3
Unitary Theories of Cognitive Architectures
CEC Joint Research Centre, Ispra, Italy
Unitary theories of cognitive architectures have been suggested as “grand theories” of human cognition. Models of human cognition have been developed in order to predict human behavior through computer simulation. This chapter review three models, each based on unitary theories: Soar (Newell, 1990), ACT* (Anderson, 1983), and the Underspecification Theory (Reason, 1990). The notion of symbol systems is used as a guideline for specification of architectures, and specific attention is paid to the models as systems for information processing. Each of the models is briefly outlined, and the main components are discussed. Examples of research on regularities applied through the various architectures are characterized, especially in relation to complex domains and human expertise.
In the last couple of decades there has been an increasing interest within cognitive psychology and cognitive science to form global theories of human cognition. The aim has been to develop an understanding of the architecture and information processing, that is necessary but also sufficient to explain complex cognitive functions. The ambition of this investigation has been to explain cognitive functions in all their variety within a single architecture. As such it has been a way of formulating “grand theories” of human cognition, and as means for developing models for computer simulation of human behavior.
We may contrast global theories with local theories for human behavior, which are based on experimental investigations. Global theories have gained their content partially from such experimental investigations for local theories and partially from the notion of human cognition as being a result of information processing. The influence from the latter has found its root’s in analogies of organizing and processing information in computers, and this bias has given it the name information processing psychology. By their very nature, global theories of human cognition just aim to be approximate, that is, to enable verification of findings on local phenomena with enough detail and evidence to justify their existence as coherent theories. As global theories provide a single framework for explanation of many aspects of human behavior, much effort has been devoted to the development of computer models for simulation studies.
Research on global theories of human cognition has paid its principal interest to the architecture of cognition. A cognitive architecture can be defined as the fixed1 structure that provides the means (and constraints) for processing information. As such cognitive architectures provide the information processing framework within which cognitive functions can be defined.
The modeling of cognitive architecture can largely be separated into two opposing movements of conception. We shall refer to these conceptions as the faculties of mind and unitary theories of cognition.
The conception of the human mind as consisting of a set of “faculties” dates back into history and was promoted in the beginning of this century by researchers such as Gall, who introduced a “phrenological diagram” of the brain. Each cognitive function is assumed to be supported by a dedicated part of the cortex. Accordingly there exists specialized centers for achieving specialized cognitive functions, for example, language. The complex behavior of humans can be explained as the invocation of faculties. Many attempts have since been made to localize complex cognitive functions to specific spatial locations of the brain. A recent example is Chomsky’s promotion of the concept of “mental organs” (Chomsky, 1980).
Luria (1973) suggests that one conceive the notion of cognitive functions as an expression of the invocation of the entire brain and not just specific tissue’s of the brain. This suggestion proposes a new interpretation of cognitive functions as being functional systems, invoking not just a limited part of the cortex, but many parts at the same time. A functional system is characterized by the flexibility by which it can achieve tasks. The tasks and the result of achieving them are invariant, but the mechanisms supporting them can vary. The concept of functional systems introduces the notion of a layered model of complex cognitive functions, in which these higher level functions can be achieved by lower level mechanisms in a flexible way.
The notion of human cognition represented as a layered model has been the basic assumption for a set of unitary theories of cognition. This line of research conceives the complexes of the human mind as being the result of a limited set of intertwined structures and processes. The basic structure forms an architecture, characterized by a number of specialized information processes. The elements of such an architecture can be combined in a flexible way and thereby enable a large variety of responses of the cognitive system.
One might describe the differences between the unitary approach and the faculties approach, as whether a primitive set of structures and processes in combination can account for complex cognitive functions, or whether there is a one-to-one mapping, with a “mental organ” for each of the higher cognitive functions.
The following sections present in more detail the characteristics of the unitary approach through a brief review of some of the most dominant unitary theories of human cognitive architecture.
SYMBOL SYSTEMS AND COGNITIVE ARCHITECTURE
When we talk about cognitive architectures, we address the properties of a system for handling symbols. Analogies with the way computers handle symbols at an appropriate level of description have provided us with the notion of physical symbol systems (Newell, 1990, 1980; Newell & Simon, 1972). The main source for this basic theory has been derived from computer science, in which the use of computers as symbol systems have a rich tradition.
