Skip to main content
Social Sci LibreTexts

3.11: A Classical Architecture for Cognition

  • Page ID
    35722
  • \( \newcommand{\vecs}[1]{\overset { \scriptstyle \rightharpoonup} {\mathbf{#1}} } \) \( \newcommand{\vecd}[1]{\overset{-\!-\!\rightharpoonup}{\vphantom{a}\smash {#1}}} \)\(\newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\) \( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\) \( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\) \( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\) \( \newcommand{\Span}{\mathrm{span}}\) \(\newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\) \( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\) \( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\) \( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\) \( \newcommand{\Span}{\mathrm{span}}\)\(\newcommand{\AA}{\unicode[.8,0]{x212B}}\)

    The physical symbol system hypothesis defines classical cognitive science. This school of thought can be thought of as the modern derivative of Cartesian philosophy. It views cognition as computation, where computation is the rule-governed manipulation of symbols. Thus thinking and reasoning are viewed as the result of performing something akin to logical or mathematical inference. A great deal of this computational apparatus must be innate.

    However, classical cognitive science crucially departs from Cartesian philosophy by abandoning dualism. Classical cognitive science instead adopts a materialist position that mechanizes the mind. The technical notion of computation is the application of a finite set of recursive rules to a finite set of primitives to evolve a set of finite symbolic structures or expressions. This technical definition of computation is beyond the capabilities of some devices, such as finite state automata, but can be accomplished by universal machines such as Turing machines or electronic computers. The claim that cognition is the product of a device that belongs to the same class of artifacts such as Turing machines or digital computers is the essence of the physical symbol system hypothesis, and the foundation of classical cognitive science.

    Since the invention of the digital computer, scholars have seriously considered the possibility that the brain was also a computer of this type. For instance, the all-or-none nature of a neuron’s action potential has suggested that the brain is also digital in nature (von Neumann, 1958). However, von Neumann went on to claim that the small size and slow speed of neurons, in comparison to electronic components, suggested that the brain would have a different architecture than an electronic computer. For instance, von Neumann speculated that the brain’s architecture would be far more parallel in nature.

    Von Neumann’s (1958) speculations raise another key issue. While classical cognitive scientists are confident that brains belong to the same class as Turing machines and digital computers (i.e., all are physical symbol systems), they do not expect the brain to have the same architecture. If the brain is a physical symbol system, then what might its architecture be like?

    Many classical cognitive scientists believe that the architecture of cognition is some kind of production system. The model of production system architecture was invented by Newell and Simon (Newell, 1973; Newell & Simon, 1961, 1972) and has been used to simulate many psychological phenomena (Anderson, 1983; Anderson et al., 2004; Anderson & Matessa, 1997; Meyer et al. 2001; Meyer & Kieras, 1997a, 1997b; Newell, 1990; Newell & Simon, 1972). Production systems have a number of interesting properties, including an interesting mix of parallel and serial processing.

    A production system is a general-purpose symbol manipulator (Anderson, 1983; Newell, 1973; Newell & Simon, 1972). Like other physical symbol systems, production systems exhibit a marked distinction between symbolic expressions and the rules for manipulating them. They include a working memory that is used to store one or more symbolic structures, where a symbolic structure is an expression that is created by combining a set of atomic symbols. In some production systems (e.g., Anderson, 1983) a long-term memory, which also stores expressions, is present as well. The working memory of a production system is analogous to the ticker tape of a Turing machine or to the random access memory of a von Neumann computer.

    The process component of a production system is a finite set of symbol-manipulating rules that are called productions. Each production is a single rule that pairs a triggering condition with a resulting action. A production works by scanning the expressions in working memory for a pattern that matches its condition. If such a match is found, then the production takes control of the memory and performs its action. A production’s action is some sort of symbol manipulation—adding, deleting, copying, or moving symbols or expressions in the working memory.

    A typical production system is a parallel processor in the sense that all of its productions search working memory simultaneously for their triggering patterns. However, it is a serial processor—like a Turing machine or a digital computer— when actions are performed to manipulate the expressions in working memory. This is because in most production systems only one production is allowed to operate on memory at any given time. That is, when one production finds its triggering condition, it takes control for a moment, disabling all of the other productions. The controlling production manipulates the symbols in memory, and then releases its control, which causes the parallel scan of working memory to recommence.

    We have briefly described two characteristics, structure and process, that make production systems examples of physical symbol systems. The third characteristic, control, reveals some additional interesting properties of production systems.

    On the one hand, stigmergy is used to control a production system, that is, to choose which production acts at any given time. Stigmergic control occurs when different agents (in this case, productions) do not directly communicate with each other, but conduct indirect communication by modifying a shared environment (Theraulaz & Bonabeau, 1999). Stigmergy has been used to explain how a colony of social insects might coordinate their actions to create a nest (Downing & Jeanne, 1988; Karsai, 1999). The changing structure of the nest elicits different nest-building behaviours; the nest itself controls its own construction. When one insect adds a new piece to the nest, this will change the later behaviour of other insects without any direct communication occurring.

    Production system control is stigmergic if the working memory is viewed as being analogous to the insect nest. The current state of the memory causes a particular production to act. This changes the contents of the memory, which in turn can result in a different production being selected during the next cycle of the architecture.

