pengowen123 / cmaes Goto Github PK

Hello! It's me again!

I've been optimizing eant2, mostly just minor stuff like amortizing allocations and increasing cache hits, parallelism, etc.

The real meat of the work of eant2 is done in CMA-ES land.

What portion of that is matrix multiplication & other parallel operations that would benefit from integration with rust-gpu and the like?

I'm interested in running "big" problems through eant2 on my consumer machine, and I'd like to understand the distribution of time spent inside cma-es. Is it known? Do you have any information on where time is spent? First thing that came to mind is GPU acceleration of course.

Thank you.

I was trying to give this a shot, but I haven't been able to puzzle out how to get net-lib to compile.

It's a bit unclear what dependencies are exactly required, and whether windows is a good idea or if it's best to go straight to linux.

The project looks absolutely fantastic, looking forward to trying it out for a physics engines meta-parameters!

I have a requirement for using cma-es with Complex<f64> arrays. How hard do you think it would be to add support for this?

Hello,
First of all, congratulations for the work you have done.
In some of my tests the objective function is quite heavy, and I can't help but notice that the solver loop uses just 1 core at once.
I can't find a setting to enable parallelism, is this my fault or is there not? And if so, is it planned?

Kind regards.

I was surprised to discovered that I couldn't find an implementation of CMA-ES on crates.io and further suprised to discover that your, apparently polished and ready for use, implementation exists but is not published there.

Is there a reason for that?

Building currently does not work on Windows due to issues with compiling netlib, but using Intel's MKL might work instead (see this comment).