On 08/30/2016 04:41 AM, Linas Vepstas wrote:
* So, the thing that we really need, is not so much a nice surface
syntax for atomese, but a nice surface syntax for inference control. Let
me explain: right now, PLN goes off and explores in some random
direction, and promptly hits a combinatorial explosion and all is lost,
and so you invent ECAN to place some reins on this beast.   But ECAN is
this opaque automated system, and we can guess at what it might be
doing, but its hard to visualize, and maybe its doing the right thing,
and maybe its not.   So I'm thinking -- lets make inference paths -- the
sequence of steps taken during a search for a proof -- lets make that
explicitly visible, a first-class object that can be seen and
manipulated and explored.  Something that can be pruned, or allow to
grow, like a good gardener tending a plant: Lets watch the many shoots
of inference, and see which seem good and which bad, and then design an
automated system to rein in, to control the directions in which
inference goes.


I totally agree with that. Actually the forward chainer creates traces of inferences in the AtomSpace as ExecutionLinks relating rules, sources and output targets.

The next step is to mine these traces and use observed patterns to improve inference efficiency. This could be done on-line (as the inference takes place) or off-line. It's not holy clear to me how we're gonna use that knowledge to drive inference control, but I guess we can go incremental, like tweaking the hebbian links so that ECAN does a better job, these hebbian links can also be linked to rules, they can be contextualized, etc. I think there's a lot of meta-learning that can take place just by tweaking the hebbian links. The challenge is gonna be to be able to keep under control the computational power allocated to meta-learning into order to speed-up learning.

I didn't mention SampleLink but I should have, here I have.

Nil

--
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 [email protected].
To post to this group, send email to [email protected].
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/57C95A38.70205%40gmail.com.
For more options, visit https://groups.google.com/d/optout.

Reply via email to