    On the other hand, production system control is usually not completely stigmergic. This is because the stigmergic relationship between working memory and productions is loose enough to produce situations in which conflicts occur. Examples of this type of situation include instances in which more than one production finds its triggering pattern at the same time, or when one production finds its triggering condition present at more than one location in memory at the same time. Such situations must be dealt with by additional control mechanisms. For instance, priorities might be assigned to productions so that in a case where two or more productions were in conflict, only the production with the highest priority would perform its action.

    Production systems have provided an architecture—particularly if that architecture is classical in nature—that has been so successful at simulating higher-order cognition that some researchers believe that production systems provide the foundation for a unified theory of cognition (Anderson, 1983; Anderson et al., 2004; Newell, 1990). Production systems illustrate another feature that is also typical of this approach to cognitive science: the so-called classical sandwich (Hurley, 2001).

    Imagine a very simple agent that was truly incapable of representation and reasoning. Its interactions with the world would necessarily be governed by a set of reflexes that would convert sensed information directly into action. These reflexes define a sense-act cycle (Pfeifer & Scheier, 1999).

    In contrast, a more sophisticated agent could use internal representations to decide upon an action, by reasoning about the consequences of possible actions and choosing the action that was reasoned to be most beneficial (Popper, 1978, p. 354): “While an uncritical animal may be eliminated altogether with its dogmatically held hypotheses, we may formulate our hypotheses, and criticize them. Let our conjectures, our theories die in our stead!” In this second scenario, thinking stands as an intermediary between sensation and action. Such behaviour is not governed by a sense-act cycle, but is instead the product of a sense-think-act cycle (Pfeifer & Scheier, 1999).

    Hurley (2001) has argued that the sense-think-act cycle is the stereotypical form of a theory in classical cognitive science; she called this form the classical sandwich. In a typical classical theory, perception can only indirectly inform action, by sending information to be processed by the central representational processes, which in turn decide which action is to be performed.

    Production systems exemplify the classical sandwich. The first production systems did not incorporate sensing or acting, in spite of a recognized need to do so. “One problem with psychology’s attempt at cognitive theory has been our persistence in thinking about cognition without bringing in perceptual and motor processes” (Newell, 1990, p. 15). This was also true of the next generation of production systems, the adaptive control of thought (ACT) architecture (Anderson, 1983). ACT “historically was focused on higher level cognition and not perception or action” (Anderson et al., 2004, p. 1038).

    More modern production systems, such as EPIC (executive-process interactive control) (Meyer & Kieras, 1997a, 1997b), have evolved to include sensing and acting. EPIC simulates the performance of multiple tasks and can produce the psychological refractory period (PRP). When two tasks can be performed at the same time, the stimulus onset asynchrony (SOA) between the tasks is the length of time from the start of the first task to the start of the second task. When SOAs are long, the time taken by a subject to make a response is roughly the same for both tasks. However, for SOAs of half a second or less, it takes a longer time to perform the second task than it does to perform the first. This increase in response time for short SOAs is the PRP.

    EPIC is an advanced production system. One of its key properties is that productions in EPIC can act in parallel. That is, at any time cycle in EPIC processing, all productions that have matched their conditions in working memory will act to alter working memory. This is important; when multiple tasks are modelled there will be two different sets of productions in action, one for each task. EPIC also includes sensory processors (such as virtual eyes) and motor processors, because actions can constrain task performance. For example, EPIC uses a single motor processor to control two “virtual hands.” This results in interference between two tasks that involve making responses with different hands.

    While EPIC (Meyer & Kieras, 1997a, 1997b) explicitly incorporates sensing, acting, and thinking, it does so in a fashion that still exemplifies the classical sandwich. In EPIC, sensing transduces properties of the external world into symbols to be added to working memory. Working memory provides symbolic expressions that guide the actions of motor processors. Thus working memory centralizes the “thinking” that maps sensations onto actions. There are no direct connections between sensing and acting that bypass working memory. EPIC is an example of sense-think-act processing.

    Radical embodied cognitive science, which is discussed in Chapter 5, argues that intelligence is the result of situated action; it claims that sense-think-act processing can be replaced by sense-act cycles, and that the rule-governed manipulation of expressions is unnecessary (Chemero, 2009). In contrast, classical researchers claim that production systems that include sensing and acting are sufficient to explain human intelligence and action, and that embodied theories are not necessary (Vera & Simon, 1993).

    It follows that there is no need, contrary to what followers of SA [situated action] seem sometimes to claim, for cognitive psychology to adopt a whole new language and research agenda, breaking completely from traditional (symbolic) cognitive theories. SA is not a new approach to cognition, much less a new school of cognitive psychology. (Vera & Simon, 1993, p. 46)

    We see later in this book that production systems provide an interesting medium that can be used to explore the relationship between classical, connectionist, and embodied cognitive science.


    This page titled 3.11: A Classical Architecture for Cognition is shared under a CC BY-NC-ND license and was authored, remixed, and/or curated by Michael R. W. Dawson (Athabasca University Press) .

    • Was this article helpful?