commons-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Gilles (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (MATH-1126) "LevenbergMarquardtOptimizer": Divergent behavior of new code
Date Thu, 12 Jun 2014 14:35:02 GMT

    [ https://issues.apache.org/jira/browse/MATH-1126?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14029202#comment-14029202
] 

Gilles commented on MATH-1126:
------------------------------

It seems that in the old implementation, "currentCost" is computed using _unweighted_ residuals,
while the new implementation uses _weighted_ residuals (cf. override of "getResiduals" in
class "DenseWeightedEvaluation").

> "LevenbergMarquardtOptimizer": Divergent behavior of new code
> -------------------------------------------------------------
>
>                 Key: MATH-1126
>                 URL: https://issues.apache.org/jira/browse/MATH-1126
>             Project: Commons Math
>          Issue Type: Bug
>    Affects Versions: 3.3
>            Reporter: Gilles
>              Labels: regression
>         Attachments: LM_cost_NEW, LM_cost_OLD
>
>
> The new implementation of "LevenbergMarquardtOptimizer" (package "o.a.c.m.fitting.leastsquares")
behaves differently from the previous one (package ".a.c.m.optim.nonlinear.vector.jacobian").
> This shows up not so much in the solutions respectively found by one and the other implementation;
there are fairly similar, but in my use-case, the number of function evaluations is quite
different. And this could explain an observed 35% performance degradation.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message