Citation
Sun, Ron. “A Tutorial on CLARION 5.0.” Department of Cognitive Science. Rensselaer Polytechnic Institute. 6 Oct. 2009. http://www.sts.rpi.edu/~rsun/sun.tutorial.pdf
Summary / Assessment
CLARION (short for Connectionist Learning with Adoptive Rule Induction ON-Line) is a cognitive architecture used to simulate human cognition and behavior. Dr Sun has led the development of CLARION at RPI. In the tutorial cited above, Dr. Sun provides an introduction to CLARION and its structure.
One important aspect of CLARION, and something that sets it apart form other cognitive architectures, is the method in which it models human knowledge. In CLARION, knowledge is split up into implicit knowledge, and explicit knowledge. The two types of knowledge are handled differently in the architecture, just as in real-life. For example, implicit knowledge is not directly accessible, but it can be used during computation (just as tacit knowledge is not easily passed on in real-life, but is used by people when solving problems). More specifically, in CLARION, implicit knowledge is modeled as a backpropagation neural network, which accurately represents the distributed and subsymbolic nature of implicit knowledge. On the other hand, explicit knowledge is represented in a symbolic (localist) way. Explicit knowledge is given a direct meaning making it more accessible and interpretable. Explicit knowledge is further divided into “rules” and “chunks”. Rules govern how an agent interacts with its environment (for example: If the stove is hot, don’t touch the stove). Chunks are combinations of implicit dimension/value pairs that are tied together to form a mental concept. For example: table-1: (size, large) (color, white) (number-of legs, four) where “table-1” is the mental concept and the dimension/value pairs are contained in parenthesis.
A layered concept is used when discussing the overall structure of CLARION. Explicit knowledge forms the top layer, while implicit knowledge forms the bottom layer. The architecture is further divided up into subsystems that handle a specific aspect of human cognition. Each subsystem has a module in the explicit layer (top layer), and a module in the implicit layer (bottom layer).
CLARION contains four subsystems: the Action-Centered Subsystem, the Non-Action-Centered Subsystem, the Motivational Subsystem, and the Meta-Cognitive Subsystem. The Action-Centered subsystem represents the part of the mind that controls an agent’s physical and mental actions. (For example, manipulating objects in the real world and adjusting goals.) The ACS uses input from the environment, and other internal information to determine the best action to take. The Non-Action-Centered Subsystem contains what can be considered as general knowledge or “semantic memory”. This type of knowledge includes ideas, objects, and facts. In the NACS, the upper layer contains connections (or associative rules) that connect declarative knowledge chunks, and the bottom layer contains implicit declarative knowledge in the form of dimension/value pair networks. The Motivational Subsystem represents the part of cognition that supplies reasons behind an agent’s actions. The MS describes the “drives” of the agent which in turn determines the agent’s goals. (For example, a drive may be to quench thirst, so the current goal structure would be centered on obtaining a source of water.) The Meta-Cognitive Subsystem is the main controller of cognition. It regulates the communication between all other subsystems. For example: it adds goals to the current goal structure, in response to the drives formulated by the Motivational Subsystem. It also transfers information learned in the ACS to the NACS.
As far as implications for design and HCI: I think that we can learn a great deal about human cognition from utilizing cognitive architectures, and we could leverage them to effectively simulate how our HCI designs will function in the real-world.
