2011 IEEE International Symposium on Intelligent Control (ISIC)

Part of 2011 IEEE Multi-Conference on Systems and Control

Denver, CO, USA. September 28-30, 2011

Norm-Optimal Control of Time-Varying Discrete Repetitive Processes

El-Sharif A. Omer and Kevin L. Moore Senior Member, IEEE
Division of Engineering, Colorado School of Mines
Golden, CO 80401, USA

Abstract— This paper extends the norm-optimal control de- with boundary conditions xk+1 (0) = dk+1 [5].
sign methodology of iterative learning control to the case of As in ILC, systems such as (3) propagate information
time-varying discrete repetitive processes. As a first contribu- in two independent directions, where the duration of one
tion, starting from a lifting-based formulation, we show a novel
iteration-domain state-space description for discrete-repetitive of these is finite (the along-the-pass or time axis, with
processes. We then pose and solve the norm-optimal control independent variable t), and another is infinite (the iteration
problem for this class of systems. We present solutions for dimension, with independent variable k, also referred to
three cases: optimal regulation, two-degree-of-freedom optimal as the iteration axis). However, unlike ILC, the trial-to-
tracking, and optimal robust servomechanism-based tracking trial memory is inherent in the plant, as opposed to the
of iteration-invariant reference signals.
controller. It has been noted that physical examples of such
Key Words: Repetitive processes, norm-optimal control, it-
processes arise in the coal mining and metal processing
eration domain, iterative learning control.
industries [5]. We further note that such systems are called
I. I NTRODUCTION unit-memory, meaning that only information from the most
Discrete-time iterative learning control (ILC) considers the recent iteration is used in the plant. The continuous version
generation of an input sequence uk (t) for the plant of (3) is called a differential repetitive process while (3) is
called a discrete repetitive process. Here we only consider
xk (t + 1) = Axk (t) + Buk (t),
(1) unit-memory discrete repetitive processes (DRPs).
yk (t) = Cxk (t),
In this paper our focus is to develop norm-optimal control
where the system is assumed to operate with repeated trials strategies for DRPs for the case when the matrices in (3)
k, over a finite trial length. On each trial we have time are time-varying. There are results in the literature on the
t ∈ [0, N − 1] and adopt a common assumption of initial analysis and control of unit-memory DRPs [5], [6], [7],
condition reset, such that xk (0) = x0 . By modifying the [8]. There is also a considerable amount of literature that
input sequence uk (t) from trial-to-trial, such as, for example, combines ILC and optimization, from one of the earliest
uk+1 (t) = uk (t) + γ(r(t) − yk (t)), publications [9] in the time domain to [3], [10], [11] in the
lifted domain (described below). A good summary of these
it can be possible to drive the error between the output yk (t) strategies up to its date of publication can be found in [12].
and the reference signal r(t) to zero. More information about However, apart from some preliminary results in [5] there
ILC can be found in [1], [2]. Here we note only that one are very few results on norm-optimal control of DRPs. We
popular ILC design methodology is norm-optimal ILC [3]. also note that most of the previous work has not considered
Norm-optimal ILC updates the input by solving, at each trial, the internal model principle in addressing the norm-optimal
the optimization problem [4] control of ILC. In this paper we will also combine the idea
uk+1 = arg min {Jk+1 (uk+1 )}, (2) of the internal model principle with the idea of norm-optimal
uk+1 control of DRPs.
subject to the dynamics (1), where the optimality criterion The remainder of this paper is organized as follows: In
used takes various forms, such as, for example, the next section we present some mathematical results and
definitions used in our derivations. Section III then focuses
Jk+1 (uk+1 ) = kek+1 k2 + kuk+1 − uk k2 .
on developing 1-D MIMO difference equation and trans-
The norm-optimal control methodology for ILC has been fer matrix models for both time-varying and time-invariant
shown to be an effective strategy for performance enhance- DRPs, based on a lifting formulation. In Section IV, a 1-D
ment in applications. MIMO state space representation of time-varying and time-
In an ILC scenario a trial-to-trial memory is put into the invariant DRPs in the lifted domain is presented and its
system via the control update. Another class of systems that advantages will be discussed. In the final technical section
also operate over multiple trials with fixed trial length is we pose and solve the norm-optimal control problem for
defined by a plant such as DRPs. We present solutions for three cases: optimal regu-
xk+1 (t + 1) = Axk+1 (t) + Buk+1 (t) + B0 yk (t), lation, two-degree-of-freedom optimal tracking, and optimal
(3) robust servomechanism-based tracking of iteration-invariant
yk+1 (t) = Cxk+1 (t) + Duk+1 (t) + D0 yk (t),
reference signals. Finally a numerical example is given to
Corresponding author: Kevin L. Moore, [email protected] illustrate the proposed algorithm.

