Nested default reasoning for user modeling
Paul Van Arragon · 1992
A model of how a computer user reasons is important for effective human/computer communication. At least two distinct levels of reasoning are involved when modeling a user: the computer system reasons about the user, and the user reasons about the domain of discourse. More levels are needed to model users who are reasoning about other agents. For accuracy and flexibility, each level of reasoning should incorporate reasoning by default as well as by deduction. Default reasoning by the system enables making assumptions about the user in case the system has incomplete knowledge. Reasoning about defaults of a user enables the system to avoid misleading the user. This thesis develops tools for reasoning by default about users who reason by default. These tools are based on Theorist, which defines explanations in terms of facts and defaults. The tools incorporate distinct levels of reasoning by having both the system and the user reason according to Theorist. Nested Theorist (NT) is a simple tool that allows default reasoning on arbitrarily-many levels. NT is expanded in two ways: Prioritized Nested Theorist (PNT) uses defaults of different priority levels to give the system's default more power and to help remove multiple extensions. Limited Nested Theorist (LNT) permits the system to reason about limitations of a user's reasoning ability. In LNT, the system assumes that the user is able to draw an inference. These assumptions can be blocked if the user's reasoning resources are limited in some way. These tools are implemented using a general technique that allows a reasoning procedure to operate on many levels. Given a single-level meta-interpreter of a reasoning procedure, an algorithm translates this meta-interpreter into an interpreter that works on multiple levels. Computational overhead is removed via program transformation. (Included is a novel implementation of Prioritized Theorist.) Several examples taken from user modeling and speech act theory illustrate how these tools are useful as a general framework for user modeling. Comparisons with related work regarding logical omniscience, metareasoning, and default reasoning illustrate the advantages obtained by tools with defaults on distinct levels.