Path: utzoo!mnetor!uunet!mcvax!ukc!strath-cs!glasgow!gilbert
From: gilbert@cs.glasgow.ac.uk (Gilbert Cockton)
Newsgroups: comp.ai
Subject: Re: Expert Systems in the Railroad Industry (is AI needed?).
Message-ID: <1042@crete.cs.glasgow.ac.uk>
Date: 27 Apr 88 08:34:38 GMT
References: <8816@agate.BERKELEY.EDU> <73@edai.ed.ac.uk> <9226@agate.BERKELEY.EDU>
Reply-To: gilbert@cs.glasgow.ac.uk (Gilbert\ Cockton)
Distribution: comp.ai
Organization: Comp Sci, Glasgow Univ, Scotland
Lines: 30
Keywords: AI, Expert Systems, Dynamic Programming, Appropriate Technology.

In article <9226@agate.BERKELEY.EDU> lagache@violet.berkeley.edu.UUCP (Edouard Lagache) writes:
>>  Why should we want to implement an expert system, when adequate techniques 
>>exist already ?
> While, the areas mentions can clearly be solved by brute force methods,
> it is unlikely that human experts employ only those sorts of stategies
> (since human cognition doesn't support large active data structures);
> thus, there may be some interesting enhancements possible on
> conventional programming techniques by learning how human experts
> perform the tasks involved.

a) Expert systems use large active data structures.
b) Knowledge representation has not yet captured human problem solving
   methods - at best they model them at the black box level.
c) We do not know how people solve problems, we only have observations
   about behaviour under certain (often fake lab.) conditions.
d) If a mathematically sound technique is slow but sound, it is
   usually preferable to heuristic-based faster but less accurate solutions.
   With some AI tuning techniques, no measure of the technique's limits
   can be given, whereas in classic Operational Research, the
   trade-offs of quick and dirty methods can be expressed quantitatively.

I know of a problem where a mathematics department said there was no
solution to the problem, so some people started to look and see if
there was an AI solution instead - i.e. a rule-based approach.  Does
anyone know of why such an approach is credible and worth pursuing?
Surely when mathematicians say a problem has no known solution, it's
rather naive to muck about with a mathematical object like a computer
to see if it gets you anywhere?  In the railway case, what grounds are
there for thinking that 4 knowledge engineers and an expert system
shell are going to outperform a decade of OR research?