978-1-4577-1061-2/11/$26.00 ©2011 IEEE 388

 II. P RELIMINARIES Let us also define the lifted “supervectors” [15] given by
A. Matrix Inversion Lemma Uk = [uk (0), uk (1), · · · , uk (N − 1)]T ,
(9)
For nonsingular A ∈ R N ×N
, the Matrix Inversion Lemma Yk = [yk (0), yk (1), · · · , yk (N − 1)]T .
states that
Proposition 1: For 0 ≤ t ≤ N − 1 and k > 0, a 1-D
−1 −1 −1 −1 −1 −1 −1 difference equation model in the lifted domain for the LTV-
(A − U C V) =A +A U (C − V A U) VA .
DRP (6) is given by
When U = I and C = I, a special case of this lemma (also
called the Woodbury matrix identity) is obtained [13]: Yk+1 = Hvu Uk+1 + Hvy Yk + Hv0 xk+1 (0), (10)

(A − V )−1 = A−1 + A−1 (I − V A−1 )−1 V A−1 . where Hvu and Hvy are lower-triangular matrices of rank N
whose elements are the Markov parameters of the system,
B. w-Transformation given by1
It is useful to introduce an operator transform that operates  
D,0 0 ··· 0
from trial-to-trial. We call this the w-transform. The w-  C1 B,0 D,1 ··· 0 
transform is similar to the standard z-transform, but it is Hvu = 

.. .. .. 

..
operating from trial-to-trial (in the iteration domain), with  . . . . 
time t fixed (as opposed to the z-transform, which operates CN −1 ΦN −1,1 B,0 ··· CN −1 B,N −2 D,N −1
along the time axis on a fixed trial). The one-sided w-
transform W (·) is given as [14] 
D0,0 0 ··· 0

∞ C1 B0,0 D0,1 ··· 0
Hvy = 
 
. X
.. .. ..
W ({uk (t)}) = uk (t)w−k . .. 
(4)  . . . .

k=0 CN −1 ΦN −1,1 B0,0 ··· CN −1 B0,N −2 D0,N −1
Note that using (4) a shift property in the iteration domain  
directly follows: C0
 C 1 A0

W ({uk−1 (t)}) = w−1 W ({uk (t)}). Hv0 =  . (11)
 
(5) ..
 . 
III. 1-D MIMO R EPRESENTATIONS OF D ISCRETE CN −1 ΦN −1,0
R EPETITIVE P ROCESSES
In this section we introduce a lifting technique that is used
This expression is readily derived using the definition of
to develop MIMO difference equations of both time-varying
supervectors (9), the transition matrix (8), and the output
and time-invariant DRPs. For each case we also show transfer
solution of the system (7).
matrix descriptions obtained by applying the w-transform
introduced above. Next, applying the w-transform (4) and its shift property
(5) to the lifted vectors Uk and Yk in (10), when initial
A. Time-Varying Case conditions are zero, gives
Consider the 2-D (SISO) linear, time-varying discrete Y (w) = w(wI − Hvy )−1 Hvu U (w), (12)
repetitive process (LTV-DRP) for 0 ≤ t ≤ N − 1 and k > 0,
= wHvu (wI − Hvy )−1 U (w),
given by
xk+1 (t + 1) = A(t)xk (t) + B(t)uk+1 (t) + B0 (t)yk (t), where U (w) = W ({Uk }) and Y (w) = W ({Yk }).
yk+1 (t) = C(t)xk (t) + D(t)uk+1 (t) + D0 (t)yk (t), Here G(w) = wHvu (wI − Hvy )−1 can be viewed as a 1-D,
(6) MIMO transfer matrix representation in the lifted domain of
the 2-D SISO time-varying representation of the DRP in (6).
with boundary conditions xk+1 (0) = dk+1 .
The output solution of the system (6) can be expressed as: B. Time-Invariant Case
t−1
X The linear, time-invariant discrete repetitive process (LTI-
yk+1 (t) = C(t)Φ(t, i + 1)B(i)uk+1 (i) + D(t)uk+1 (t) DRP) (3) can be considered as a special case of the LTV-DRP
i=0 (6), where all the matrices (A, B, C, D, B0 , and D0 ) are
t−1
X time-invariant. In this case, the transition matrix is
+ C(t)Φ(t, i + 1)B0 (i)yk (i) + D0 (t)yk (t)  t−t
i=0 A 0 , t > t0
Φ(t, t0 ) = (13)
+ C(t)Φ(t, 0)xk+1 (0), (7) I, t = t0 .

