Ph.D. Thesis on Neural/Fuzzy Control available.

Thu Oct 1 00:07:56 EDT 1998

Hi,

My recently awarded Ph.D. thesis is now available from the following
web-page:
  http://www.elec.uq.edu.au/~annis/papers/HansThesis/theCOEM.html

The abstract and other details are included at the bottom of this
message.

----------------------------------------------------------------------

The Controller Output Error Method
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Ph.D. Thesis by:
  Hans Christian Asminn Andersen
Supervised by:
  Dr Louis Westphal
in the field of:
  Electrical Engineering
at the:
  Department of Computer Science and Electrical Engineering,
  University of Queensland, Brisbane, Australia.

Abstract:

This thesis proposes the Controller Output Error Method (COEM) for
adaptation of neural and fuzzy controllers. Most existing methods of
neural adaptive control employ some kind of plant model which is used to
infer the error of the control signal from the error at the plant
output. The error of the control signal is used to adjust the controller
parameters such that some cost function is optimized. Schemes of this
kind are generally described as being indirect.

Unlike these, COEM is direct since it does not require a plant model in
order to calculate the error of the control signal. Instead it
calculates the control signal error by performing input matching. This
entails generating two control signals; the first control signal is
applied to the plant and the second is inferred from the plant's
response to the first control signal. The controller output error is the
difference between these two control signals and is used by the COEM to
adapt the controller.

The method is shown to be a viable strategy for adaptation of
controllers based on nonlinear function approximation. This is done by
use of mathematical analysis and simulation experiments. It is proven
that, provided a given controller is sufficiently close to optimal at
the commencement of COEM-adaptation, its parameters will converge, and
the control signal and the output of the plant being controlled will be
both bounded and convergent. Experiments demonstrate that the method
yields performance which is comparable or superior to that yielded by
other neural and linear adaptive control paradigms. In addition to these
results, this thesis shows the following:
 * The convergence time of the COEM may be greatly reduced by    
   performing more than one adaptation during each sampling period.
 * It is possible to filter a reference signal in order to help ensure 
   that reachable targets are set for the plant.
 * An adaptive fuzzy system may be prevented from corrupting the 
   intuitive inter-pretation upon which it was originally designed.
 * Controllers adapted by COEM will perform best if a suitable sampling 
   rate is selected.
 * The COEM may be expected to work as well on fuzzy controllers as it 
   does on neural controllers.  Furthermore, the extent of the 
   functional equivalence between certain types of neural networks and 
   fuzzy inference systems is clarified, and a new approach to the 
   matrix formulation of a range of fuzzy inference systems is 
   proposed.