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.

Many Artificial Intelligence systems depend on the agent's updating its beliefs about the world on the basis of experience. Experiments constitute one type of experience, so scientific methodology offers a natural environment for examining the issues attendant to using this class of evidence. This paper presents a framework which structures the process of using scientific data from research reports for the purpose of making decisions, using decision analysis as the basis for the structure and using medical research as the general scientific domain. The structure extends the basic influence diagram for updating belief in an object domain parameter of interest by expanding the parameter into four parts: those of the patient, the population, the study sample, and the effective study sample. The structure uses biases to perform the transformation of one parameter into another, so that, for instance, selection biases, in concert with the population parameter, yield the study sample parameter. The influence diagram structure provides decision theoretic justification for practices of good clinical research such as randomized assignment and blindfolding of care providers. The model covers most research designs used in medicine: case-control studies, cohort studies, and controlled clinical trials, and provides an architecture to separate clearly between statistical knowledge and domain knowledge. The proposed general model can be the basis for clinical epidemiological advisory systems, when coupled with heuristic pruning of irrelevant biases; of statistical workstations, when the computational machinery for calculation of posterior distributions is added; and of meta-analytic reviews, when multiple studies may impact on a single population parameter.

BPS, the Bayesian Problem Solver, applies probabilistic inference and decision-theoretic control to flexible, resource-constrained problem-solving. This paper focuses on the Bayesian inference mechanism in BPS, and contrasts it with those of traditional heuristic search techniques. By performing sound inference, BPS can outperform traditional techniques with significantly less computational effort. Empirical tests on the Eight Puzzle show that after only a few hundred node expansions, BPS makes better decisions than does the best existing algorithm after several million node expansions

An efficient algorithm is developed that identifies all independencies implied by the topology of a Bayesian network. Its correctness and maximality stems from the soundness and completeness of d-separation with respect to probability theory. The algorithm runs in time O (l E l) where E is the number of edges in the network.

Stochastic simulation approaches perform probabilistic inference in Bayesian networks by estimating the probability of an event based on the frequency that the event occurs in a set of simulation trials. This paper describes the evidence weighting mechanism, for augmenting the logic sampling stochastic simulation algorithm [Henrion, 1986]. Evidence weighting modifies the logic sampling algorithm by weighting each simulation trial by the likelihood of a network's evidence given the sampled state node values for that trial. We also describe an enhancement to the basic algorithm which uses the evidential integration technique [Chin and Cooper, 1987]. A comparison of the basic evidence weighting mechanism with the Markov blanket algorithm [Pearl, 1987], the logic sampling algorithm, and the evidence integration algorithm is presented. The comparison is aided by analyzing the performance of the algorithms in a simple example network.

We describe a mechanism for performing probabilistic reasoning in influence diagrams using interval rather than point valued probabilities. We derive the procedures for node removal (corresponding to conditional expectation) and arc reversal (corresponding to Bayesian conditioning) in influence diagrams where lower bounds on probabilities are stored at each node. The resulting bounds for the transformed diagram are shown to be optimal within the class of constraints on probability distributions that can be expressed exclusively as lower bounds on the component probabilities of the diagram. Sequences of these operations can be performed to answer probabilistic queries with indeterminacies in the input and for performing sensitivity analysis on an influence diagram. The storage requirements and computational complexity of this approach are comparable to those for point-valued probabilistic inference mechanisms, making the approach attractive for performing sensitivity analysis and where probability information is not available. Limited empirical data on an implementation of the methodology are provided.

Bayesian inference systems should be able to explain their reasoning to users, translating from numerical to natural language. Previous empirical work has investigated the correspondence between absolute probabilities and linguistic phrases. This study extends that work to the correspondence between changes in probabilities (updates) and relative probability phrases, such as "much more likely" or "a little less likely." Subjects selected such phrases to best describe numerical probability updates. We examined three hypotheses about the correspondence, and found the most descriptively accurate of these three to be that each such phrase corresponds to a fixed difference in probability (rather than fixed ratio of probabilities or of odds). The empirically derived phrase selection function uses eight phrases and achieved a 72% accuracy in correspondence with the subjects' actual usage.

Plan recognition does not work the same way in stories and in "real life" (people tend to jump to conclusions more in stories). We present a theory of this, for the particular case of how objects in stories (or in life) influence plan recognition decisions. We provide a Bayesian network formalization of a simple first-order theory of plans, and show how a particular network parameter seems to govern the difference between "life-like" and "story-like" response. We then show why this parameter would be influenced (in the desired way) by a model of speaker (or author) topic selection which assumes that facts in stories are typically "relevant".

As the technology for building knowledge based systems has matured, important lessons have been learned about the relationship between the architecture of a system and the nature of the problems it is intended to solve. We are implementing a knowledge engineering tool called BART that is designed with these lessons in mind. BART is a Bayesian reasoning tool that makes belief networks and other probabilistic techniques available to knowledge engineers building classificatory problem solvers. BART has already been used to develop a decision aid for classifying ship images, and it is currently being used to manage uncertainty in systems concerned with analyzing intelligence reports. This paper discusses how state-of-the-art probabilistic methods fit naturally into a knowledge based approach to classificatory problem solving, and describes the current capabilities of BART.

This extended abstract presents a logic, called Lp, that is capable of representing and reasoning with a wide variety of both qualitative and quantitative statistical information. The advantage of this logical formalism is that it offers a declarative representation of statistical knowledge; knowledge represented in this manner can be used for a variety of reasoning tasks. The logic differs from previous work in probability logics in that it uses a probability distribution over the domain of discourse, whereas most previous work (e.g., Nilsson [2], Scott et al. [3], Gaifinan [4], Fagin et al. [5]) has investigated the attachment of probabilities to the sentences of the logic (also, see Halpern [6] and Bacchus [7] for further discussion of the differences). The logic Lp possesses some further important features. First, Lp is a superset of first order logic, hence it can represent ordinary logical assertions. This means that Lp provides a mechanism for integrating statistical information and reasoning about uncertainty into systems based solely on logic. Second, Lp possesses transparent semantics, based on sets and probabilities of those sets. Hence, knowledge represented in Lp can be understood in terms of the simple primative concepts of sets and probabilities. And finally, the there is a sound proof theory that has wide coverage (the proof theory is complete for certain classes of models). The proof theory captures a sufficient range of valid inferences to subsume most previous probabilistic uncertainty reasoning systems. For example, the linear constraints like those generated by Nilsson's probabilistic entailment [2] can be generated by the proof theory, and the Bayesian inference underlying belief nets [8] can be performed. In addition, the proof theory integrates quantitative and qualitative reasoning as well as statistical and logical reasoning. In the next section we briefly examine previous work in probability logics, comparing it to Lp. Then we present some of the varieties of statistical information that Lp is capable of expressing. After this we present, briefly, the syntax, semantics, and proof theory of the logic. We conclude with a few examples of knowledge representation and reasoning in Lp, pointing out the advantages of the declarative representation offered by Lp. We close with a brief discussion of probabilities as degrees of belief, indicating how such probabilities can be generated from statistical knowledge encoded in Lp. The reader who is interested in a more complete treatment should consult Bacchus [7].