Published by MIT Press/Bradford Books, Cambridge, MA, 1995.
Click here for ordering information.
Cover and jacket design by Jeffrey Kalin.
Among the most significant new themes over this decade is the exploration of how the learner's goals and prior knowledge drive the learning process. Before 1980, the vast majority of research on learning focussed on the task of estimating or hypothesizing an unknown function given only a sample of its inputs and outputs, with no explicit notion of the learning goal. This research grew out of the tradition of earlier work on statistical pattern recognition, and early studies in psychology on learning "nonsense" concepts such as abstract geometric patterns. But over time many researchers became convinced that this was too simple a formulation of the learning problem to model the surprising learning skills exhibited by humans.
To see the issue here, consider learning to play a game such as chess. Suppose you have just lost your queen, and wish to learn the general pattern of board pieces that led to this failure. Inductive learning methods require hundreds or thousands of such failures (and non-failures), to hypothesize which of the many board features distinguish the positions in which you lose your queen from those in which you do not. But people are typically able to learn such concepts from only a handful of examples. How? People appear to direct learning towards concepts relevant to their goals -- by explaining the cause of their failure, and thereby noticing the relevant board features (e.g., the opposing knight that was simultaneously threatening your King and Queen) while ignoring the irrelevant (e.g., the three pawns in the second row). The oversimplification of early work on learning was that it omitted any reference to the learner's goals, and therefore could not model this kind of learning by explaining.
During the 1980's researchers began exploring the role of explanations, goals, and explicit prior knowledge in the learning process. Explanation-based learning algorithms were developed that generalized more accurately than earlier inductive approaches, by explicitly taking into account the learning goal (e.g., to avoid losing the queen) and related prior knowledge (e.g., the legal moves of chess). The key insight underlying this work was that learning is much easier to understand and duplicate in computers if the learning goal and related knowledge are explicitly manipulated by the learning algorithm.
This initial work led to a flurry of research on goal-driven learning, exploring more broadly the ways in which learning processes are influenced by learning goals. Whereas initial research considered how learning goals influence the process of generalizing from examples, more recent work has considered how learning goals drive other processes as well, such as the process of experimenting to collect new training data, and the process of generating useful learning subgoals. Whereas the initial research produced simplified, brittle algorithms for using the learning goal to guide generalization, more recent work has produced significantly more robust and practical methods and has raised questions concerning the origins of learning goals, the role of goals in guiding other learning tasks, and how to choose appropirate learning strategies to achieve learning goals.
Current research in goal-driven learning deals with a wide range of issues dealing with how and when learning goals arise and the ways in which goals influence a broad range of learning processes. These issues, which are now being addressed in machine learning, cognitive psychology, and education, are the focus of this volume, which summarizes recent work on goal-driven learning, and presents a number of new research results in this area. As you will see from many of the chapters, the field is actively exploring new directions, and many of the approaches are still under development. Although we still lack a full understanding of how to best use goals to guide the learning process, it seems obvious that if we are to progress in understanding learning, then we must take into account more and more of the the rich context in which it occurs. Taking learning goals into account is one essential step, and it is difficult to imagine a future for machine learning or cognitive science in which this does not play a central role.
Tom M. Mitchell, Carnegie Mellon University, Pittsburgh, PA
The fundamental tenet of goal-driven learning is that learning is largely an active and strategic process in which the learner, human or machine, attempts to identify and satisfy its information needs in the context of its tasks and goals, its prior knowledge, its capabilities, and environmental opportunities for learning. It is increasingly evident that investigation of goal-driven learning can benefit from a multidisciplinary effort employing diverse perspectives on a common research agenda. To this point, however, research in goal-driven learning has largely been confined to isolated efforts, with little framework to connect related results and to aid in their analysis. The purpose of this book is to establish such a framework, to collect and solidify existing results on goal-driven learning, and to point the way for future investigations of goal-driven learning.
The book begins with a discussion of fundamental questions for goal-driven learning: the motivations for adopting a goal-driven model of learning, the basic goal-driven learning framework, the specific issues raised by the framework that a theory of goal-driven learning must address, the types of goals that can influence learning, the types of influences those goals can have on learning, and the pragmatic implications of the goal-driven learning model (chapter 1). The remainder of the book is divided into two parts. The first is a collection of recent research papers that serve as case studies in goal-driven learning. Each paper addresses a piece of the goal-driven learning puzzle, reflecting a particular research perspective from one of the several disciplines that have been investigating this area in recent years. These works address issues such as the justification of goal-driven learning models through functional arguments about the role and utility of goals in learning (chapters 2--4), the justification of such models through cognitive results (chapters 5, 6, 14), goal-based processes for deciding what to learn (chapters 7, 8) and for guiding learning and the learning process (chapters 4, 7, 9--13), and pragmatic implications of goal-driven learning for design of instructional environments (chapters 14, 15).
The second part of the book is based on the Symposium on Goal-Driven Learning organized by David Leake and Ashwin Ram at the Fourteenth Annual Conference of the Cognitive Science Society in Bloomington, Indiana, in 1992. It presents an overview of the workshop discussion and a collection of papers from the symposium panelists representing their individual perspectives on fundamental issues and their proposals for fruitful future directions in goal-driven learning research.
The works in this volume reflect both the diversity of goal-driven learning research and the fundamental relationship of different approaches within the broader goal-driven learning framework. Together, they provide a comprehensive overview of recent research in goal-driven learning and illuminate on-going investigations and open issues to provide a foundation for future study of goal-driven learning.
Ashwin Ram, Georgia Tech, Atlanta, GA
David Leake, Indiana University, Bloomington, IN