where Φ(t, i0 ), for t ≥ i0 , is the transition matrix, given as: 1 To distinguish between B, D and B , D , in the matrices in (11), it
0 0
 should be noted that B,0 , B(0), D,0 , D(0) while B0,0 , B0 (0),
A(t − 1)A(t − 2) · · · A(t0 ), t > t0 D0,0 , D0 (0). For the other variables , the subscript means the value of
Φ(t, t0 ) = (8)
I, t = t0 . that variable at that instant of time.

389
 Proposition 2: For 0 ≤ t ≤ N − 1 and k > 0, the 1-D Proof: Applying the w-transform (4) and its property
difference equation model in the lifted domain for the LTI- (5) to the lifted state space system (16), when initial condi-
DRP (3) is given by tions are zero, gives
Yk+1 = H u Uk+1 + H y Yk + H 0 xk+1 (0), (14) Y (w) = {Hvu + Hvy (wI − Hvy )−1 Hvu }U (w). (18)
where H u and H y are in this case lower-triangular Toeplitz If we let A = Hvy , V = wI and C = U = I in the
matrices of rank N whose elements are the Markov param- lemma (4) and then expand the term (wI − Hvy )−1 in (18),
eters of the system, given by the following transfer matrix in the lifted domain is obtained:
Y (w) = {Hvu − Hvy {Hvy −1 + Hvy −1 (I
 
D 0 ··· 0
 CB D ··· 0 − wHvy −1 )−1 wHvy −1 }Hvu }U (w),
Hu = 
 
.. .. .. .. 
 . . . . which can be simplified as follows:
CΦN −1 B ··· CB D
Y (w) = {Hvu − {I + (I − wHvy −1 )−1 wHvy −1 }Hvu }U (w),
···
 
D0 0 0
CB0 D0 ··· 0 Y (w) = {Hvu − {Hvu + w(Hvy − wI)−1 Hvu }}U (w),
Hy = 

.. .. .. .. 
 . . . .  Y (w) = w(wI − Hvy )−1 Hvu U (w).
CΦN −1 B0 ··· CB0 D0
The last equation is the transfer matrix (12).
Similarly applying the w-transform (4) and its property (5)
 
C
 CA  to the lifted state space system (17), when initial conditions
H0 =  . (15)
 
.. are zero, gives
 . 
CAN −1 Y (w) = {Hvu + (wI − Hvy )−1 Hvy Hvu }U (w). (19)

Again if we let A = Hvy , V = wI and C = U = I in the

This expressions follows directly by using the definition lemma (4) and expand the term (wI − Hvy )−1 in (19), the
of “supervectors” (9), the transition matrix (13), and the following transfer matrix in the lifted domain is obtained:
output solution similar to (7). Note that here the matrices Y (w) = {Hvu − {Hvy −1 + Hvy −1 (I
are Toeplitz, in contrast to the time-varying case, where the
system matrices are only lower-triangular. − wHvy −1 )−1 wHvy −1 }Hvy Hvu }U (w),
Next, following a similar procedure as in the previous which can be simplified as follows:
section, one can show that G(w) = wH u (wI − H y )−1 is
the 1-D, MIMO transfer matrix model in the lifted domain Y (w) = {Hvu − {I + Hvy −1 (I − wHvy −1 )−1 w}Hvu }U (w),
of the 2-D SISO time invariant representation of the DRP in Y (w) = {Hvu − {Hvu + w(Hvy − wI)−1 Hvu }}U (w),
(3).
Y (w) = w(wI − Hvy )−1 Hvu U (w).
IV. 1-D MIMO S TATE S PACE R EPRESENTATION OF
The last equation is the transfer matrix (12) and this com-
DRP S IN THE L IFTED D OMAIN
pletes the derivation.
To this point, a 1-D MIMO difference equation model and
a 1-D MIMO transfer matrix model for 2-D SISO LTV-DRP A. Properties of 1-D MIMO State Space Representation of
and LTI-DRP systems were developed in the lifted domain. LTV-DRPs in the lifted domain
In this section a 1-D MIMO state space model in the lifted It was reported in [5] that the commonly-used 2-D Roesser
domain is given for both linear, time-varying and linear, time- and Fornasini-Marchesini state-space models [16], [17] re-
invariant discrete repetitive processes. This representation is spectively, are not a very convenient framework to address
one of the main contributions of the paper. problems of controllability and observability. Here we show
Theorem 1: For 0 ≤ t ≤ N − 1 and k > 0, a 1-D MIMO that the developed 1-D state space model gives a more con-
state space model in the lifted domain of the dynamics of venient framework to address these problems for DRPs. We
2-D SISO linear, time-varying discrete repetitive process (6) consider separately controllability/observability and stability.
is given by: 1) Controllability and Observability: Consider the DRP
Xk+1 = Hvy Xk + Hvu Uk , representation in (16). Using the properties of lower-
(16) triangular matrices, it is straightforward to show that the
Yk = Hvy Xk + Hvu Uk ,
matrix
or
2
C = Hvu Hvy Hvu · · · Hvy N −1 Hvu ∈ RN ×N , (20)
 
Xk+1 = Hvy Xk + Hvy Hvu Uk ,
(17)
Yk = Xk + Hvu Uk ,
which consists of blocks of lower triangle matrices, has full
where Hvu and Hvy are defined as in (11). row rank (row-rank(C)=N ), provided that D(t) 6= 0 for

390
t = 0, which implies that (16) is controllable. Similarly the In the following subsections we consider three variants
matrix of this problem for the case of a quadratic cost function:
 y 
Hv optimal regulation, two-degree-of-freedom optimal tracking,
 Hvy 2  2
and an optimal robust servomechanism for tracking iteration-
O =  .  ∈ RN ×N , (21) invariant reference signals.
 
.
 . 
Before proceeding, we note that the output of the system
Hvy N (16) has a direct link. For this reason, the optimal state
which consists of blocks of lower triangle matrices, has full feedback control laws developed here will have a different
column rank (column-rank(O)=N ), provided that D0 (t) 6= 0 form of solution than typically found in standard optimal
for t = N − 1, which implies that (16) is observable. It control derivations.
should be noted that, the same results can be obtained if A. Optimal Regulator Problem:
(17) is used.
Theorem 2: Consider the LTV-DRP (16). The optimal
2) Stability: From the properties of lower-triangular ma-
state feedback control law that minimizes the criterion
trices, it is also straightforward to show that the system (16)
L−1
or (17) is asymptotically stable if, and only if, X
J= (YkT QYk + UkT RUk ), (25)
ρ(D0 (t)) < 1 for all 0 ≤ t ≤ N − 1. (22) k=0

where ρ(·) is the spectral radius of a matrix. This result where Q = Q > 0 and R = RT > 0, is given by
T

