Prioritizing Multiple, Contradictory Sources in Common-sense Learning by Being Told; Or, Advice-taker Meets Bureaucracy

LIMITED DISTRIBUTION NOTICE This report has been submitted for publication outside of IBM and will probably be copyrighted if accepted for publication. It has been issued as a Research Report for early dissemination of its contents. In view of the transfer of copyright to the outside publisher, its distribution outside of IBM prior to publication should be limited to peer communications and speciic requests. After outside publication, requests should be lled only by reprints or legally obtained copies of the article (e.g., payment of royalties). Abstract An important common-sense capability is to learn by \taking advice", or \being told", from a variety of sources of information: messages from various other agents, reading various texts, etc.. Such learned or \assimilated" knowledge provides an important basis for one's actions. Yet a basic feature of life is that one cannot believe everything that one is told. Not only may advice be incorrect; e.g., it may be contradicted by direct experience. Worse, diierent sources of advice information may contradict each other, or even themselves. How is an advice-taking agent to maintain a consistent, yet usefully actionable, set of beliefs, then? Here, we address this problem, which has not received much previous attention in the AI literature. We illustrate, with an actual example , how it arises in the domain of bureaucracy. A common human strategy is to treat advice as working belief, which can be overridden by other advice and direct experience. This suggests the rst element of our approach: to represent advice as default-status knowledge. A crucial issue in logically formalizing common-sense reasoning is then how to represent precedence among multiple, potentially contradictory sources of information. We observe that bases for such precedence include reliability and authority. We observe furthermore that precedence relationships may themselves be inferrable and even defeasible. We show how, for the rst time, to represent advice-taking, with such precedence among sources, as reasoning in a non-monotonic logical formalism. We employ a generalized variant of circumscription that we have recently developed: Defeasible Axiomatized Policy (DAP) circumscription. (See attached paper detailing DAP circum-scription.) Precedence among sources is represented explicitly via the mechanism of prioritization in DAP circumscription. DAP circumscription is the rst non-monotonic logical formalism to express defeasible prioritization-type precedence. Our approach to representation of precedence among sources has a number of advantages over other non-monotonic logical formalisms. It treats precedence more explicitly than all previous approaches except perhaps Lifschitz' 1988 …