Physical symbol systems (or symbol systems for short) can be characterized in a number of ways. At an abstract level, they can be described as systems, which take inputs based on patterns in outside structures (outside of the system) and provide for some interpretation (coding) into internal symbol structures. The result of internal processing provides for some response (again through interpretation) outside the system (decoding). Symbols are stored internally in memory, upon which operations are being performed. These are the basic components of symbol systems and these are the components that, in an abstract way, have provided a basic theory for the development of cognitive architectures.
Symbol systems take their definition within an architecture. The architecture provides the fixed structure within which functions for realizing the symbol system are achieved. We can view the distinction between symbol systems and architecture as one of content and structure.
The notion of symbol systems and architecture has been applied to provide theories of the cognitive architecture of the human mind. The architecture, which realizes the symbol system of human cognition, is not supposed to be equivalent to the physical structure of the human brain in the literal sense. It does however specify a functional system in which the basic symbol processing of the human mind can be described.
The relation between architecture and symbol systems provides us with a guidance for how higher level cognitive functions are brought about. The relation between the two carries the notion of two separate levels of description, one above the other. The architecture provides the means to achieve the symbol level. This conception, known within computer science as virtual machines, provides us with a theory for the way higher level cognitive functions can be achieved. In effect, the separation into levels of virtual machines defines distinct system levels, each characterized by different properties and each realized through the level beneath.
The notion of virtual machines on top of one another provides a way in which the approximation of rational and adaptive behavior of a cognitive system can be achieved. This orientation has been described by Rasmussen (1986) for representing system levels in an abstraction hierarchy.2
THREE THEORIES OF COGNITIVE ARCHITECTURES
As a way of illustrating some characteristics of cognitive architectures, the following sections briefly discuss three theories of cognitive architectures: SOAR (Newell, 1990), ACT* (Anderson, 1983), and the Underspecification Theory (Reason, 1990).
All of the theories take as a basic assumption the unitary approach and all strive to describe the variety of cognitive functions in one framework. Each of them may be conceived as representing state of the art in modeling cognition and major contributions to the field. The theories of Soar and ACT* are closely linked to the synergy created in the area formed between cognitive psychology and computer science, and have mutually more in common than the Underspecification Theory.
Soar
Soar is a computer based embodiment of a unified theory of cognition set forth by Allen Newell (1990), defining an architecture capable of approximating rational and adaptive behavior. The notion of physical symbol systems has been formulated by Newell and is the underlying principle of Soar.
Historically Soar is the present result of research starting with the development of the Logic Theorist (Newell et al., 1957) and later the General Problem Solver (Ernst & Newell, 1969). Much of the theory forming the basis for the latter is collected in a general theory for human problem solving (Newell & Simon, 1972). The present architecture was initially developed in 1982 by Laird, Rosenbloom, & Newell (Laird et al., 1986) and collects much of the theory developed since the early days.
Soar is characterized by a uniform architecture. All tasks done by Soar are formulated by problem spaces; there is only one long-term memory consisting of productions; productions are matched with memory elements of working memory and all production that matches are applied. Subgoals are generated as a result of insufficient knowledge in the present problem space, and learning is achieved through chunking information as a result of subgoaling.
Soar specifies a basic organization of memory and primitive processing functions, called the symbol level (Figure 1). It does however go beyond this level of representation and build two abstraction levels above, called problem space level, and knowledge level. Each of the three levels in Soar is organized as a virtual machine and each formulate a coherent computational model.
The symbol level, which is the lowest level, contains four major mechanisms: recognition memory (long-term memory), working memory (shortterm memory), decision procedure (selector of next action), and chunker (providing the means for learning).
The decision procedure generates new elements in working memory by matching productions (from recognition memory) with the existing contents of working memory and asserting those that match into working memory (assertions may trigger other productions). On the basis of preferences a problem space, state, or an operator is selected.
