Plan Recognition For Intelligent Interfaces (NOT KNOWN IF PUBLISHED BECAUSE AUTHOR HAS LEFT AT&T)

Plan recognition algorithms use knowledge of likely plans to infer the intent behind particular input actions. The use of plan recognition to enhance both natural language and other interface systems has been widely documented in the artificial intelligence literature [1,3,4,7,9,16,17,18,23,21,24]. Although plan recognition inherently involves multiple agents, a set of simplifying assumptions is typically made to eliminate any potential disparities of planning knowledge. For example, by assuming that every agent's planning knowledge is valid and complete, a pre-defined library of plans common to all agents can be used, and plan recognition reduced to processes of search or parsing. In our work we relax the assumptions that the system's knowledge of the user's plans is complete, and that the user's plan communication is error-free. We propose a more constructive view of plan recognition (CPR) to resolve the resulting class of disparities. CPR provides greater robustness to plan- based systems, by using plan library disparity to either guide the recognition of novel plans or to resolve various types of plan miscommunication. The utility of CPR is illustrated in the context of a plan- based, computer-aided design system.