generalizes that obtained in [5] for time-invariant discrete

unit-memory linear repetitive process. Uk = −Kk Xk − R2−1 R12
T
Xk , (26)
B. 1-D MIMO State Space Representation of LTI-DRPs where Kk satisfies
Corollary 1: For 0 ≤ t ≤ N − 1 and k > 0, the 1-D Kk = {R2 + Hvu T [R1 − R12 R2−1 R12
T
+ Pk+1 ]Hvu }−1 Hvu T ·
MIMO state space model in the lifted domain of the dy- [R1 − R12 R2−1 R12
T
+ Pk+1 ][Hvy − Hvu R2−1 R12
T
],
namics of 2-D SISO linear, time-invariant discrete repetitive (27)
processes (3) is given by
and
Xk+1 = H y Xk + H u Uk ,
(23) Pk = [Hvy − Hvu R2−1 R12
T T
] [R1 − R12 R2−1 R12
T
+ Pk+1 ][Hvy
Yk = H y Xk + H u Uk ,
or − Hvu R2−1 R12
T
− Hvu Kk ], (28)
Xk+1 = H y Xk + H y H u Uk , where
(24)
Yk = Xk + H u Uk ,
R1 = Hvy T QHvy ,
where H u and H y are defined as in (15).
The models (16), (17) and (23), (24) are now in the R2 = Hvu T QHvu + R,
standard state space representation which makes it possible R12 = Hvy T QHvu . (29)
to directly apply existing results from state space control Proof: Substituting for Yk from (16) into (25) yields
theory to discrete repetitive processes, including: L−1
X
1) Optimal controller design. J= (XkT Hvy T QHvy Xk + XkT Hvy T QHvu Uk
2) Pole placement state feedback control (Uk = −KXk ) k=0
with convergence condition |Hvy − Hvu K| < 1, as long + UkT Hvu T QHvy Xk + UkT Hvu T QHvu Uk + UkT RUk ).
as the system (16) is controllable. (30)
3) State observer with convergence condition |Hvy −
Since Q and R are symmetrical positive definite matrices,
Hvu L| < 1, as long as the system (16) is observable.
equation (30) can be simplified to be
4) etc.
L−1
V. N ORM -O PTIMAL C ONTROL OF THE L IFTED LTV
X
J= (XkT R1 Xk + 2XkT R12 Uk + UkT R2 Uk ), (31)
D ISCRETE R EPETITIVE P ROCESSES k=0
Definition 1: The norm-optimal discrete repetitive pro- where the weighting matrices R1 , R2 and R12 are defined
cesses (NO-DRP) problem in the lifted domain is to find as in (29). The minimization of the cost function (31) can be
an input sequence Uk ∈ RN that solves the following directly extended from standard result found in any optimal
optimization problem control text, such as [18], by including the direct link of the
min ||J(Uk )|| output in the steps of the derivation, which results in the
Uk ∈U control law (26).
subject to It should be noted that the difference equation (28) is
Xk+1 = Hvy Xk + Hvu Uk , conveniently solved backward, where first Kk is computed
from the terminal condition Pk+1 = Q through (27), and
Yk = Hvy Xk + Hvu Uk ,
then Pk is computed from Pk+1 and Kk through (28), etc.
for a given J. Equation (28) is the equivalent of the Riccati equation.

391
B. Two-Degree-of-Freedom Optimal Tracking Control
Theorem 3: Consider the LTV-DRP (16). The following
control law minimizes the cost function (25) while producing
zero steady-state tracking error:
−1
Uk = −KXk − R2−1 R12
T
Xk + Gcl (1) R, (32)
where Gcl (w) = {Hvu + (Hvy − Hvu K){wI − Hvy +
Hvu K}−1 Hvu } is the closed loop transfer matrix, K = Fig. 1. The DRP optimal controller structure in the lifted domain.
K + R2−1 R12 T
, K is the steady-state feedback gain ma-
trix, in terms of the original system variables and R =  
[r(0), r(1), . . . , r(N −1)]T is the iteration-invariant reference Xk
where X k = ∈ R2N is the state vector of the
input. Ωk
We comment that this is to be considered an open- augmented system.
loop tracking controller as it consists of a feedback part Theorem 4: Consider the augmented system (36). If the
involving the matrices K and R2−1 R12 T
and a feedforward system described by (16) is completely state controllable,
−1 then the optimal state feedback control law that minimizes
portion involving Gcl (1) . This is also commonly called
a two-degree-of-freedom controller. It should be noted that the criterion
−1 L−1
Gcl (1) exists if and only if the open-loop transfer matrix X
Gol (w) = {Hvu + Hvy (wI − Hvy )−1 Hvu } has no zeros at J= (YkT QYk + UkT RUk ), (37)
w = 1. k=0
Proof: Let us consider any asymptotically control law where Q = Q > 0 and R = RT > 0, is given by
T
of the form
Uk = −KXK + Ur . (33) Uk = −F¯k Xk , (38)
At the steady state we want Yk → R. The closed-loop system ¯ −1 T
 
