MOSES: BScore -> Knobs

[Nil Geisweiller]

Hi (esp. Ben and Moshe),

what about learning a map between bscores (behavioral scores) and knob
settings during deme search?

bscores are generally represented as R+^n

and knob settings as R^m

Such a map could be obtained by linear, SVM or whatever regression (or
perhaps some continuous probabilistic modeling as it's obviously not a
one-to-one mapping). Then one would generate new candidates by
sampling those corresponding to low bscores (recall that lower is
better), or perhaps biasing the local search or who knows what...

Have you already considered this? Tried (in Plop perhaps)? Or rejected
for whichever reason?

Thanks for your feedback,
Nil

[Ben Goertzel]

Never tried it, but thought before that doing this using SVM (or MOSES
itself) would be a good idea...

ben

> --
> You received this message because you are subscribed to the Google Groups "plop-dev" group.
> To post to this group, send email to plop...@googlegroups.com.
> To unsubscribe from this group, send email to plop-dev+u...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/plop-dev?hl=en.
>
>

--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
CTO, Genescient Corp
Vice Chairman, Humanity+
Adjunct Professor of Cognitive Science, Xiamen University, China
Advisor, Singularity University and Singularity Institute
b...@goertzel.org

"My humanity is a constant self-overcoming" -- Friedrich Nietzsche

[Nil Geisweiller]

On Sat, Aug 27, 2011 at 10:54 PM, Ben Goertzel <b...@goertzel.org> wrote:
> Never tried it, but thought before that doing this using SVM (or MOSES
> itself) would be a good idea...

OK, perhaps Jekin, the student who's gonna work on SVM-based
optimization, might try that as well if time permits...

Nil