Goal-Driven Learning: Fundamental Issues: A Symposium Report

implicit in the user’s choice of training examples); and they address the question of how to learn by applying a single, fixed learning method. Although such systems provide a useful test bed for examining individual learning mechanisms, they are inadequate for use as real-world learners. The problem is that realworld situations offer countless opportunities for learning, and each of these opportunities licenses the learning of infinitely many concepts, few of which are actually useful. Consequently, an indiscriminate learning system will expend enor■ In AI, psychology, and education, a growing body of research supports the view that learning is a goal-directed process. Psychological experiments show that people with varying goals process information differently, studies in education show that goals have a strong effect on what students learn, and functional arguments in machine learning support the necessity of goalbased focusing of learner effort. At the Fourteenth Annual Conference of the Cognitive Science Society, a symposium brought together researchers in AI, psychology, and education to discuss goaldriven learning. This article presents the fundamental points illuminated at the symposium, placing them in the context of open questions and current research directions in goal-driven learning. Learning is a central area of study for researchers interested in human cognition as well as those interested in machine intelligence. Its study has benefited greatly from the multiple perspectives provided by disciplines such as psychology, AI, and education. In AI, machine-learning research has developed a rich repertoire of learning mechanisms. However, less attention has been given to understanding the issues involved in applying these methods—when learning should occur, what knowledge should be learned, and which learning strategies are appropriate in a given context. Standard machinelearning systems address the question of when to learn by attempting to learn in response to every input; they address the question of what to learn by learning a user-supplied target concept (either explicit in the input provided to the system or operationalization criteria (Mitchell, Keller, and Kedar-Cabelli 1986). In most systems, these focusing criteria are fixed (contrast Utgoff 1986). However, as circumstances change, the need for learning changes as well. Because inappropriate learning might actually degrade system performance (Minton 1988), effective performance depends on assuring that what is learned actually furthers system goals. Goal-driven learning takes system goals as a starting point in the learning process. The idea of goal-driven learning is that because the value of learning depends on how well it satisfies system goals, system goals should direct decisions of when and what to learn. In this way, goal-driven learning follows the spirit of research on failure-driven learning systems, in which learning is motivated by deficiencies in system performance (Sussman 1975; Riesbeck 1981; Schank 1982; Birnbaum et al. 1990; Hammond 1989; Ram and Cox 1993; Schank and Leake 1989). Likewise, goal-driven learning is in the spirit of explanation-based learning research into forming useful target concepts (Kedar-Cabelli 1987) and judging the utility of learning (Keller 1987; Minton 1988). Goal-driven learning, however, takes a broader view, examining the relationships between the many possible motivations for learning and the many strategies to achieve it. The effectiveness of goal-driven learning depends on being able to make good decisions about when and what to learn and on selecting the best strategies for achieving the desired learning. Unlike the passive and static process used in many learning systems, goal-driven learning is itself a planful process in which selection of target concepts and learning strategies is guided by desires and needs for knowledge (Hunter 1990). Recent research provides growing support for goal-driven approaches to learning, both on cognitive and on functional grounds. In psychology, learner goals have been shown to have strong effects on the human learning process (Barsalou 1991; Goal-Driven Learning: Fundamental Issues A Symposium Report