Bayesian Belief Networks have been largely overlooked by Expert Systems practitioners on the grounds that they do not correspond to the human inference mechanism. In this paper, we introduce an explanation mechanism designed to generate intuitive yet probabilistically sound explanations of inferences drawn by a Bayesian Belief Network. In particular, our mechanism accounts for the results obtained due to changes in the causal and the evidential support of a node.

We formulate Dempster Shafer Belief functions in terms of Propositional Logic using the implicit notion of provability underlying Dempster Shafer Theory. Given a set of propositional clauses, assigning weights to certain propositional literals enables the Belief functions to be explicitly computed using Network Reliability techniques. Also, the logical procedure corresponding to updating Belief functions using Dempster's Rule of Combination is shown. This analysis formalizes the implementation of Belief functions within an Assumption-based Truth Maintenance System (ATMS). We describe the extension of an ATMS-based visual recognition system, VICTORS, with this logical formulation of Dempster Shafer theory. Without Dempster Shafer theory, VICTORS computes all possible visual interpretations (i.e. all logical models) without determining the best interpretation(s). Incorporating Dempster Shafer theory enables optimal visual interpretations to be computed and a logical semantics to be maintained.

Measures of uncertainty and divergence are introduced for interval-valued probability distributions and are shown to have desirable mathematical properties. A maximum uncertainty inference procedure for marginal interval distributions is presented. A technique for reconstruction of interval distributions from projections is developed based on this inference procedure. They may represent collections of confidence intervals derived from frequency data, imprecisely stated subjective probabilities, known linear equality or inequality constraints, etc. Thus, interval distributions sometimes provide a more realistic characterization of uncertainty than do real-valued probability distributions.

Many writers have observed that default logics appear to contain the "lottery paradox" of probability theory. This arises when a default "proof by contradiction" lets us conclude that a typical X is not a Y where Y is an unusual subclass of X. We show that there is a similar problem with default "proof by cases" and construct a setting where we might draw a different conclusion knowing a disjunction than we would knowing any particular disjunct. Though Reiter's original formalism is capable of representing this distinction, other approaches are not. To represent and reason about this case, default logicians must specify how a "typical" individual is selected. The problem is closely related to Simpson's paradox of probability theory. If we accept a simple probabilistic account of defaults based on the notion that one proposition may favour or increase belief in another, the "multiple extension problem" for both conjunctive and disjunctive knowledge vanishes.

During the summer of 1989, the Forecast Systems Laboratory of the National Oceanic and Atmospheric Administration sponsored an evaluation of artificial intelligence-based systems that forecast severe convective storms. The evaluation experiment, called Shootout-89, took place in Boulder, and focussed on storms over the northeastern Colorado foothills and plains (Moninger, et al., 1990). Six systems participated in Shootout-89. These included traditional expert systems, an analogy-based system, and a system developed using methods from the cognitive science/judgment analysis tradition. Each day of the exercise, the systems generated 2 to 9 hour forecasts of the probabilities of occurrence of: non significant weather, significant weather, and severe weather, in each of four regions in northeastern Colorado. A verification coordinator working at the Denver Weather Service Forecast Office gathered ground-truth data from a network of observers. Systems were evaluated on the basis of several measures of forecast skill, and on other metrics such as timeliness, ease of learning, and ease of use. Systems were generally easy to operate, however the various systems required substantially different levels of meteorological expertise on the part of their users--reflecting the various operational environments for which the systems had been designed. Systems varied in their statistical behavior, but on this difficult forecast problem, the systems generally showed a skill approximately equal to that of persistence forecasts and climatological (historical frequency) forecasts. The two systems that appeared best able to discriminate significant from non significant weather events were traditional expert systems. Both of these systems required the operator to make relatively sophisticated meteorological judgments. We are unable, based on only one summer's worth of data, to determine the extent to which the greater skill of the two systems was due to the content of their knowledge bases, or to the subjective judgments of the operator. A follow-on experiment, Shootout-91, is currently being planned. Interested potential participants are encouraged to contact the author at the address above.

This paper presents some ideas and results of using uncertainty management methods in the presence of data in preference to other statistical and machine learning methods. A medical domain is used as a test-bed with data available from a large hospital database system which collects symptom and outcome information about patients. Data is often missing, of many variable types and sample sizes for particular outcomes is not large. Uncertainty management methods are useful for such domains and have the added advantage of allowing for expert modification of belief values originally obtained from data. Methodological considerations for using belief functions on statistical data are dealt with in some detail. Expert opinions are Incorporated at various levels of the project development and results are reported on an application to liver disease diagnosis. Recent results contrasting the use of weights of evidence and logistic regression on another medical domain are also presented.

In two recent papers, I have proposed a description of decision analysis that differs from the Bayesian picture painted by Savage, Jeffrey and other classic authors. Response to this view has been either overly enthusiastic or unduly pessimistic. In this paper I try to place the idea in its proper place, which must be somewhere in between. Looking at decision analysis as defeasible reasoning produces a framework in which planning and decision theory can be integrated, but work on the details has barely begun. It also produces a framework in which the meta-decision regress can be stopped in a reasonable way, but it does not allow us to ignore meta-level decisions. The heuristics for producing arguments that I have presented are only supposed to be suggestive; but they are not open to the egregious errors about which some have worried. And though the idea is familiar to those who have studied heuristic search, it is somewhat richer because the control of dialectic is more interesting than the deepening of search.

We show an approach to automated control of machine vision systems based on incremental creation and evaluation of a particular family of influence diagrams that represent hypotheses of imagery interpretation and possible subsequent processing decisions. In our approach, model-based machine vision techniques are integrated with hierarchical Bayesian inference to provide a framework for representing and matching instances of objects and relationships in imagery and for accruing probabilities to rank order conflicting scene interpretations. We extend a result of Tatman and Shachter to show that the sequence of processing decisions derived from evaluating the diagrams at each stage is the same as the sequence that would have been derived by evaluating the final influence diagram that contains all random variables created during the run of the vision system.

It is suggested that an AI inference system should reflect an inference policy that is tailored to the domain of problems to which it is applied -- and furthermore that an inference policy need not conform to any general theory of rational inference or induction. We note, for instance, that Bayesian reasoning about the probabilistic characteristics of an inference domain may result in the specification of an nonBayesian procedure for reasoning within the inference domain. In this paper, the idea of an inference policy is explored in some detail. To support this exploration, the characteristics of some standard and nonstandard inference policies are examined.

This paper argues that the principal difference between decision aids and most other types of information systems is the greater reliance of decision aids on fallible algorithms--algorithms that sometimes generate incorrect advice. It is shown that interactive problem solving with a decision aid that is based on a fallible algorithm can easily result in aided performance which is poorer than unaided performance, even if the algorithm, by itself, performs significantly better than the unaided decision maker. This suggests that unless certain conditions are satisfied, using a decision aid as an aid is counterproductive. Some conditions under which a decision aid is best used as an aid are derived.