Professional Documents
Culture Documents
2001 Songchon&Longman The Waterbed Effect and Stability PDF
2001 Songchon&Longman The Waterbed Effect and Stability PDF
This manuscript has been reproduced from the microfilm master. UMI films the
text directly from the original or copy submitted. Thus, some thesis and
dissertation copies are in typewriter face, while others may be from any type of
computer printer.
The quality o f this reproduction is dependent upon the quality o f the copy
submitted. Broken or indistinct print, colored or poor quality illustrations and
photographs, print bleedthrough, substandard margins, and improper alignment
can adversely affect reproduction.
In the unlikely event that th e author did not send UMI a complete manuscript and
there are missing pages, these will be noted. Also, if unauthorized copyright
material had to be removed, a note will indicate the deletion.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The Waterbed Effect and Stability in Learning/Repetitive Control
Thuanthong Songchon
COLUMBIA UNIVERSITY
2001
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
UMI Number: 3005799
__ ___ __®
UMI
UMI Microform 3005799
Copyright 2001 by Bell & Howell Information and Learning Company.
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
©2001
Thuanthong Songchoii
All Right Reserved
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A bstract
Thuanthong Songchon
feedback control systems. It says that if the frequency response o f a feedback control
system has the property that it can significantly attenuate the effects o f disturbances
in one frequency range, then it must amplify disturbances in some other frequency
range. Learning and repetitive control often aim to eliminate all periodic disturbances
o f a fixed period, i.e. for a fundamental frequency and all harmonics. This suggests
that one must pay for this elimination by amplifying errors that occur between these
evenly spaced frequencies. In this thesis, learning and repetitive control are studied
using linear phase lead as well as low pass filtering for a frequency cutoff o f the
learning for purposes o f stabilization. It is shown that learning control has the ability
to bypass the waterbed effect, and the range o f mechanisms for doing so are
described. Technically, in finite time problems such as learning control one is never
in steady state response. Hence, the waterbed conclusions apply to that part o f the
trajectory beyond a settling time o f the system. In the repetitive control problem, time
shown that for real time repetitive control, one is subject to the waterbed effect. On
the other hand, it is possible to bypass this effect by making batch updates of the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
repetitive control signal, at least for periodic inputs and disturbances o f the period
being addressed. Sufficiently slow updates will result in the situation that applies to
learning control. Methods are developed to predict the disturbance to error steady
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
T ab le o f C o n ten ts
Chapter 1 1
Introduction
1.1 Background 1
1.2 Dissertation Outline 2
Chapter 2 5
Stability Conditions of Learning and Repetitive Control
2.1 Introduction 5
2.2 A Frequency Response Based Stability Condition 6
2.3 The True Stability Boundary for ILC and RC 8
2.4 Stability o f First Order Systems 10
2.5 Stability o f Second Order Systems 13
2.6 Stability o f Third Order Systems 15
2.7 Conclusions 16
Chapter 3 23
Design Trade-offs in Learning/Repetitive Controllers Using Zero-
Phase Filtering
3.1 Introduction 23
3.2 Sources o f Tracking Error in Feedback Control Systems 23
3.3 Requiring Good Learning Transients in Learning and 25
Repetitive Control
3.3.1 Learning Control 25
3.3.2 The Good Transient Condition in Learning Control 28
3.3.3 Repetitive Control 30
3.4 An Example System for Numerical Investigations 32
3.5 What Do We assume We Know About The System When 33
We Design The Learning or Repetitive Controller
3.6 The Design Approach 34
3.6.1 Compensator Design 35
3.6.2 Zero-Phase Low Pass Filtering 36
3.6.3 Purpose o f This Chapter 39
3.7 The Steady State Error Level When Filtering The Total Command 39
Signal Going Into The Feedback Controller
3.7.1 Iterative Learning Control 39
3.7.1.1 A Simple Special Case 41
3.7.1.2 Numerical Example Using a Zero-Phase 42
Butterworth Filter
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
3.7.2 Repetitive Control 43
3.7.3 Periodic filtering in Learning Control 45
3.8 The Steady state Error level W hen The Desired Trajectory Part 45
o f The Command Is N ot Filtered
3.8.1 Iterative Learning Control 45
3.8.2 Comparison W ith the Previous Case 47
3.8.3 Repetitive Control 48
3.9 Conclusions 49
Chapter 4 57
The Waterbed Effect and Iterative Learning Control Using Zero-
Phase Filtering
4.1 Introduction 57
4.2 The Standard W aterbed 60
4.3 A Robot Example 62
4.4 Generalized Bode Integral Theorem 62
4.5 Finite Time Trajectories and The Frequencies in GBI Theorem 66
4.6 Implications for Causal ILC Laws 66
4.7 GBI Theorem and N oncausal ILC 68
4.8 Low-Pass Filter Learning C utoff 70
4.9 GBI Theorem and Zero-Phase Filtering 72
4.10 Conclusions 74
Chapter 5 11
The Waterbed Effect and Repetitive Control with Filtering
5.1 Introduction 77
5.2 Iterative Learning Control and Repetitive Control Background 78
5.3 Batch Zero-Phase Low-Pass Filtering 81
5.4 Real Time Low Pass Filtering 85
5.5 Some Waterbed Conclusions 86
5.6 The Need for Zero-Phase Filtering in RC 92
5.7 Understanding The M echanism for Amplification 93
5.8 Comments on Stability Conclusions and Stability Boundary 95
5.9 Frequency Response o f Periodic Update Repetitive Control 99
5.10 Conclusions 110
Chapter 6 121
Conclusions
Refferences 123
ii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A c k n o w le d g m e n ts
Longman for his support, encouragement, and guidance w hile working on this
Pleum jitr Songchon, for their constant love, support, and education. Also, I would
like to thank my sisters, M s.Krisana Songchon and M s.Bavom nan Songchon for their
iii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
1
C hapter 1
Introduction
1.1 Background
fields o f learning and repetitive control, apply to situations in which the same
command is given to a feedback control system many tim es, or the command is
periodic, and these controllers examine the tracking error o f the system from each
repetition or period, and adjust the command the next repetition or period in order to
try to make the error go to zero. Thus, by simply changing the com m and given to a
control system that has a desired output that is constant, but it is subject to periodic
disturbances.
produces zero tracking error following a desired trajectory. The system is restarted
from the same initial condition each time a command is given. Repetitive control
(RC) applies to systems with a periodic desired output, or with a constant desired
output and there are periodic disturbances, and the command to the closed loop
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
system is adjusted from one period to the next in order to converge to zero tracking
error.
appeared simultaneously in the early 1980s. Arimoto et al. [1]. Casalino and Bartolini
[4] and Craig [5] are independent developments o f similar ideas that year, with
Uchiyama [35] being one o f the precursors. Middleton et al. [23], subm itted in 1984.
was another independent developm ent motivated by robotics but using repetitive
control. The origin o f repetitive control (RC) had different motivation, and early
works include Inoue et al. [16], O m arta et al. [26], Hara et al. [10,11], Nakano and
Hara [24], and Tomizuka et al. [34], These two fields appear to be very different in
the literature, but we show that in practical use they are not so different.
and develops an understanding o f the relationships between the waterbed effect and
for stability for both learning and repetitive control. In learning control the stability
boundary is presented as given in Phan and Longman [27], For repetitive control the
methods are different. One condition for defining the true stability boundary, makes
use o f the method o f Pierre [28], and the second condition is an approximate
frequency response based monotonic decay condition o f Elci et al. [7], In this chapter.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
these stability conditions are compared for first order, second order, and third order
systems for both learning and repetitive control, [31]. It is shown that for practical
using a zero-phase filter. In previous works o f Elci et al.[6,7] and H sin et al.[l3],
zero-phase infinite impulse response (HR) low-pass filtering is used. In this chapter
we develop m ethods o f predicting the steady state error levels that will be reached for
control. The original work o f Bode [2] used analytic function theory to examine the
properties o f a continuous feedback loop in the frequency domain. Later, many others
such as, in Seron et al. [30], studied the analogous results for the discrete time case.
The discrete time version is o f interest in here. All the methods are restricted to the
case o f all poles o f the sensitivity transfer function being inside the unit circle, and the
sensitivity transfer function being a proper rational function. In learning and repetitive
control, improper sensitivity functions often occur, especially when one uses zero-
phase filtering. In this chapter we extend the Bode integral theorem to the more
general cases needed for learning and repetitive control systems [32].
Chapter 5 studies the waterbed effect and repetitive control using filtering for
a frequency cut o ff o f the learning for purposes o f stabilization. It is show n that for
real-time repetitive control, one is subject to the waterbed effect. On the other hand, it
is possible to bypass this effect for the frequencies being learned by m aking batch
updates o f the repetitive control signal. Sufficiently slow batch updates in repetitive
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
4
control results in the same situation that applies to iterative learning control for the
frequencies being learned, where it is possible for certain classes o f learning control
laws to bypass the w aterbed effect. Methods are developed to predict the disturbance
to error frequency response characteristics for the range o f possible batch update rates
[33].
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5
Chapter 2
2.1 Introduction
Stability is one o f the major concepts in control systems. Without stability, the
goal o f the control cannot be achieved; with stability the system can at least be used.
In both iterative learning control (ILC)-and repetitive control (RC) cases it is common
derived to suggest convergence based on decay o f the steady state frequency response
components o f error. Elci et al. [7] and Longman [19] show that this condition is a
sufficient condition for asymptotic stability o f the learning process, but suggests that
its real use is as a technique for ensuring good transients during the learning process.
