The role of truth maintenance in model-based reasoning (abstract only)

Abstract

Knowledge system developers now have available to them well-engineered software environments such as the KEE system that provide integrated collections of “power tools” to support their work. In particular, these environments provide tools for acquiring and representing models of a task domain, reasoning with those models, and constructing user interfaces. Experience in complex application domains has shown that object-oriented structural models often play a central role in the architecture of a knowledge system. The domain knowledge embodied in such a model can be used by the system for many purposes including various kinds of problem solving, question-answering, training, and simulation. Such multi-task models and their associated reasoning facilities in current knowledge systems provide power and generality not available in the collections of task-specific production rules that characterized first generation expert systems.Object-oriented modeling facilities typically provide a repertoire of reasoning services, including inheritance of prototype object descriptions, automatic maintenance of structural integrity constrains, and retrieval operations that involve deduction from intentional descriptions. In our recent work with the KEE system, we have been developing various ways of extending these reasoning services by using de Kleer's assumption-based truth maintenance system (ATMS). Those extensions are the focus of this talk. For example, an ATMS provides facilities for storing with a derived fact the assumptions (i.e., primitive facts) on which its derivation is based and for including the derived fact in the model only when the justifying assumptions are also in the model. Thus, the ATMS can be used to provide a facility for managing derived results as a model changes during problem solving. In addition, we will describe our use of the ATMS to provide facilities for exploring alternative hypothetical action scenarios, eliminating undesirable states during problem solving, inheritance of object attributes across arbitrary relations (e.g., the owner of each subpart is the same as the owner of the whole), explanations, and reasoning with defaults.