search

ray-project / ray

Ray is a unified framework for scaling AI and Python applications. Ray consists of a core distributed runtime and a set of AI Libraries for accelerating ML workloads.
https://ray.io
Apache License 2.0
32.95k stars 5.58k forks source link

[Tune] Support NLopt search algorithms #31492

Open nikosavola opened 1 year ago

nikosavola commented 1 year ago

Description

NLopt support could be added to the search algorithms to increase the amount of available optimisation algorithms and suitability for general black-bok optimisation.

There exists a Python API with a common interface through nl.opt to which support could be added.

Use case

Enabling using Ray Tune as a manager for black-box optimisation with NLopt.

richardliaw commented 1 year ago

Seems cool! Would you be interested in making an initial PR? We'd be happy to shepherd it after :)

jon-chuang commented 1 year ago

general black-box optimisation.

This would probably have to be the global and local derivative-free algorithms. The ones that require gradients probably won't work without them, and it's hard to supply those gradients to Tune (you'd have to have a setup similar to meta learning to collect e.g. d loss / d hyperarameter).

It's also unclear to what degree algorithms like NLOpt would be for Tune. Algorithms like BayesOpt are designed to be sample efficient, by contrast, I am a little afraid that NLOpt algorithms may assume evaluation of the objective function f is cheap.

It is also a little unclear to me to what degree the constraints will be useful, especially non-linear constraints. Perhaps I am intending to set a non-linear bound on the complexity of my models?

e.g. channels^3 + layers^2 < 500

Anw, I think I can give this a shot.