Path: utzoo!attcan!uunet!seismo!sundc!pitstop!sun!decwrl!purdue!bu-cs!mirror!rayssd!raybed2!linus!mbunix!bwk
From: bwk@mitre-bedford.ARPA (Barry W. Kort)
Newsgroups: comp.ai
Subject: Re: <7749@klaatu.rutgers.edu>
Summary: Let's get down to the elements of intelligence.
Keywords: Knowledge Representation and Reasoning Methods
Message-ID: <42739@linus.UUCP>
Date: 9 Dec 88 03:21:07 GMT
References: <9038@klaatu.rutgers.edu> <193600003@trsvax>
Sender: news@linus.UUCP
Reply-To: bwk@mbunix (Barry Kort)
Organization: IdeaSync, Inc., Chronos, VT
Lines: 57

In article <193600003@trsvax> don@trsvax.UUCP writes:

 > Perhaps more insight would result if we could somehow come up
 > with new words for machine intelligence which avoided comparisons
 > with human intelligence.

Perhaps it would be constructive if we itemized some of the structural
elements of human intelligence.  We can start with the simpler
components and work our way up to higher cognitive functions over time.

First there is the issue of knowledge representation.  We know from
Lisp that a tree is isomorphic to a list of nested lists, and a great
deal of knowledge can be represented this way.  Roget's Thesaurus is
one of the largest collection of ideas arranged in outline form.  The
Dewey Decimal system is another.

More elaborate than the tree topology is the semantic network, which
can have loops.  We know from Hypercard that the semantic network is
a useful structure for navigating through a knowledge base.  And the
success of Infocom's text adventure games suggests that humans enjoy
wandering through Markov Processes, and visiting every node.

Humans also store knowledge as metaphors, parables, and analogies,
but I have yet to understand how analogical knowledge is represented.

Humans also engage in deductive and inductive reasoning, and we know
from rule-based expert systems that networks of cause and effect
relationships can be traversed like a squirrel searching a tree
for his acorn.  Forward-chaining from hypothesis to conclusion is
the easier path.  Goal-directed backward chaining is the more
interesting challenge for diagnostic expert systems.  The Resolution
Theorem Prover provides the algorithm for Prolog and related languages.

Reasoning by analogy (model-based reasoning) will probably be the next
method to succumb to the silicon thinker.  I imagine the simplest
way a computer can recognize an analogy is by comparing the topological
structure of its knowledge bases.  If two trees or two semantic
networks bear a family resemblance, it may be possible to match them
node for node and complete the analogy.  Pattern matching of large
structures may be a formidable technical challenge, but in principle
there appears to be no theoretical obstacles.

Beyond analogy we can look forward to inferential reasoning
(discovering previously unknown cause and effect pairs), and
visual reasoning (useful for ambulatory robots).  I know that
I can walk around without bumping into walls, but I'm not
sure I can explain to a computer just how I'm doing it.
Still, if MIT can have artificial insects crawling around the
labs, I suppose others don't suffer my inability share such
knowledge.

So far, nothing I have mentioned is beyond the realm of silicon.

And as far as I can tell, the human cognitive faculties which we
value for their elegance are all fair game for the silicon Golem.

--Barry Kort