Regarding the pattern mining challenge:
It seems it would be valuable to support embedded subtrees. For instance
given ground terms
Evaluation
Predicate "P"
List
Concept "A"
Evaluation
Predicate "P"
List
List
Concept "A"
...
we have currently no way to express queries like:
atoms starting with Evaluation P ... and containing Concept "A"
somewhere down their tree.
One can do that by inserting a scheme function call in the query but
that's neither convenient nor efficient, and ultimately unnatural to
build, which the pattern miner needs to do.
Regarding the guided inference challenge:
Unless we have an obscene number of rules (sketch-based or other), I
don't think it should be too problematic because decision is cheap (at
least as currently implemented as Thomson bandit, close to linear with
the number of rules I believe, assuming rules are cheap to evaluate
though). However, I'm not seeing precisely how resources allocated to
each sketch would be managed, it feels that it already requires to jump
to another meta-level. I don't know.
I agree with Alexey that good proof sketching requires good
conceptualization. I suppose it's kinda of a chicken-egg problem, isn't
it? Yet more turtles...
Nil
On 01/08/2018 03:27 PM, Alexey Potapov wrote:
Ben,
here are some of my random thoughts on this.
Indeed, loose reasoning over generalized concepts should be very
important for AGI, and proof sketching seems an interesting analogy
here. However, there are others. E.g. in Heuristic Search, there were
attempts to generalize states and transitions between them, and
to search in this greatly reduced search space first. Unfortunately, I
don't know any general and interesting solution here. In deep
reinforcement learning, there also was a paper on learning both a space
of generalized states and a policy for it. I don't believe that such
deep learning models will scale up to complex symbolic domains, but
'theorem proving' approach might also be too restrictive...
I have been thinking about this topic for a while recently, and I
believe that inference control should be tightly connected with
conceptualization. We rarely can find patterns in inference trees per
se, but humans usually construct new concepts, in terms of which they
can describe (domain-specific) inference rules. E.g. in the Go game,
players use quite abstract notions ('wall', etc.) and reasoning over
them (building a wall here will protect the territory and spread the
influence). Such rules and concepts can be mined not in the inference
trees, but in historical data of agent-environment interactions...
So,
- Most inference rules are domain-specific rules, and they involve
concepts constructed specifically to be used in these rules (one can go
further and say that most of our concepts are inference control
concepts, but it sounds too radical)
- There are just a few general inference rules (e.g. entities, which are
similar w.r.t. some properties, might be similar w.r.t. other
properties). These rules involve general concepts (e.g. similarity),
which can be either pre-defined, or which can also be constructed
together with these rules for these rules to work (e.g. similar entities
are entities for which this inference rule works). Such rules based on
predefined abstract concepts and relations can be found by Pattern
Miner, but this is of limited interest.
- Inference/reasoning is an abstracted simulation/prediction. There
should be no huge difference in constructing higher-level concepts from
experience and from inference trees.
- Generalization is an extremely non-trivial task. And what I see is
that OpenCog is very refined in the part of reasoning, but it uses very
simplistic Pattern Miner for generalization. Obviously, we cannot use
anything heavier at the scale of the whole Atomspace, but for isolated
domains, this should necessarily be done. Well, there is also MOSES in
OpenCog, but it is also somewhat specialized, and not deeply integrated...
Well... 'Proof sketching' for inference control is the step in the right
direction, but we should focus much more on a stronger generalization...
-- Alexey
2018-01-07 13:52 GMT+03:00 Ben Goertzel <[email protected]
<mailto:[email protected]>>:
Nil, Zar, Alexey, Eddie, Mike, anyone else interested,
Attached file "inference-sketch-notes.pdf" outlines some speculative
thinking i've been doing regarding using "proof sketching" as a means
of PLN inference control...
(the other attached file contains, toward the end, an example bio-AI
inference that I use as an example in the document...)
Nil, I am throwing these ideas out here now in part because they
present a potentially important use-case for the integration of
rule-engine inference into the pattern miner, as you're in the midst
of working on...
-- Ben
--
Ben Goertzel, PhD
http://goertzel.org
"In the province of the mind, what one believes to be true is true or
becomes true, within certain limits to be found experientially and
experimentally. These limits are further beliefs to be transcended. In
the mind, there are no limits.... In the province of connected minds,
what the network believes to be true, either is true or becomes true
within certain limits to be found experientially and experimentally.
These limits are further beliefs to be transcended. In the network's
mind there are no limits." -- John Lilly
--
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]
<mailto:opencog%[email protected]>.
To post to this group, send email to [email protected]
<mailto:[email protected]>.
Visit this group at https://groups.google.com/group/opencog
<https://groups.google.com/group/opencog>.
To view this discussion on the web visit
https://groups.google.com/d/msgid/opencog/CACYTDBejdBtiWWTKpdAq3M3RbomTCccSESfqQ-eBMHhWH9DR2w%40mail.gmail.com
<https://groups.google.com/d/msgid/opencog/CACYTDBejdBtiWWTKpdAq3M3RbomTCccSESfqQ-eBMHhWH9DR2w%40mail.gmail.com>.
For more options, visit https://groups.google.com/d/optout
<https://groups.google.com/d/optout>.
--
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/b095ba43-f191-1a08-fbdd-d7bb8c94cdd4%40gmail.com.
For more options, visit https://groups.google.com/d/optout.
