Can ACT-R Realize "Newell's Dream"?

Can ACT—R Realize “Newell’s Dream”? Matthew F. Rutledge-Taylor (mrtaylo2@connect.carleton.ca) Institute of Cognitive Science, Carleton University, 1125 Colonel By Drive Ottawa, Ontario, KIS 5B6 Canada Abstract In “The Atomic Components of Thought”, John Anderson and Christian Lebiere claim that ACT—R (4.0) realizes “Newell’s Dream” of a unifying theory of cognition. In this paper it is suggested that each ACT—R model can account for only a finite set of cognitive processes, and cannot therefore be used to model an unbounded whole mind. It is suggested that this is due to an inherent context dependence of ACT—R models. This limitation runs counter to the intuitive criterion that a unifying theory of cognition ought to be able to provide an account of the mind as a system not bound to any particular context. It is suggested that thought in cognitive models ought to be conceived as temporary context specific operations based on persistent context independent knowledge. The basis for a new cognitive architecture, which differentiates thought from knowledge is proposed. This new architecture combines ACT—R with elements of Lawrence Barsalou‘s situated simulation theory. Keywords: ACT—R; cognitive architectures; knowledge representation; situated cognition; perceptual symbols systems. Introduction In 1972, at the Carnegie Symposium on Cognition, Allan Newell raised a concern about the course of research in psychology. He delivered a paper entitled “You can’t play 20 questions with nature and win”, in which he lamented the fact that there was very little that unified the wealth of knowledge that had been accumulated about individual human cognitive processes (l973a). In a paper published separately in the proceedings of the symposium, Newell suggested that production systems might serve as detailed models of the human control structure (1973b; 1990). Eighteen years later, Newell published a book entitled “Unified Theories of Cognition” in which he proposed that cognitive architectures hold the key to unifying psychology (1990). There are many different cognitive architectures used to produce cognitive models of psychological phenomena, including Newell’s own SOAR architecture, which was first released in 1982 (Laird & Rosenbloom, 1996). However, the most popular architecture is ACT—R (Anderson, 1993; Anderson & Lebiere, 1998). This popularity is by no means accidental. The theory of cognition it implements has been well developed, and hence, has allowed a wide variety of researchers to produce theoretically grounded models of various cognitive phenomena. Additionally, and perhaps most significantly, ACT—R models typically fit the human experimental data they are designed to model, quite well. 1895 This paper considers whether ACT—R, as it currently exists, realizes “Newell’s Dream” of a unifying theory of cognition, or not. It is suggested that given appropriately strict criteria ACT—R may be inadequate. N ewell’s Criteria According to Newell: a theory is an explicit body of knowledge, from which answers to questions of a predictive, explanatory, or prescriptive type can be given; theories are approximate; theories cumulate; and, theories develop iteratively (1990, pp. 13-14). Newell defines a unified theory of cognition as “a single set of mechanisms for all of cognitive behavior” (1990, p. 15). He specifies these mechanisms as a prioritized list of areas of cognitive phenomena to be covered. They are, in order: problem solving, decision making, and routine action; memory, learning, and skill; perception, and motor behaviour; language; motivation, and emotion; and, imagining, dreaming, and, daydreaming. Thus, a complete unified theory of cognition, should account for all of these cognitive phenomena. However, given Newell’s views on theory development, an acceptable strategy would be to begin with a unified theory of the phenomena at the top of the list, and slowly augment the theory so as to accommodate successive items. Background Theory Ubiquitous in cognitive science is the View that cognitive systems can be analysed from a variety of perspectives. The tri-level hypothesis is that there are three basic levels of analysis (Dawson, 1998). Various researchers apply their own labels to these three levels. Newell divided them into the biological, cognitive, and rational (Newell, 1990); Zenon Pylyshyn makes use of the physical, syntactic, and semantic (Pylyshyn, 1999); and, Michael Dawson, the implementational, algorithmic, and computational levels (Dawson, 1998). Despite the difference in terms, there is arguably an equivalence between these hierarchies. The biological, physical, and implementational levels are, in the case of humans, the levels of description that (typically) appeal primarily to neural processes. The cognitive, syntactic, and algorithmic levels, describe human cognition in terms of operations on syntactic (or, otherwise, formal) structures. The rational, semantic, and computational levels are those at which the cognitive system is described in terms of its knowledge (i.e., goals, beliefs, and perceptions etc.). Pylyshyn asserts that a fundamental hypothesis in cognitive science is that this knowledge level is an autonomous (or, at least, partially autonomous) level of