Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!agate!pasteur!ames!lll-winken!uunet!mcvax!unido!uklirb!shell
From: rehbold@uklirb (Robert Rehbold AG Richter)
Newsgroups: comp.ai.shells
Subject: Expert System Complexity
Message-ID: <4638@uklirb.UUCP>
Date: 23 Apr 89 17:40:02 GMT
Sender: shell@uklirb.UUCP
Lines: 37
Approved: shell@uklirb.uucp

When I visited a workshop on diagnostic expert systems some weeks ago the
discussion came to the topic of a "good and expressive" measure for the size 
and complexity of an expert system. Most participants disliked the idea to
use the number of rules as the only characterization even for purely
rule-based systems, since some systems use variables in their rules and
some don't, thus needing a rule for each possible value of each variable.

Some other interesting criterions may be
- number and complexity of the database objects
- in diagnosis: number of symptoms used and diagnoses included
- ability to explain
- ...

In the March issue of the CACM V.E.Barker and D.E.O'Connor give a case study on
"XCON and beyond". As a characterization of the X... expert systems the
mention not only the rule counts and type (OPS-5), but also the average number
of condition and action elements per rule. Moreover they give the average
number of attributes per condition element and the size and attribute-per-
object complexity of the database. This description seems sufficient, since
most readers will know OPS-5 and its expressiveness. Difficulties may arise
when using not-so-common rule systems or expert systems not mainly based
on rules, such as model-based systems (i.e. systems that draw their
conclusions from knowledge of structure, function and behavior).

I would like to start a discussion on what YOU think a "good" measure for
expert system complexity. Maybe we should stick to rule-based systems first,
but ideas for a more general measurement are welcome too! Since this is
comp.ai.shells, we should also discuss the influence of different shells on the
solutions of a given problem, which may be quite different in approach and
complexity.

Robert Rehbold

| e-mail:    rehbold@uklirb.UUCP     or       ..!mcvax!unido!uklirb!rehbold  |
| real name: Robert Rehbold, University of Kaiserslautern                    |
|            Department of Computer Science                                  |
|            PO-Box 3049, D-6750 Kaiserslautern, West-Germany                |