Re: [tlug] genetic algorithm/optimization framework

[Lars Kotthoff <lists@example.com>]

> You can always map the efficiency frontier at the cost of multiple
> optimization runs by optimizing various linear combinations of the
> objectives, typically starting with the "corners" (each objective by
> itself) and the "middle" (equal weight on all objectives).  That might
> be too expensive in Attila's use case, of course, but it's a simple
> approach when throwing machine cycles at the problem is cheaper than
> burning neurons on it.

Right. The problem with applications usually is that it's not clear how the
different objectives should be weighted relative to each other :)

Starting from the extreme points you can compute the Pareto frontier and
hopefully that will be something relatively simple (although none of the
packages I've mentioned explicitly support computing this).

Depending on the application, it may also make sense to have a portfolio of
different parameter settings instead of a single one, i.e. in one case you
really care about bandwidth and want to use the optimal configuration for that
and in another case you care more about something else. This would certainly
make the optimisation easier (and there are good methods for working with such
portfolios).

Of course it's also possible (albeit unlikely) that, when optimising for each
objective in turn, the same parameter settings (or at least very close settings)
come out at the top.

Cheers,

Lars