[
https://issues.apache.org/jira/browse/MATH-1128?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gilles updated MATH-1128:
-------------------------
Attachment: MATH-1128.patch
Here the proposed patch.
Mainly, it introduces a "LazyUnweightedEvaluation" where the computation of the
model and Jacobian are deferred until one or the other is actually requested.
Also, there is a new "LocalMultivariateJacobianFunction" class that replaces
the anymous class that was created when calling the "model" method.
The "evaluate" method of "LocalLeastSquaresProblem" instantiates either
"UnweightedEvaluation" or "LazyUnweightedEvaluation", as requested by the user
of the "create" factory method.
Any objection to applying this patch?
> Lazy evaluation needed in "o.a.c.m.fitting.leastsquares"
> --------------------------------------------------------
>
> Key: MATH-1128
> URL: https://issues.apache.org/jira/browse/MATH-1128
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 3.3
> Reporter: Gilles
> Assignee: Gilles
> Labels: regression
> Fix For: 3.4
>
> Attachments: MATH-1128.patch
>
>
> In "LocalLeastSquaresProblem" (private inner class defined in
> "o.a.c.m.fitting.leastsquares.LeastSquaresFactory"), the "evaluate" method
> computes the values of both the model and the Jacobian at creation of the
> "Evaluation" instance.
> Optimizers ("LevenbergMarquardtOptimizer" in particular) may not need both
> for all of the evaluated points. And this can lead to too many evaluations of
> the model which in some applications is the costliest part.
> In my use-case, the current code in "o.a.c.m.fitting.leastquares" leads to a
> performance degradation of about 20% w.r.t. the implementation in
> "o.a.c.m.optim.nonlinear.vector.jacobian".
--
This message was sent by Atlassian JIRA
(v6.2#6252)
