• Business Analyst - Integration
• Technical Support / Network...
• IT MANAGED SERVICES - BDM -...
• CHANNEL ACCOUNT MANAGER $ 210K...
• Wintel Systems Specialist

Hybrid cognitive architectures

Dr Ron Sun, Information Age

14/10/2005 15:07:32

1 What is a cognitive architecture?

A cognitive architecture is a broadly-scoped, domain-generic computational cognitive model, capturing the essential structure and process of the mind. It may be used for a broad, multiple-level, multiple-domain analysis of cognition (Newell 1990,Sun2002).

An analogy is useful here. The architecture for a building consists of its overall framework and its overall design, as well as roofs, foundations, walls, windows, floors, and so on. Furniture and appliances can be easily rearranged and/or replaced and therefore they are generally not considered part of the architecture. By the same token, a cognitive architecture includes overall structures, essential divisions of modules, relations between modules, basic representations, essential algorithms, and a variety of other aspects (Sun 2004). In general, a cognitive architecture includes those aspects of a cognitive system that are relatively invariant across time, domains, and individuals. In particular, it deals with componential processes of cognition in a structurally and computationally well defined way.

In relation to understanding the human mind (ie, in relation to cognitive science), a cognitive architecture provides a concrete framework for more detailed computational modelling of cognitive phenomena (through specifying essential structures, divisions of modules, relations between modules, and so on).Its function is to provide an essential framework to facilitate more detailed modelling and understanding of various components and processes of the mind. Research in computational cognitive modelling explores the essence of cognition and various cognitive functionalities through developing detailed, process-based understanding by specifying computational models. It embodies descriptions of cognition in computer algorithms and programs. That is, it produces runnable computational models. Detailed simulations are then conducted based on the computational models. In this enterprise, a cognitive architecture may be applied, especially for broad, multiple-level, multiple-domain analyses of cognition.

In relation to building applied intelligent systems, a cognitive architecture specifies the underlying infrastructure for building such intelligent systems, which includes a variety of capabilities, modules, and subsystems. On that basis, application systems presumably can be more easily developed. A cognitive architecture carries also with it theories of cognition and understanding of intelligence gained from studying the human mind. Therefore, the development of intelligent systems can be more cognitively grounded, which may be advantageous in many circumstances.

2 Why are cognitive architectures important?

Psychologically oriented cognitive architectures are particularly important because (1) they are "intelligent" systems that are cognitively realistic(relatively speaking) and therefore they are more human-like in many ways, (2) they shed new light on human cognition and therefore they are useful tools for advancing the science of cognition, (3) furthermore, they may (in part) serve as a foundation for understanding collective human behaviour and social phenomena(Sun and Naveh 2004).

Specifically, for cognitive science, the importance of such cognitive architectures lie in the fact that they are enormously useful in terms of understanding the human mind. In understanding cognitive phenomena, the use of computational simulation on the basis of cognitive architectures forces one to think in terms of process, and in terms of detail. Instead of using vague, hand waving conceptual theories, cognitive architectures force theoreticians to think more clearly.

Thus they are critical tools in the study of the mind. Researchers who use cognitive architectures must specify a cognitive mechanism in sufficient detail to allow the resulting models to be implemented on computers and run as simulations. This approach requires that important elements of the models be spelled out explicitly, thus aiding in developing better, conceptually clearer theories.

Furthermore, an architecture serves as an initial set of assumptions to be used for further modelling of cognition. These assumptions, in reality, may be based on either available scientific data (for example, psychological or biological data), philosophical thoughts and arguments, or ad hoc working hypotheses (including computationally inspired such hypotheses). An architecture is useful and important precisely because it provides a comprehensive initial framework for further modelling in a variety of task domains.

Cognitive architectures also provide a deeper level of scientific explanation of cognitive phenomena and data. Instead of a model specifically designed for a specific cognitive task (often in an ad hoc way), using a cognitive architecture forces modellers to think in terms of the mechanisms and processes available within a generic cognitive architecture that are not specifically designed for a particular task, and thereby to generate explanations of the task that is not centred on superficial, high-level features of a task; that is, explanations of a deeper kind.

To describe a task in terms of available mechanisms and processes of a cognitive architecture is to generate explanations centred on primitives of cognition as envisioned in the cognitive architecture, and therefore such explanations are deeper explanations.

Because of the nature of such deeper explanations, this style of theorising is also more likely to lead to unified explanations for a large variety of data and/or phenomena, because potentially a large variety of tasks, data, and phenomena can be explained on the basis of the same set of primitives provided by the same cognitive architecture. Therefore, using cognitive architectures leads to comprehensive theories of the mind (Newell 1990, Anderson and Lebiere1998, Sun2002).

