[
https://issues.apache.org/jira/browse/MATH-177?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12664968#action_12664968
]
Luc Maisonobe commented on MATH-177:
------------------------------------
The Brent minimizer class has been added in trunk in the new minimization
package as of r735475.
Some tests have been added, but they fail unless very loose threshold are used.
The algorithm
also seems to be very slow for now. This seems to be due to the fact the
parabolic fitting is in fact
never used during the tests, only the golden section part is used leading to a
very large number of
iterations (between 45 and 50) and low accuracy.
I didn't investigate the problem. This may be due to the refactoring I did to
fit the new interfaces and
our current coding style.
Gilles, could you review what I have done ?
> Provide a general minimizing package with a classical Gauss-Newton algorithm
> ----------------------------------------------------------------------------
>
> Key: MATH-177
> URL: https://issues.apache.org/jira/browse/MATH-177
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 2.0
> Reporter: Mick
> Assignee: Luc Maisonobe
> Fix For: 2.0
>
> Attachments: BrentMinimizer.java, UnivariateRealSolver.java,
> UnivariateRealSolverImpl.java
>
>
> Currently the math API provides least squares only method for minimizing
> (solving). The limitation to least-squares problems comes from the
> Levenberg-Marquardt algorithm. A more general minimizer (not for quadratic
> forms) could be implemented by refactoring this with a classical GN, steepest
> descent and also conjugate gradient. We could use them as a basis for some
> least-squares solvers (and also keep the very efficient and specialized
> Levenberg-Marquardt too).
> Based on email exchange with Luc Maisonobe entitled [math] Minimizer on
> 1/15/08.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
