Login
Computer Science @ Rochester
Friday, April 15, 2011
1:00 PM
CSB 601
Phillip Michalak
University of Rochester
Task Model Reasoning
Knowledge representation plays a central role in many Artificial Intelligence approaches and techniques. Examples include rule-based and expert systems, semantic networks, frame-based systems, belief-desire-intention (BDI) agents, planning, and plan recognition systems. All of these manipulate and reason about a symbolic formulation of their problem space. In the past decade, researches have noted and begun to address the fact that this type of symbolic knowledge is quite difficult to come by.

Programming By Demonstration (PBD) has emerged as a means of decreasing the cost of acquiring task models for automated execution. PBD tools are designed to acquire task models simply by observing how a (non-programmer) user performs a task. Some PBD inspired systems go even further and allow users to annotate what is happening as the task is being demonstrated, which can decrease the number of training examples required to learn an executable task model. Even so, it is still an expensive endeavor to acquire large repositories of task models.

In this talk I describe some techniques for making better use of task models that have been acquired previously with the Procedure Learning On the Web (PLOW) system. I show how a plan-based analysis can detect and eliminate two classes of systematic error in the learned task models and how models can be combined to create more general abstraction task models.