The CLARION Cognitive Architecture: A Tutorial

The CLARION Cognitive Architecture: A Tutorial Sebastien Helie (helies@rpi.edu) Cognitive Science Department, Rensselaer Polytechnic Institute 110 Eighth Street, Troy, NY 12180, USA Nick Wilson (wilson3@rpi.edu) Cognitive Science Department, Rensselaer Polytechnic Institute 110 Eighth Street, Troy, NY 12180, USA Ron Sun (rsun@rpi.edu) Cognitive Science Department, Rensselaer Polytechnic Institute 110 Eighth Street, Troy, NY 12180, USA level, implicit associations are learned through gradual trial- and-error learning. In contrast, learning of explicit knowledge is often one-shot and represents the abrupt availability of explicit knowledge (following “explicitation” or through newly acquired linguistic information in the top level). The emphasis on bottom-up learning (i.e., the transformation of implicit knowledge into explicit knowledge) is, in part, what distinguishes CLARION from other cognitive architectures. Nevertheless, top-down learning is also included in CLARION. Abstract This full-day tutorial introduces participants to CLARION, a dual-process/dual-representation cognitive architecture that focuses on the distinction between explicit and implicit cognitive processes. CLARION is also integrative, involving cognition, motivation, metacognition, and so on. This tutorial presents a detailed description, along with many simulations, advanced topics, and formal results. Although some prior exposure to cognitive architectures and artificial neural networks can be helpful, prior understanding of these areas is not required, as the full-day format allows a detailed presentation of basic, as well as advanced, topics related to cognitive modeling using CLARION. This tutorial will enable participants to apply the basic concepts, theories, and computational models of CLARION to their own work. The Action-Centered Subsystem In the Action-Centered Subsystem, the top level contains simple “State AE Action” rules, while the bottom level uses multi-layer perceptrons to associate states and actions. Learning in the bottom level is captured by a reinforcement learning algorithm (with backpropagation), while rule learning in the top level is mostly “one-shot” and can be performed bottom-up or independently. The ACS has been used to model navigation in mine fields (Sun et al., 2001). In addition, because CLARION focuses on the dichotomy between explicit and implicit knowledge, benchmark psychological tasks used to show implicit learning were also successfully captured and explained (Sun et al., 2005). Overview CLARION is a cognitive architecture composed of four main subsystems: the Action-Centered Subsystem (ACS), the Non-Action-Centered Subsystem (NACS), the Meta- Cognitive Subsystem (MCS), and the Motivational Subsystem (MS). The ACS is used mainly for action decision-making. The NACS is usually a slave system to the ACS and is used to store declarative and episodic knowledge. This subsystem is also responsible for reasoning in CLARION. The MS is responsible for determining motivational drive levels (which in turn lead to the setting of goals). The MCS is responsible for cognitive monitoring and parameter setting in both the ACS and NACS, and makes the goal setting determinations based on drive levels reported from the MS. In addition to the aforementioned subsystem structure, CLARION is based on two other basic assumptions: representational differences and learning differences of two different types of knowledge: implicit versus explicit (Sun, Merrill, & Peterson, 2001; Sun, Slusarz, & Terry, 2005). The main difference between these two types of knowledge is accessibility. In each subsystem, the top level contains explicit knowledge (easily accessible) whereas the bottom level contains implicit knowledge (harder to access). The second assumption in CLARION concerns the different learning processes in the top and the bottom level of each subsystem (Sun et al., 2001, 2005). In the bottom The Non-Action-Centered Subsystem In the Non-Action-Centered Subsystem, the top level contains simple logical rules while the bottom level uses a nonlinear neural network. Learning in the bottom level is captured by associative (e.g., contrastive Hebbian) learning. Rule learning in the top level is mostly “one-shot” (similar to the ACS). The NACS in CLARION has been used mainly to simulate reasoning. In particular, CLARION was able to capture data showing the mixed effect of rule-based reasoning and similarity-based reasoning when judging the likelihood/strength of simple deductive forms. Other reasoning phenomena can also be naturally explained using CLARION (e.g., inheritance-based reasoning, reasoning from incomplete information, etc).