Advantages and a Limitation of Using LEG Nets in a Real Time Problem

After experimenting with a number of non-probabilistic methods for dealing with uncertainty many researchers reaffirm a preference for probability methods (Lemmer 1977) (Cheeseman 1985), although this remains controversial. The imporatance of being able to form decisions from incomplete data in diagnostic problems has highlighted probabilistic methods (Lemmer 1983) which compute posterior probabilities from prior distributions in a way similar to Bayes Rule, and thus are called Bayesian methods. This paper documents the use of a Bayesian method in a real time problem which is similar to medical diagnosis in that there is a need to form decisions and take some action without complete knowledge of conditions in the problem domain. This particular method has a limitation which is discussed. Background and Scope It has been shown (Lemmer and Barth 1982) (Lemmer 1983) that it is not necessary to form a prior distribution over the joint events of all variables in a given event space to produce a posterior probability for a given event; Component Marginal Distributions (CMDs) can be formed over the joint events of smaller related groups of variables called Local Event Groups (LEGs). Each LEG is related to other LEGs by common variables called intersection sets. The LEGs thus form a network of CMDs, related by their intersection sets, called a LEG Net. By specifying known relationships from statistical measurements and estimating higher order effects from expert knowledge (Lemmer and Norton 1986), prior probabilities can be calculated for the CMDs in a LEG Net (subject to some limitations on the structure of the Net) which are consistent with an assumed underlying distribution over all variables in the LEG Net. This estimation technique makes minimum information assumptions when constraints on a distribution are not fully specified. Once consistent distributions are made, updating techniques (Lemmer and Barth 1982) reflecting the introduction of data for known values of variables in a specific case, guarantee continued consistency of the CMDs posterior to the introduction of the data. The new distributions represent a minimum information update of the prior distributions (ibid), and the values of the unspecified variables represent posterior probabilities given the introduced data. Several knowledge engineering tools have been developed at PAR Government Sy�tems Corporation consistent with the LEG Net method of dealing with uncertainty. These tools provide the knowledge engineer with the ability to specify LEG Nets and CMDs in each LEG. They also provide the ability to test scenarios within the LEG Net, so specified, to allow a level of confidence in the resultant system (Barth and Norton 1986). There is even a tool which allows explanation of