|
From: | Patrick Alken |
Subject: | Re: [Help-gsl] GSL 2.3 test failures on OBS |
Date: | Mon, 3 Apr 2017 15:55:35 -0600 |
User-agent: | Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.8.0 |
Hello,Sorry for the late reply. I wouldn't worry about the box3d failure - the failure seems to occur only for the LM+accel algorithm and only when using finite differences to compute the directional derivative vector. I know from some experience that finite differences don't work very well to calculate that vector, so in some sense this is a known problem. I will log this into the bug tracker so I can take a look when I find some more time, and possibly contact the original author of that algorithm to see what he thinks.
But you can sleep well - there are around 41 different test problems for that module, which thoroughly test the various algorithms. I guess if you're really worried then if you use the LM+accel algorithm, be sure to use an analytic expression for the directional derivative vector.
Thanks, Patrick On 02/20/2017 09:03 AM, Adam Majer wrote:
On 02/11/2017 07:11 PM, Patrick Alken wrote:Also for the multifit_nlinear box3d failure, I am aware of this problem - it also fails on an old macOS system I have. I currently don't have a fix for it, but it seems that recent compilers don't have an issueHow recent? This is using GCC 6.3.1 - Adam
[Prev in Thread] | Current Thread | [Next in Thread] |