MOSES+RNN update 8/19

[Joel Lehman]

I've updated my wiki:
http://www.opencog.org/wiki/Extending_MOSES_to_evolve_Recurrent_Neural_Networks#Current_Progress
And branch:
https://code.launchpad.net/~lehman-154-gmail/opencog/moses-rnn

In summary, I tested MOSES+RNN on the hardest variant of
pole-balancing, and unfortunately, the reductions
did not help performance. I'll be polishing things this final week,
but will continue (time permitting) to try and
figure out why the reductions are not helping and eventually trying
out a different optimization routine as it
becomes available.

[Ben Goertzel]

Hi Joel,

You've done a great job on the project and it's too bad you didn't get
to test the obvious hypothesis as to why two-pole-balancing isn't
getting solved effectively [i.e. that the univariate search is
inadequate...]. Hopefully you or someone else will be able to follow
up on this during the fall...

I think that **if** annealing or some other multivariate search makes
a big difference in the two-pole balancing performance (or some other
interesting improvement makes that difference), then we'll have the
basis for a nice journal paper here...

It actually wouldn't be too hard to apply MOSES-RNN to some practical
dataset... we have a lot of biology datasets lying around for which
one can do classification or nonlinear curve-fitting and compare with
other algorithms ... which is useful for generating a publication.
But, I agree with you that the failure (so far) on the relatively
simple 2-pole-balancing task indicates a shortcoming in the approach
as currently implemented, that needs to be remedied...

So, keep this in mind when you finish the documentation and clean up
the code ;-) .... If I do recruit someone else to follow up on this
work (or help you follow up on the work), the odds of them actually
doing anything are way higher if the code and docs are usable by
them!! I do have one enthusiast who is interested in working
part-time on this during the fall...

-- Ben

--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
Director of Research, SIAI
b...@goertzel.org

"Truth is a pathless land" -- Jiddu Krishnamurti

[Nil Geisweiller]

BTW, contin for hillclimbing will likely be ready this week. SA (which
will support contin as well) is ready too but needs some bug fixing
and improvements.

Nil