STAR@LAVALVM1.BITNET (Spencer Star) (03/14/88)
First I saw a surprising reference to Spiegelhalter quoting him as saying that probabilites were inappropriate for representing uncertainty in expert systems, and then in the March 8th AI LIst, a reference to the original quote. I just don't believe Spiegelhalter made those remarks, or at least they should be put in a context that reflects his views on probabilities. Here's why. > We shall not attempt to review the long, and sometimes acrimonious, > debate as to whether probability theory is an appropriate tool in > this context [expert systems];...Finally the 'statistical/engineering' > model adheres to the probability calculus, justified both from a > theoretical perspective (Lindley, 1982, 1987) and from the pragmatic > claim that it alone provides flexible and operational means of assessment, > criticism and learning (Cheeseman, 1985; Spiegelhalter, 1987). Pearl (1986a) > also argues for probabilistic structuring in expert systems as > providing a good model for human understanding and memory. S.L. Lauritzen and D.J. Spiegelhalter "Local computations with probabilities on graphical structures and their application to expert systems" Oct. 1987 The referenced articles include Lindley, "Scoring rules and the inevitability of probability" Internat. Stat. Review, 50, 1982. and Cheeseman, "In defense of probability" AAAI-85. To put it simply, Spiegelhalter is one of the researchers most committed to putting a probabilistic approach to work in expert systems. My own view is that a subjective probability approach appears to be a better choice than either fuzzy sets or Dempster-Shafer' belief functions because it is the only approach that has the characteristics of 1. Being based on a few simple, acceptable axioms. 2. Being able to connect directly with decision theory (Dempster-Shafer can't) 3. Having efficient algorithms for computation (The Laruitzen-Spiegelhalter paper cited above gives one; Pearl gives another) 4. Being well understood. (Look at what people are doing with Dempster-Shafer belief functions or fuzzy sets. People are not agreed as to what their fundamental theory says. However, if someone prefers one of the other approaches, fine. It's really a question of whether someone wants to work on the mainstream approach, which is Bayesian subjective probabilities or Bayesian decision theory, or if a more experimental approach is preferred, such as fuzzy sets or belief functions. Spencer Star (Bitnet: star@lavalvm1)