[EvA] Question regarding to EvA2

[Marcel Kronfeld]

TSE GUAN TAN wrote:
> I am using the EVA2 to evolve the weights of Neural Network.
> However, the weights may evolve to be any size, they are not constrained in
> any way.
> How can I set the real valued vector problem without any range?
> (unconstraint assignment problem)
> or any suggestion?

Hi there,

there are basically two ways to do this. Either you select the problem
range to be very large (near Double.MAX_VALUE) or you use the range as
initialization range and allow the optimizer to leave the initial range.

The first variant will be problematic for many target functions since
for typical population sizes, the chance to initialize far away from any
sensible value is too high. If you choose to try this option, please
make sure to checkout the latest version of EvA2 (2.035 as of today)
which clears a bug with very large numbers in the GUI.

We would therefore suggest to use the second variant, select the most
feasible region as initial range and use optimizers without range
constraints. This can be easily done in several optimizers (among others
DE, PSO and Tribes) by unchecking the option called "checkRange" (or
sometimes "checkConstraints") from within the GUI or in analogy use
setCheckRange(false) from the API.

However, not every optimizer supports this directly, for example with
ES, it depends on the specific mutation operater you choose. Yet CMA,
for example, allows this option as well.

With option one you might also think about reimplementing the
initPopulation method (from AbstractProblemDouble) in your problem class
and handle individual initialization by yourself without regarding the
very large range.

Kind regards,
Marcel Kronfeld

-- 
 Dipl.-Inform. Marcel Kronfeld
 Wilhelm-Schickard-Institute for Computer Science
 University of Tuebingen
 Sand 1 - A305, 72076 Tuebingen, Germany

 Phone: (+49/0)-7071-29-78987, Fax: (+49/0)-7071-29-5091
 EMail: marcel.kronfeld at uni-tuebingen.de