Controlling a nonlinear spring-mass system with a cerebellar model

摘要: The highly regular anatomy of the cerebellum and its involvement in motor control have inspired a number of theories relating its structure to function (Marr, 1969; Albus, 1971; Kawato et al., 1987; Houk, 1987). Recent anatomical and physiological studies suggest that the neural circuitry of the cerebellum is organized into small interconnected modules (Ito, 1984; Gibson et al., 1987). These findings have motivated the development of the Adjustable Pattern Generator (APG) theory of the cerebellum (Houk, 1987; Houk & Barto, 1992). In this theory, each module is capable of issuing motor commands which will generate some simple parameterized class of movements. The simple movements that the APGs generate act as components of more complex movements. Thus a collection of APGs working together can command a rich set of motor behaviors.The aim of the study presented in this paper is to understand the nature of the learning required so that an APG can control a dynamic plant. In previous APG models (Houk et al., 1990; Berthier et al., 1993), the input patterns were very simple, learning took place on just one synapse of the modeled Purkinje cell (PC) at a time, and the models were used to control kinematic plants. The cerebellar model developed in the present work learns to perform position control of a single degree–of–freedom, nonlinear, dynamic spring–mass system. It employs mossy fiber input codings modeled on neurophysiological data and a large granule cell layer that recodes the mossy fiber pattern into a much higher dimensional space. In this paper, we present a model that consists of a single cerebellar module in order to …