Novelty Search improvement about sferes2 HOT 2 CLOSED

Hi Ahn

Thank you for your feedback.

On Jan 23, 2015, at 17:52, Anh Nguyen [email protected] wrote:

• In this file: sferes/modif/novelty.hpp, after lowering the rho_min, I think we need to reset the _not_added to be 0; otherwise every generation after this, the accumulated _not_added will always be larger than Params::novelty::stalled_tresh, resulting in _rho_min keeps getting lowered.

// update rho_min

if (_not_added > Params::novelty::stalled_tresh)//2500

{
_rho_min
*= 0.95
;
_not_added
= 0; // <--- suggested change

}

Done. Thank you. Note that you may want to have a lower threshold in this case: if you start with rho_min that is too high, you will need a lot of evaluation before reaching the “right value”.

• I find that after satisfying this check (n > _rho_min || misc::rand() < Params::novelty::add_to_archive_prob) (in the code below), one individual will be added directly to the archive. However, there is a chance that this individual is exactly the same as one already in the archive or very close to one in the archive (distance ~ 0).

if (n >
_rho_min

|| misc::rand() < Params::novelty::
add_to_archive_prob) {
_archive
.push_back(ea.
pop()[i]);
_not_added
= 0
;

++
added;
}

So I'm wondering if adding another check condition would help the search more efficient.
What do you think?

That’s a good point but it would also add a new parameter (the threshold) and I don’t think it is in the original novelty search paper.

best regards,
— Jean-Baptiste Mouret
http://pages.isir.upmc.fr/~mouret/

from sferes2.

Note that you may want to have a lower threshold in this case: if you start with rho_min that is too high, you will need a lot of evaluation before reaching the “right value”.

This piece of advice is very useful, thank you!

from sferes2.