Explaining the Link between Causal Reasoning and Expert Behavior

摘要

Causal reasoning can be a powerful tool, but expert diagnosticians don't seem to use it extensively in everyday practice. Yet, being able to provide the causal rationale that underlies a diagnosis or other medical decision seems to be critical in providing satisfying explanations and justifications of that decision. Thus, expert systems are presented with a paradox. It appears that they should reason non-causally in most circumstances, but still have access to the causal rationale behind their decisions for providing explanations. In this paper, we present a paradigm for expert system construction that provides that capability. In our approach, causal reasoning that is performed while the expert system is being designed does not appear in the expert system itself. But because the design process is recorded in a machine readable form, explanation routines have access to that causal reasoning and thus can justify an expert system's behavior with a causal argument. We present three increasingly sophisticated frameworks that embody this approach, XPLAIN and two versions of the Explainable Expert Systems framework.