Search

A number of quantitative and qualitative approaches to defeasible reasoning have appeared in the artificial intelligence literature in the past decade. However, probability-based methods often languished, due to the supposed complexity of representation in terms of joint distributions. More recently, probabilistic representations based on networks have been shown to be quite compact. In expert systems especially, the use of probability theory for defeasible reasoning is now becoming more widespread. This dissertation describes how the statistical causal modeling technique called path analysis can be used as a probabilistically correct method for defeasible reasoning. Path analysis also makes use of networks, but the quantification of arcs is done by regression rather than by the conditional probabilities typical of other probabilistic methods. It is shown that the regression coefficients of path analysis in fact form a compact and computationally efficient representation for probabilistic reasoning. In addition to deriving the relationship between an extended version of path analysis and probability propagation on networks, the new method is illustrated by treating a number of problems which have appeared in the defeasible reasoning literature. Finally one of these so-called "benchmark" problems is analyzed in detail, and compared with a number of other well-known methods. It is shown that extended path analysis makes explicit the assumptions that all defeasible reasoning strategies must make, but seldom announce.