Expert systems?

David Vaughan dvk at dvkconsult.com.au
Tue Oct 8 19:57:01 EDT 2002


Alex

I pulled these two comments off a 1998 Metacard digest on the web, 
responding to an enquirer who wanted to identify fish (as another 
writer said, probably better off with a neural network than a rules 
system). Anyway, FWIW, a couple of people you might be able to contact 
or advice you might prefer. Perhaps I should cautiously clarify that I 
was indeed joking (with smiley) in a previous mail when I indicated we 
might write it. My business sympathies are more with Grant's view 
below; my personal with finding something with which to play around, as 
ever.

cheers
David


This problem can be conveniently solved with a expert system. The main
advantage of an expert system is that it is a symbolic problem solver, 
wich
can eventually explain its behaviour (the reasons of the conclusion). I
remember a (very) old publication in Byte vulgarizing the principle of
expert system with the support of the identification of animals. It is
possible to write an ES with metacard, but it is not the best language 
for
this aim.
As ES are based on logic, your first objective is to determine the data
representation : propositional logic or first order logic. The 
consequences
are important for the engine of the system. In fact, the engine is 
entirely
conditioned by this option.
Then you must determine the behaviour of the engine : forward chaining
(saturation) or backward chaining (demonstration of hypothesis).
At last you can connect the elements of the rules with elements of the
interface (graphical choice of characteristics).

If you need more information, you can contact me (I'm assistan 
professor in
medical computer science, in the domain of artificial intelligence).

Sincerely,

H. Chaudet

- --------------------------------------------------------
Herve Chaudet  &  Liliane Pellegrin-Chaudet
   e-mail : [PRIVACY PROTECTION]
   Laboratoire de Biomathematiques, Informatique Medicale
       Faculte de Medecine - 27, Bd Jean Moulin
       13385 Marseille cedex 5 - France
       Tel (33) 0 491 791 910 - Fax (33) 0 491 794 013
   Service de l'Information Medicale des Hopitaux Nord
       Hopital Nord - Chemin des Bourrelly
       13326 Marseille cedex 15 - France
       Tel (33) 0 491 968 022 - Fax (33) 0 491 968 025


Subject: Re: Ruled-based, decision-tree expert system via xtalk?

on 1998/10/13 14:47 Dave Jones said:

»snip=B9

 >Does anyone have any suggestions, ideas, examples, pointer, scripts, 
etc.
 >that would lead me in the right direction to produce a rule-based,
 >decision-tree expert identification system using xtalk???

That sounds like a fascinating project.  However, you might wish to
consider whether a commercial system might already do the job, and
cheaper.  I realize that university researchers' time is valued in
today's rubles -- worthless, and shrinking fast -- but I believe there
are systems out there that have solved these sorts of problems, to one
degree or another of satisfaction, and if one of them would do the job
the world might be better served if you did those sorts of research 
which
the rest of us can't do.

Unfortunately, I'm long out of touch with what's available.  I can only
point to the university hospital where I used to work, which decided to
roll its own (in nextsteP) for a platelet transfusion  expert system.  
It
came up, and worked; but it (and to be fair, other parts of a 
physician's
workstation project) took so many people and so much time (several times
as much of each as budgeted) that the project was killed by
administrators; people, including many borrowed from other projects, 
were
laid off; those other projects outsourced, ... and that became just part
of a major IS implosion which became part of the decision to sell the
whole hospital (!) to a private chain which eviscerated it and (IMO)
turned it from a major academic medical center into a scrounging branch
office.  The roll-our-own effort was just an unquantifiably small part 
of
the overall disaster, but the lesson was clear: researchers should
research what has never been done before, in their own disciplines, and
buy the rest.

Yes, an expert-system package of the required flexibility would be
expensive, probably very expensive.  There's a reason for that: it's
very, very hard.  However much it costs, it would cost you more, and 
take
you away from your unique contributions to ichthyology.

That said, it might very well be possible for experts in knowledge
engineering to use MetaCard to create such a package.  A programming
language -- MetaTalk -- even a little closer than C++ to the language of
the problem domain (("er, um, well, let's see .. no, maybe ...") is
always an advantage, as is an interface closer to the target machines'.
Evaluating rules, especially quantified rules, is computationally
expensive; but I suspect that with just a dozen characteristics the
evaluations wouldn't make a modern system so much as hiccup, even in an
interpreted language.  The logic to evaluate properly how ever many
characteristics there are, especially non-Boolean (Bayesian) ones -- ah,
there's the rub.  There's where you want a package that's the product of
experts in *that* sort of research, so you can apply the taxonomic
expertise, developed over a whole career, that only you have.

Let me put it close to home.  I've been programming for 35 years -- 33 
of
them professionally.  I've been doing xTalk stuff for 10 years, and love
it.  I have a bachelor's in Math and Phil, with coursework in
epistemology and symbol manipulation systems.  I have a 
more-than-passing
interest in biology; and though I didn't do that transfusion system I
hung out with the folks who did, and heard their stories, and worked on
other parts of the workstation system.  And if you were my boss; and I
were a programmer in your department, chafing and squirming in COBOL; 
and
you said, "How'd you like to do an expert system in MetaCard?  We have a
year, and I can hire some additional help" -- I'd say, "Boss, you'd be
better off to buy it.  Let's you and I do something else."

And with MetaCard (and other whateverCards) there is so much else that
only you could do.  An expert system package, for anyone not 
specializing
in that kind of programming, is the La Brea tar pits.

Grant Schampel
2191 Carter Ave.
St. Paul, MN   55108-1710
U. S. A.
(651) 644-5796


-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/enriched
Size: 6687 bytes
Desc: not available
URL: <http://lists.runrev.com/pipermail/use-livecode/attachments/20021008/5a3fe9f5/attachment.bin>


More information about the Use-livecode mailing list