Building plan recognition systems on arbitrary applications: the spying technique

There are today a lot of different techniques for performing plan recognition in various domains, such as Intelligent Tutoring Systems, Human-Computer Interaction, or multi agent Systems. In order to bring theory into practice, we claim that experiments should be carried out a larger scales than what is done today. To do so, we stress on the importance of building incomplete plan recognition systems that may run on arbitrary existing applications. We propose a technique for grafting plan recognition systems onto arbitrary object-oriented applications, without modifying their code. This technique is based on the notion of spy a particular object, that may be inserted in object-oriented systems in a non-intrusive manner, and may track incoming messages to arbitrary objects. We show how spies may be defined, installed automatically to produce low-level information about a system's behavior. Information produced by spies may then be fed to plan recognition systems that perform various tasks such as advice-production in tutorial systems or program introspection and analysis.