Re: [tlug] genetic algorithm/optimization framework

[Lars Kotthoff <lists@example.com>]

> The goal is to optimize a multi-band, circular polarized antenna for
> bandwith, radiation pattern, axial ratio (over the whole sphere),
> stability against material change,... etc
> 
> I am not aware of any other optimization method that would work with
> such a setup as there is no easy way to describe (ie linear, polynomial
> or exponential...) relationship between the parameters of the antenna
> and its properties.

Interesting! What comes to mind apart from GAs is stochastic local search
(https://en.wikipedia.org/wiki/Local_search_(optimization)). It sounds like that
would do quite well here and it's ridiculously easy to program.

That said, the current state of the art in parameter tuning (black-box parameter
optimisation, i.e. the algorithm fiddles with the parameters without knowing
what they affect) it to use model-based methods. Roughly, the method builds a
model of the parameter-response surface and uses that to determine where to
evaluate next. This is a much more targeted exploration of the search space and
uncertainty is taken into account as well (e.g. I may want to evaluate this set
of parameters not necessarily because I think it'll be best, but because it will
give me the most useful additional information for the parameter-response
model).

There are several packages that allow you to do this, e.g.
- Hyperopt https://jaberg.github.io/hyperopt/
- Spearmint https://github.com/JasperSnoek/spearmint
- SMAC http://www.cs.ubc.ca/labs/beta/Projects/SMAC/
- irace http://iridia.ulb.ac.be/irace/ (ok, this isn't really model-based in the
  same sense as the others...)

Most of these will only optimize for a single objective, but you may be able to
combine your multiple objectives into a single one. These methods have been
applied with success to problems with difficult parameter spaces (most
prominently optimising parameters of machine learning algorithms).

I do work in this area and I'd be happy to answer any questions.

Cheers,

Lars