Combining power with tractability in beliefmodels

Belief models are used in AI computer applications in a great many roles, both to keep track of the beliefs of users and to simulate the belief states of agents modelled by the system. Epistemic logic is often used as a paradigm for the behaviour of such models, but has problems of computational complexity, which are compounded by the need to include nonmonotonic inferences and change over time. Systems based on standard logics also sanction inferences which human reasoners would be unlikely to make. In some aspects of reasoning, simpliications can be made which both reduce the computational problems and restrict inferences to those most plausible in humans. However, in order to reap the computational advantages of restricting inferences, a model-building approach similar to a compiler must be used. A description of the domain and the agents' knowledge is entered in a formal language, and translated into a restricted representation which allows a reasoning process embodying the restrictions and other eeciency features to be used. The process of compilation and execution is described in detail, and compared with models which use diierent representations for diierent beliefs while remaining complete in terms of standard logic. Relationships between computational simpliications and improvements in psychological plausibility are discussed.