Using Plan Recognition in Human-Computer Collaboration
In order for a collaborative agent to assist a user, the agent must know something about the user's goals. We have developed a plan recognition component which infers users' goals from their actions in the context of an ongoing collaboration between a human user and a computer agent. Adding plan recognition to human-computer collaboration significantly reduces the need for users to explictly explain their goals to the agent, allowing for much more natural and efficient interaction. Our plan recognition component has been included in Release 1.1 of COLLAGEN, a Java middleware system for building collaborative interface agents.
Background & Objective: Release 1.0 of COLLAGEN was delivered in November, 1998. Release 1.1 of COLLAGEN includes a plan recognition component. Software developers can use COLLAGEN to implement a collaborative interface agent for any Java application.
Technical Discussion: Technical discussion: An important trend in recent work on interactive computer systems has been to view human-computer interaction as a kind of collaboration. In this approach, the human user and the computer (often personified as an "agent") coordinate their actions toward achieving shared goals. A common setting for collaboration, which is the focus of this work, is when two participants can both communicate with each other and observe each other's actions on some shared artifact. Although plan recognition is a well-known feature of human collaboration, it has proven difficult to incorporate into practical human-computer collaboration systems due to its inherent intractability in the general case. Our investigation, however, has shown that we can exploit three properties of the collaborative setting in order to make plan recognition practical. The first property we exploit is the focus of attention. We can limit the search required by plan recognition based on the observation that consecutive utterances or actions will refer to the same aspect of the plan, unless a shift of the focus of attention is signaled. The second property we exploit is that, in contrast to the general case for plan recognition, we do not need to infer the user's entire plan. Instead, during collaboration, the plan recognizer need only incrementally elaborate a hierarchical description of the current plan to account for new observations. The third property we exploit is the ability of the agent to occasionally ask users questions about their plans and goals.
Outside Collaborations: This research has been done jointly with Lotus Development Corporation (IBM).
Contact: Richard (Dick) Waters
Technology Area: Artificial Intelligence
Modification Date: September 12, 2007
