internal/ceres/trust_region_step_evaluator.cc

   1 // Ceres Solver - A fast non-linear least squares minimizer
   2 // Copyright 2016 Google Inc. All rights reserved.
   3 // http://ceres-solver.org/
   4 //
   5 // Redistribution and use in source and binary forms, with or without
   6 // modification, are permitted provided that the following conditions are met:
   7 //
   8 // * Redistributions of source code must retain the above copyright notice,
   9 //   this list of conditions and the following disclaimer.
  10 // * Redistributions in binary form must reproduce the above copyright notice,
  11 //   this list of conditions and the following disclaimer in the documentation
  12 //   and/or other materials provided with the distribution.
  13 // * Neither the name of Google Inc. nor the names of its contributors may be
  14 //   used to endorse or promote products derived from this software without
  15 //   specific prior written permission.
  16 //
  17 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  18 // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19 // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20 // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
  21 // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  22 // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  23 // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
  24 // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
  25 // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  26 // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  27 // POSSIBILITY OF SUCH DAMAGE.
  28 //
  29 // Author: sameeragarwal@google.com (Sameer Agarwal)
  30
  31 #include <algorithm>
  32 #include "ceres/trust_region_step_evaluator.h"
  33 #include "glog/logging.h"
  34
  35 namespace ceres {
  36 namespace internal {
  37
  38 TrustRegionStepEvaluator::TrustRegionStepEvaluator(
  39     const double initial_cost,
  40     const int max_consecutive_nonmonotonic_steps)
  41     : max_consecutive_nonmonotonic_steps_(max_consecutive_nonmonotonic_steps),
  42       minimum_cost_(initial_cost),
  43       current_cost_(initial_cost),
  44       reference_cost_(initial_cost),
  45       candidate_cost_(initial_cost),
  46       accumulated_reference_model_cost_change_(0.0),
  47       accumulated_candidate_model_cost_change_(0.0),
  48       num_consecutive_nonmonotonic_steps_(0){
  49 }
  50
  51 double TrustRegionStepEvaluator::StepQuality(
  52     const double cost,
  53     const double model_cost_change) const {
  54   const double relative_decrease = (current_cost_ - cost) / model_cost_change;
  55   const double historical_relative_decrease =
  56       (reference_cost_ - cost) /
  57       (accumulated_reference_model_cost_change_ + model_cost_change);
  58   return std::max(relative_decrease, historical_relative_decrease);
  59 }
  60
  61 void TrustRegionStepEvaluator::StepAccepted(
  62     const double cost,
  63     const double model_cost_change) {
  64   // Algorithm 10.1.2 from Trust Region Methods by Conn, Gould &
  65   // Toint.
  66   //
  67   // Step 3a
  68   current_cost_ = cost;
  69   accumulated_candidate_model_cost_change_ += model_cost_change;
  70   accumulated_reference_model_cost_change_ += model_cost_change;
  71
  72   // Step 3b.
  73   if (current_cost_ < minimum_cost_) {
  74     minimum_cost_ = current_cost_;
  75     num_consecutive_nonmonotonic_steps_ = 0;
  76     candidate_cost_ = current_cost_;
  77     accumulated_candidate_model_cost_change_ = 0.0;
  78   } else {
  79     // Step 3c.
  80     ++num_consecutive_nonmonotonic_steps_;
  81     if (current_cost_ > candidate_cost_) {
  82       candidate_cost_ = current_cost_;
  83       accumulated_candidate_model_cost_change_ = 0.0;
  84     }
  85   }
  86
  87   // Step 3d.
  88   //
  89   // At this point we have made too many non-monotonic steps and
  90   // we are going to reset the value of the reference iterate so
  91   // as to force the algorithm to descend.
  92   //
  93   // Note: In the original algorithm by Toint, this step was only
  94   // executed if the step was non-monotonic, but that would not handle
  95   // the case of max_consecutive_nonmonotonic_steps = 0. The small
  96   // modification of doing this always handles that corner case
  97   // correctly.
  98   if (num_consecutive_nonmonotonic_steps_ ==
  99       max_consecutive_nonmonotonic_steps_) {
 100     reference_cost_ = candidate_cost_;
 101     accumulated_reference_model_cost_change_ =
 102         accumulated_candidate_model_cost_change_;
 103   }
 104 }
 105
 106 }  // namespace internal
 107 }  // namespace ceres