Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!uunet!lll-winken!iggy.GW.Vitalink.COM!widener!dsinc!netnews.upenn.edu!msuinfo!news
From: sticklen@pleiades.cps.msu.edu
Newsgroups: comp.ai
Subject: Re: Unified model for knowledge representation? (Impossible)
Message-ID: <1991Jun3.110003.13773@msuinfo.cl.msu.edu>
Date: 3 Jun 91 11:00:03 GMT
References: <1991May31.164608.238@nmt.edu> <5255@syma.sussex.ac.uk>
Sender: news@msuinfo.cl.msu.edu
Organization: Comp Sc, Mich State Univ
Lines: 103

In article <5255@syma.sussex.ac.uk> aarons@syma.sussex.ac.uk (Aaron Sloman)
writes:
>francia@nmt.edu (Guillermo A. Francia) writes:
>
>> Date: 31 May 91 16:46:08 GMT
>> Organization: New Mexico Tech
>>
>>
>> In as much as there are a number of means to represent (or code)
>> knowledge in AI, I would like to know if anyone has done any
>> research on a UNIFIED (mathematical) model of knowledge representation.
>> Does anyone out there know of any work done on this?
>>
>> --------------------------------------------------------------------
>> francia@jupiter.nmt.edu                     Guillermo A. Francia III
>>                                           P.O. Box 2335 CS
>> francia@minos.nmt.edu                     New Mexico Tech     
>
>Some of the people who have worked on first order logic have thought
>it could serve as the universal (= unified?) notation for
>representing knowledge. I don't know if this is what you mean by a
>model.
>
>However, it seems pretty obvious from the history of science and
>culture that different formalisms are useful for different purposes,
>including, logic, algebra, pictures, maps, tables, flow-charts,
>musical notation, 3-D models, natural languages, etc.
>
>One reason for this is that different kinds of notations have
>different kinds of variability, which limits their expressive power
>in different ways. This can sometimes be important when exploring
>large search spaces. If the structure of the notation is such that
>it won't let you express things that would only be rejected anyway,
>it can have great heuristic power.
>
>Equally if the structure of the notation is such that certain
>frequently used forms of inference can be done with simple
>algorithms, that is also useful. (E.g. roman numerals are very good
>for addition: simply concatentate the numerals, provided that the
>numbers you are adding total no more than III ! Arabic numerals are
>an ad-hoc but very useful notation, good for a wider range of
>operations, but not all that powerful when it comes to taking square
>roots, dividing large numbers that are not powers of 10, etc.)
>
>AI has so far explored only a tiny subset of possible forms of
>knowledge representation, mostly the ones that are easiest to
>manipulate using current computers and programming languages.
>
>I don't think that anyone will ever find a
>    "UNIFIED (mathematical) model of knowledge representation"
>unless it takes the form of a sort of meta-theory explaining why
>different models are needed for different knowledge domains
>and different kinds of uses of knowledge.
>
>If anyone claims to have a unified model, ask him/her if it will
>serve for the purposes of representing knowledge about the contents
>of the current optic array in a robot's visual system, and for
>transforming that knowledge in the process of discovering what's out
>there (e.g. discovering binocular disparities for stereo vision) or
>for fine control of posture and actions, etc.
>
>Aaron Sloman,
>School of Cognitive and Computing Sciences,
>Univ of Sussex, Brighton, BN1 9QH, England
>    EMAIL   aarons@cogs.sussex.ac.uk
>or:
>            aarons%uk.ac.sussex.cogs@nsfnet-relay.ac.uk
>


I'd like to comment on Aaron's thrust that "unified mathematical models of
representation" may be misguided (my word, not Aaron's) and that if some
unification is possible it will take place at the level of an architecture that
selects problem solving methods for given use.

The point of what in knowledge-based systems is being called "task specific
architectures" (TSAs) is to suggest that problem solving is best analyzed and
"mimiced/implemented" by using primitives that are used in the domain of the
problem solving. For example, in a diagnostic domain, it is natural to both
analyze diagnostic problem solving, and to build computer versions of
diagnostic problem solving, by using the concept of "diagnostic hypothesis". 
Specializations of the TSA concept like Chandrasekaran's generic tasks (GTs)
take one additional step of saying that there exists a finite set of problem
solving strategies that are generally useful. These GTs are defined by giving
both a knowledge representation template, and a control strategy. For example,
there is a GT for classification problem solving, one for simple ("routine")
design, one for function-based model level reasoning... 

At the same time as researchers seek to extend the capability of TSAs like
those in the generic tasks, grand architectures for problem solving are being
developed to handle any problem solving situation - SOAR may be the easiest
example. SOAR generates problem spaces appropriate for a given problem.
Although not a developed capability yet, the SOAR architecture may one day be
able to generate very tailored problem spaces after analyzing a given problem
situation - perhaps problem spaces not unlike GTs.  

If that happens, then it will be a unification. But a unification along the
lines that Arron pointed to, not a unification at the base level of saying "one
representation fits all." 


  Jon Sticklen
  AI/KBS Laboratory