A linear-complexity reparameterisation strategy for the hierarchical bootstrapping of capabilities within perception-action architectures

摘要

Perception-action (PA) architectures are capable of solving a number of problems associated with artificial cognition, in particular, difficulties concerned with framing and symbol grounding. Existing PA algorithms tend to be 'horizontal' in the sense that learners maintain their prior percept-motor competences unchanged throughout learning. We here present a methodology for simultaneous 'horizontal' and 'vertical' perception-action learning in which there additionally exists the capability for incremental accumulation of novel percept-motor competences in a hierarchical fashion.rnThe proposed learning mechanism commences with a set of primitive 'innate' capabilities and progressively modifies itself via recursive generalising of parametric spaces within the linked perceptual and motor domains so as to represent environmental affordances in maximally-compact manner. Efficient reparameterising of the percept domain is here accomplished by the exploratory elimination of dimensional redundancy and environmental context.rnExperimental results demonstrate that this approach exhibits an approximately linear increase in computational requirements when learning in a typical unconstrained environment, as compared with at least polynomially-increasing requirements for a classical perception-action system.