Scaling Up Goal Recognition

Plan recognition is the task of determining what a person is doing based on a partial view of their behavior. For example, a plan recognizer might observe a computer user executing a sequence of ls, cd, and compress commands in a Unix shell and infer that the computer user is compressing all large les that she owns. Plan recognition is often studied in domains in which people's behavior is characterized by fewer than 100 plans and goals, such as the tra c [13] and student-scheduling [3] domains. The objective of this thesis to apply goal recognition, a special case of plan recognition, to domains in which there are tens of thousands of possible goals and hundreds of plans per goal. Goal recognition is essential to many applications involving communication between people and machines, including natural language understanding [2], student modeling [3], and user help systems [5]. Recognizing a person's goals allows many useful \interventions." For example, an intelligent agent can o er to complete the task that a person is currently working on. Alternatively, an intelligent agent might o er people advice about how to better achieve their tasks in the future. The key factor in scaling up plan or goal recognition is the size of the plan library, which contains the plans and goals the actor might have. Massive plan libraries arise because the library can be exponential in the number of primitive actions and goal predicates. Just as many sentences can be formed from a few words, many plans can be formed out of a few actions and many goals can be formed out of a few predicates. Large libraries arise in the domains in which we have tested our goal recognizers; namely Unix, Microsoft Windows, and a simulated kitchen domain. We address the following four challenges to scaling up goal recognition [7,8,9,10]: