Learning Hierarchical Task Models by Demonstration
We have developed machine learning techniques for inferring a hierarchical task model from partially-annotated examples of task-solving behavior. Hierarchical task models are used in many artificial intelligence applications, such as intelligent tutoring, plan recognition, planning, and decision theory. Our approach is based on the intuition that it easier for people to demonstrate and discuss concrete examples of how to accomplish tasks than to formalize task model abstractions directly.
Background & Objective: This work is motivated in part by the COLLAGEN (COLLaborative AGENt) system, which uses hierarchical task models to support a wide range of collaborative human-computer interaction. COLLAGEN embodies a general theory of collaboration, but requires the development of a hierarchical task model for each domain in which it is used. Developing these task models, a special case of the well-known "knowledge acquisition bottleneck," is a significant engineering burden that we hope to lessen by employing machine learning techniques.
Technical Discussion: In our approach, a domain expert first demonstrates one or more sequences of primitive actions required to achieve typical tasks in the application domain. (The primitive actions may be performed directly on an instrumented graphical interface, if such an interface already exists, or by specifying simulated actions for an as-yet-unimplemented application). The expert then reviews and annotates the recorded example sequences. Finally, our learning algorithm constructs a hierarchical task model that generalizes the given examples. Most previous work on learning task models has dealt only with non-hierarchical models. Our hierarchical learning algorithm also extends previous work by inferring propagators, which enforce relationships, such as equalities, between actions at different levels in the task hierarchy. We have also identified a fundamental search problem, called alignment, faced by any task-model learning algorithm, and have developed both restrictive and preference biases in order to make the alignment search tractable. Our learning algorithm has been proved both sound and complete.
Contacts:
Joseph Katz
Technology Area: Artificial Intelligence
Modification Date: January 22, 2007
