Belief change in knowledge-based agents

点击翻译按钮获取中文摘要

In Knowledge Representation in Artificial Intelligence, an agent's knowledge will generally be incomplete, inaccurate, and will evolve over time. This will be the case whether the agent is a robot in a physical domain, or a software agent on the web, or a diagnostic agent dealing with a corpus of medical information. On the one hand, an agent must be able to incorporate new information, perhaps correcting or filling in its erroneous or incomplete knowledge. On the other hand, an agent must deal as best it can with the information it has, by drawing plausible conclusions or otherwise dealing with incomplete information.My research program focusses on these issues, specifically on managing belief change and on the closely-related issues of defeasible and preferential reasoning. While a great deal of work has been carried out in these areas, significant challenges remain: there are fundamental issues that have not been fully addressed or that are not fully understood; applications of belief change theories in practical settings are limited; and work remains in developing belief change theories within a dynamic environment.Research in the previous grant period focussed on belief change, with emphasis on non-classical logics and fragments of classical logic, and on belief change in an action domain. For the next grant period, I will focus primarily on belief change, while also addressing the related topic of using defeasible reasoning for dealing with incomplete knowledge. Three interrelated projects will be undertaken. The first covers formal aspects of belief change and its application in realistic settings. I will look at the specific problems of dealing with (ir)relevant information and in revising in a first-order setting; these are arguably the two major unaddressed (or incompletely addressed) problems in belief change. I also will continue ongoing work in applying belief change in description logics (essentially, fragments of first-order logic), where the goal will be to develop an application for updating a large knowledge base (tentatively SNOMED CT) based on "commonsense" principles. The second project will address topics in qualitative, defeasible, reasoning: a new approach to nonmonotonic reasoning will be developed that is intended to be suitable for description logics; and second, a preference-based approach to deontic reasoning, dealing with instructing an agent what it should do, will be developed and implemented. The third topic continues work from the previous grant period, where the goal is to provide a qualitative, high-level specification of an agent in an action domain where actions may fail or have unanticipated or nondeterministic effects, sensors may not be reliable, and the agent's beliefs may be incorrect.