv::out0() << scprintf("%15.10lf", values.front()) << endl;
    values.pop_front();
  }

  if(descent) { 
    ExEnv::out0() << indent << "Using step " << using_step << endl;
   
    // use next to last step if value went up
    if( using_step != took_step ) {
      function()->set_x( function()->get_x() - backtrack );
      Ref<NonlinearTransform> t = function()->change_coordinates();
      apply_transform(t);
    }
  }
  else {
    function()->set_x( initial_x_ );
    Ref<NonlinearTransform> t = function()->change_coordinates();
    apply_transform(t);
  }

  // returning 0 only if search direction is not descent direction
  return acceptable;
}

/////////////////////////////////////////////////////////////////////////////

// Local Variables:
// mode: c++
// c-file-style: "CLJ"
// End: