[0b990d] | 1 | //
|
---|
| 2 | // newton.cc
|
---|
| 3 | //
|
---|
| 4 | // Copyright (C) 1996 Limit Point Systems, Inc.
|
---|
| 5 | //
|
---|
| 6 | // Author: Curtis Janssen <cljanss@limitpt.com>
|
---|
| 7 | // Maintainer: LPS
|
---|
| 8 | //
|
---|
| 9 | // This file is part of the SC Toolkit.
|
---|
| 10 | //
|
---|
| 11 | // The SC Toolkit is free software; you can redistribute it and/or modify
|
---|
| 12 | // it under the terms of the GNU Library General Public License as published by
|
---|
| 13 | // the Free Software Foundation; either version 2, or (at your option)
|
---|
| 14 | // any later version.
|
---|
| 15 | //
|
---|
| 16 | // The SC Toolkit is distributed in the hope that it will be useful,
|
---|
| 17 | // but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
| 18 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
| 19 | // GNU Library General Public License for more details.
|
---|
| 20 | //
|
---|
| 21 | // You should have received a copy of the GNU Library General Public License
|
---|
| 22 | // along with the SC Toolkit; see the file COPYING.LIB. If not, write to
|
---|
| 23 | // the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
|
---|
| 24 | //
|
---|
| 25 | // The U.S. Government is granted a limited license as per AL 91-7.
|
---|
| 26 | //
|
---|
| 27 |
|
---|
| 28 | #ifdef __GNUC__
|
---|
| 29 | #pragma implementation
|
---|
| 30 | #endif
|
---|
| 31 |
|
---|
| 32 | #include <math.h>
|
---|
| 33 | #include <float.h>
|
---|
| 34 |
|
---|
| 35 | #include <math/optimize/newton.h>
|
---|
| 36 | #include <util/keyval/keyval.h>
|
---|
| 37 | #include <util/misc/formio.h>
|
---|
| 38 | #include <util/state/stateio.h>
|
---|
| 39 |
|
---|
| 40 | using namespace std;
|
---|
| 41 | using namespace sc;
|
---|
| 42 |
|
---|
| 43 | /////////////////////////////////////////////////////////////////////////
|
---|
| 44 | // NewtonOpt
|
---|
| 45 |
|
---|
| 46 | static ClassDesc NewtonOpt_cd(
|
---|
| 47 | typeid(NewtonOpt),"NewtonOpt",1,"public Optimize",
|
---|
| 48 | 0, create<NewtonOpt>, create<NewtonOpt>);
|
---|
| 49 |
|
---|
| 50 | NewtonOpt::NewtonOpt(const Ref<KeyVal>&keyval):
|
---|
| 51 | Optimize(keyval)
|
---|
| 52 | {
|
---|
| 53 | init();
|
---|
| 54 |
|
---|
| 55 | accuracy_ = keyval->doublevalue("accuracy",KeyValValuedouble(0.0001));
|
---|
| 56 | print_x_ = keyval->booleanvalue("print_x");
|
---|
| 57 | print_hessian_ = keyval->booleanvalue("print_hessian");
|
---|
| 58 | print_gradient_ = keyval->booleanvalue("print_gradient");
|
---|
| 59 | }
|
---|
| 60 |
|
---|
| 61 | NewtonOpt::NewtonOpt(StateIn&s):
|
---|
| 62 | SavableState(s),
|
---|
