|
Russell's PhD Thesis
I finished my Ph.D. in July 1998, in the department of Electrical and
Electronic Engineering at the University of Auckland, New Zealand.
The thesis title is:
Intelligent Motion Control with an Artificial
Cerebellum.
Abstract
This thesis describes a novel approach for adaptive optimal control and
demonstrates its application to a variety of systems, including motion
control learning for legged robots. The new controller, called ``FOX'',
uses a modified form of Albus's CMAC neural network. It is trained to
generate control signals that minimize a system's performance error.
A theoretical consideration of the adaptive control problem is used to
show that FOX must assign each CMAC weight an ``eligibility'' value
which controls how that weight is updated. FOX thus implements a kind
of reinforcement learning which makes it functionally similar to the
cerebellum (a part of the brain that modulates movement). A highly efficient
implementation is described which makes FOX suitable for on-line control.
FOX requires a small amount of dynamical information about the system
being controlled: the system's impulse response is used to choose the
rules that update the eligibility values. A FOX-based controller design
methodology is developed, and FOX is tested on four control problems:
controlling a simulated linear system, controlling a model gantry crane,
balancing an inverted pendulum on a cart, and making a wheeled robot
follow a path. In each case FOX is effective: it associates sensor
values with (and anticipates) the correct control actions, it compensates
for system nonlinearities, and it provides robust control as long as
the training is comprehensive enough.
FOX is also applied to the control of a simulated hopping monoped, and
a walking biped. FOX learns parameters that fine tune the movements of
pre-programmed controllers, in a manner analogous to the cerebellar
modulation of spinal cord reflexes in human movement. The robots are
successfully taught how to move with a steady gait along flat ground,
in any direction, and how to climb and descend slopes.
|