Huang and Longman [15] suggest that the discrepancy between this frequency
response based condition and the true stability boundary is very large for ILC, but for
RC the condition will normally be very close to the true stability boundary. It is the
purpose of this chapter to investigate the distinction between the frequency response
based condition and the true stability boundary, for the set o f all first order systems,
all second order systems, and all third order systems with no zero. This is done for the
most basic form o f ILC and RC, integral control based learning. In the case o f ILC
this adjusts the command at time step k o f repetition j according to Uj-i(kT) = Uj(kT)
+<fiej((k+l)T), where <p is the learning gain, T is the time step interval, and e, is the
tracking error (desired output minus actual). The one step time shift in the error
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
6
accounts for the usual one step delay between a change in the input to a change in the
output in a digital system . In repetitive control the change in repetition num ber j is
replaced by a tim e shift o f p time steps corresponding to one period. W ritten in terms
o f z-transform s this becomes 2?U(z) = U{z) + (jrzE(z). The associated repetitive control
system block diagram has the desired output as the command, uses unity feedback,
has a repetitive controller box with transfer function <f/zl(zp - I ) , and the plant is the
transfer function o f the closed loop feedback control systems w hose input is being
A com m on stability condition for both iterative learning control and repetitive
( 2 . 1)
for all frequencies co up to Nyquist. where T is the sample tim e interval, and Gc(z) is
the z-transfer function o f the associated closed loop control system . It can be shown
that this condition is a sufficient condition for convergence to zero tracking error for
both ILC and RC (Longm an [19], Elci et al. [7], Huang and Longm an [15]). These
references suggest that its real importance is as an approximate condition for assuring
To see this for ILC, suppose that the output o f the closed loop control system
is Y{z) = Gc{z) + W(z) where W represents any disturbance that appears every time the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
7
command is given. Write this for repetitions /+ 1 and j, and take the difference.
Express the result in terms o f a difference o f errors, and use the learning control law-
described in the previous section to produce the input in terms o f the error. The result
can be w ritten as Ej+\(z) = [1 -<fizGc(z)]Ej(z) (note that the initial condition on E is zero
for ILC). By satisfying (2.1). it is guaranteed that the steady state frequency response
components o f the error will decay monotonically with repetitions. However, since
Nevertheless, w hen the trajectory is significantly longer than a few time constants o f
the system, this condition makes a good condition to satisfy in order to ensure good
For RC, find the transfer function from the periodic desired trajectory' YJ^z) to
the associated error, for the block diagram described in the previous section. This
due to the periodicity with period p o f the desired trajectory and the disturbance. This
note that the multiplication by z p is a shift one period forward in time. This suggests
that if (2.1) is satisfied, there will again be monotonic decay o f each frequency
component o f the error. This time there is a quasi-static assumption made, in order to
If one chooses to satisfy (2.1) in order to obtain good transient o f the learning
process, it is o f interest to know how much more restrictive satisfying (2.1) is than
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
8
simply satisfying the if and only if condition for stability. In this paper we show how
these differ for ILC and RC for first, second, and third order systems.
In ILC the true stability boundary when using integral control based learning
where B and C are from the discrete-time state-space model x((/t+l)7~) = Ax(kT) +
In repetitive control, the most natural way to express the true stability
boundary is in terms o f the Nyquist stability criterion. The repetitive control loop
contains the closed-loop feedback control system Gc( z ) and the repetitive control
law <pzl(zp - \ ) together with unity feedback, so that the characteristic polynomial
convenient for us to separate the DC g a i n o f the feedback control system from the
rest o f its transfer function according to Gc (z) = KcG ( z ) , and then define the product
o f the DC gain with the learning gain as K = Kc<f>, and then this serves as our gain
the discrete time Nyquist criterion is inconvenient because o f the p roots on the unit
circle. Following Huang and Longman [15], we apply the method o f Pierre [28] to
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
9
The troublesome poles on the unit circle become zeros o f O, simplifying the plotting.
Plot Q{e ' ° ) for 0 going from 0 to 180 deg, deleting any points for which O is singular
(one does not have to go around these points). For any gain K as a point on the real
Z = ( - W /180°) + Pq 4 - ( « ’/ 2) (2.4)
where
W is the angle swept by the vector pointing from point (AT, 0) to the moving
point O(e ‘0) for 6 going from 0 to 180 deg with singularities deleted. Clockwise is
counted as positive.
Pq is the total number o f poles o f O outside the unit circle, finite poles plus
poles at infinity.
Z is the number o f zeros o f -K + Q(z) = 0 that are outside the unit circle.
In using (2.4), one normally knows the values for Pq and r i , W is determined from
the plot, and hence Z is known. The system is stable for all K that produce Z equal
zero. Note that this stability condition which represents the true stability boundary,
depends on the number o f time steps p in a period, whereas the previous approximate
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
10
condition (2.1) and the true stability boundary (equation (2.2) for ILC and (2.4) for
RC) for all possible first order systems. Start with a continuous tim e transfer function
G v(s) = K ca /(s + a) where K c is the DC gain. When fed by a zero order hold, this
take the z-transform o f the function represented in the square bracket. Then the Gc(z)
There are three parameters whose values may affect these stability conditions: the
gain K = <pKc which is the product o f the learning gain with the DC gain o f the
system, the value o f aT related to the time constant o f the system and the sample time,
and the num ber o f time steps p in the desired trajectory or period.
Concerning Limits on the Parameters: In order for a discrete tim e control system to
function well,one should have the sample rate be such that there are several time
steps in a tim e constant o f the system. In this case the time constant is1/a. so a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
11
generous upper limit on the value o f a T is unity. W hen we get to second order
system s s 2 + 2^a>0s + col - we need at least one sample per time constant when the
roots are real, and when the roots are complex we again ask for one sam ple per time
constant for the real part o f the root, and at least two sample times per period for the
oscillatory part o f the root (this lim its values o f co0T to a maximum o f n r / -Jl - C 2 ).
condition (2.1) can be found by plotting <pe'°Gc(e‘e ) for £?going from 0 to 180 deg,
and seeing how large K can be before the curve goes outside the unit circle centered
at +1. This happens first w hen 9 = 0 . and produces the inequality 0 < K < 2.
param eter a T as well. As stated above, this is a sufficient condition for stability for
The True Stability Boundary f o r Learning Control: The discrete tim e state variable
respectively. Then the stability boundary is given by 0 < K < 2/(1 —e~a l). This is
independent o f p, and the stable range on K tends to infinity as the sam ple time tends
to zero. The boundary is show n in Fig. 2.1. It is always larger than the sufficient
condition (2.1) as it must be, and it becomes arbitrarily larger as the sample time
approaches zero.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
12
The True Stability Boundary f o r Repetitive Control: For conciseness, denote sine by
A typical plot o f Q(e'°) is given in Fig. 2.2, p — 10 and a T = 0.88. Applying the
modified Nyquist plot rules above to any point (K.0) between 0 and the first time the
plot crosses the positive real axis produces, W = 5 • 180°, Pq — 10, n' = 0 with Z = 0.
Hence, all gains K between zero and this first crossing o f the positive real axis
correspond to stability. For this first order system, it happens that the first loop is
always the one determining stability, but for the second and third order systems
discussed below, this is not necessarily the case. Then, the procedure for determining
the maximum stable gain Kmax as the parameters o f the system are run through their
(i) Set the imaginary part o f 0 ( e ' ° ) equal to zero. For the case o f equation (2.8)
e~aTc{{p - 2 ) 9 / 2 ) - c ( p 9 / 2 ) = 0 (2.9)
(ii) Substitute each solution for 9 into O (ei0) to find the associated value o f K
according to (2.3). The minimum o f these values is the gain for the stability
boundary, K max.
Figure 2.3 shows the results o f this procedure for the first order system, giving
Kmax for a T in the range o f reasonable values from 0 to I, and for various values o f
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
13
the number o f tim e steps in a period, p. As m ust be the case, the stability boundary is
always above the value K = 2 given by the monotonic decay condition (2.1).
However, this difference is only substantial w hen the num ber o f time steps in a period
is quite sm all, e.g. for a p o f 10. In typical digital control systems with sam ple rates
like 100 or 1000 Hz, any reasonable length period for the desired m otion will have a
p sufficiently large that the distinction between the true stability boundary in
repetitive control, equation (2.4), and the approximate monotonic decay condition
(2.1) is close to a requirement, even though it does not correspond to the true stability
boundary. T he difference between (2.1) and (2.4) is insignificant for typical length
using (2.4).
N ow consider the set o f all stable strictly proper second order system s. The
transfer function in continuous time is G f(s) = K cco] (ds + l ) / 0 r + 2Ccoos -hco2) . and
we consider th at it is fed by a zero order hold, and then convert to the associated z-
transfer function. This time the parameters that can influence stability are K , coaT ,C ,
d/T. and p .
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
14
a, = ^ , T - A = c o j 4 \ - C 1 , r = cotJ , S = d !T (2.10)
a 2 =( o 0t {c + 4 ^ - - \ \ J32 = c o j { c - 4 C - 1) (2-11)
A3 = ( r 2J - r - l ) e r + e 2r , A = 1+ ( r - r 2^ - l)er (2.12)
These Gc(z) are used in (2.1) for the approximate monotonic decay condition.
The conversion o f Gc(z) to G(z) for use in (2.4) is analogous to equations (2.5) and
( 2 .6 ).
The True Stability Boundary f o r Learning Control: For the Gc(z) above for the
underdamped, overdam ped and critically damped cases, we can convert the
associated second order scalar difference equation into state variable form, and
0<K<2/Al
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
15
The results are shown in Fig. 2.4 for the case d = 0, where again the range on K goes
The Approximate Monotonic Decay Condition: It is not possible for second order
systems to obtain a simple analytical expression for the range o f K as was done for
the first order case. Figures 2.5 and 2.6 give the results for the cases o f d!T = 0 and
d/T = 1 respectively. Again K is lim ited by 2, the diameter o f the unit circle the plot
o f (pzGfz) should stay within. But this time it is not only along the real axis that one
might start leaving this circle when the gain is too large, but when there is a resonant
peak it can easily leave earlier at some frequency other than zero. Hence, as C
decreases, the range o f K is made smaller and smaller. All results are, o f course,
independent o f p , and the monotonic decay condition is vastly different than the true
The True Stability Boundary f o r Repetitive Control: Figures 2.7 and 2.8 give the
true stability boundary using equation (2.4). It is seen that for overdamped systems
with a very small number o f time steps p per period, it is possible to go slightly above
the K equal 2 value. Under other circumstances, the distinction between the
monotonic decay condition (2.1) and the true stability boundary is very slight.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
16
f 1 \
Gx{s) = K cf d 1 (2.14)
+ d y U 2 +2£a>0s + a ) ; )
After discretizing, the parameters are K, conT , dT, C . and p. The equations involved
are much more complicated than those o f the second order system and are not shown.
