Acquiring Rules for Rules: Neuro-Dynamical Systems Account for Meta-Cognition

Abstract

Both animals and humans use meta-rules in their daily life, in order to adapt their behavioral strategies to changing environmental situations. Typically, the term meta-rule encompasses those rules that are applied to rules themselves. In cognitive science, conventional approaches for designing meta-rules follow human hardwired architectures. In contrast to previous approaches, this study employs evolutionary processes to explore neuronal mechanisms accounting for meta-level rule switching. In particular, we performed a series of experiments with a simulated robot that has to learn to switch between different behavioral rules in order to accomplish given tasks. Continuous time recurrent neural networks (CTRNN) controllers with either a fully connected or a bottleneck architecture were examined. The results showed that different rules are represented by separate self-organized attractors, while rule switching is enabled by the transitions among attractors. Furthermore, the results showed that neural network division into a lower sensorimotor level and a higher cognitive level enhances the performance of the robot in the given tasks. Additionally, meta-cognitive rule processing is significantly supported by the embodiment of the controller and the lower level sensorimotor properties of environmental interaction.