MSIMS@RUTGERS.ARPA (03/26/84)
From: Michael Sims <MSIMS@RUTGERS.ARPA> [Forwarded from the Rutgers bboard by Laws@SRI-AI.] MACHINE LEARNING BROWN BAG SEMINAR Speaker: Richard Keller Date: Wednesday, March 28, 1984 - 12:00-1:30 Location: Hill Center, Room 254 Placing Learning in Context: SOURCES OF CONTEXTUAL KNOWLEDGE FOR CONCEPT LEARNING (Alternatively titled: The Mysterious Origins of LEX's Learning Goal) In this talk, I will describe a new source of knowledge for concept learning: knowledge of the learning context. Most previous research in machine learning has failed to recognize contextual knowledge as a distinct and useful form of learning knowledge. Contextual knowledge includes, among other things, knowledge of the purpose for learning and knowledge of the performance task to be improved by learning. The addition of this meta-knowledge, which describes the learning process, provides a broader perspective on learning than has been available to most previous learning systems. In general, learning systems that omit contextual knowledge have an insufficient vantagepoint from which to supervise learning activity. Both AM [Lenat-79] and LEX [Mitchell-83], for instance, were limited by an inability to adapt to changes in their respective learning environments, even when the changes were a result of their own learning behavior. This limitation is not particularly surprising; neither of these systems contained an explicit representation of the task they were performing (specifically, mathematical discovery and integral calculus problem solving, respectively). Nor did these systems contain any knowledge about the relationship between learning and the task performance. Before it is reasonable to expect a learning system to adapt to changes in the task environment, it is necessary to represent task knowledge and to incorporate this knowledge into learning procedures. My research, therefore, focuses on the representation and use of contextual knowledge -- including task knowledge -- as guidance for concept learning. In this talk, I will describe a learning framework that incorporates the use of contextual knowledge. In addition, I will introduce various alternative methods of representing contextual knowledge, and sketch the design of some learning algorithms that utilize contextual knowledge. Examples will be drawn, in large part, from my work on incorporating contextual knowledge within the LEX learning system.