Computational modelling of the human motor control system: Nonlinear enhancement of the adaptive model theory through simulation and experiment
Thesis DisciplineElectrical Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Adaptive Model Theory (AMT) proposes that the brain forms and adaptively maintains inverse models of the world around it for adaptive feedforward control. This leading motor control theory unites principles of neurobiology, psychology and engineering. A modified version of AMT was developed with the capacity to control nonlinear systems, to predict signals with nonlinear statistical characteristics, and to perform simultaneous feedback and feedforward adaptive control. The modified version is called nonlinear Adaptive Model Theory or nAMT. An experimental study was also performed investigating inverse model formation in the human motor control system, the results of which were then compared with the nAMT model. A nonlinear dynamic system identification method was developed for nAMT to replace the linear structures employed by AMT. This method employs a neurobiologically-inspired locally-recurrent neural-network structure. A multi-layer adaptation algorithm was also developed specifically for this structure. Nonlinear AutoRegressive Moving-Average (NARMA) adaptive predictor structures replace the linear Moving Average (MA) predictor circuits used in AMT. Adaptive feedback control is augmented using a nonlinear dynamic forward model observer to improve the quality of the estimated response signal. Nonlinear dynamic inverse models are formed by placing the forward model in an internal feedback loop in which the gain function is adjusted to maintain stability. The internal inverse model motor-control hypothesis was tested experimentally in a study looking at human open-loop performance in a tracking task. The study was aimed at directly demonstrating the formation of an internal inverse model of a novel visuomotor relationship for feedforward control in the brain. The study involved 20 normal adult subjects who performed a pursuit random tracking task with a steering wheel for input. During learning the response cursor was periodically blanked, removing all feedback about the external system (i.e., about the relationship between hand motion and response cursor motion). Results showed a transfer of learning from the unblanked runs to the blanked runs for a static nonlinear system (14% median improvement between first 4 and last 4 runs, p = .001) thereby demonstrating adaptive feedforward control in the nervous system. No such transfer was observed for a dynamic linear system, indicating a dominant adaptive feedback control component. The observed open-loop responses showed a high-pass frequency response which could not be explained using traditional control-systems motor control models. Experimental results were compared with simulated results from the nAMT model. Results from the experimental study were used to verify and tune the computational model. The resulting simulations produced effects that mirrored the closed- and openloop characteristics of the experimental response trajectories. This supports the claim that an internal feedback loop is used for the inversion of external systems in the human brain. Other control-systems models (both AMT and feedback-error learning) would require substantial ad hoc modification to reproduce the observed disparity between closed- and open-loop results. In contrast, nAMT naturally reproduced the effect as a consequence of its novel nonlinear inversion method. In nAMT an inverse model is formed by embedding a forward model in an internal feedback loop incorporating a low derivative gain. The derivative loop-gain caused the inverse model to be relatively inaccurate at low frequencies, for which the feedback control loop was adequate, but to be increasingly accurate at higher frequencies. Maintenance of the loop-gain at the lowest possible levels maximizes the internal stability of the inverse. The simulation work confirmed that the nAMT model is capable of reproducing human behaviour under a wide range of conditions.