[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Seeking professional opinion on PC-based "expert system shells"
Date: Mon, 30 Jul 90 13:38 EDT
From: pan@ATHENA.PANGARO.dialnet.symbolics.com (Paul Pangaro)
... for colleague who wants to get feet wet. Does the Symbolics user
community have a sense of the market, in a range from "toy to play with"
up to "not-too-expensive but can do some good experimentation and not
learn bad habits"? Any advice (pro- or con-) appreciated.
My own perception of the pc-shell market (I'm calling a pc expert system
shell any program that incorporates inferencing) goes from
< $100 = (possibly (?) educational) toy to play with, e.g.
turbo-prolog or some other minimal rule based system
$7000 => $10000 = software capable of doing useful work
(though the hardware may not be). These are often pc versions
of the shells that cost several tens of thousands for
workstation or mainframe versions. (e.g. KEE, AION )
Many people have written books and/or review articles on this subject.
I don't know which are best but lot's of people seem to use a book
written by Harmon and King (whose title escapes me now but I could look
it up if needed). Also I believe Feigenbaum has written something.
There are many issues involved in choosing such a shell, including core
issues such as knowledge representation, inferencing paradigms, truth
maintenance, explanation facilities plus "peripheral" issues such as
user interface and foreign language interface.
You pretty much get what you pay for. The cheaper shells have less
functionality and a short learning curve. The pricey stuff has
extensive functionality and a very steep learning curve.
How much can you learn by playing with the cheap shells? Good question.
Even the cheapest shells illustrate some aspects of knowledge representation
and inference, and I believe much can be learned by using them as long as
one avoids the misperceptions I describe next. (In particular, there are
some nice versions of OPS5 for production systems at ~$200 ?)
there are two ways one might draw incorrect conclusions about
expert systems by learning from these shells:
(1) NON-LINEAR SCALEUP - You can buy prolog, put in a few rules, and in
short order you have a program that will name your Uncle. However,
it's probably not obvious that creating a program to drive your car will
require not only lots more Horn clauses but also a much more
sophisticated architecture employing a variety of knowledge
representation techniques. It is the design and implementation of this
architecture which is the most difficult part of building an expert
system, and no sense of this difficulty is present in the cheaper
(2) INCOMPLETE IMPLEMENTATIONS OF INFERENCE PARADIGMS - For example,
turbo prolog requires typing of the predicate arguments. Thus one of
the most powerful elements of complete prolog, general unification, is
missing. Another example I've seen is the TI Personal Consultant shell
(~$500 ?). This shell does backward chaining, but it turns out it can't
backtrack. I.e., If the backward chaining reaches a dead end you want
the system to unravel it's variable bindings in a forward direction
until it gets back to a new node from which it can again try backward
chaining, possibly reusing rules with new bindings. It turns out the
TI Personal Consultant shell can't reuse it's rules.