where Fk = Fk + R2 R12 = Fx −Fr and F satisfies
with the control law (33), becomes T T
Fk = {R2 + B [R1 − R12 R2−1 R12
T
+ Pk+1 ]B}−1 B [R1
Xss = (Hvy − Hvu K)Xss + Hvu Ur ,
(34) − R12 R2−1 R12
T
+ Pk+1 ][A − BR2−1 R12
T
], (39)
R = (Hvy − Hvu K)Xss + Hvu Ur .
Substituting for Xss from the state equation in (34) into the and
output equation and solving for Ur gives (32) and this shows Pk = [A − BR2−1 R12
T T
] [R1 − R12 R2−1 R12
T
+ Pk+1 ][A
that the desired steady-state is reached. It then becomes
straightforward to show that this controller also optimizes − BR2−1 R12
T
− BFk ], (40)
the cost function, again following any standard derivation where y   u 
from optimal control theory. Hv 0 2N ×2N Hv 2N ×N
A= u ∈R ,B= u ∈R , and
−H
 yv  I −H v
C. Optimal Robust Servomechanism C = Hv 0 ∈ R N ×2N
.
In order to achieve robust tracking of an iteration-invariant
input reference (e.g., step functions in iteration) and to limit Here the weighting matrices R1 , R2 and R12 are defined
the effect of plant parameter variations that might occur in as follows:
practice, an integrator in the iteration domain (35) must be T
inserted in the feedforward path between the tracking error R1 = C QC
comparator and the DRP. This is shown in Figure 1. This R2 = Hvu T QHvu + R
is typically called a robust servomechanism and exploits the R12 = C QHvu
T
(41)
well-known internal model principle. To develop an optimal It should be noted that all optimization equations mentioned
controller we first embed the model of the reference in the above are easily extended to the LTI-DRP case simply by
loop and build an augmented plant for which we design an replacing Hvu and Hvy with H u and H y , respectively.
optimal regulator.
Let us define VI. N UMERICAL E XAMPLE
Consider the 2-D (SISO) linear time varying discrete
Ωk+1 = (R − Yk ) + Ωk , (35) repetitive process (LTV-DRP) for 1 ≤ t ≤ 6 and 1 ≤ k ≤
where Ω ∈ RN is the state vector of the integrator. Assuming 20 :
 
all states are measurable and by combining (16) with (35) −0.25t 0
xk+1 (t + 1) = x (t)
0 0.05t k
the following augmented system is defined as    
 y   u    0.5 0.08
Hv 0 Hv 0 + uk+1 (t) + y (t),
X k+1 = X k + U + R, 0.5 0.4 k
−Hvy I −Hvu k I  
yk+1 (t) = 1 1 xk (t) + cos(t)uk+1 (t) + sin(t)yk (t),
Yk = Hv 0 X k + Hvu Uk ,
 y 
(36) (42)

392
 Following the procedure presented above, the 2-D (SISO) the iteration-varying reference input is considered [6]. In this
DRP (42) can be converted to 1-D (MIMO) state space case we use the following dynamics equation:
system (16) with: n−1
X

0.5403 0 0 0

0 0 Ωk+n = (R − Yk ) + ai Ωk+i , (43)
 1.0000 −0.4161 0 0 0 0  i=0
−0.2000 1.0000 −0.9900 0 0 0 
Hvu =  0.1950 −0.3000 1.0000 −0.6536 0 0 
n
Pn−1
where the polynomial ( w − i=0 ai wi ) is the denominator
 
−0.1860 0.3900 −0.4000 1.0000 0.2837 0 polynomial of the reference input in the iteration domain. In
0.2347 −0.4650 0.6500 −0.5000 1.0000 0.9602
future work we will develop a more general norm-optimal
controller that can handle such iteration-varying references.
 
