Reasoning about evidence in causal explanations

Abstract

Causal models can provide richly-detailed knowledge bases for producing explanations about behaviors in many domains, a task often termed interpretation or diagnosis. However, producing a causal explanation from the model can be time-consuming. This paper describes a system that solves a new problem by recalling a previous, similar problem and modifying its solution to fit the current problem. Because it is unlikely that any new problem will exactly match a previous one, the system evaluates differences between the problems using a set of evidence principles that allow the system to reason about such concepts as alternate lines of evidence, additional supporting evidence, and inconsistent evidence. If all differences between the new situation and the remembered situation are found to be insignificant, the previous causal explanation is adapted to fit the new case. This technique results in the same solution, but with an average of two orders of magnitude less effort. The evidence principles are domain independent, and the information necessary to apply them to other domain models is described.