Learning from neural control in motor systems

Abstract

Tremendous efforts have been made to interpret the learning mechanism of motor systems. Based on the experimental results, this paper presents a biologically-plausible computational model for motor learning and control system by using a newly developed deterministic learning theory. In the computational model, the localized Gaussian neural network is employed. By analyzing the properties of the network structure, motor learning ability is implemented dynamically during the process of controlling repeatable movements, and internal model for external force field acting upon the limbs is constructed in a local region along the periodic trajectory. The significance of this paper is that truly learning ability of the internal model for motor control is demonstrated in a dynamic and deterministic manner. Theoretical analysis and numerical simulation show that this novel model can not only provide sound explanations to many experiments of motor learning and control, but also offer a reasonable and feasible solution for the development of smarter robots.