chrisstroemel / simple Goto Github PK

Thank you for publishing this great algorithm (terrible name for online visibility but great idea)!

I wrote a Rust port in order to experiment with it, it currently features a mapping from the unit simplex to the unit hypercube in order to take an hypercube as parameter and a reformulation of the exploration_preference into an exploration_depth parameter which, I hope, has a clearer semantic.

Hi! I'm doing some ~10--20 dimensional parameter optimization for an astrophysical model and came across your package whilst researching efficient algorithms. My model takes ~2 mins to run for each call and I do not have access to gradient information for the underlying PDF, so Simple appears like it might be perfect for my needs.

Whilst I'd love to try this package, I am having some trouble working out how to initialize a suitable nd-simplex that covers the entire hyper-rectangle corresponding to my allowed parameter space.

I'm imagining that this would be a common requirement for a lot of people wishing to make use of Simple, and so I was wondering if you had any plans to add a utility function for generating an initial nd-simplex with constraints?

Thank you for the great tool! Baysian optimization really is very slow.

One thing I'm unsure about from the readme is whether Simple assumes that calling f(x) once gives the true value of the function, or if it still works if f(x) is heavily noisy, say returns a random value 95% of the time?

What if you mapped the simplex into the cube
(e.g. https://math.stackexchange.com/questions/384713/looking-for-a-nonlinear-map-from-n-dimensional-cube-to-an-n-dimensional-si)

Then you could support rectangular domains easily. Is this something you've considered?

Incidentally, you may be aware of https://stefan-endres.github.io/shgo/ and if not you may be well placed to appreciate it.

Thanks for this, it appears to be nearly as effective as the bayesian optimization method I was using for my problem yet with much less overhead. I had a few "theoretical " questions before I start trying to look deeper and expand on this:

• Do you see any issue with limiting some parameters to integer values?
• You say: "Simple does possess a hard requirement of needing to sample dim+1 corner points before optimization can proceed,"
  -- To do Bayesian Optimization you could hotstart using a small random search and/or hand picked values. Would it make sense to incorporate a step like this into simple at some point of the process? If so, when? What about other schemes to avoid local minima like checking a random parameter set every n steps?