Figures 2.9 and 2.10 give the monotonic decay condition (2.1) results, and Figures
2.11 and 2.12 give the true stability boundary. The implications o f these plots are
2.7 Conclusions
This chapter has shown that the approximate monotonic decay condition (2.1).
a sufficient condition for stability, is sufficiently close to the stability boundary for
repetitive control that in practical applications one should aim to satisfy it. Condition
(2.1) differs from the true stability boundary substantially for very small p . and for
first order systems, and otherwise the difference becomes negligible. As discussed in
Longman [19] it is important to satisfy condition (2.1) for ILC in order to obtain
reasonable learning transients, in spite o f the very big difference between (2.1) and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
17
20
Fig. 2.1 The stability boundary o f learning control applied to the first order system.
a.
real part
Fig. 2.2 Plot o f 0(e 0), with # fro m 0-180°, o f the first order system w hen p =10 and
aT = 0.88.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Fig. 2.3 True stability boundary o f repetitive control applied to the first order system.
450
400
350
300
250
^ 200
150 = 0.2
100
= 0.2
Fig. 2.4 The stability boundary o f learning control applied to the second order system
when d /T = 0 and <£"= 0.2, 0.5, 0.7, 1. and 2 from left to right.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
19
0.7
= 0.5
0.5 = 0.2
Fig. 2.5 Monotonic decay boundary o f repetitive control applied to the second order
= 0.7
= o.
0.5
= 0.2
Fig. 2.6 Monotonic decay boundary o f repetitive control applied to the second order
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
20
2.5
1.5
r = 0.5, p= 50
- ->m
0.5
Fig. 2.7 True stability boundary o f repetitive control applied to the second order
r = i,p = 5 0
s 200
A 10.
r= 0.7, p=50
* onn
0.5
Fig. 2.8 True stability boundary o f repetitive control applied to the second order
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
21
£ = 0.2
0.5
Fig. 2.9 M onotonic decay boundary o f repetitive control applied to the third order
0.5
3.5
Fig. 2.10 Monotonic decay boundary o f repetitive control applied to the third order
system when dT = 1.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
22
2.5
0.5 C= 2, p= 50
- 200
4.5
Fig. 2.11 True stability boundary o f repetitive control applied to the third order
2.5
1.5
0.5
Fig. 2.12 True stability boundary o f repetitive control applied to the third order
system when dT = 1.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
23
C hapter 3
Phase Filtering
3.1 Introduction
zero tracking error is not enough. One needs to impose conditions to ensure good
decay o f the errors with repetitions o f the task, or with periods o f the periodic
command, and this can be done in the frequency domain. However, it is very difficult
to satisfy this monotonic decay condition for all frequencies. In previous work
summarized in [18], some simple learning and repetitive control design methods were
requiring the adjustment o f a few parameters. They achieve the desired mono tonic
decay o f the errors, but do so at the expense o f not fixing the errors at all frequencies.
In this paper we develop m ethods for the learning or repetitive control system
designer to use in picking his learning law or the parameters o f his learning law, in
order to achieve both good learning transients, and good final values o f the error.
particular form uses a digital input, and the feedback m easurem ent device gives a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
24
digital output. The command input U(z) to a feedback control system is usually the
desired trajectory YJ^z). W e choose to look at the continuous time output only at the
sam ple tim es, so we include an A/D converter on the output. The K(s) is a
disturbance input appearing in the most com m on location, ju st before the plant
equivalent output disturbance. We will do it here for this typical digital control
system , but it can be done in an analogous w ay for m ost any more general system
w ith disturbances in any location. When a continuous tim e system Gjc(s) is fed by a
zero order hold with input M(z), the outputs at the sam ple times are given by using
disturbance on the output o f the plant block, but since the disturbance does not go
through a hold, we m ust solve the differential equation with the disturbance as its
input, and then add the result V(z) = Z(Gic{s) KG)) to the output, as in the top o f Fig.
3.2. From the block diagram, solving for Y(z) in terms o f the input U(z) and the
V(z) (3.2)
N o w G(z) represents the closed loop discrete time system dynamics (closed loop
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
25
For normal feedback control, we make the com m and equal the desired output,
U{z) = YJjz), and then we see from equation (3.1) that the actual output Y(z) will
1. The inverse transform o f the first term in (3.1) is a convolution sum o f the inverse
transform s o f G(z) (the unit pulse response) and the desired output YJz). We can also
include initial conditions into this particular solution if w e wish. The convolution sum
o f a function with yd(k) will almost never equal yjjc), and hence there is a
determ inistic tracking error for almost all desired trajectories. We discuss the use o f
learning and repetitive control to try to eliminate this deterministic source o f error
2. The disturbance W{z) is a second source o f error. This disturbance can be from
random causes, and we will not concern ourselves with this type o f disturbance in this
paper. It can also be from deterministic sources. There are many situations in which
the same determ inistic disturbances occur every time you give a specific command.
An exam ple is the disturbance torque on a robot link as it follows a given path
to perform the same trajectory many times, always starting from the same initial
condition. Based upon the error observed in the previous repetition the command to
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
26
the feedback controller is adjusted in the next repetition to try to decrease the error.
To m ake a general formulation for linear learning control (as in [27]) we can convert
the discrete tim e closed loop system G(z) to m odem control form
There is a p step desired trajectory, y j j c +1). The initial conditions each time are r(0).
and the same output disturbance w(k +1) appears every repetition. We assume that the
time delay o f the input-output map o f the digital feedback control system is one time
step. This is the expected delay for a digital control system, and when the delay is
different, one would usually know the value and adjust the mathematics accordingly.
Denote the variables for repetition j by using a subscript / , and write <5j in front of
any variable to form the difference: the value o f the variable at repetition j minus that
repetition j , starting with yj( 1) at the top. Let Uj be a column vector o f the
corresponding inputs starting with w/0) (note the one time step shift between these
two colum n vectors corresponding to the one step delay from discrete time input to
discrete tim e output). In terms o f these definitions, equation (3.4) can be written as
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A general linear learning control law takes the form
In other words, the command to the feedback control system at repetition j is changed
from that at repetition j - 1 by adding a m atrix o f learning gains L tim es the error
taken place, we simply apply the desired output as the command to the feedback
controller. This means that the control history applied to repetition j + 1 is the sum
j
\LJ+x = l L j + L e J = y j + Y JLe l (3.7)
i=0
Reference [18] discusses the various form s o f the matrix L for different learning
laws. Noting that 8} y = - 8 j e . the error histories from one repetition to the next obey
e ^ \ = { I ~ P ,I ) e , (3.8)
and will converge to zero tracking error if and only if all eigenvalues o f U ~ are
condition. For example, integral control based learning control uses a diagonal matrix
L. with the same number <p in each entry on the diagonal. The convergence condition
becomes |l - CsBs<j>\ < 1 which can be satisfied by a sufficiently small <f> o f the right
sign, and is independent o f the dynam ics o f the system in the matrix Asl Although
such a law has this very impressive stability robustness, being independent o f the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
28
system dynam ics, it can be impractical because o f poor transients [15]. Consider an
example. For the third order plant described in the exam ples section below, we use a
sample rate o f 100 Hz, and ask for a desired trajectory o f y d (0 = sin(2^r) on the time
interval from 0 to 1. A learning gain o f <f>= 1 is used. The result is shown in Fig. 3.3.
The root mean square o f the tracking error decreases from 0.4330 in repetition 0
w hen the desired trajectory is given as the command to the feedback controller, to a
m inim um at repetition 7 o f 0.1402. The mathematics above proves that the error will
converge to zero, but after repetition 7 the error starts to grow, reaching a maximum
RM S error level o f 1.1991 x 1051 at repetition 62,132. Then the error decreases and
ultimately reaches a numerical zero. For example, at repetition 3 x 103 the error has
reached 1.3145 x 10“*8. In order for the digital computer to be able to handle this
problem, we had to simplify it, by choosing a shorter trajectory, and using 100 Hz
instead o f 400 Hz as the sample rate. Otherwise, overflow occurs, and the iterations
between the amplitudes o f the frequency components o f the error from one repetition
to the next. This means that we are characterizing the response in each repetition by
the steady state frequency response components. This assum ption is valid provided
that the trajectory is relatively long compared to a settling tim e o f the feedback
control system, so that the majority o f the trajectory is not influenced by transient
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
29
We will also restrict the choice o f the learning law L slightly, in order to make
it correspond to a time invariant law. This requires that all entries on the diagonal be
identical, all entries on the subdiagonal be identical, etc. And, as in the case o f the
initial condition transients, we assume that the matrix L is sufficiently large that edge
effects in the matrix can be ignored. Then, we convert the equations to the z-
transform domain, and to get the frequency response, we substitute z = e'mT where T
is the sample time, and co is the radian frequency under consideration. Then, the
Here the G(z) is the z-transfer function o f equation (3.1), and the L(z) is the
transform o f the learning law associated with each row o f L. For exam ple, L(z) = $rz
in the case o f the integral control based learning control. If the L were similar except
that the nonzero entries were on the super diagonal, shifted one entry to the right from
L(z) = <pzr . When there are entries in more than one diagonal, the L(z) is formed
Equation (3.11) says that going from one repetition to the next, the amplitude
For any frequency for which this number is less than unity, the error will decay
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
30
monotonically, and any frequency for which the num ber is greater than unity, the
(3.12)
for all co up to the N yquist frequency, then all amplitudes o f the frequency
components o f the error decay monotonically. This statem ent applies to all portions
o f the trajectory for w hich steady state frequency response thinking can be applied.
This thinking suggests that (3.12) defines the stability boundary, but this is not the
convergence to zero tracking error [6]. Nevertheless, when the desired trajectory is
sufficiently long com pared to the settling time o f the system that steady state
frequency response thinking can be applied, then satisfying (3.12) assures mono tonic
eliminate tracking errors that appear each period. A special case o f particular
importance is when the command is a constant (periodic with any frequency) and one
wants to eliminate the effects o f a periodic disturbance w(k). We use the same form o f
the control law, except that repetitions become periods, where one period o f the
periodic command or disturbance has p time steps. The relationship of equation (3.10)
becomes
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
31
The block diagram o f applying this repetitive control to the closed loop feedback
From the block diagram we can solve for the error in terms o f the two inputs
to obtain
A fter clearing fractions, this represents a difference equation. Note that the right hand
side represents the difference o f values shifted forward in time by p time steps minus
the current values. Since y j k ) and w(k) are both periodic with period p time steps,
this difference is zero and the right hand side o f (3.14) is zero. Therefore, the
tracking error occurs if all roots of the characteristic polynomial have magnitude less
than one.
