Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!mcgill-vision!snorkelwacker!bloom-beacon!eru!luth!sunic!mcsun!unido!uklirb!shell
From: timm@runxtsa.runx.oz.au (Tim Menzies)
Newsgroups: comp.ai.shells
Subject: Re: Dempster-Shafer theory
Message-ID: <6442@uklirb.informatik.uni-kl.de>
Date: 15 Aug 90 13:52:00 GMT
References: <6390@uklirb.informatik.uni-kl.de>
Sender: shell@uklirb.informatik.uni-kl.de
Organization: RUNX Unix Timeshare.  Sydney, Australia.
Lines: 40
Approved: shell@informatik.uni-kl.de
Posted-Date: Fri Aug 17 07:05:17 GMT 1990


In article <6390@uklirb.informatik.uni-kl.de> sapong@dvinci.usask.ca (Kofi Sapong) writes:
>If anyone has any information on expert system shells that use the
>Dempster-Shafer approach to handle uncertainty, could he/she kindly
>pass it on to me. On the other hand if you have or know of any program
>that implements any form of a propagation scheme using the above 
>approach, I will interested in having it.

I got real excited about D-M a while back. I chased up the original D-M article
"Evidential Reasoning in a Hierarchy" by Gordon J. and Shortliffe E.H. in
Artificial Intelligence 26 1985, 323-357. On page 350 of that article I read:

	We have previoulst suggested in fact that the details of a model
	of evidential reasoning in an AI system may be relatively
	unimportant since the careful semantic structuring of a domain's
	knowledge seems to blunt the sensitivity of the inferences to the
	values of the numbers used.

Translation: topology and not clever combination rules make for a smart system.

I've heard anecdoctal stories of people performing sensitivity analysis on
systems where they tinkered with the certaintity factor combination rules,
threseholds of belief, etc, and found that that their semantic nets were
less sensitive to the numbers than to the connections within the net. I've
seen myself a study in which someone ran a set of rules for black-jack and
deduced what stragegies to follow based on various ways of combining
numeric certainty factors. They tried four ways, including straight averaginh
of certainty factors for the conclusions. This averaging method was included
as the "straw-man". It was intended to show how poor this method was compared
to (say) the sophistication of the MYCIN-style certainty factor kludge. 
Surprise, surprise, straight averaging woked just as well as anything else.

So, I'm not so interested in D-M anymore. Now I seek ways to organise the
topology of my categorical and-or nets.

--
 _--_|\  Tim Menzies (timm@runxtsa.oz)      PH: 02 9297729, 61 2 4280200 (fax)
/      \ HiSoft Expert Systems Group,       "Software should be altered to 
\_.--._/ 2-6 Orion Rd Lane Cove, NSW,       accomodate the quirks of human
      v  Australia, 2066                    thought, and not vice versa."