| 63 | Optimize(s)
|
---|
| 64 | {
|
---|
| 65 | s.get(accuracy_);
|
---|
| 66 | s.get(maxabs_gradient);
|
---|
| 67 | s.get(print_hessian_);
|
---|
| 68 | s.get(print_x_);
|
---|
| 69 | s.get(print_gradient_);
|
---|
| 70 | }
|
---|
| 71 |
|
---|
| 72 | NewtonOpt::~NewtonOpt()
|
---|
| 73 | {
|
---|
| 74 | }
|
---|
| 75 |
|
---|
| 76 | void
|
---|
| 77 | NewtonOpt::save_data_state(StateOut&s)
|
---|
| 78 | {
|
---|
| 79 | Optimize::save_data_state(s);
|
---|
| 80 | s.put(accuracy_);
|
---|
| 81 | s.put(maxabs_gradient);
|
---|
| 82 | s.put(print_hessian_);
|
---|
| 83 | s.put(print_x_);
|
---|
| 84 | s.put(print_gradient_);
|
---|
| 85 | }
|
---|
| 86 |
|
---|
| 87 | void
|
---|
| 88 | NewtonOpt::init()
|
---|
| 89 | {
|
---|
| 90 | Optimize::init();
|
---|
| 91 | maxabs_gradient = -1.0;
|
---|
| 92 | }
|
---|
| 93 |
|
---|
| 94 | int
|
---|
| 95 | NewtonOpt::update()
|
---|
| 96 | {
|
---|
| 97 | // these are good candidates to be input options
|
---|
| 98 | const double maxabs_gradient_to_desired_accuracy = 0.05;
|
---|
| 99 | const double maxabs_gradient_to_next_desired_accuracy = 0.005;
|
---|
| 100 | const double roundoff_error_factor = 1.1;
|
---|
| 101 |
|
---|
| 102 | // the gradient convergence criterion.
|
---|
| 103 | double old_maxabs_gradient = maxabs_gradient;
|
---|
| 104 | RefSCVector xcurrent;
|
---|
| 105 | RefSCVector gcurrent;
|
---|
| 106 |
|
---|
| 107 | // get the next gradient at the required level of accuracy.
|
---|
| 108 | // usually only one pass is needed, unless we happen to find
|
---|
| 109 | // that the accuracy was set too low.
|
---|
| 110 | int accurate_enough;
|
---|
| 111 | do {
|
---|
| 112 | // compute the current point
|
---|
| 113 | function()->set_desired_gradient_accuracy(accuracy_);
|
---|
| 114 |
|
---|
| 115 | xcurrent = function()->get_x();
|
---|
| 116 | gcurrent = function()->gradient().copy();
|
---|
| 117 |
|
---|
| 118 | // compute the gradient convergence criterion now so i can see if
|
---|
| 119 | // the accuracy needs to be tighter
|
---|
| 120 | maxabs_gradient = gcurrent.maxabs();
|
---|
| 121 | // compute the required accuracy
|
---|
| 122 | accuracy_ = maxabs_gradient * maxabs_gradient_to_desired_accuracy;
|
---|
| 123 |
|
---|
| 124 | if (accuracy_ < DBL_EPSILON) accuracy_ = DBL_EPSILON;
|
---|
| 125 |
|
---|
| 126 | // The roundoff_error_factor is thrown in to allow for round off making
|
---|
| 127 | // the current gcurrent.maxabs() a bit smaller than the previous,
|
---|
| 128 | // which would make the current required accuracy less than the
|
---|
| 129 | // gradient's actual accuracy and cause everything to be recomputed.
