[
https://issues.apache.org/jira/browse/MATH-1282?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14977136#comment-14977136
]
Thomas Neidhart commented on MATH-1282:
---------------------------------------
Thanks for providing an isolated test case, but I fear that this will not be
enough to track down the problem. Investigating the test case shows that the
gradientAtTrustRegionCenter variable has numbers of very large magnitude
(10e100 and larger). Subsequent calculations involving this parameters result
in Infinity and NaN, which might be the final cause for the infinite loop.
Looking at the unit tests, I could not observe similar large values for the
gradient. Would it be possible for you to trace the evolution of this parameter?
> BOBYQA trsbox infinite loop
> ---------------------------
>
> Key: MATH-1282
> URL: https://issues.apache.org/jira/browse/MATH-1282
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.2
> Reporter: Tillmann Gaida
>
> Apparently the BOBYQA optimizer can get stuck in the trsbox method. This
> happened to us out of the blue after millions of successful optimizations.
> Unfortunately our target function has too many dependencies, so I can't
> provide a unit test for the full optimization. I have however created a unit
> test which isolates the call to trsbox: http://pastebin.com/wYPVS3SC
> I took a quick look at the code, but it looks like one can't understand a
> thing without having read the original paper. Can someone help?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
