Release checklist

[abelsiqueira]

• Fix the specialized interface (Specialized interface #38)
• Generate the documentation
  • Core and specialized interface (Documentation #32)
  • Update the README.md
  • Make an read the docs online documentation.
• Bindeps again #105
• Don't allocate gradient-like structure for specialized functions unless required #107
• Change develop to master
• Decide the version number: v0.1.0

What else we should do to release a stable version?

[timholy]

Add a deps/build.jl script that (in an ideal world) works on all major platforms?

[dpo]

Technically we don't support Windows. It's doable for OSX and Linux.

[dpo]

Done for OSX and Linux (in the develop branch).

[timholy]

I just tested the build script. Looks really promising, but (1) I had to install libgsl-dev manually (Ubuntu), and (2) it ends with an error; I've posted the log in this gist.

[dpo]

Thanks for the feedback! The build script still needs improvements. What version of gsl did you install? You need gsl1. Perhaps you installed gsl2?!

[timholy]

Looks like Ubuntu 16.04 doesn't offer gsl1; only gsl2 is available:

$ apt-cache search gsl
libgsl-dbg - GNU Scientific Library (GSL) -- debug symbols package
libgsl-dev - GNU Scientific Library (GSL) -- development package
libgsl2 - GNU Scientific Library (GSL) -- library package
gambas3-gb-gsl - Gambas GNU Scientific Library component
gsl-bin - GNU Scientific Library (GSL) -- binary package
gsl-ref-html - GNU Scientific Library (GSL) Reference Manual in html
gsl-ref-psdoc - GNU Scientific Library (GSL) Reference Manual in postscript
...

[dpo]

I'm not a linux user so I have to ask @abelsiqueira how he does it.

[abelsiqueira]

Sorry for the delay, I'll investigate this now.

[abelsiqueira]

For future reference:

• gsl updated to 2.0 in 2015
• Ubuntu LTS picked it up on 2016 (The last LTS version was 14.04 on 2014)
• Travis is using 12.04, so I wasn't aware of the issue.
• Homebrew is delaying using gsl 2.0 until all packages have migrated

Current solution is to

Download gsl-1.16 from http://mirror.nbtelecom.com.br/gnu/gsl/gsl-1.16.tar.gz
tar -zxf gsl-1.16.tar.gz
cd gsl-1.16
./configure
make
sudo make install

Additionally, you have to create a link to lgfortran:

sudo ln -s /usr/lib/x86_64-linux-gnu/libgfortran.so.3 /usr/lib/libgfortran.so

or something similar. Should work after this. I'll update the installation steps.

[dpo]

Thanks. Importantly, I don't think CUTEst builds against gsl2, so gsl1 is really needed here.

[timholy]

Awesome, this worked for me.

I spent some time scanning through the source and the CUTEr/CUTEst documentation, but to my surprise I don't see a way to search (or even list) the available problems---it seems you have to know the name of the problem you want. Is there a good way to do that? Ideally I'd love to be able to loop over all problems fitting a particular description (e.g., all problems with continuous 2nd derivatives and fewer than n variables). Also, I presume some problems have a known optimum, is there a good way to extract that?

[abelsiqueira]

That's an open PR: #73
Don't know if there's a way to get the known optimum.

Outside of Julia, there is the CUTEst select tool, which I actually never used, and I started a Python tool but never finished.

[dpo]

You'll be scared when you see what the current best way to select problems is. There's a tool named select in $SIFDECODE/bin. It'll ask you a few questions and write the resulting list of problems to a file. It's very old school (think Fortran 77). #73 is supposed to improve on that for CUTEst problems. Further down the road, we'd like to be able to select problems written as in https://github.com/JuliaSmoothOptimizers/OptimizationProblems.jl.

Documenting the CUTEst problems has been an long-standing wish. There's a blurb at the beginning of each SIF file explaining where the problem comes from, but it's necessary to scour the papers and books to discover whether a solution is known. Many of those problems are nonconvex, so it's not even clear what a "solution" should be. Almost all solvers are only guaranteed to find a stationary point, not a local minimum (though they sometimes find a local minimum). For academic problems, it may be possible to find out what the/an optimum is. For the rest, we could record a "best known solution".

[timholy]

Don't know if there's a way to get the known optimum.

It turns out there are many SIF files that have a line something like

*LO SOLTN               0.00102734

which records the intended solution. But many are 0 when my best minimum isn't 0 (could be a dummy value?), and at least in 32 cases I've found a lower value than the one in the file. Could be that there are multiple minima and I've just found a better one.

[abelsiqueira]

Will use version number 0.1.0