|
---|
| 130 | accurate_enough = (
|
---|
| 131 | function()->actual_gradient_accuracy()
|
---|
| 132 | <= accuracy_*roundoff_error_factor);
|
---|
| 133 |
|
---|
| 134 | if (!accurate_enough) {
|
---|
| 135 | ExEnv::out0().unsetf(ios::fixed);
|
---|
| 136 | ExEnv::out0() << indent
|
---|
| 137 | << "NOTICE: function()->actual_gradient_accuracy() > accuracy_:\n"
|
---|
| 138 | << indent
|
---|
| 139 | << scprintf(
|
---|
| 140 | " function()->actual_gradient_accuracy() = %15.8e",
|
---|
| 141 | function()->actual_gradient_accuracy()) << endl << indent
|
---|
| 142 | << scprintf(
|
---|
| 143 | " accuracy_ = %15.8e",
|
---|
| 144 | accuracy_) << endl;
|
---|
| 145 | }
|
---|
| 146 | } while(!accurate_enough);
|
---|
| 147 |
|
---|
| 148 | if (old_maxabs_gradient >= 0.0 && old_maxabs_gradient < maxabs_gradient) {
|
---|
| 149 | ExEnv::out0() << indent
|
---|
| 150 | << scprintf("NOTICE: maxabs_gradient increased from %8.4e to %8.4e",
|
---|
| 151 | old_maxabs_gradient, maxabs_gradient) << endl;
|
---|
| 152 | }
|
---|
| 153 |
|
---|
| 154 | RefSymmSCMatrix hessian = function()->hessian();
|
---|
| 155 | RefSymmSCMatrix ihessian = function()->inverse_hessian(hessian);
|
---|
| 156 |
|
---|
| 157 | if (print_hessian_) {
|
---|
| 158 | hessian.print("hessian");
|
---|
| 159 | }
|
---|
| 160 | if (print_x_) {
|
---|
| 161 | int n = xcurrent.n();
|
---|
| 162 | ExEnv::out0() << indent << "x = [";
|
---|
| 163 | for (int i=0; i<n; i++) {
|
---|
| 164 | ExEnv::out0() << scprintf(" % 16.12f",double(xcurrent(i)));
|
---|
| 165 | }
|
---|
| 166 | ExEnv::out0() << " ]" << endl;
|
---|
| 167 | }
|
---|
| 168 | if (print_gradient_) {
|
---|
| 169 | int n = gcurrent.n();
|
---|
| 170 | ExEnv::out0() << indent << "gradient = [";
|
---|
| 171 | for (int i=0; i<n; i++) {
|
---|
| 172 | ExEnv::out0() << scprintf(" % 16.12f",double(gcurrent(i)));
|
---|
| 173 | }
|
---|
| 174 | ExEnv::out0() << " ]" << endl;
|
---|
| 175 | }
|
---|
| 176 |
|
---|
| 177 | // take the step
|
---|
| 178 | RefSCVector xdisp = -1.0*(ihessian * gcurrent);
|
---|
| 179 | // scale the displacement vector if it's too large
|
---|
| 180 | double tot = sqrt(xdisp.scalar_product(xdisp));
|
---|
| 181 | if (tot > max_stepsize_) {
|
---|
| 182 | double scal = max_stepsize_/tot;
|
---|
| 183 | ExEnv::out0() << endl << indent
|
---|
| 184 | << scprintf("stepsize of %f is too big, scaling by %f",tot,scal)
|
---|
| 185 | << endl;
|
---|
| 186 | xdisp.scale(scal);
|
---|
| 187 | tot *= scal;
|
---|
| 188 | }
|
---|
| 189 |
|
---|
| 190 | RefSCVector xnext = xcurrent + xdisp;
|
---|
| 191 |
|
---|
| 192 | conv_->reset();
|
---|
| 193 | conv_->get_grad(function());
|
---|
| 194 | conv_->get_x(function());
|
---|
| 195 | conv_->set_nextx(xnext);
|
---|
| 196 |
|
---|
| 197 | // check for convergence before resetting the geometry
|
---|
| 198 | int converged = conv_->converged();
|
---|
| 199 | if (converged)
|
---|
| 200 | return converged;
|
---|
| 201 |
|
---|
| 202 | ExEnv::out0() << endl << indent
|
---|
| 203 | << scprintf("taking step of size %f", tot) << endl;
|
---|
| 204 |
|
---|
| 205 | function()->set_x(xnext);
|
---|
| 206 |
|
---|
| 207 | // make the next gradient computed more accurate, since it will
|
---|
| 208 | // be smaller
|
---|
| 209 | accuracy_ = maxabs_gradient * maxabs_gradient_to_next_desired_accuracy;
|
---|
| 210 |
|
---|
| 211 | return converged;
|
---|
| 212 | }
|
---|
| 213 |
|
---|
| 214 | /////////////////////////////////////////////////////////////////////////////
|
---|
| 215 |
|
---|
| 216 | // Local Variables:
|
---|
| 217 | // mode: c++
|
---|
| 218 | // c-file-style: "ETS"
|
---|
| 219 | // End:
|
---|