As in the learning control case above, suppose that each period o f p time steps
is long compared to the settling time o f the system. Then the response during each
period can be characterized by the steady state response to the errors of the previous
z pE { z ) = [ \ - L { z ) G { z ) \ E { z ) (3.15)
The z pE ( z ) is a time shift o f the error signal p steps ahead. Equation (3.15) suggests
that w ithin the assumptions here, the frequency com ponents o f the error one period
ahead are related to the error components o f the current period according to equation
(3.11), now using j to denote the period rather than the repetition. Hence, for both
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
learning and repetitive control we wish to satisfy the same condition (3.12) for
In the following sections we develop methods for use by the control system
designer to aid him in handling tradeoffs in satisfying the monotonic decay condition
on which to apply the ideas discussed and observer the performance. We pick a
closed loop digital control system G(z) which is the discrete time version o f the
f 2 \
( d 1 -) 6)0
U -t-c/J U " + 2£co0s + co; )
fed by a zero order hold with sampling at 100 Hz. The constants are d = 8.8
corresponding to a break frequency at 1.4 Hz, £ = 0.5. and con = 3 7 rad /sec
good representation o f the closed loop response o f each joint o f a Robotics Research
behavior by many learning control laws and instability by many repetitive control
laws.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
y j ( k T ) = Y^an[\-zos((DnkT)~\\Q < k T <10 (3.17)
n=i
where the con are 0.2/z\2;r.4;z\6/r....J00/T . and the amplitudes are a„ = 80e_<u" . This
3.5 W hat Do W e Assume We Know About The System W hen We Design The
An issue in learning and repetitive control is, how much information o f what
type do we assum e we know about the system so that a design can be made that will
have good learning transients? Here we say that it is reasonable to have the same
information that the control system designer has, since it may be the same person.
More specifically, we say that it is reasonable to assume that we have the following
system and taking data to use in designing the learning or repetitive system:
I . We assume that we have a Bode plot o f the closed loop system behavior, i.e. we
know the steady state response of the control system to a sinusoidal input for all
frequencies up to Nyquist, both the amplitude change from input to output given by
|G(e'"r )|, and the phase change from input to output given by the angle m ade by
G(e'mT) with the positive real axis. Equivalently, we know G(e'a / ) for each
input, then our knowledge o f G(e'wT) for the higher frequencies may be somewhat
less certain. Even when direct measurements are made o f each frequency separately,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
34
2. We can also have a m easurem ent o f the output error when w e applied the desired
applying the desired trajectory as the command. Then the error in the output
EFB(z) = Ya( z ) - Y ( z )
= { \ - G ( <z))YJ { z ) - W { z ) (3-18)
The final expression converted to its frequency com ponents is one that we will need
in evaluating final error levels in one o f the options discussed below. And if we want
to know the frequency com ponents o f W(z) for a different option discussed below, we
can compute them because we know the frequency components o f YJ^z) and EFB{z).
determination o f the frequency response needs to be done correctly. One can input a
signal such as white noise and record the data. Then input the same command
multiplied by some factor. Subtract the two inputs and the two outputs to eliminate
the repeating disturbances from the input-output relationship. One can then use this
We have stated that we wish to satisfy the good learning transients condition
(3.12), and if w e do so for all frequencies, it will result in convergence to zero final
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
35
value o f the error. However, it is very difficult to satisfy (3.12) for all frequencies.
And even if we think that we have satisfied it. according to our model, it can easily
happen that there is extra phase lag at high frequencies that could not be determined
One first designs a compensator that attem pts to keep the plot o f the complex
the unit circle centered at +1 defined by (3.12). One m ight try to cancel the system
dynamics, but this is very often difficult. One can design compensators to influence
the phase. One very simple and effective method used in [36,12,14] is to use a linear
phase lead. The L(z) = (pz/ for y > 1 is a linear phase lead, and choice o f this simple
number can easily bring the plot o f L{emT)G(e,mT) inside the unit circle for nearly all
frequencies (see [22,39] for discussion o f automatic tuning methods for this
parameter). Figure 3.6 shows plots of <f>(emTY G ( e ”°T) for various values o f y . The
single learning gain <j) is equal to 1 in this figure. A lthough it is not visible to plotting
accuracy, the curves all go outside the unit circle. In the case o f y = 5 , the maximum
distance from +1 is 1.0051, going outside the unit circle by only 0.0051. This will
poor transient in learning control. Figure 3.7 shows the RMS tracking error versus
roughly 80 repetitions, and then it starts to diverge. In [18], several methods to handle
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
36
this persistent situation are discussed. These include: 1. use o f zero-phase low-pass
matched basis functions [25,37,38]. Here we develop the zero-phase filtering option.
and [12] does so in repetitive control. The objective is to cut out the frequencies that
go outside the unit circle from the learning process. By making the learning blind to
so we no longer aim at zero tracking error, and in exchange we produce good learning
We can use a Butterworth low-pass filter to cut out the chosen frequencies. A
normal Butterworth filter will attenuate the frequencies above the cutoff, but will
introduce phase lags as well, and these disturb the plot sending it outside the unit
circle earlier. A zero-phase filter is created by filtering the signal o f interest forward
in time, then putting the output into the same filter but with the time sequence
reversed. The first application puts in phase lags, and the second puts in
compensating phase leads. There are some subtleties in extending the trajectories to
allow time for the filter transients to die away before reaching the trajectory segment
We denote the zero-phase low-pass filter by F(z). Suppose that the original
Butterworth filter has transfer function Bw(z) with the amplitude Bode plot coming
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
37
from M F{e ",ir') -|£ » -(e,<ur) |. Then the frequency response o f the zero-phase version
phase.
Suppose that we choose to filter the error to cut out the unw anted frequencies
before they go to the learning control law (3.7). The symbol Uj in equation (3.7) is an
accumulated signal associated w ith the sum o f all errors through repetition j - 1 (and
analogously for repetitive control for periods through period j - 1). I f the low-pass
filtering is not perfect, then some small amount of error in the unwanted frequency
range can get through. And this m eans that eventually this part o f the signal will grow
must apply the filter to the accumulated signal to prevent such slow growth.
However, the accumulated signal contains the desired trajectory as the command
given in the Oth repetition. One may ask should this be filtered or not. We examine
both possibilities. In the first, we make a learning control law o f the form
In the second, we separate out the desired trajectory part o f the sum in (3.7) according
to
U j ( z ) = Yd (z) + U t j( z ) (3.20)
and filter only the signal that is added to yJkT) by the learning process to form the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
38
more com plicated. It requires that a time signal be filtered both forward and backw ard
Apply the repetitive control law version o f (3.7) for m repetitions. Then freeze the
repetitive control signal applying it each repetition until the zero-phase filtered result
becomes available. Starting with the last o f the m repetitions and continuing perhaps
through two repetitions after freezing the signal, we apply the zero-phase filter. By
applying the filter to three repetitions in a row with the same signal we can pick out
the results for the middle period, using the first and last for decay o f forward and
reverse filter transients. Once we have the filtered result corresponding to the right
hand side o f (3.19) or (3.20), we apply the result at the start o f the next period. Again,
timing belt drive system reported in [12], the instability resulting from frequencies
going outside the unit circle, did not appear until around repetition or period 2650. In
this case, one m ight entertain applying the zero-phase low-pass filter every m = 2,000
periods.
The way in which the two options (3.19) and (3.20) appear in repetitive
(3.19), we can use the block diagram as in Fig. 3.4. The block containing
L(z)f(z p - 1 ) is a difference equation needing initial conditions for the signal u(kT)
for time steps from k = I to p. We can use the actual command applied during the 0th
period w hich is the desired trajectory, and then the block produces the equivalent o f
(3.19). O n the other hand we could use zero initial conditions, in which case the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
39
accumulated signal does not contain y^JcT), and then we still have to add it on
In the above design process, w e are giving up zero tracking error to get
decay o f the error, and do so with as little final error as possible when the learning
process is complete. The answ er to this depends on the frequency ranges o f the errors
that one needs to eliminate. Here we develop methods that allow the designer to
predict the final error levels: 1. For each choice o f compensator, 2. For different
choices o f zero-phase low-pass filters, and filter orders, 3. For filtering the total
command signal as in (3.19) or filtering the command with the desired trajectory
removed as in (3.20), and 4. For different choices o f how often one does the zero-
phase filtering.
3.7 The Steady State E rror Level W hen Filtering The Total Command Signal
In order to predict error levels for the filtered learning control law (3.19), we
first find the difference equation for the error. Use equation (3.1) in the definition o f
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
40
We now wish to com bine this with (3.19), and eliminate the U dependence so that we
have an equation for the error in terms o f the two inputs, Y^z) and W(z). To do this.
The filter passes frequencies below the cutoff, essentially unchanged, so that
perfect cut off, at frequencies above the cutoff, F(z) is zero, and the forcing function
on the right is sim ply Yt/( z ) —JV(z). This remaining part o f the forcing function
can be written as
for all co up to N yquist frequency. It is clear that it is easy to pick a filter M } ( e ,a>r)
Once this decay condition is satisfied, we then wish to know what the final
error level will be. W hen this final steady state error level is reached, F ,(z ) = £ ;_,(z)
which we can call F vv( z ) . The transfer function from Yt/( z ) - W ( z ) to £„v(z) with z
replaced by e,a>r is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
41
, % /2 \ r t
1 - M], (e'mT)[1 - G{e ioir )L{e,a“ )]
0 - 2 5 )
To find the steady state error level for each frequency com ponent, we decompose
Yi/(z) —W(z) into its frequency components and then m ultiply these input amplitudes
for each frequency by \$i (e “r )| for the corresponding frequency, to determine the
amplitude o f the steady state error at that frequency. W hen we are only concerned
with response to Ya{z) it is easy to decompose it into its frequency components. When
there is a repetitive disturbance PF(z), we are not likely to know what it is directly. In
this case we can determine its frequency com ponents as described below equation
(3.18).