On the other hand, for the fields of artificial intelligence and computational intelligence (AI/CI), the importance of cognitive architectures lies in the fact that they support the central goal of AI/CI - building artificial systems that are as capable as human beings.

Cognitive architectures help us to reverse engineer the only truly intelligent system around -the human being, and in particular, the human mind. They constitute a solid basis for building truly intelligent systems, because they are well motivated by, and properly grounded in, existing cognitive research.

The use of cognitive architectures in building intelligent systems may also facilitate the interaction between humans and artificially intelligent systems because of the similarity between humans and cognitively based intelligent systems.

3 An example of a cognitive architecture

For instance, CLARION is an integrative architecture (Sun 2002), consisting of a number of distinct subsystems, with a dual representational structure in each subsystem (implicit vs explicit representations).

Its subsystems include the action-centred subsystem (the ACS), the non-action-centred subsystem (the NACS), the motivational subsystem (the MS), and the meta-cognitive subsystem (the MCS).

The role of the action-centred subsystem is to control actions, regardless of whether the actions are for external physical movements or for internal mental operations. The role of the non-action-centred subsystem is to maintain general knowledge, either implicit or explicit. The role of the motivational subsystem is to provide underlying motivations for perception, action, and cognition, in terms of providing impetus and feedback (eg, indicating whether outcomes are satisfactory or not).The role of the meta-cognitive subsystem is to monitor, direct, and modify the operations of the action-centred subsystem dynamically as well as the operations of all the other subsystems.

Each of these interacting subsystems consists of two levels of representation (ie, a dual representational structure): Generally, in each subsystem, the top level encodes explicit knowledge and the bottom level encodes implicit knowledge. The distinction of implicit and explicit knowledge has been amply argued for before (see Reber 1989, Sun 2002).

The two levels interact, for example, by cooperating in actions, through a combination of the action recommendations from the two levels respectively, as well as by cooperating in learning through a bottom-up and a top-down process (see Sun 2002).

Essentially, it is a dual process theory of mind. Symbolic and distributed representation are used in the two levels, respectively. It is therefore a hybrid connectionist-symbolic system.

CLARION has been successful in accounting for, and explaining, a variety of psychological data. For example, a number of well known skill learning tasks have been simulated using CLARION that span the spectrum ranging from simple reactive skills to complex cognitive skills. See, for example, Sun (2002) for some details.

CLARION may also serve as a model for building autonomous intelligent systems. We tried to apply CLARION to a few reasonably interesting tasks in this regard, including learning Tower of Hanoi, and learning minefield navigation. See, for example, Sun and Peterson (1998) for some details.

4 A challenge

In general, building cognitive architectures is an extremely difficult task, because (1) a cognitive architecture needs to be compact but yet comprehensive in scope, (2) it needs to remain reasonably simple yet capture empirical data accurately,

(3) it needs to be computationally feasible but also consistent with psychological theories (4) it needs somehow to sort out and incorporate the myriad of incompatible psychological theories in existence, and so on. However, we can expect that the field of cognitive architectures will have a profound impact on cognitive science, as well as a profound impact on AI/CI.

Their impact may be both in terms of better understanding cognition and in terms of developing better artificially intelligent systems. As such, this research area should be considered a grand challenge and correspondingly a significant amount of collective research effort should be put into it.

For further details of this research area, in particular CLARION, see: Dr Ron Sun is Professor of Cognitive Science at Rensselaer Polytechnic Institute (Troy, New York) and Professor of Computer Science at the University of Missouri-Columbia. He was a plenary speaker at the 9th Knowledge-Based Intelligent Information and Engineering Systems Conference (KES 2005) in Melbourne in September. The ACS was a sponsor.

References J. Anderson and C. Lebiere, (1998). The Atomic Components of Thought. Lawrence Erlbaum Associates, Mahwah, NJ. A.Newell,(1990). Unified Theories of Cognition. Harvard University Press, Cambridge, MA. A.Reber,(1989).Implicit learning and tacit knowledge. Journal of Experimental Psychology:General.118(3),219-235. R.Sun, (2002).Duality of the Mind. Lawrence Erlbaum Associates, Mahwah, NJ. R.Sun, (2004).Desiderata for cognitive architectures. Philosophical Psychology,17(3),341-373. R. Sun and T. Peterson, (1998). Autonomous learning of sequential tasks: experiments and analyses. IEEE Transactions on Neural Networks, 9 (6),1217 1234.