Knowledge-Based Probabilistic Reasoning from Expert Systems to Graphical Models

An important research enterprise for the Artificial Intelligence community since the 1970s has been the design of expert or “knowledge-based” systems. These programs used explicitly encoded human knowledge, often in the form of a production rule system, to solve problems in the areas of diagnostics and prognostics. The earliest research/development program in expert systems was created by Professor Edward Feigenbaum at Stanford University (Buchanan and Shortliff 1984). Because the expert system often addresses problems that are imprecise and not fully proposed, with data sets that are often inexact and unclear, the role of various forms of probabilistic support for reasoning is important. The 1990s saw radical new approaches to the design of automated reasoning/diagnostic systems. With the creation of graphical models, the explicit pieces of human knowledge (of the expert system) were encoded into causal networks, sometimes referred to as Bayesian belief networks (BBNs). The reasoning supporting these networks, based on two simplifying assumptions (that reasoning could not be cyclic and that the causality supporting a child state would be expressed in the links between it and its parent states) made BBN reasoning quite manageable computationally. In recent years the use of graphical models has replaced the traditional expert system, especially in situations where reasoning was diagnostic and prognostic, i.e., extending from concrete situations to the best explanations for their occurrence. This type reasoning is often termed abductive. In this chapter we first (Section 1) present the technology supporting the traditional knowledge-based expert system, including the production system for reasoning with rules. Next (Section 2), we discuss Bayesian inference, and the adoption of simplifying techniques such as the Stanford Certainty Factor Algebra. We then (Section 3) introduce graphical models, including the assumptions supporting the use of Bayesian belief networks (BBN), and present an example of BBN reasoning. We conclude (Section 4) with a brief introduction of a next generation system for diagnostic reasoning with more expressive forms of the BBN.