During problem solving, situations may arise when Soar has incomplete or missing knowledge about which problem space, state, or operator to choose to solve a problem. In that case a new (sub)goal is defined and immediately processed. When the (sub)problem has been solved and the subgoal has been reached, Soar chunks the knowledge that was used to solve the problem of the subgoal (Laird et al., 1985; Rosenbloom, 1983). This knowledge is coded as productions and immediately added to the long-term recognition memory.
FIGURE 1. The Soar architecture described at three levels of computational models (adapted from Newell, 1990).
The problem space level (the level above the symbol level) is based on the Problem Space Hypothesis (Newell, 1980), which argues for the existence of four model elements: goals, problem spaces, states, and operators. A task to solve is given through a goal. By choosing a problem space the environment for problem solving is selected, and a set of states can be explored by applying operators to these states. At any given time is has to be decided which problem space, state, or operator to use. When knowledge is unavailable for making this selection, Soar automatically sets up a new goal (a subgoal) of resolving this problem.
At the knowledge level (Newell, 1982) Soar can be described as a goal-oriented knowledgeable agent in an environment. It has perceptions of the environment and on the basis of a body of knowledge about the environment it can take actions. The body of knowledge is like a memory, but with no details of how it is encoded (this is implemented at the lower levels). Behavior at the knowledge level is based on the principle of rationality: If an agent has knowledge that one of its actions will lead to one of its goals, then the agent will select that action.
Soar has been applied to a large variety of tasks. The dominant search for psychological validity has been within problem solving, learning, and immediate-response tasks. It has been applied to syllogistic reasoning, natural language comprehension, algorithmic design, and so on. However, no major applications have been made with Soar to the modeling of human behavior in complex working environments.
ACT*
Although ACT* clearly belongs to the unitary approach, John Anderson has taken a critical stand concerning the simplicity that can be applied in the formulation of theories of cognitive architectures. The basic framework of ACT* can however be stated in simplified (although only approximate) terms, as done in the following.
The theory upon which ACT* (Anderson, 1983) is based is the most recent embodiment of a series of theories starting with ACTE (Anderson, 1976), and later ACTF (Anderson et al., 1977). As a computational framework ACT* specifies a set of invariant properties about its structure and some rules for ways in which to operate on the structure. One of the cornerstones in the theory of ACT* is the structure of memory.
The framework provided by ACT* is formed by three memories interconnected through a set of information processes (Figure 2). ACT* has a working memory, which contains the information that the system can utilize immediately. The information is provided by retrieval from the declarative memory, from temporal structures provided from the outside world by encoding processes, and by the application of productions from the production memory. The declarative long-term memory contains memory elements, each of which carry a degree of activation, in a semantic net. Productions in the production memory have a set of conditions in the working memory, that on activation create new memory elements in the working memory.
FIGURE 2. The ACT* cognitive architecture (adapted from Anderson, 1983).
The processes interconnecting the different kinds of memory form a unified set of primitives for processing information. The set of processes consists of some encoding processes, which provide the working memory with information about the outside world. Actions to the outside world are supported by performance processes based on commands in the working memory. The content of the declarative memory is formed and updated through storage processes on the basis of units of knowledge in the working memory: Likewise, the content of the working memory is provided by retrieval processes, based on units in the longterm declarative memory.
Memory elements in working memory are brought into contact with the productions of the production memory through match processes, and execution processes bring the actions of matched productions into working memory. The acquisition of new productions is done through application processes.
Anderson (1983) introduces a tri-code theory for representing knowledge. The theory assumes three different types of representation: temporal strings, spatial images, and abstract propositions. Temporal strings are created to record the sequential structure of events, spatial images preserve the configuration of elements of a spatial array, and abstract properties encode meaning.
As a theory, ACT* covers a large amount of regularities; at the immediate-response level it covers properties such as priming, fact retrieval, and memory effects. At higher levels of cognition, ACT* can account for elementary programming, geometry proofs, and language learning. Like Soar, ACT* has not been applied to model human behavior in complex environments.
The Underspecification Theory
Whereas the theories of Soar and ACT* have emerged in a scientific environment in which the analogy to the processing of information in computers has been a major source for the formulation of theories, the development of the Underspecification Theory (UT) has come in a different scientific environment, in which the purposes are slightly different.