At first it would appear that a very sim ple design m ethod would simply use a
constant for the filter. As noted above, the values o f [1 —G(e ,a>T)L(e,a>T)] reaches a
maximum o f 1.0051 for the example o f (3.6) and (3.7), and using L(z) = <pz/ with
for learning or repetitive control. Figure 3.9 shows the result. Due to the scaling in the
figure, one cannot see w here the curve starts, but the initial RMS error with feedback
only on repetition 0 is 2.5659. The RMS error at repetition 2 is 0.2605, and the final
value is 0.31445. This very small change from F(z) = 1 in Fig. 3.7 to 1/1.0051 in this
figure has made a very substantial degradation in the minimum value of the error
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
42
reached, but it did prevent the apparent divergence in Fig. 3.7. Figure 10 shows how
s', (e'“r ) attenuates or am plifies the various frequency components in YJ^z) - W(z) in
producing the final value o f the error, for several different values o f constant
F ( r ) = M j,( e ,a,r) . We observe that this function is very sensitive to small changes in
the constant value, and can easily produce large am plifications o f errors in certain
frequency ranges. We conclude, that although the use o f a simple constant for the
filter F{z) at first appears attractive, that it is not likely to be effective at producing
We now use a 5th order Butterworth filter w ith a cutoff frequency o f 100
rad/sec (15.9 Hz). The filter cuts o ff before the system goes outside the unit circle in
Fig. 3.4. Since the filter is applied twice, the Mfr(e,a>r) attenuates above this cutoff
like a 10th order filter. Figure 3.11 shows the RMS error versus repetitions using this
learning law. As before the RMS error starts at 2.5659. This time it decays to a steady
error level o f 2.4684* 10-4 . The growth in the error in Fig. 3.7 has been eliminated,
with only a small increase in the steady error level above the minimum observed in
Fig. 3.7. Figure 3.12 shows how the choice o f the order o f the Butterworth filter and
the choice o f the learning gain influence the values o f j^i (e <ur) |, and this information
can be used to help pick these values in tuning the learning control law.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
43
characterize the response o f the system in the current repetition to the repetitive
control action taken in the previous repetition. This assumes that the settling time o f
And this assum ption means that we can characterize the response in each period as if
it were a separate function from that in the previous period, and label the variables for
period j with the subscript j, just as we did in learning control. And this makes
repetitive control look very similar to learning control. The substantial difference is
that in repetitive control we cannot do the zero-phase filtering in real time, and use
Suppose that we apply the repetitive control law L(z) for m repetitions, and
that there are r - m repetitions with the frozen signal before we apply the filtered
result and start the learning process again. Thus the process repeats every r
repetitions. Suppose we start with the error at repetition j = l r . Then the updates
U ,r+I
(z) = U ,r (z) + L(z)E,r(z),...,Ufr+nt_ x (z) = U ,r+(„ _ 2 (z) + L(z )E(r+m_2 (z ) (3.26)
And then we wish to filter the next command signal, and it gets applied at repetition
j = V + l)r ,
(z) + ( 2 )]
U lM)r (z) = F(z){ U „ (z) + £ (z )£ ,„ , (z) + £ ( z ) £ ,„ , (z) + ••• + £ (z )£ „ „ „ , (z)] <3'27)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
44
During the first m repetitions, equation (3.15) applies which indicates that error at
" l - d - G ( z ) L ( z ) ) lr+m'
= E { z ) L ( z) E,r{z)
G (z)L(z)
(3.29)
that the left hand side o f (3.19) appears. Then substitute the right hand side for this
This equation is analogous to equation (3.22). This time the frequency components
outside the unit circle are allowed to grow for m repetitions, and the filter F(z) must
be picked to attenuate this accumulated growth to a number less than one. The
and satisfaction o f this produces monotonic decay o f the frequency components o f the
error for the errors after each filtering, at &*,(£ + l)r,(£ + 2 ) r ,.... The frequency
£ (e™) = I-M W *)
2 1 —Mj-(e'“7 )[1 —G{e,iaI )£(e""7)]"'
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
45
M ultiply the m agnitude o f this for any frequency tim es the am plitude o f the
com ponent o f Yd (z) —W (z) at that frequency, and you get the am plitude o f the steady
filter every repetition. Experiments show that the error level m ay stay low for many
repetitions, and one w ould not feel the need to apply the filter every repetition.
Instead one m ight apply it every 100 or 1000 repetitions, as appropriate, to keep the
RMS error from starting to grow. In order to predict the steady state error levels for
this process for any choice o f the number of repetitions between applications, the
3.8 The Steady State Error Level When The Desired Trajectory Part o f The
In the previous section we filtered the entire comm and signal that goes to the
differences when we use equation (3.20) instead, taking out the desired trajectory
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
46
Following the sam e procedure as before, this time to elim inate U L dependence, we
use (33) in EJ{z) —F { z ) E /_l{z) in order to create a term U Lj (z) —F( z ) U Lj_x{z) and
equation (3.23) and (3.24), but the steady state error addressed in (3.25) has som e
1. Suppose that we have made an experimental run giving the desired trajectory' as the
com m and to the feedback controller, and measured the error in the output. Then
m ultiplied by 1 - £(z) to form the forcing function in (3.34). Then we can find the
Sx(e,a>T) in equation (3.25) for that frequency component, and obtain the amplitude o f
the steady state error for that frequency. Thus, the frequency transfer function
S ,(e'“r ) applies directly to the measured error using feedback only, whereas in the
previous case we had to use the measured error and com pute W(z) in order to know
the input Y\,{z) —W ( z ) . Thus we use the same S l(e"u l) in both cases, but we apply it
to different signals, and hence we will obtain different steady state errors.
2. Suppose our main interest is the ability o f the system to eliminate the effects o f
disturbances W(z), and perhaps YJz) = 0, as often happens in repetitive control. Then,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
47
equation (3.34) and (3.25) become indistinguishable, and it does not m atter which
3. Suppose that we do not expect any disturbances that repeat each repetition or
period, and our concern is to know the steady state error in following the desired
trajectory. Then we find the amplitudes o f the frequency components o f Yd(z). and
this time m ultiply by the magnitude o f the following frequency transfer function for
the same frequencies, in order to know the steady state error for each frequency
cw -'T x [ l - M ; - ( e ^ ) ] [ l —G { e - T)}
\ - M l . { e ,0lT) \ \ - G { e ,0,T)L(e‘mT)} ’
Figure 3.13 shows the result o f using this learning control approach o f equation (3.20)
substantially better than the steady state error level for the approach o f (3.19). It is
very close to the minimum error level observed in Fig. 3.7. So the approach here is
able to reach essentially the same minimum level as without the filtering, and then
One now wants to know which approach is better, to filter the signal including
was the second approach. But the answer depends on the situation. If the point o f the
makes no difference which one uses. If the point is to eliminate the deterministic
errors o f the feedback controller in following a tracking command, i.e. the error in the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
48
particular solution, then which one is better depends on the dom inant frequency range
(3.36)
If the dom inant frequency content o f the com m and is in a range for which this plot is
above unity, then it is best to filter the desired trajectory as in (3.19). If the dom inant
error is a frequency range where this plot is below one, then separate the desired
N ow consider the repetitive control version o f the filtering with the desired
trajectory removed. We can parallel the developm ent o f equations (3.26) through
(3.32). The only change is that again we ad ju st £ y(z) according to equation (3.33).
(1 - G ( z ) ) Y j ( z ) (and replace £\(z) by E [ f {z). but these get eliminated). Then the
Again these equations also apply when u stn g periodic filtering in learning control
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
49
3.8 Conclusions
In this chapter we have developed analytical methods that guide the designer
repetitive situations, and getting the best final value o f the error when the learning
process has converged to steady state. The frequency transfer functions in (3.25).
(3.32), and (3.35-3.37) predict the steady state error levels for different learning and
repetitive control approaches. These can be used to adjust the compensator used, such
as making the choice o f the linear phase lead y , and to pick the learning law to use or
to adjust a learning gain such as $ . And these equations allow the designer to make
the best choice o f whether to include the desired trajectory in the signal to be filtered.
In the case o f repetitive control the equations are used to decide how often one must
apply a filter in order to keep the RMS error from growing. The approach developed
here is practical in the sense that all the learning or control system designer needs to
know is experim entally measured closed loop frequency response information for an
existing feedback controller, and the error it produces in following the desired
trajectory w hen it is given as the command. From this he can design a learning or
repetitive controller that will substantially improve the tracking performance over that
o f the feedback control system, by iteratively adjusting the command given to the
feedback controller.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
50
Vds)
Ftz)
D/A
A/D
ZO H
A/D
viz)
♦O
W{z)
Y(z)
*o
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
51
60
40
RMS tracking error
to-'
•20
,-40
,-60
0.5 2.5
Repetition number
Fig. 3.3 The learning transients during the convergence to zero tracking error in
integral control based learning.
mz)
Fig. 3.4 Repetitive control applied to the closed loop control system G(z).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
52
100
o
oo 60
C3
40
20
Time(sec.)
Fig. 3.5 The desired trajectory for the numerical simulations of learning control.
0.8
0.6
0.4
G
.
G3 0.2
0.2
-0.4
- 0.6
- 0 .;
-0.5 0.5
real part
Fig. 3.6 Nyquist plot of product of the learning law and the closed loop transfer
function, (fc rG(z), for y = 1.2. 3. 4. 5. and 6 .
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
53
o
C3
i o '5
0 100 200 300 400 500
Repetition number
Fig. 3.7 RMS tracking error versus repetitions using a learning controller with a
linear phase lead of y = 5 .
VJ.Z) W(Z)
YAz) + E(z)
Uz)
G(z)
4I zT-l
Fig. 3.8 Repetitive control with the desired trajectory removed from the
accumulated signal by setting the repetitive controller initial conditions to zero.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
54
5o 1CT
C3
in
2
o'
to'M_ _ _ _ _ _ _ _ i_ _ _ _ _ _ _ _ i_ _ _ _ _ _ _ _ c_ _ _ _ _ _ _ _ ,_ _ _ _ _ _ _ _
0 100 200 300 400 500
Repetition number
Fig. 3.9 Learning control using the simplest zero phase filter, constant F(z) =
1/1.0051.
