A Straightforward Implementation of Sensorimotor Abstraction in a Two-Layer Architecture for Dynamic Decision-Making

Cognitive and sensorimotor functions are usually assessed separately and therefore also modeled individually although they are strongly intertwined. One way to link these two levels conceptually is sensorimotor abstraction. It is the simplification of complex sensorimotor experiences, and it might enable goal-directed planning in situations with high uncertainties. We propose a computational model for dynamic decision-making that employs two distinct layers, a (lower) sensorimotor control layer holding sub-symbolic information, and a (higher) cognitive control layer holding abstracted information as symbols. In this two-layer architecture information about action control is passed upwards in the hierarchy, abstracted, and used to generate explicit action intentions which are passed downwards again. The hierarchization of model components is intended to represent the different levels of regulatory control (automated vs. fully conscious). We also use different forms of modeling for the individual layers. We employ predictive coding for sensorimotor and ACT-R for cognitive control. An agent equipped with the two-layer architecture is situated in a grid world and tasked to reach a finish line. However, the environment poses challenges on motor control by causing perturbations in the action execution of traversal reflecting varying uncertainty encountered in the real world. Here we describe a straightforward approach to the multi-layer architecture and relate it to the embodied cognition perspective. We also discuss possible extensions that we plan to introduce which depict fundamental cognitive functions such as representing the visual environment in varying granularity.