ITERATIVE LEARNING CONTROLLER DESIGN FOR
MULTIVARIABLE SYSTEMS
Manuel Olivares
1
Pedro Albertos Antonio Sala
Dept. of Electronics, Univ. Técnica F. Santa María
P.O. Box 110-V, Valparaíso, Chile
mos@elo.utfsm.cl
Dept. of Syst. Eng. and Control, Univ. Politécnica de Valencia
P.O. Box 22012, E-46071 Valencia, Spain
pedro@aii.upv.es asala@isa.upv.es
Abstract: In this paper, a novel expression for the convergence of an iterative learning control
algorithm for sampled linear multivariable systems is stated. The convergence analysis shows
that, applying this algorithm, the input sequence converges to the system output inverse
sequence, specified as a finite-time output trajectory, with zero tracking error on all the
sampled points. Also, it gives insight on the learning gain matrix selection to act on the
convergence speed or the decoupling of inputs, allowing for an easy tuning using methods
from modern control theory. The results are illustrated by some examples, showing a number
of options to be investigated. Copyright ©2002 IFAC
Keywords: Iterative learning control, multivariable systems, feedforward control,
decoupling.
1. INTRODUCTION
Iterative learning is a well known methodology allow-
ing to incorporate the acquainted experience in the use
of preliminary designs to get the final one. Since the
seminal work of Arimoto, (Arimoto et al., 1984), this
has been successfully applied in the field of control,
mainly for repetitive working conditions, like it is the
case in many manufacturing applications, with a finite
time control horizon.
The iterative feature is very attractive from the user
viewpoint but a number of formal problems arise, the
main one being the existence of a final design, in
some sense defined as optimal, to which the iterative
solution converges. Some sufficient conditions have
been proved, (Moore, 1993; Chen and Wen, 1999),
for special cases. Another important problem, if this
approach is intended to be applied in a more general
range of working conditions, is the need of general-
1
Partially supported by projects USM 23.01.11 and GV00-100-14
ization, that is, to get a final design able to control
the plant under operating conditions differing to those
in which the learning process was carried out. This
can be partially solved in the case of affine in control
systems, where some sort of scaling can be applied.
One of the drawbacks of the approach is that the
controller tends to implement the inverse model of the
plant, restricting its application to stable and inverse
stable open loop plants. Again, the a priori knowledge
of the plant model is a usual requirement that also
limits the range of applications.
The way the iterative learning approach is imple-
mented, usually leads to a feedforward controller. In
any control application, an additional control loop
would be required to guarantee some kind of distur-
bance rejection (Amann et al., 1996b), bounded steady
state error and stability margin, as basic controlled
system features.
In the literature, like in (Shoureshi, 1991; Amann
et al., 1996a), there are many proposed algorithms
Copyright © 2002 IFAC
15th Triennial World Congress, Barcelona, Spain