F(r) =1/1.0051
1. 6 - ... F(z) =1/1.01
— F{z) =1/1.015
.4-
1.2
0.8
0. 6 -
0.4-
0.2
40
Frequency (Hz)
Fig. 3.10 The values o f l-SiCe"^)! versus co for different choices of F(z). F{z) =
1/1.0051, I/1.0I,and 1/1.015.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
0 100 200 300 400 500
Repetition number
Fig. 3.11 RMS tracking error versus repetitions when the entire command each
repetition goes through the zero-phase low pass filter.
0.8
04
0 .4 -
0.2
40
Frequency (Hz)
Fig. 3.12 The values o f )| versus co when using the n th order Butterworth
zero-phase filter on the entire command with different learning gains (f>. Curves,
left to right. («, $ : (5, 0.25), (5. 1). (10. 0.25), (10, 1).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
56
Fig. 3.13 RMS tracking error versus repetitions when the desired trajectory part of
the command is not filtered.
0.8
■
0.6
0.4
0.2
Frequency (Hz)
Fig. 3.14 Comparison o f including or not including the desired trajectory in the
filtered signal.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
57
C hapter 4
Phase Filtering
4.1 Introduction
experience. The original motivation was for eliminating errors o f robots in executing
the same trajectory many times. There are two classes o f errors. One is the fact that
for nearly all tracking commands, the response o f a feedback controller is not equal to
the command. The second is that in many situations, there are repeating disturbances
every time the same command is executed, for example, the gravity torque on a robot
which the system, e.g. the robot, returns to the same starting point before each
execution o f the command. Looking at the error in following the trajectory in the
decrease the tracking error, and it is possible to have learning control converge to
result, the Bode integral theorem [30], which can be interpreted as the waterbed
effect. The waterbed effect was first recognized by Francis and Zames [9] as they
studied the feedback constraints imposed by open right h alf plane zeros o f the plant
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
58
typical feedback control system, for the transfer function from com m and to resulting
error, the waterbed effect says that if the error is attenuated substantially by the
feedback control law over some frequency range, then it w ill be amplified
substantially over some other frequency range. In learning control we seek to have
zero tracking error at all frequencies, although to make design easy and
cut off the learning process above some frequency. In this paper w e investigate in
what way the Bode integral theorem type o f thinking applies to large classes o f
learning control laws. We investigate how ILC laws are able to bypass the waterbed
error over a range o f frequencies without paying for the improvement by amplifying
errors in some other frequency range. This includes introducing more zeros than
poles, modifying the gain in the sensitivity transfer function, and introducing poles
learning control is concerned with executing a finite time desired trajectory. Hence,
the results we obtain here will relate to the execution of the desired trajectory for that
portion that is independent o f initial conditions, i.e. for that part o f the trajectory after
a number o f time constants have elapsed. This means that we need not be concerned
Consider a digital feedback control system with closed loop transfer function
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
59
w here tV(z) is any repeating disturbance that appears every tim e we execute the
trajectory. Subscript j indicates the repetition number. The desired trajectory is Yd (z)
and define the tracking error as Ey(z) = Yd (z) - Y ( z ) . In this chapter we follow ILC
Write (4.1) for j +1 and for J, take the difference, and apply the learning law
(4.2), to obtain the error update equation from one repetition to the next
The square bracket contains a transfer function from the error at one repetition to the
error at the next repetition. Substituting z = exp(icoT) (T is the sam ple time) produces
the frequency transfer function, whose magnitude at any frequency determines the
attenuation or am plification o f the amplitude o f the error at that frequency from one
(4.4)
for all frequencies up to Nyquist. then all amplitudes o f the error decay monotonically
with repetitions, for the part o f the trajectory governed by steady state response. In
[19,6] it is show n that this is a sufficient condition for stability o f the learning
process, and convergence to zero tracking error for all time steps o f the trajectory.
Because o f its monotonic decay property for the steady state part o f the trajectory, it
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
60
Consider a typical digital feedback control system with unity feedback. After
converting the plant Laplace transfer function fed by a zero order hold to a z-transfer
function, and com bining it with the controller transfer function to form Go[{z) in the
forward part o f the loop, the transfer function from command to error is
S(z) = 1/(1 + Gol( z ) ) . This transfer function has two special properties: the degree o f
the num erator polynomial and the denom inator polynomial are the same (call it n),
Theorem 1 (Bode integral theorem): Let S(z) be a transfer function having the
system. Define ln*|/?| = max(0,ln|/?j) for any com plex num ber J3, and let the zeros of
(4.5)
A proof is given in [30] making use o f the Poisson integral that requires
stability. Below we generate a proof o f a m ore general theorem that has this result as
In (4.5) one can let go = 0 / T to represent the radian frequency. Hence, (4.5)
says that the logarithm o f the amplitude Bode plot o f the sensitivity transfer function
S(z) , averaged over all frequencies up to N yquist is given by the right hand side. If
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
61
there are no zeros outside the unit circle, then the right hand side is zero. This
produces the waterbed effect. If one decreases the error substantially in a frequency
range, then there m ust be some other frequency range for which the error is amplified.
The result is in logarithm space. To illustrate the implication, suppose that the error
up to one h alf Nyquist frequency were decreased by a factor o f 10, then the error
above this frequency m ust be increased by a factor o f 10. It is not the error itself that
averages to zero, it is the factor multiplying the error that averages to zero.
Perhaps the majority o f digital control system s will have zeros outside the unit
circle. W hen one creates the discrete equivalent o f a plant fed by a zero order hold,
and there are at least two more zeros than poles in continuous time, then there will be
at least one zero outside the unit circle if one samples fast enough. The presence o f
such a zero in the Bode integral simply increases the right hand side to some positive
number. Hence, zeros outside the unit circle m akes matters worse, there will be more
In any case, there is no way in which one can get a negative number on the right
hand side. This means that typical feedback controllers improve performance by
occurs, and amplifying in parts o f the frequency spectrum where there is very little
It is the purpose o f this paper to investigate in what ways the Bode integral
theorem type o f result applies to iterative learning control. More specifically, we seek
to understand how it is possible for linear ILC laws to produce large attenuation of
error over most or all o f the frequency range up to Nyquist, without being subject to a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
62
waterbed effect, i.e. without having to pay for this benefit by amplifying errors
somewhere else.
In order to illustrate the developments o f the follow ing sections, a third order
Corporation robot is used [6,7]. The continuous tim e m odel is given in equation
(3.16).
Note that one o f the transfer functions that w e w ould like to consider with
respect to a Bode integral theorem type of result, is the transfer function in (4.3).
Suppose that (f>= 1 and y = 1 (and G(z) has one more pole than zero). This
corresponds to the most common learning control law based on pure integral control
action applied at every time step in the discrete repetition domain. The transfer
function in (4.3) then has the same number o f poles as zeros, but there is a coefficient
in front depending on the learning gain. The presence o f this coefficient means that
the conditions o f the Bode integral theorem are not satisfied. Consider a / > 1. This
produces more zeros than poles, and again the conditions o f the theorem are not
satisfied.
necessary to create a much more general version o f the Bode integral theorem.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
63
This time we allow the degree o f the numerator polynom ial to be different
than the degree o f the denom inator polynomial, n is not necessarily equal to m. And
in fact we will allow more zeros than poles. Also, any non zero gain K r is allowed, so
that even if m was equal to n, the limit as z goes to infinity is not necessarily one. And
finally, we do not require that all poles be inside the unit circle. We create the
analogous result to the Bode integral theorem, for this general situation.
Theorem 2 (Generalized B ode integral theorem, GBI): For the R(z) o f equation
(4.6)
1 ~ m n
- J l n | « ( e " ) |e / ^ = l n | ^ | + £ l n * | 2/| - £ l n * | / » l| (4.7)
^ 0 »=i '=1
As before, zeros outside the unit circlemake this logarithmic average o f the
On the other hand, having a K r that is less than unity in magnitude decreases the
value o f the integral, and having poles outside the unit circle decreases the values o f
the integral. In the following sections we will see how these two effects are used by
learning control laws to get improved performance over a broad band o f frequencies.
Proof: Because o f the possibility o f poles outside the unit circle, we need to
use a different approach to obtaining a proof, making use o f a mean value theorem in
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
64
Substitute z = exp(/<9) and integrate each term from —jt to JrTC, and divide by 2 k .
Because j/?(e'°)| and |/?(e~'s )|are equal, the stated integral o f the left hand side of
(4.8) becom es the left hand side o f (4.7). The integration o f the ln|^fR| term leaves it
unchanged, equal to the first term on the right o f (4.7). It remains to demonstrate that
.T
(4.9)
m aking use o f concepts in [17]. This dem onstration is divided into three cases, when
P is inside the unit circle, on the unit circle, or outside the unit circle.
(4.10)
where CQ is the unit circle about the origin. Note that In(l + z) = f - i f 2+ i f 3 — - is
convergent for |i| < 1 (and convergent on the unit circle except at f = - 1). which
implies that the integrand is analytic in the disk o f radius \z\ < 1/ b larger than the unit
circle. Hence, the integral in (4.10) is equal to zero. Using z = exp {id) on the unit
circle, and converting the variable o f integration from 6 to a new 9 equal to 9 +cp.
~ jt Z
0 = / Jln(l - be'6 )de = i |R e[ln(l - be'0 )]c/ 0 - Jlm [ln(l - be'0 )]d9 (4 . 11)
The first integral on the right is the imaginary part o f (4.10) which must be zero.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
65
(4.12)
w hich indicates that its integral is the negative o f the integral o f interest. But the
integral is equal to zero which completes the proof for this case.
since the logarithm series does not converge. In place of (4.10) consider the integral
(4.13)
(
around the contour in Fig. 4.1. To form this integral we have absorbed the phase (p
into 6 as in the previous case. Again, we know this is equal to zero. To evaluate the
first integral, substitute z = I + se‘° and dz = iee'°dO . Taking the limit as s —> 0 . and
noting that s l n s - > 0, again produces the result that the integral (4.9) is zero for this
case.
jzj < 1+ e for some e > 0 . The real and imaginary parts o f an analytic function are
harmonic functions, i.e. ln |z -/? | is harmonic in this disk. The mean value theorem
1
(4.14)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
66
N ote that the generalized B ode integral theorem (as well as the Bode integral
up to the Nyquist frequency. W hen applying digital learning control, the desired
trajectory has a finite number o f tim e steps, and this implies that the error record for a
repetition contains information only about a finite number o f frequencies, i.e. the
frequency associated with the total time o f the trajectory as period, and all harmonics
up to Nyquist. Hence, in practice, the only meaningful frequencies are a finite set of
discrete frequencies. As the num ber o f time steps in the trajectory is increased, the
num ber o f discrete frequencies up to N yquist that are contained in the sam pled data
becomes larger, tending to a countable infinity. The mathematics o f the developm ent
o f the integral theorem does not have any information about the length o f the finite
Technically, the theorem average only applies as the trajectory length tends to
infinity, so that all frequencies are present. We must expect such a lim itation, since
the theorem is concerned with steady state frequency response, and technically no
system reaches steady state except in the limit as the trajectory time tends to infinity.
