M.Sc Thesis, 1996, Electrical Engineering, Technion - Israel Institute of Technology
Amir
Karniel
MODEL OF HUMAN MOTOR CONTROL USING ARTIFICIAL NEURAL NETWORKS
Supervisor:
Professor Gideon Inbar.
Abstract
The human nervous system controls hundreds of muscles and dozens of joints to create posture and smooth movement. The human motor control is a modular hierarchic system that includes various levels of control that act in serial and in parallel. The main levels are the spinal cord, the cerebellum (and other mid brain structures), and some specific areas in the cerebral cortex. Reaching-movement is a fast movement towards a given target. The main characteristics of such a movement are straight line movement and a bell shape speed profile. Reaching-movement is a very fast movement in comparison to the delays and the time constants of the biological system. That is the reason why it is considered to be a ballistic movement, that is, with an open loop control.
In this work a mathematical model for the control of the human arm during reaching-movement is presented. The model of the arm contains the kinematics, and dynamics of a planar manipulator with two degrees of freedom (2DOF), and a mechanical nonlinear model of 6 muscles. The arm model is taken from the literature with minor changes. The spinal cord is modeled by a linear unit for each muscle. Each linear unit corresponds to a motor neuron. The inputs are the sensory input from the peripheral nervous system, and from the central nervous system (CNS). The model includes an option to control the weights of the inputs. This option can be explained as a pre-synaptic inhibition or as the gamma system which controls the sensitivity of the spindles. The cerebellum is modeled as an Adjustable Pattern Generator (APG) that creates the control signals to the muscles. In the model the signals are rectangular pulses activated at various amplitudes and timings, which are determined according to the given target. These amplitudes and timings are the parameters which should be related to each target. The central nervous system includes a neural net which maps any given target to the parameters of the APGs. In order to train this net, the CNS model also includes a sensitivity model to transform the error from the arm coordinates to the parameter coordinates, and an algorithm to learn the weights of the neural nets and the sensitivity model.
In analyzing the model performance in mimicking 2DOF human movements, the following results where observed: (I) A nonlinear model of the muscle can be an advantage. In the presented model the viscosity was a function of the force and of the velocity, according to Hill’s model. This nonlinearity enabled a bell shape velocity in fast movements without overshoot. This conclusion fits the general idea that the motor control is a combination of muscle dynamics with nervous control. (II) A simple feedback control can’t perform fast movements given the long delays in the biological system. (III) A typical reaching-movement can be performed with the typical rectangular signals to the muscles, and in this way the dimension of the control problem is reduced dramatically. (IV) For one joint the artificial network can learn the complete domain; for two joints the performance is good only when the area used for training and performance is limited. (V) The performance of the system is sensitive to the choice of parameters and proper selection of the parameters can make the learning phase shorter and the performance better.