To take the similarities between the different approaches first: Both traditions have urged for a theory for the basic cognitive architecture, and both have chosen the unitary approach as the basic standard. In the case of the UT, however, one main design criterion was the “design for a fallible machine”. This emphasis was introduced due to interest in human error. To quote from James Reason’s book about the UT (Reason, 1990, p. 125), the question raised was “… what kind of information-handling device could operate correctly for most of the time, but also produce the occasional wrong responses characteristic of human behavior?”
In the UT, the cognition architecture is presumed to possess two distinct structural components: a limited, serial, but computational powerful working memory (WM), and a virtually limitless knowledge base (KB).
The KB holds a vast set of structural knowledge encoded in frames or schema including both declarative and procedural knowledge. The knowledge of these structures captures the experience of an individual.
The WM represents a working area for the cognition processes, the management and the temporal storage of data. The role of the WM is: (1) to assess the external and internal cues of data from the environment to match with similarities in the KB; (2) to decide whether the hypothesis is appropriate, and if not (3) to reiterate the search with revised cues.
The cognitive architecture has two primitives for selecting stored knowledge units and bringing them into WM (Figure 3). These primitives, similarity-matching (SM) and frequency gambling (FG), operate in a parallel, distributed and automatic fashion within the KB. SM matches perceived cues coming from the work environment with corresponding diagnostic cues described in the KB. FG solves possible conflicts between partially matched hypotheses selected by SM, in favor of the more frequent, already encountered and well-known item of knowledge.
A third primitive, direct inference (DI), operates in WM when the SM and FG have not successfully identified an already compiled plan of actions contained in the knowledge base. DI is based on analogical and inductive/deductive reasoning. Based on the UT, a cognitive model named COSIMO (Cognitive Simulation Model) has been developed to simulate human decision making and behavior in complex working environments (Cacciabue et al., 1989, 1992). The model is used to study human behavior in simulated accident situations and to identify suitable safety recommendations as well as reliability and effectiveness of procedures. The domain in which the model has been applied is supervision and control of nuclear installations.
In the COSIMO model it is assumed that, given the emergency situations in which it is to be applied, no sophisticated reasoning, such as direct inference, is performed and that the operator resorts to previous experience on similar cases. Consequently, COSIMO focuses on the KB of the dual architecture and on the cognitive primitives of SM and FG only. The combination of the capabilities of the KB component and the processing primitives of SM and FG supports a rapid selection of a frame for action.
As a computer model, the UT has been applied to the skill-based behavior in complex environments. Specifically the model has been applied to control a simulation of an auxiliary system of a nuclear-power plant (Cacciabue et al., 1992). Additionally the UT has been used to model the retrieval of incomplete knowledge (Marsden, 1987).
FIGURE 3. The architecture supporting the Underspecification Theory (Cacciabue et al., 1992).
DISCUSSION
Each of the theories of cognitive architectures discussed are based on the unitary approach, and as such represents the state of the art within research in this direction. The basic view held by the approaches discussed is that human cognition can be described as an information processing system. The models all call upon a basic arrangement of memory elements acted upon through a limited array of elementary information processes. The systems react upon stimulus from the environment according to rules guarding the cognitive architecture. They all focus on the elements of cognition and build complexes of cognitive functions from that. This characterization captures the strengths and weakness of the models.
An aspect of specific importance in this context, is how human expertise can be modeled and explained. Although the theories discussed do address the differences in strategy and behavior between experts and novices, they provide only little account of the structure of knowledge utilized by humans confronting complex systems. The models are rich on general control theory, but lack specifications on how human expertise in complex and rich domains develops and how it may be properly formalized.
The work on Soar and ACT* have dedicated little interest to the specific requirements of humans in complex domains, while the Underspecification Theory in its computer embodiment of COSIMO has been applied to model human supervisory control of nuclear power stations.
The information processing approach has been around for some time, and it has provided profound knowledge on cognitive behavior. Questions are however being raised as to whether theories of cognition should address the cognitive architecture as its basic principles. John Anderson has partially abandoned his approach taken in the work on ACT*, because “… it is just not possible to use behavioral data to develop a theory of the implementation level …” (Anderson, 1990, p. 24).
