Abstract

Plan recognition is the task of ascribing intentions about plans to an actor, based on observing the agent's actions or utterances. The plan recognition problem appears in three different forms: Plan recognition when the actor is aware and actively co-operating to the recognition, for example by choosing actions that make the task easier (intended plan recognition), plan recognition when the actor is unaware of or indifferent to the plan recognition process (keyhole plan recognition), or plan recognition when the actor is aware of and actively obstructs the plan recognition process (obstructed plan recognition). I consider a specific application of plan recognition: that when a computer system ascribes intended plans to {\it human users interacting with the system}. In computer interfaces, intended plan recognition becomes an almost trivial task, similar to the task of interpreting a command language, whereas keyhole plan recognition can be hard, or even impossible, to achieve. In this thesis, I formulate ways to combine intended and keyhole plan recognition in human-computer interaction applications.

• I define a principle for designing interfaces, ``co-operative task enrichment'' that employ plan recognition, in which users can inspect and control the results of keyhole plan recognition. This design principle achieves a way to use the results from keyhole plan recognition to encourage intended plan recognition in the interface design. The design principle has been utilised in an information seeking application, where the answers to user queries are adapted to the user's task.
• The main reason that keyhole plan recognition may be hard or impossible to achieve in interfaces, is that the user may {\it not} be following a pre-planned behaviour, and for this reason seem to frequently change his or her plan. For some applications, this problem can be circumvented by restricting the system's ``memory'', so that plan recognition only takes into account a number of the most recent actions instead of all observed user actions. I describe a very simple way to implement such a plan recognition mechanism, the ``intention guessing'' approach.
• I describe a reasoning framework for plan recognition, in which keyhole and intended plan recognition can be integrated. The main feature of the integrated reasoning framework is that several components that usually are kept implicit in frameworks for plan recognition are here represented explicitly. The integration of keyhole and intended plan recognition requires that multiple definitions of intentions can be formulated and coexist.
• The reasoning framework is based on a definition of abductive explanation, ``PID-abduction'' that allows an explicit representation of the purpose of the plan recognition process. This is crucial in formulations of plan recognition for collaborative answer generation, and for addressing plan recognition in domains where user's change their intentions over time. PID-abduction is defined within the theory of Partial Inductive Definitions.