One large class o f learning control laws use a causal learning law. By this we
mean that the degree o f the num erator in (pzrG { z ) is less than or equal to the degree
o f the denominator. This implies that the learning law does not make use o f error
terms in the previous repetition that are in the future to the current time in the present
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
67
repetition, but adjusted forward by the usual one time step delay o f the system. Causal
learning controllers are formed by using a causal filter or com pensator with the
num ber o f poles greater than or equal to the number o f zeros, and by picking the
value o f y less than or equal to the sum o f delays through the feedback system and
compensator.
Consider the standard integral control based learning applied to the robot
example, using a learning gain o f (p = 1 and y = 1 to look at the error one step ahead
in the previous repetition. Computing the transfer function o f (4.3) gives 3 zeros and
3 poles, all within the unit circle, and it has gain K r for equation (4.6) equal to
1 - <pKa , where K G is the corresponding Kr value for the discrete time zero order
hold version o f (3.16). The result is that the integral o f the GBI theorem (4.5) is equal
to ln|/CR| = -0 .00178. Although the standard Bode integral theorem does not apply,
which would have made the integral zero, the logarithmic average represented by the
integral is still only slightly negative. And it is not sufficiently negative to keep from
Although the learning law is asymptotically stable, this am plification makes very
poor transients during the learning process [19,21]. Note that the value o f K c is
0.0018. Thus, a very large learning gain would be required to significantly reduce the
integral and reduce the waterbed effect. Going significantly above a learning gain o f
unity is not advisable. A learning gain o f unity totally corrects the DC error in the
next repetition, since the DC gain o f (3.16) is unity. A learning gain between 1 and 2
would still produce convergence o f the DC error, but with an alternating sign each
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
68
repetition. And a learning gain greater than 2 would cause the DC com ponent o f the
error to grow each repetition. This statement, like the generalized Bode integral
result, applies to the steady state part o f the trajectory. It implies bad transients during
learning, but it does not imply instability [19], a gain approaching 100 is allowed
before there is mathematical instability. We see that although this learning control
law is not subject to the usual Bode integral theorem constraint, the usual waterbed
Learning control is fortunate that the data processed for the update o f the
control action comes from the previous repetition. The time step o f m ost importance
in the previous repetition is that one corresponding to the current tim e step, but
shifted forward one step. But it is possible to make the learning law depend on steps
future to this in the past repetition. Reference [36] discussed the fact that this allows
one to use noncausal digital signal processing techniques, including using the linear
phase lead y . Such learning laws have the property that (p zf has more zeros than
poles, with the excess being larger than one (when G(z) has the usual property of
one less zero than pole). In place o f the 1 — < pK G o f the previous section for the K r of
equation (4.6), we get — <f>Ka . N ote that this can be made arbitrarily small if we are
willing to learn sufficiently slowly. This gives one considerably more control over the
logarithmic average o f the frequency response o f equation (4.7). M aking this product
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
69
o f gains sufficiently small can make the logarithmic average go toward negative
infinity, or the frequency response amplitude go toward zero for all frequencies.
(4.2) with a learning gain 0 = 1. a phase lead o f y = 5 (including one step delay o f the
system), and 0 = 1, applied to the robot problem. Then 1—<pzr G{z) has three poles
and seven zeros. Four o f these zeros are outside the unit circle. The logarithmic
average o f equation (4.7) is —0.05998. Again, the usual waterbed result does not apply
to this learning control law, and we can have a negative logarithmic average.
One might want this average to be more negative. The learning gain used was
the gain to 0 = 0.1 makes this term equal -8.6322, and the logarithmic average o f
error every repetition. Further decrease o f the learning gain to K R =0.01 makes
ln|/ffl | = —10.9348, but the logarithmic average o f (4.7) becomes worse, equaling -
0.0005222. To understand the observed behavior, note that the learning gain is part o f
the coefficient o f the highest power in the zeros polynomial. As this coefficient tends
toward zero, at least one root tends to infinity. Therefore, decreasing the learning gain
improves the first term on the right o f (4.7), but it can make the second term worse.
Hence, from the point o f view o f the logarithmic average, there is an optimum value
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
70
Experience indicates that it is not easy to satisfy the good transients condition
(4.4) all the way to Nyquist frequency. There will normally be some high frequency
interval in which the magnitude o f the transfer function (4.3.4.4) exceeds unity. This
can cause a slow or fast long-term growth o f the error before convergence occurs
[19].
range that violates (4.4) (see also, [19,29,21]). This is also a technique for producing
robustness o f ILC laws. Since ILC normally aims for zero tracking error, the
m odeling errors at high frequencies, no matter how small are not ignored, and they
can be extrem ely important to the convergence properties. This fact is in contrast to
normal feedback control, where m odel errors are unimportant when they are at high
frequencies w ith small output, where the frequencies are far above the zero dB
crossing line where stability is determ ined. Modeling errors at high frequency are
normally associated with parasitic poles, or unm odeled modes, and m athematically
can be described as singular perturbations. The filtering can be applied to cut out
There are various types o f filtering. Here we consider IIR filters, and in
particular, Butterworth filters. To im plem ent such a filter, one applies a chosen order
Butterworth filter to the data set for one repetition, and then applies the same filter to
the resulting output sequence, but do so running through the sequence backwards in
time. Each application produces attenuation above the cutoff, and the reversed time
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
71
application introduces a phase lead that cancels the phase lag o f the original
application.
(4.15)
V'
1+
l + z -I
where n is the order o f the original filter, c -cot{cocTI T) , and coc is the filter cutoff
frequency. For exam ple, when n is 3 the above equation is produced from
1
H{ z) = (4.16)
1+ c
1—z
l + z
_-i -t 'N
Reference [21] gives two options for application o f the filtering. In the first,
the control signal in (4.2) is filtered before application, producing the learning control
law
with U0( z) = Yj ( z ) . In the second case, the desired trajectory is separated from the
learning control adjustm ents, and only the learning control adjustm ents are filtered
The choice o f w hich option can be optimized based on the error amplitude frequency
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
72
H aving introduced zero-phase filtering, we are no longer asking for zero error
in the limit. Hence, in addition to the transfer function o f the error from one repetition
to the next, it is o f interest to look at the transfer functions from the command to
resulting error, and from the repeating disturbance to error, to see what is implied by
the logarithmic average (4.7). The robot examples below use a 5th order Butterworth
the definition o f error, equations (4.1) and (4.17), in such a way as to eliminate t/y (z) .
This produces
Paralleling the thinking in [6], we can sum the geometric series involved to obtain
To design a learning controller, one adjusts the cutoff frequency o f the zero-phase
filter to make [/?, ( e ,a)‘ )| < 1 for all frequencies up to Nyquist. Note that H(z) is
analytic within a disk about the origin, which contains the unit disk. And therefore
H(z~l ) isanalytic outside some disk with radius less than unity.This makes the
product analytic in an annulus containing the unit circle. Figure 4.2 plots RL(z) fo rz
going around circles o f different radii, and shows that it is less than one in magnitude
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
73
for radii in the range 0.9701 < |z| < 1.1499. Once convergence is obtained in such an
In the limit as the repetitions go to infinity, the first term on the right will go
to zero for all z in the annulus. Then Sjfz) represents the transfer function from
command to the resulting error after learning is complete, and it also is the transfer
and there is a modified sensitivity transfer function from com m and to error
5 3( z ) = [ 1 - G ( z ) ] 5 , ( z ) (4 .2 2 )
while the transfer function from disturbance to error remains the same. Figure 3 .1 4 in
frequency spectrum, one o f the two approaches will be best for the final error level.
The logarithmic average for the two sensitivity functions S j(z) and S~(z) are
identical. To see this, note that the log o f a product is the sum o f the logs, that the two
transfer functions differ by the factor l - G ( z ) , and that this factor satisfies the
conditions o f the standard Bode integral theorem. The logarithmic average for both o f
control method that is subject to the standard Bode integral theorem. Although both
integrals give the same logarithmic average, the error is distributed differently with
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
74
frequency in accordance with Fig. 3.14. Figure 4.3 gives a plot for S', (r) as a function
o f frequency. The plot o f S3(z) is too sim ilar to show. The plot shows attenuation to
'‘zero” below the cutoff, and “no” am plification above the cutoff -- the perform ance
one aim s to achieve. As one might expect, examining the numbers from the com puter
shows that the part o f the curve above the cutoff is actually slightly larger than one.
N ow consider how this perform ance is accomplished. For S,(z) there are 13
zeros and 17 poles. O f the zeros, three are clearly inside the unit circle and the
rem aining 10 numerically differ from one, either larger or smaller, only in the 9th or
10th decimal place. So there is no real adverse effects from the zeros term in (4.7).
The new aspect o f the filtering approach is that there are now poles outside the unit
circle. Since the system is not causal, this does not mean instability. O f the 17 poles, 9
are outside the unit circle. As a result, the poles term in equation (4.7) contributes -
15.2572. On the other hand, the gain term is given by i n | = 12.3754, making the
4.10 Conclusions
In this chapter we have shown that iterative learning control law's are able to
bypass the waterbed effect that applies to nearly all feedback control systems. This is
other than unity for the appropriate sensitivity transfer function, from com m and or
disturbance to error, or from error at one repetition to error at the next. The second is
that ILC can use noncausal filtering, and this allows one to have poles outside the unit
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
75
here to handle the new situations introduced by ILC, having m ore zeros than poles,
poles outside the unit circle, and sensitivity function gains other than unity. The
restricted to be zero or positive, and this generalized theorem gives one the ability to
determine how negative this average can be, i.e. how much overall attenuation is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
76
-0.5
-0.5 0.5
Fig. 4.1 The contour used in evaluating the Fig. 4.2 Convergence in an annulus — plot
integral o f (ln (l- z ) ) / z . o f R(z) when z = r e x p ( /0 ) for 0 < Q < k .
