Applications of connectionist modeling techniques to simulations of motor control systems

Abstract

The performance of animal motor control systems is unmatched by current robot motor control systems. Attempts to improve robot control systems through a better understanding of animal motor control systems have been defeated by the complexity of the structure of animal control systems. We suggest that difficulties with current models of animal motor control systems are due to the lumped nature of most current approaches and the little isomorphism between model and anatomy. We propose the use of connectionist modeling, a highly parallel, distributed computational framework, to model the fine structure of animal motor control systems. Connectionist techniques have been used in the recent past to model several intelligent processes. These processes have all been Recognition Processes that do not require continuous interaction with a changing environment. There have been no previous connectionist models of Control Processes that require the system to interact continuously with its environment. The primary objective of this research is to examine the effectiveness of connectionist modeling as a framework for simulating neural control processes. We wish to see if connectionist simulations of neural control systems give us computational insights that are difficult to derive from conventional models. A secondary objective to this research is to see if connectionist models of neural motor control systems lead to connectionist controllers for robot systems. In order to so examine the utility of connectionist modeling, we build a model of a part of the Oculomotor (OM) system of the primate. The purpose of the exercise is to see if the connectionist model of the OM system is useful to those wishing to understand the fine structure of the OM system. The connectionist model is successful in that it closely approximates animal behaviour in a wide variety of situations. The model makes specific predictions about the fine structure of the OM system. For example, the model shows how neurons might be interconnected to produce the extended time constant behaviour of the Vestibuloocular Reflex (VOR) and the Optokinetic (OKS) subsystems of the OM system. The model also shows how neurons might be connected to generate the correct interactions between the VOR and pursuit subsystems of the OM system. The model points out anomalies in the gain behaviour of the OM system and also shows neural interconnections that might lead to the adaptive behaviour of the VOR and OKS. The study of the OM system leads to the concept of channel structured control systems in which several identical control systems act in parallel, each computing a small part of the total dynamic range, to effect the final behaviour. . . . (Author's abstract exceeds stipulated maximum length. Discontinued here with permission of author.) UMI