NoisyRprop Class Reference
Rprop-like algorithm for noisy function evaluations. More...
#include <NoisyRprop.h>
Public Member Functions | |
| NoisyRprop () | |
| Constructor. | |
| ~NoisyRprop () | |
| Destructor. | |
| void | init (Model &model) |
| initialization with default values | |
| void | initUserDefined (Model &model, double delta0P=0.01) |
| user defined initialization | |
| void | initUserDefined (Model &model, const Array< double > &delta0P) |
| user defined initialization with coordinate wise individual step sizes | |
| double | optimize (Model &model, ErrorFunction &errorfunction, const Array< double > &input, const Array< double > &target) |
| optimization step | |
Protected Attributes | |
| Array< double > | delta |
| individual step sizes | |
| int | minEpisodeLength |
| minimal episode length | |
| Array< int > | EpisodeLength |
| current episode length | |
| Array< int > | sample |
| fraction 1 to 100 of the episode length | |
| Array< int > | sampleIteration |
| next 1/100 of the episode length | |
| Array< int > | iteration |
| iteration within the current episode | |
| Array< int > | position |
| normalized position within the episode | |
| Array< int > | toTheRight |
| "step to the right" event statistic | |
| Array< int > | leftmost |
| leftmost visited place | |
| Array< int > | rightmost |
| rightmost visited place | |
| Array< double > | leftsum |
| sum of positive derivatives | |
| Array< double > | rightsum |
| sum of negative derivatives | |
| Array< int > | current |
| current error evaluation | |
| Array< int > | numberOfAverages |
| number of error evaluations per iteration | |
| Array< double > | gradient |
| accumulated gradient for each coordinate | |
Detailed Description
Rprop-like algorithm for noisy function evaluations.
- The Rprop algorithm (see Rprop.h) is a very robust gradient descent based optimizer for real-valued optimization. However, it can not deal with the presence of noice.
- The NoisyRprop algorithm is a completely novel algorithm which tries to carry over the most important ideas from Rprop to the optimization of noisy problems. It is, of course, slower than the Rprop algorithm, but it can handle noisy problems with noisy gradient information.
Definition at line 66 of file NoisyRprop.h.
Constructor & Destructor Documentation
| NoisyRprop::NoisyRprop | ( | ) |
| NoisyRprop::~NoisyRprop | ( | ) |
Member Function Documentation
| void NoisyRprop::init | ( | Model & | model | ) | [virtual] |
initialization with default values
Implements Optimizer.
Definition at line 54 of file NoisyRprop.cpp.
References initUserDefined().
| void NoisyRprop::initUserDefined | ( | Model & | model, | |
| const Array< double > & | delta0P | |||
| ) |
user defined initialization with coordinate wise individual step sizes
Definition at line 66 of file NoisyRprop.cpp.
References current, delta, EpisodeLength, Model::getParameterDimension(), gradient, iteration, leftmost, leftsum, minEpisodeLength, numberOfAverages, position, rightmost, rightsum, sample, sampleIteration, and toTheRight.
| void NoisyRprop::initUserDefined | ( | Model & | model, | |
| double | delta0P = 0.01 | |||
| ) |
user defined initialization
Definition at line 59 of file NoisyRprop.cpp.
References Model::getParameterDimension().
Referenced by init().
| double NoisyRprop::optimize | ( | Model & | model, | |
| ErrorFunction & | errorfunction, | |||
| const Array< double > & | input, | |||
| const Array< double > & | target | |||
| ) | [virtual] |
optimization step
Implements Optimizer.
Definition at line 103 of file NoisyRprop.cpp.
References current, delta, EpisodeLength, ErrorFunction::errorDerivative(), Model::getParameter(), Model::getParameterDimension(), gradient, Model::isFeasible(), iteration, leftmost, leftsum, minEpisodeLength, numberOfAverages, position, rightmost, rightsum, sample, sampleIteration, Model::setParameter(), and toTheRight.
Member Data Documentation
Array<int> NoisyRprop::current [protected] |
current error evaluation
Definition at line 127 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<double> NoisyRprop::delta [protected] |
individual step sizes
Definition at line 91 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::EpisodeLength [protected] |
current episode length
Definition at line 97 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<double> NoisyRprop::gradient [protected] |
accumulated gradient for each coordinate
Definition at line 133 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::iteration [protected] |
iteration within the current episode
Definition at line 106 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::leftmost [protected] |
leftmost visited place
Definition at line 115 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<double> NoisyRprop::leftsum [protected] |
sum of positive derivatives
Definition at line 121 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
int NoisyRprop::minEpisodeLength [protected] |
minimal episode length
Definition at line 94 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::numberOfAverages [protected] |
number of error evaluations per iteration
Definition at line 130 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::position [protected] |
normalized position within the episode
Definition at line 109 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::rightmost [protected] |
rightmost visited place
Definition at line 118 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<double> NoisyRprop::rightsum [protected] |
sum of negative derivatives
Definition at line 124 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::sample [protected] |
fraction 1 to 100 of the episode length
Definition at line 100 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::sampleIteration [protected] |
next 1/100 of the episode length
Definition at line 103 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
Array<int> NoisyRprop::toTheRight [protected] |
"step to the right" event statistic
Definition at line 112 of file NoisyRprop.h.
Referenced by initUserDefined(), and optimize().
The documentation for this class was generated from the following files: