Recognizing the Plans of a Replanning User

When applying plan recognition to Human Computer Interaction, one must cope with users exhibiting a large amount of reactive behaviour: users that change tasks, or change strategies for achieving tasks. Most current approaches to keyhole plan recognition do not address this problem. We describe an application domain for plan recognition, where users exhibit reactive rather than plan-based behaviour, and where existing approaches to plan recognition do not perform well. In order to enable plan recognition in this domain, we have developed an extremely simplistic mechanism for keyhole plan recognition, ”intention guessing”, and applied it to this domain. The algorithm is based on descriptions of observable behaviour, and is able to recognize certain instances of plan failures, suboptimal plans and erroneous actions. At run-time, the algorithm only keeps track of a limited number of the most recent actions, which makes the algorithm ”forgetful”. This property makes the algorithm suitable for domains where users frequently change strategies.