 |
Image Component Library (ICL)
|
|
Image Component Library (ICL)
|
The StochasticOptimizer is a tiny frame-work for simple stochastic optimization processes. More...
#include <StochasticOptimizer.h>
Public Types | |
| typedef StochasticOptimizerResult< T > | Result |
| Result structure. More... | |
Public Member Functions | |
| StochasticOptimizer (int dataDim) | |
| create a stochastic optimizer with given data dimension More... | |
| Result | optimize (int maxTimeSteps) |
| start optimization process with given step count More... | |
| Result | optimize (T minError, int maxSteps) |
| start optimization process with given step count and minimal error stop-criterion More... | |
Protected Member Functions | |
| virtual T * | getData ()=0 |
| must return the current data vector More... | |
| virtual T | getError (const T *data)=0 |
| must return current error value (>=0) More... | |
| virtual const T * | getNoise (int currentTime, int endTime)=0 |
| returns a noise vector (of size dataDim, which was passed to the constructor) More... | |
| virtual void | reinitialize ()=0 |
| this function is called before the optimization is started More... | |
| virtual void | notifyProgress (int t, int numSteps, int startError, int currBestError, int currError, const T *data, int dataDim) |
| a pure utility function, which can be implemented in derived classes to notify optization progress (somehow) More... | |
Private Attributes | |
| int | m_dataDim |
| internal data-dimension variable More... | |
The StochasticOptimizer is a tiny frame-work for simple stochastic optimization processes.
Stochastic optimization is minimization of an error function, which depends on a set of latent variables (here: called data). The naive approach origins from starting configuration, which must be created by the pure virtual reinitialize function. Now in each optimization step, the current data vector is changed slightly using an additive noise vector which must be provided by the also pure virtual getNoise-function. If the error arising from the changed data is less then the current minimal error, then the change data vector use hold, otherwise the last change of data is reverted. This procedure is iterated until either a maximum number of iterations is reached or until a given minimal error is reached.
NEW** now this class is a template (defined for float and double)
| typedef StochasticOptimizerResult<T> icl::math::StochasticOptimizer< T >::Result |
Result structure.
| icl::math::StochasticOptimizer< T >::StochasticOptimizer | ( | int | dataDim | ) |
create a stochastic optimizer with given data dimension
|
protectedpure virtual |
must return the current data vector
data is un-const, as it is changed in each step if the reached error is less then the current best error. Returned pointer must have at least the length of dataDim passed to the constructor
|
protectedpure virtual |
must return current error value (>=0)
returns the error measurement dependent on given data vector
|
protectedpure virtual |
returns a noise vector (of size dataDim, which was passed to the constructor)
optionally, the noise-strength might depend on the current time-progress. Therefore, currentTime and endTime is also passed to this functions
|
protectedvirtual |
a pure utility function, which can be implemented in derived classes to notify optization progress (somehow)
|
inline |
start optimization process with given step count
| Result icl::math::StochasticOptimizer< T >::optimize | ( | T | minError, |
| int | maxSteps | ||
| ) |
start optimization process with given step count and minimal error stop-criterion
|
protectedpure virtual |
this function is called before the optimization is started
Internally data must be initialized
|
private |
internal data-dimension variable
1.8.15