----- Original Message ----
From: Mike Tintner <[EMAIL PROTECTED]>
To: [email protected]
Sent: Thursday, March 27, 2008 5:30:12 PM
Subject: Re: [agi] Microsoft Launches Singularity
DIV {
MARGIN:0px;}
Steve,
Some odd thoughts in reply. Thanks BTW for
article.
1. You don't seem to get what's implicit in the main point - you can't reliably
work out the sense of an enormous number of words by any kind of word lookup
whatsoever. How do you actually work out how to "handle the object" - the
slimy, slippery twisted ropey thing-y, or whatever? By looking at it. By
looking at images of it - either directly or by entertaining them mentally -
not consulting any kind of dictionary or word definitions at all. By imagining
what parts of the object to grip, and how to configure your hands to grip it.
Steve: Sorry that I missed that. But your clarifying issue is quite
interesting. Let me try to tease appart your scenario and explain how the
envisioned Texai system would process the command "handle the object". I
assume that you agree that an AGI designed to our mutual satisfaction should in
principle be able to process that particular command with at least the same
competence as a human. So the issue for me is to explain in brief how Texai
might do it.
First I assume that Texai has a body of commsense knowledge about, and skills
applicable to, the kinds of objects that can be handled. If not, then there
is a knowledge acquisition phase, and skill acquistion phase, that must be
completed beforehand.
Second, I assume that the linquistic concepts are expressed internally by the
system as symbolic terms. Many terms, for example objects that can be handled,
are grounded to the real world by an abstraction hierarchy. Descending down
this hierarchy, objects are represented less and less as symbols in logical
statements, and more and more as clustered feature vectors, and perhaps, at the
lowest levels, as no internal state at all - just sensors and actuators in
contact with the real world.
Thirdly, I distinguish between the understanding the command "handle the
object" and generating the behavior required to perform the command. I think
that you are conflating these two notions to make the scenario more difficult
that it otherwise would be. Perhaps as you know, Texai is a hierarchical
control system. I expect that skills will be present to handle various kinds
of objects, so for me the issue is to determine the correct skill to invoke in
order to perform the given command. As I explained in my previous post, Fluid
Construction Grammar does not determine semantics by word lookup, rather it
looks up constructions, which might be words, but often are not.
Given these assumptions of mine, your scenario suggests that the object to be
handled is one for which the system has no previous skill, or for which the
existing skill cannot be recognized as applicable to the given object. Because
I now building a bootstrap dialog system, that is motivated entirely by the
need to process novel situations, I am tempted to say that the system should
simply ask the user to teach it how to handle the novel object, or to ask if an
existing skill can be applied to the given object. However, lets move beyond
this approach, and I'll explain how the system uses existing perception and
planning skills to handle the given object.
By way of simplification, I'll assume your intent when asking the system to
"handle the object" means to pick it up with some physical actuator. And I'll
preface my explanation of this step by stating without proof that this task is
analogous to those already solved by state-of-the-art, urban, driverless cars,
e.g. "drive yourself to location X", where the driverless car has never been to
X. Rather than a futile attempt to explain all cases that come to mind, I'll
discuss a couple to give a flavor my approach.
Case 1 The system can sense that the novel object is not dangerous and cannot
be easily destroyed by its actuators. Then I propose that the first strategy
tried should be to pick it up in the most direct fashion, and compensate in
subsequent attempts for failure modes that resulted from from the earlier
attempts. This is like the pole balancing task that can be accomplished by
connectionist methods and no symbolic planning.
Case 2 The system senses that the actions to pick up the object are not subject
to experimentation, but must be performed correctly on the first attempt. For
this task, the system must observe all the object state that it can to remove
uncertainty. It must create a symbolic model of the object and its dynamics at
the right level of abstraction, and perform planning using symbolic
representions of its possible actions in order to create a trajectory that
satisfies the command to "handle the object". Then it must execute the plan,
repairing the plan as needed as problem state evolves that was not planned in
advance for (e.g. the object starts slipping from the system's grasp). At
lower abstraction levels, reactive behavior can substitute for planning (e.g.
when slippage is detected by a sensor, tighen the gripping actuator).
