Theoretical elements of hierarchical control in vertebrate motor systems

Abstract

The complexity of the control problem for vertebrate movement may lead investigators to assume that the controller which solves it must be similarly complex. This assumption implies that electrophysiological measurements from intermediate points in the functioning controller may be difficult to interpret due to their involvement in poorly understood computations. However, if it can be shown that a controller based on simple principles is capable of solving the problem, then simpler models of the biological computation can be proposed which may help to guide physiological research. In this thesis, I describe models which use artificial neural networks to learn certain computations involved in the adaptive control of large-scale nonlinear systems. These models find compact intermediate representations of motor and sensory variables which allow the creation of a hierarchy of subunits. Such a hierarchy can divide the control problem into smaller pieces that can be learned efficiently. Central to these ideas is the concept of optimal unsupervised motor learning, which specifies criteria for choosing an internal representation. I develop algorithms which can compute the optimal internal representation using only observations of sensory and motor variables. The purpose of the theoretical work is to demonstrate the existence of simple algorithms which share some of the impressive motor learning behaviors of vertebrate motor systems. The feasibility of these algorithms is demonstrated using simulations as well as a real two-joint planar robot arm. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)