Symbol-Processing Theories and the SOAR Architecture

Abstract

Allen Newell believes that psychology is data-rich and theory-poor, and argues that it is time to redress this imbalance (Newell, 1990). Newell's framework for cognition and his particular proposal, SOAR, are here evaluated in the context of the present state of the art. Although SOAR has not yet shown it can avoid the charge of being programmed for each new task nor demonstrated human-scale power in problem solving or learning, it seems possible that it will be able to overcome these problems, but not soon. However, Ne well's fundamental assumption of the necessity for a uniform representational medium and single learning process raises the deeper issue of accounting for the development of specialized representations for space, time, the physical world, and even simple mental models. The ability of SOAR to learn these things or do without them is an open question. Other unsolved problems include the representation of continuous processes, homeostatic goals, avoidance of negative interference among tasks, computational efficiency, and the relations among cognition, perception, and motor systems. At the very least we must credit SOAR with making these important issues explicit and presenting a worthy if ambitious candidate. For these reasons, Newell's work will be a watershed in psychology.