Research on modeling of cognition has taken new directions, in which the notion of cognition as being the result of cognitive primitives has been partially abandoned. Instead the modeling of the overall behavior of human’s and their activities is suggested (Hollnagel, 1993; Hollnagel & Woods, 1983). In this direction to model human behavior and expertise in complex domains, the main emphasis is on the structure of the tasks performed and the competencies to accomplish that.
REFERENCES
Anderson, J. R. (1976). Language, memory, and thought. Hillsdale, NJ: Lawrence Erlbaum Associates.
Anderson, J. R. (1983). The architecture of cognition. Cambridge, MA: Harvard University Press.
Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ: Lawrence Erlbaum Associates.
Anderson, J. R., Kline, P. J., & Beasley, C. M. (1977). A theory of the acquisition of problem-solving skill. In J.R. Anderson (Ed.), Cognitive skills and their acquisition (pp. 45-55). Hillsdale, N.J: Lawrence Erlbaum Associates.
Cacciabue, P. C., Decortis, F., Mancini, G., Masson, M., & Nordvik, J. P. (1989, September). A cognitive model in a black board architecture: Synergism of AI and Psychology. Paper presented at the Second European Conference on Cognitive Science Approaches to Process Control. Siena, Italy.
Cacciabue, P. C., Decortis, F., Drozdowicz, B., Masson, M., & Nordvik, J. P. (1992). COSIMO: A cognitive simulation model of human decision making and behaviour in accident management of complex plants. IEEE Transactions on Systems, Man, and Cybernetics, 22, 1058-1074.
Chomsky, N. (1980). Rules and representation. Behavioral and Brain Sciences, 3, 1-61.
Ernest, G. W., & Newell, A. (1969). GPS: A case study in generality and problem solving. New York: Academic Press.
Hollnagel, E. (1993). Reliability of cognition: Foundations of human reliability analysis. London: Academic Press.
Hollnagel, E., & Woods, D. D. (1983). Cognitive Systems Engineering: New wine in new bottles. International Journal of Man-Machine Studies, 18, 583-606.
Kjær-Hansen, J., Cacciabue, P. C., & Drozdowicz, B. (1991, June). A framework for cognitive simulation. Paper presented at the CADES meeting, Paris.
Laird, J. E., Rosenbloom, P. S., & Newell, A. (1985). Towards chunking as a general learning mechanism. Pittsburgh, PA: Carnegie-Mellon University, Department of Computer Science.
Laird, J. E., Rosenbloom, P. S., & Newell, A. (1986). Universal Subgoaling and chunking: The automatic generation and learning of goal hierarchies. Berlin: Kluwer Academic Publishers.
Luria, A. R. (1973). The Working Brain: An introduction to neuropsychology. Harmondsworth: Penguin Books.
Marsden, P. (1987). The actual frequency of encounter of American Presidents. Manchester: University of Manchester, Department of Psychology.
Newell, A., Shaw, J. C., & Simon, H. A. (1957, Sept.). Empirical explorations of the logic theory machine: A case study in heuristics. Paper presented at the Western Joint Computer Conference. Washington, DC.
Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall.
Newell, A. (1980). Reasoning, problem-solving and decision processes: The problem space as a fundamental category. In N. Nickerson (Ed.), Attention and performance VIII (pp. 105-120). Hillsdale, NJ: Lawrence Erlbaum Associates.
Newell, A. (1982). The knowledge level. Artificial Intelligence, 18, 53-70.
Newell, A. (1990). Unified theories of cognition. Cambridge, MA: Harvard University Press.
Reason, J. (1990). Human error. Cambridge, UK: Cambridge University Press.
Rosenbloom, P. S. (1983). The chunking of goal hierarchies: A model of practice and stimulus-response compatibility (Tech. Rep. No. 83-148). Pittsburgh, PA: Carnegie-Mellon University, Computer Science Department.
1Although fixed, we may envision an architecture to change gradually over time as a result of adaption.
2The application of distinct cognitive levels on a hierarchy of abstractions has been described in relation to The Underspecification Theory (Reason, 1990) in Kjær-Hansen et al. (1991).