almond@alta.stat.washington.edu (Russell Almond) (10/10/90)
Archive-name: belief/03-Oct-90 Original-posting-by: almond@alta.stat.washington.edu (Russell Almond) Original-subject: Re: Dempster-Shafer theory Archive-site: hustat.harvard.edu [128.103.28.101] Reposted-by: emv@math.lsa.umich.edu (Edward Vielmetti) On shells available using Dempster-Shafer theory: I am the developer of a package of Common Lisp functions for implementing the belief function (Dempster-Shafer) theory. The program is called BELIEF and it operates in both the belief function and ordinary (Bayesian) probability modes (in the latter case it implements the Lauritzen and Spiegelhalter algorithm (JRSS, Series B, 1988). It is available via anonymous ftp from hustat.harvard.edu. It is still currently in a test state (as a matter of fact, I have corrected some bugs which I have not yet posted), but it works fairly well. The operator end of the user interface is very primitive (I'm a statistician, not a computer scientist), but the tools available for specifying knowledge bases are well designed. A TeX version of the operator's manual can be gotten from the same source. It can also be optained from Harvard University, Departement of Statistics (Technical Report S-128). A detailed description of the algorithms involved is in my disseration (Tech. Report S-130). As I am interested in pursing this research and having people use the program, I will offer support over the Network. On the applicability of the theory of belief functions: All of the "problems" I've seen with applications of the theory of belief functions has come down to the following scenario. An expert system builder, usually without formal training in statistics, tries to apply the theory without making formal assumptions in the model. The informal assumptions he makes are incorrect, and the result is a paradox. This has always been a problem in statistics, and usually the first few weeks of any statistics course involves describing the pitfalls that careless assumptions produce. A statistician knows that she will need to carefully justify each assumption she makes (especially independence assumptions). The situation is no different in the more general belief function theory. Unfortunately, sloppy thinking leads to sloppy answers. This is true if the model you are using is first order logic, probability, belief function, fuzzy logic or the human brain. In order to get a good expert system using belief functions, one must apply the same model critisism/refinement procedures that one would apply to any other knowledge base. Work is only starting on those ideas for Bayesian and belief function approaches. I would be happy to discuss specific examples. --------------------- Russell Almond Dept. of Statistics, GN-22 U. of Washington Seattle, WA 98195 (206) 543-4302 almond@stat.washington.edu