Control in Act-R and Soar

The last decade has seen the emergence of a variety of cognitive architectures. This is good news, in general, for cognitive modeling, because architectures provide a readymade set of tools and theoretical constraints that can— according to architectural research methodology—assist the cognitive modeling enterprise by constraining the possible models of a set of phenomena or even making the “right” model an obvious consequence of the architectural constraints (Newell, 1990). Just as special purpose programming tools, such as spreadsheets, give us the right language and tools for attacking specialized tasks, cognitive architectures are meant to give us the right constraints for building cognitive models. Although the various cognitive architectures often apply to overlapping cognitive capacities, and share some similarities, they also make many different theoretical distinctions, which can of course have a major influence on the nature of cognitive models supported by each architecture. Despite this, very little work has been done to compare alternative architectures. This paper attempts to rectify this by offering an initial comparison of two of the most well-known cognitive architectures: Act-R (Anderson, 1993; Lebiere, 1996) and Soar (Laird, Rosenbloom & Newell, 1986; Laird, Newell & Rosenbloom, 1987; Newell, 1990), two of the most wellknown cognitive architectures. This comparison has two goals. The first is to identify the similarities and differences between Act-R and Soar by listing each architecture’s fundamental theoretical distinctions. These distinctions are loosely organized around control, long-term memory, working memory, learning, and latency derivation. The second goal of this paper is to assess the empirical support for the major differences in the architectures’ control mechanisms. Clearly, a complete evaluation will require a similar comparison between the remaining categories; however, this is beyond the scope of the present paper. This is not the first effort to compare and evaluate Soar and Act. Newell, Rosenbloom and Laird (1989) compared the then current version of Soar to Act*, Act-R’s precursor. Their goal, however, was more to use Soar and Act as two different examples of cognitive architectures, not to critically evaluate and compare them. Much of the theoretical effort on Soar has been on using it in specific domains, rather than directly testing its theoretical distinctions. To my knowledge, the only serious effort to critically evaluate Soar is Cooper and Shallice’s (1995) evaluation of Soar as both a psychological theory and an example of the methodology of unified theories of cognition. Their general conclusion is that Soar fairs poorly as a psychological theory and that the unified theory methodology (at least as exemplified by Soar research) does not offer any advantages over traditional psychological research methodology and possibly falls prey to some of the same problems that plagued the grand theories of the 30’s and 40’s. In contrast to the Soar approach, Anderson and his colleagues have regularly tested Act-R both by applying it to specific domains and by directly testing its theoretical assumptions. Many of these results can help us discriminate between Soar and Act-R.