0.8415 0 0 0 0 0
R EFERENCES
 0.4800 0.9093 0 0 0 0 
 0.0000 0.4800 0.1411 0 0 0
Hvy = 0.0360
 [1] H.-S. Ahn, K. Moore, and Y.-Q. Chen, Iterative Learning Control:
 0.0000 0.4800 −0.7568 0 0 
 Robustness and Monotonic Convergence for Interval Systems. London:
−0.0288 0.0720 0.0000 0.4800 −0.9589 0 Springer-Verlag, 2007.
0.0378 −0.0720 0.1200 0 0.4800 −0.2794 [2] D. A. Bristow, M. Tharayil, and A. G. Alleyne, “A survey of itera-
tive learning control: A learning-based method for high-performance
tracking control,” IEEE Control Systems Magazine, vol. 26, no. 3,
pp. 96–114, 2006.
[3] N. Amann, D. Owens, and E. Rogers, “Iterative learning control for
discrete-time systems with exponential rate of convergence,” in IEE
Proceeding of Control Theory and Applications, pp. 217–244, 1996.
[4] D. H. Owens, E. Rogers, and K. L. Moore, “Analysis of linear
iterative learning control schemes using repetitive process theory,”
Asian Journal of Control, Special Issue on ILC, vol. 4, pp. 90–98,
March 2002.
[5] R. Rogers, K. Galkowski, and D. Owens, Control Systems Theory and
Applications for Linear Repetitive Processes. Springer, 2007.
[6] K. L. Moore and E. A. Omer, “Internal model principle for discrete
repetitive processes,” in Proceedings of the 48th IEEE Conference on
Decision and Control, (Shanghai, China), pp. 446–451, 2009.
[7] K. L. Moore and F. Lashhab, “Iteration-domain closed-loop frequency
response shaping for discrete-repetitive systems,” in Proceedings of
2010 American Control Conference, (Baltimore, MD), pp. 1284–1289,
June 2010.
[8] K. L. Moore and E. A. Omer, “Model reference adaptive control of
discrete repetitive processes in the iteration domain,” in Proceedings
of the IEEE Multi- Conference on systems and Control, (Yokohama,
Japan), pp. 386–371, 2010.
Fig. 2. The associated closed-loop poles and zeros (number of poles = [9] M. Togai and O. Yamano, “Analysis and design of an optimal learning
12). control scheme for industrial robots: a discrete system approach,” in
Proceedings of the 24th IEEE Conference on Decision and Control,
(Florida, USA), pp. 1399–1404, 1985.
[10] D. H. Owens and S. Daley, “Iterative learning control- monotonicity
Figure 2 depicts the iteration-invariant desired signal to and optimization,” Int. J. Appl. Math. Comput. Sci, vol. 18, no. 3,
be tracked, the tracking error and the associated closed-loop p. 279293, 2008.
[11] S. Gunnarsson and M. Norrlof, “On the design of ilc algorithms
poles and transmission zeros of the (MIMO) system (36). using optimisation,” IEEE Transaction on Automatica Control, vol. 37,
It should be noted that there are twelve poles in this figure pp. 2011–2016, 2001.
located as follows: six repeated poles at w ≈ 0.3820 and six [12] J. Hatonen, “Issues of algebra and optimality in iterative learning
poles at w ≈ 0. It remains to investigate the properties of [13] control,” PhD thesis, University of Oulu, 2004.
N. J. Higham, Accuracy and Stability of Numerical Algorithms, 2nd
the pole locations for norm-optimal DPR problems. It can ed. SIAM, 2002.
be seen that a fast and perfect tracking is achieved for k < 6 [14] K. L. Moore and M. H. A. Verwoerd, “l1 -optimal robust iterative
learning controller design,” in Proceedings of 2008 American Control
trials. Conference, (Seattle, WA), pp. 3881–3886, June 2008.
[15] K. L. Moore, “Multi-loop control approach to designing iterative
VII. C ONCLUSION learning controllers,” in Proceedings of the 37th IEEE Conference on
Decision and Control, (Tampa, Florida, USA), pp. 666–671, 1998.
In this paper we have presented a state space represen- [16] R. P. Roesser, “A discrete state space model for linear image process-
ing,” IEEE Transactions on Automatic Control, vol. AC-20, pp. 1–10,
tation that can be used for time-varying and time-invariant 1975.
discrete repetitive processes. We then used this representa- [17] E. Fornasini and G. Marchesini, “Doubly-indexed dynamical systems:
tion to develop solutions to several norm-optimal control state space models and structural properties,” Mathematical Systems
Theory, no. 12, pp. 59–72, 1978.
problems, including the optimal regulator, a two-degree-of- [18] H. Kwakernaak and R. Sivan, Linear Optimal Control Systems. New
freedom optimal tracking controller, and an optimal robust York: John Wiley and Sons, Inc., 1972.
servomechanism for iteration-invariant references that used
the internal model principle. We note that the internal model
principle described by the dynamics equation (35) is a
special case of the general internal model principle when

393