Re: [opencog-dev] Deep Learned Crossover

[Nil Geisweiller]

Hi Fabian,

On 12/20/2017 11:18 AM, Fabian Klopfer wrote:
> 1. Have there been a try in feeding the crossover/genetic operator
> results in the inner MOSES loop into some deep net to figure out
> relevant features, advanced genetic operators or some hints on how to
> exactly cross over/ mutate the genome(s)?

No. The most advanced thing we've done with MOSES in terms of
meta-learning for feature selection is described in this paper

https://arxiv.org/abs/1703.06990

MOSES also allows to use MOSES itself dynamically perform feature selection.

> 2. Is there something like a minimal example to PLN (based reasoning) or
> sth like a formal intro to PLN in a nutshell?

See

https://github.com/opencog/atomspace/tree/master/examples/rule-engine

then move onto

https://github.com/opencog/opencog/tree/master/examples/pln

(not all pln examples work, I think amusing-friend,
chicken-feet-or-pizza, crazy-happy and which-product don't)

There's also the tutorial

https://wiki.opencog.org/w/Hands_On_With_OpenCog#Pattern_Matching
https://wiki.opencog.org/w/Hands_On_With_OpenCog#Probabilistic_Logic_Networks_.28PLN.29

>
> 3. How are MOSES and the pattern miner connected/related? Does MOSES use
> the pattern miner?

For now they aren't (well both are connected in spirit but not in code).
The idea is to have both rely on the URE
https://wiki.opencog.org/w/Unified_rule_engine

URE-based pattern mining is in the work, URE-based will code after.

The reason for moving as much intelligence on the URE is because it is
controllable by "control rules" that can ultimately be learned, see

https://github.com/opencog/opencog/tree/master/examples/pln/inference-control-learning

> 4. How are moses procedures linked to the atomspace/where are they saved?

You can either invoke MOSES with the right option to output scheme
programs that can be loaded into the atomspace. Or convert combo problem
into scheme, see

https://github.com/opencog/moses/blob/master/moses/comboreduct/main/combo-fmt-converter.cc

or moses --help for outputting directly in scheme format.

Note that output don't reflect all information are would be required to
reason about the models, such as the definitions of the training data,
fitness function and probably more.

Nil

>
> Sorry if those questions have already been answered somewhere else and
> thanks in advance
> FK
>
> --
> You received this message because you are subscribed to the Google
> Groups "opencog" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to opencog+u...@googlegroups.com
> <mailto:opencog+u...@googlegroups.com>.
> To post to this group, send email to ope...@googlegroups.com
> <mailto:ope...@googlegroups.com>.
> Visit this group at https://groups.google.com/group/opencog.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/opencog/b321a93f-2d95-4f7c-a266-85a9b3576c5b%40googlegroups.com
> <https://groups.google.com/d/msgid/opencog/b321a93f-2d95-4f7c-a266-85a9b3576c5b%40googlegroups.com?utm_medium=email&utm_source=footer>.
> For more options, visit https://groups.google.com/d/optout.