Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!uunet!tut.cis.ohio-state.edu!ucbvax!bloom-beacon!eru!hagbard!sunic!dkuug!iesd!kgo From: kgo@iesd.auc.dk (Kristian G. Olesen) Newsgroups: comp.ai Subject: Re: info wanted on learning probabilities Message-ID: <1991Feb4.113103.29237@iesd.auc.dk> Date: 4 Feb 91 11:31:03 GMT References: <1991Jan30.213055.25485@cs.ucla.edu> <15562@milton.u.washington.edu> Sender: news@iesd.auc.dk (UseNet News) Organization: CS and Math, University of Aalborg, Denmark Lines: 34 A short version of Spiegelhalter and Lauritzens work on learning probabilities can be found in: Spiegelhalter and Lauritzen: Techniques for Bayesian Analysis in Expert Systems. Annals of Mathematics and Artificial Intelligence, 2 (1990) 353-366. As pointed out by Almond there are some problems involved with e.g. inclomplete data. At Aalborg university we're currently working on a prototype implementation of the scheme where uncertainty on conditional probabilities are modelled as Dirichlet distributions. A series of experiments with this prototype aims at a clarification of the strengths and limits of the method. We're dealing with learning as well as adaption of conditional probabilities as cases become known. Currently four factors are being investigated: 1. Significance of the precision of original conditional probabilities. 2. Observational schemes (incomplete data). 3. Different types of prior distributions. 4. Learning schemes (learning, adaption). The results so far seems promising, but it's still to early to conclude. I do, however, feel confident that practically applicable methods will turn up. If succesfull the method will be integrated in the HUGIN shell. Kristian G. Olesen Aalborg University Institute of Electronic Systems Frederik Bajers Vej 7 D 9220 Aalborg Ost Denmark Phone +45 98 15 85 22, 4960 E-mail: kgo@iesd.auc.dk