2. This discussion brings up an interesting
question. I suspect that there is a great deal of selectivity going into what
texts NLP chooses to process - and that they don't include how-to,
instructional
texts, like recipe books (and most educational texts), which tell you to
do things - like "take a cup," "add water etc" - and deal with a real world
situation, in-the-world. (?) If you're dealing more in historical
texts, - "the cat sat on the mat" etc - you don't have to confront the
open-ended nature of words, quite so violently. Hey, the cat did some kind of
sitting - as long as that's possible, who cares exactly what kind it was? But
if
you're a cool cat told to "sit" on a real mat that happens to be full of
objects
- , and you have to put those instructions into deeds rather than more words, -
you care, and words' open-endedness becomes apparent.
Steve: I agree with your insight. Much of NLU research is now focused on
either information / document retrievel, or machine translation. My main gripe
while at Cycorp was that Cyc, in the same fashion you describe, concentrated on
being taught facts and rules and then deductively answering queries. But what
could Cyc do beyond that? An AGI aspiring system should be capable of
representing skills (e.g. codelets or procedures), acquiring them by being
taught, and able to perform them as commanded, or on its own initiative. I
speculate that it will be easier to ground linguistic symbolic terms in the
rather precise world of computer programming and algorithms, but that truth
remains to be seen. (e.g. Texai, compile and run the unit tests for the program
that we wrote yesterday).
3. While philosophically, intellectually, most people dealing with this area
may expect words to have precise meanings, they know practically and
intuitively that this is impossible and work on the basis that words can have
different meanings according to who uses them - and that they themselves keep
shifting their usage of words. Philosophers, for example may argue
philosophically that words can and should have precise meanings and be treated
as true or false, but know in practice that pretty well all the major
words/concepts in philosophy, like "mind"/"consciousness"/"determinism" - have
multiple, indeed endless definitions. Or just think about AGI'ers and
"intelligence."
Steve: Actually at Cycorp, at one time we had dozens of Ph.D. philosophers
whose responsibity was to add precise symbolic concepts to the Cyc knowledge
base. The company likewise had a smaller staff of Ph.D. computational
linguists whose job was to interface NLP to the rather precise Cyc concepts.
My experiences at Cycorp with their parsers (i.e. Link Grammar, HPSG, Stanford
Parser & Charniak Parser) also have strongly influenced my choice to embrace
Fluid Construction Grammar. Despite the current lack of English coverage in
FCG, there is much less impedence mismatch between sytactic form and semantics.
IOW any general intelligence that wants to successfully use language must have
a metacognitive/ metalinguistic level of thought - where it asks explicitly,
as we do, "what does that word mean?"/"do I like that definition?" / "is it
reliable?" / "how should I use/order words?" / "what is the best kind of
diction when talking about this subject?".Life's complicated!
Steve: Given this statement, you might agree with my bootstrap English dialog
approach, in which metalinguistic skills are the first ones hard-coded.
P.S. If you haven't read, I recommend Lakoff's Case Study on "Over" at end of
"Women, Fire and Dangerous Things" - shows vast number of meanings and schemas
that can be attached to that word - and amplifies this discussion.
Steve: No I actually do not yet have this text by Lakoff, but I have some
recent experience with another preposition "on". In my first use case "the
book is on the table", I accomodate the following alternative interpretations
in order to test my design for disambiguation:
book - a bound book copybook - a sheath of paper, e.g. match bookis - has as an
attributeis - situation described ason - an operational device"on the table" -
subject to negotiation [ a multiword construction ]
on - located on the surface ofI hope you don't mind me using your issues to
explain how Texai should work.
-Steve
Stephen L. Reed
Artificial Intelligence Researcher
http://texai.org/blog
http://texai.org
3008 Oak Crest Ave.
Austin, Texas, USA 78704
512.791.7860
____________________________________________________________________________________
Looking for last minute shopping deals?
Find them fast with Yahoo! Search.
http://tools.search.yahoo.com/newsearch/category.php?category=shopping
-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
http://www.listbox.com/member/?member_id=8660244&id_secret=98558129-0bdb63
Powered by Listbox: http://www.listbox.com
