Academic journal article Human Factors

Production Compilation: A Simple Mechanism to Model Complex Skill Acquisition. (Special Section)

Academic journal article Human Factors

Production Compilation: A Simple Mechanism to Model Complex Skill Acquisition. (Special Section)

Article excerpt


In such diverse environments as air traffic control and nuclear power plant operation, researchers in human factors have accumulated extensive empirical knowledge of human performance in complex and dynamic tasks. However, the development of detailed computational models that can explain how people are able to perform and learn such tasks has lagged behind. One reason for this disparity is that theoretical investigations of skill acquisition and empirical investigations of complex and dynamic task performance have existed largely as separate and independent areas of research, with theoretical development of skill acquisition primarily focusing on models of learning simple tasks. Although valuable insights have been gained from studying simple tasks, in order to move toward a more complete theory of skill acquisition, researchers need to develop and test their models against complex and dynamic tasks that are more typical of human learning in the real world.

The goal of this article is to show how the theoretical gap between learning simple tasks and performing complex tasks can be bridged. Our approach offers a new way to conceptualize and validate task analyses and provides insights into the nature of human skill acquisition. We believe our approach can help human factors researchers in developing applications in which learning is an integral aspect of the task. Our approach is to use the ACT-Rational (ACT-R) cognitive architecture, which is grounded in psychological theory, to model learning and performance in complex tasks. The key aspect of the architecture is production compilation, a computational account of skill acquisition that combines aspects of theories proposed by Anderson (1982, 1987) and Newell and Rosenbloom (1981). We use production compilation to develop a detailed model of learning in a simulated air traffic control task.

Mechanisms for Skill Acquisition

Skill acquisition has traditionally been viewed as going through three stages: the cognitive, the associative, and the autonomous (Fitts, 1964). In the cognitive stage performance is slow and prone to errors, whereas in the autonomous stage performance is fast and error free. Anderson (1982) posited that these three stages could be understood as a shift from using declarative knowledge to using procedural knowledge. For the cognitive stage, he argued that the knowledge to perform a task is mostly declarative and needs to be interpreted. The process of interpreting declarative knowledge is slow and can lead to errors, especially if the relevant knowledge cannot be retrieved from declarative memory when needed. In the autonomous stage, the knowledge to perform a task is mostly procedural. Procedural knowledge is compiled and therefore fast and free of errors. The associative stage reflects a transitional period during which knowledge is partly declarative and partly procedural.

To model this shift from using declarative knowledge to procedural knowledge, Anderson proposed a learning mechanism he called knowledge compilation, which became part of the ACT* cognitive architecture (Anderson, 1983), a predecessor of ACT-R. Knowledge compilation is a computational theory of learning that consists of a two-step process: proceduralization and composition. During proceduralization, domain-specific declarative knowledge is inserted into the procedures, replacing general-purpose knowledge. During composition, multiple-step procedures are collapsed into a single procedure. To further refine procedural knowledge, two more mechanisms are used: generalization (dropping conditions or replacing constants by variables) and specialization (adding conditions or replacing variables by constants). However, Anderson later abandoned knowledge compilation (Anderson & Lebiere, 1998) because of a lack of empirical evidence and problems with what he called "computational misbehavior," which is caused by learned procedural knowledge that brings the system to an endless loop or causes it to abort prematurely. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.