Autonomous learning in psychologically-oriented cognitive architectures: A survey

Abstract

Abstract This survey paper discusses the topic of autonomous learning in psychologically-oriented cognitive architectures and reviews some of the most popular cognitive architectures used in psychology, namely ACT-R, Soar, and Clarion. Autonomous learning is critical in the development of cognitive agents, and several learning-related desiderata useful for ‘psychological’ cognitive architectures are proposed. This article shows that all the reviewed cognitive architectures include some form of explicit (‘symbolic’) and implicit (‘subsymbolic’) learning. Additionally, ACT-R and Clarion are shown to include a top–down learning algorithm (from explicit to implicit), and Clarion also includes a bottom–up learning process (from implicit to explicit). Two simulation examples are presented with each cognitive architecture to illustrate the autonomous learning capacities of each modeling paradigm. While Clarion is more autonomous (requiring less a priori knowledge), Soar and ACT-R have so far been used in more complex tasks. The presentation concludes with some general considerations for future work.