/sci/ - Science & Math » Thread #13340862

24KiB, 240x240, gurobi.png

View Same Google iqdb SauceNAO

Optimizers

Anonymous Wed 30 Jun 04:04:46 2021 No.13340862 View Reply Original Report

Quoted By: >>13340893 >>13340911

Why is the top optimization software all proprietary? In my experience, the teams working on these projects are usually pretty small - I'd imagine some research groups or NEETS could get together and make some FOSS that could compete.

Anonymous

Anonymous Wed 30 Jun 2021 04:16:49 No.13340893 Report

Quoted By:

>>13340862
based, do it

Anonymous

Anonymous Wed 30 Jun 2021 04:25:32 No.13340911 Report

Quoted By: >>13340937

>>13340862
Yea it's very doable.

There's usually some secret sauce that is domain specific in these software packages, but as far as I know there aren't any secret algorithms floating around out there. But the secret sauce is usually just hyperparameters or heuristics. Well tuned hyperparameters make a huge difference and it can be a pain in the ass to figure that out yourself.

I've written a number of these things myself over the years, its not particularly difficult if you have the research to work from.

Anonymous

Anonymous Wed 30 Jun 2021 04:39:42 No.13340937 Report

Quoted By: >>13340951

>>13340911
This might sound retarded, but if it really is domain specific, why not just explore the optimizer hyperparamter space with a fat design of experiments on a given problem dataset, then have some shitty sklearn algo train to reduce the error of it.

It would overfit the output RSM algo to that domain, but who cares - people only care about their domain anyways.

Anonymous

Anonymous Wed 30 Jun 2021 04:44:12 No.13340951 Report

Quoted By: >>13341003

>>13340937
yea thats pretty much how you do it.

I forgot to mention, I imagine these proprietary packages are also highly compute optimized. This is another bag of tricks that is maybe even more esoteric than the hyperparameter shit. To really make these things go fast you need to be optimizing on the register/cache level, minimizing cache misses etc. That's a very specialized skill set even among experienced programmers.

If there's custom or specialized hardware involved that could be a factor to watch out for too.

Anonymous

Anonymous Wed 30 Jun 2021 05:03:08 No.13341003 Report

Quoted By: >>13341025

>>13340951
I never thought of the hardware requirements. Guess that means finding someone who can write Fourtran or not be afraid of looking at assembly in this day and age. No wonder these companies are the only ones that develop optimizers - they're probably the last bastion of the old timers who learned those languages and never moved on.

Anonymous

Anonymous Wed 30 Jun 2021 05:15:50 No.13341025 Report

Quoted By:

>>13341003
Yea. These guys are sitting on some epic greybeard wizard knowledge. I've never worked with proprietary stuff but I imagine they charge what they do for a reason. If you need that level of performance, you just need it.

Anonymous

Anonymous Wed 30 Jun 2021 05:19:23 No.13341031 Report

Quoted By:

any of you use an orthogonal diagonalizer in your research?

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches