inference sketching as a means of inference control
55 views
Skip to first unread message
Ben Goertzel
Jan 7, 2018, 5:52:19 AM1/7/18
to opencog, Nil Geisweiller, Alexey Potapov, Zarathustra Goertzel, Eddie Monroe, Michael Duncan
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
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
Alexey Potapov
Jan 8, 2018, 8:27:07 AM1/8/18
to ope...@googlegroups.com, Nil Geisweiller, Alexey Potapov, Zarathustra Goertzel, Eddie Monroe, Michael Duncan
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
--
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+unsubscribe@googlegroups.com.
To post to this group, send email to 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/CACYTDBejdBtiWWTKpdAq3M3RbomTCccSESfqQ-eBMHhWH9DR2w%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.
Nil Geisweiller
Jan 8, 2018, 11:27:09 AM1/8/18
to Alexey Potapov, ope...@googlegroups.com, Nil Geisweiller, Alexey Potapov, Zarathustra Goertzel, Eddie Monroe, Michael Duncan
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
> <mailto:b...@goertzel.org>>:
> send an email to opencog+u...@googlegroups.com
> <mailto:opencog%2Bunsu...@googlegroups.com>.
>
>
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
> <mailto:opencog%2Bunsu...@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
> <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>.
> 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>.
>
>
Ben Goertzel
Jan 8, 2018, 11:38:43 AM1/8/18
to opencog, Alexey Potapov, Nil Geisweiller, Alexey Potapov, Zarathustra Goertzel, Eddie Monroe, Michael Duncan
Alexey: Yeah the pattern miner is a crude tool but it's still a
valuable one ... sometimes statistically pruned brute force search
(which, in an Atomspace pattern context, is the pattern miner) is all
you can do...
As you know one of our TODOs for this year is to fully integrate
MOSES, PLN, Concept Blending and Pattern Miner within Atomspace using
the URE ... this will give us a much richer toolset for concept
creation and generalization ...
Sometime in a few weeks I will write up some details on my ideas/plans
for integrating all these things (I think Nil gets it already
though...)... for the next few weeks it seems I'll be swamped with
SingularityNET organizational stuff ...
-- Ben
> email to opencog+u...@googlegroups.com.
> To post to this group, send email to ope...@googlegroups.com.
> Visit this group at https://groups.google.com/group/opencog.
> For more options, visit https://groups.google.com/d/optout.
valuable one ... sometimes statistically pruned brute force search
(which, in an Atomspace pattern context, is the pattern miner) is all
you can do...
As you know one of our TODOs for this year is to fully integrate
MOSES, PLN, Concept Blending and Pattern Miner within Atomspace using
the URE ... this will give us a much richer toolset for concept
creation and generalization ...
Sometime in a few weeks I will write up some details on my ideas/plans
for integrating all these things (I think Nil gets it already
though...)... for the next few weeks it seems I'll be swamped with
SingularityNET organizational stuff ...
-- Ben
> To post to this group, send email to 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/b095ba43-f191-1a08-fbdd-d7bb8c94cdd4%40gmail.com.
> For more options, visit https://groups.google.com/d/optout.
Ben Goertzel
Jan 8, 2018, 11:39:46 AM1/8/18
to opencog, Alexey Potapov, Nil Geisweiller, Alexey Potapov, Zarathustra Goertzel, Eddie Monroe, Michael Duncan
> It seems it would be valuable to support embedded subtrees.
Yes, definitely...
> To post to this group, send email to 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/b095ba43-f191-1a08-fbdd-d7bb8c94cdd4%40gmail.com.
> For more options, visit https://groups.google.com/d/optout.
Reply all
Reply to author
Forward
0 new messages