Solid line r - 1, dash line r = 0.9701, dash-
dot line r = 1.1499.
0.8
0.6
0.2
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
77
C hapter 5
5.1 Introduction
Repetitive control is a relatively new field that uses control theory based
a feedback control system, or to leam to elim inate determ inistic error in a feedback
controller in following a periodic command. The repetitive control law uses error
inform ation from the previous period to update the com m anded trajectory to the
feedback controller in the current period. The field normally considers one periodic
disturbance. This means that there can be a fundam ental frequency having this period,
and then there can be any number o f harm onics going up to Nyquist frequency. This
can form any arbitrary periodic disturbance w ith the given period.
One large class o f spacecraft uses a m om entum wheel for attitude stiffening.
A small imbalance in the wheel produces vibrations in the spacecraft structure, w ith a
fundamental period associated w ith the tim e for one rotation o f the wheel, and then
there can be harmonics. There are other classes o f spacecraft with one primary source
o f vibration, such as a cryo pump. W hen there is a fine pointing instrument on board,
it may be necessary to isolate the instrument from there vibrations. Passive methods
are simple, but may not give good enough perform ance. Typical feedback control can
do better, but such a system does not take advantage o f the knowledge that the jo b o f
the controller is to eliminate a periodic disturbance, i.e. one that can be predicted.
Hence, feedback control designs approach each new period o f the periodic
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
78
reacting to it, and not be able to totally cancel its effect on the output. Repetitive
control (RC) addresses this issue, and with sufficient design effort can entirely
The waterbed effect in feedback control says that if the control attenuates the
some other frequency range [30]. This implies that feedback control works because it
attenuates in a range where there are significant disturbances, and amplifies in a range
where the disturbances are small. In chapter 4 and [32] the waterbed effect was
studied as it relates to iterative learning control (ILC). ILC differs from repetitive
control in that instead o f a periodic disturbance or command, the system is given the
same command repeatedly, but it is restarted from the same initial condition each
time. It is shown there how ILC is able to bypass the waterbed effect, and get better
performance than any feedback controller could do. ILC and RC are very similar,
w ith the m ajor difference being the restarting in the case o f ILC. It is the purpose o f
this chapter to study the relationship o f the waterbed effect to repetitive control.
slightly different form. The solution to equations (3.3) and (3.4) can be packaged in
matrix form as
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
79
y - p v M + w + ^ , „ X V( 0 ) ; ^ vo (5.1)
U.J+1 = ^ ( m 7 + ^ £ y) (5.2)
where L is the learning gain matrix, and F is a chosen filter which we normally use as
a zero-phase low-pass filter in this chapter, although we also consider non-zero phase
matrices to describe the error each repetition. If each repetition is sufficiently long
that we can ignore the effects o f initial conditions, we can switch to a z-transfer
function representation, using G(z),L(z),F(z) for the plant, learning control law, and
produces the difference equation describing the error history as a function o f j. and
the sensitivity transfer function from command minus periodic output disturbance to
function on the right when considering it. Then the coefficient matrix for
appears as a transfer function from the error in one repetition to the error in the next.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
80
If we ask that the associated frequency transfer function have amplitude less than
this suggests m onotonic decay o f the error with repetitions. Here. T is the sampling
time interval. The above developm ent is heuristic, based on steady state frequency
response thinking, but it can be shown that it is in fact a sufficient condition for
asymptotic stability o f the learning process [19]. In practice, without using a filter F.
it is hard to create a learning law L that is guaranteed to satisfy this inequality. So the
design process is to pick a n L that satisfies it up to some frequency, and then use a
zero phase filter F to cut o f f the learning above that frequency to obtain stability at
the expense o f not attem pting to cancel high frequency errors above the cutoff [19].
Now consider real tim e repetitive control. Converting equation (5.2) to the
transform domain, and recognizing that going back one repetition is now going back
z pU { z ) = F(z)\U(z) + L(z)E(z)]
[zp - F ( z ) + G(z)FCz)L(<z)]E(z) = [ z p - F{z)]\Ytl( z ) - W { z ) \ (5.5)
z p —F(z)
S r (Z) = z p - F{ z ) [ \ - G{ z ) L { z ) ]
and the homogeneous equation whose solution determines stability o f the process can
the error in the next period, we have the same heuristic argum ent for stability as
before, and again ask that equation (5.4) be satisfied. As before, it is possible to
demonstrate that this condition is in fact a sufficient condition for stability [19].
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
81
Form f o r Im plem entation: N ow consider the form of the zero phase filter F used in
ILC. and also used in RC when the update is done as a periodic batch process. Pick a
causal low pass filter such as a Butterworth filter. In state space form
where we allow a D term which is needed for most such filters. If this is used to filter
a chosen length set o f data v(k), the result is to attenuate the am plitude o f frequency
components above the cutoff and leave the amplitudes below the cutoff close to their
original values, but considerable phase delay in introduced. Then the same filter is
This doubles the attenuation o f frequency components above the cutoff, and cancels
the phase lag o f the previous filtering by a corresponding phase lead. A fter filtering
both forward and backward through a data set, the v{k) is the zero-phase low-pass
filtered value o f v ( k ) . One must pick initial conditions at time 0 for the forward
computation, and initial condition at the final time for the reverse computation. There
are various ways tc do this as discussed in [29], together with ways to extend the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
82
M atrix Form: Use the underbar as before to represent the whole history o f the
associated variable. Suppose that we are using the zero phase filter on a set o f data
points starting with time step ka and going through time step kj-. so we make the
underbar vector start with this first time step and progress to this final tim e step. The
v = V v + Aiex(k0) (5.8)
Here the P ,Aic are analogous to the Pv,Am in equation (5.1) w ith the A.B.C o f the
filter substituted and also with the influence o f D taken into consideration. This
means that the diagonal entries in P become D, and the other entries are pushed down
one. And the powers o f A in the Aic become one less than before. Let Ir be a matrix
like the identity matrix, but the opposite diagonal in the matrix contains the nonzero
entries. Then l rv_ puts the result o f the forward filtering from (5.8) into time reversed
order. Sending it through the same filter again, and then returning to the original order
produces
V= Ir[PIrY- + A,ck(kf )\
= + V rPIr)At x(.kJ + I,A,Akf) (x9)
One can show that / rP / r = P 7 for matrices o f the form o f P, so that the overall result
If we want to pick the initial condition for the reverse time filtering as the final state
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
83
x { k f ) = c [ A c j x ( k 0) + Prv]
C = [0 - 0 /]
where Alc t ,Pr indicate the matrices Ak ,P with the C replaced by the identity matrix.
When the trajectory is extended in various ways, one can develop the corresponding
formulas analogously.
Convolution Sum Representation: Denote the unit pulse response o f the forward
*/—*» kr~K,
+ k) = f U - k )vU + k„) + X W ~ k )CA'x (ko) + CAk,~K~kx(kf ) (5.12)
J=k i—k
k r - k 0- k kf -ka- j
initial conditions to zero, which drops the second and third terms on the right of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
84
Look at the P r P in (5.10) and write the product for a matrix large enough that the
and look at a tim e step k sufficiently far from the endpoints that the upper limits can
sufficiently far from the beginning or end o f the m atrix that end effects are not
corresponds to the infinite limits on f { j , k ) . Then exam ine the product H{z~x) H{ z)
and identify by the power o f z the output at different times, and one sees that they
match. Hence
The z-transfer function o f the zero-phase filter is Fr (z) when implemented in this
manner. Note that if there is a pole inside the unit circle in H(z), then there will be a
corresponding pole outside the unit circle in H(z~x). One might think that this
suggests instability, but in the finite time zero-phase filtering problem, the filtering is
noncausal, and poles outside the unit circle do not im ply a stability problem. Hence,
when studying the waterbed phenomenon for ILC with zero phase filtering we can
need the generalization to poles outside the unit circle given by the GBI Theorem.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
85
Zero phase filtering is not a causal operation under normal circumstances, but
with a som ew hat heavy com putation burden each tim e step, we can create a real time
filter for repetitive control [20]. Repetitive control laws such as the linear phase lead
Law only require one zero phase filtered value each tim e step. A t time step n it wants a
zero phase filtered value o f u(n —p) + <pe(n —N + y ) , where p is the period o f the
linear phase lead chosen to keep the frequency response plot o f (5.5). as used for
equation (5.4), satisfied up to as high a frequency as possible before using F to cut off
the learning. Let N = p —y be the number o f tim e steps back from the current time
step to the step o f the estim ated quantity. Then at tim e step n the forward computation
o f v(«) is as in equation (5.6). The needed backward zero-phase filtered value at time
Provided one can perform this sum within one tim e step, one can implement a real
time zero phase filter for repetitive control. This form ula ignores the initial condition
at time zero, since it will get arbitrarily far back in time as time progresses. But the
other initial condition used on the backward filter starting back at time step «, x ( n ) , is
included in the equation in case the number o f time steps N o f decay is not
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
86
To create a transfer function representation for (5.15). we ignore the first term
Then
This differs from the previous result (5.15) because the H(z~x) expansion has been
truncated to N terms. Note that the / / v(z_l) only introduces zeros in the transfer
function, and hence the poles o f this zero phase filter are simply the poles o f the
forward low pass filter. This time there will be no poles outside the unit circle, as is
One way to pick the initial condition for the backward filter is to make its
initial state equal the final state o f the forward filter, x(ri) = x ( n ) . Doing this and
recognizing the time argument k being used, produces the different z-transfer
function
Still different version would result if we chose to extend by repeating this state or by
In learning and repetitive control, we can apply the Bode or generalized Bode
integral theorem to two different transfer functions, (i) the appropriate sensitivity
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
87
transfer function or (ii) the transfer function from one repetition to the next
F (z)[l —G( z ) L( z ) \ .
Learning Control: Reference [32] gives exam ples o f this case. First consider ILC
laws that converge to zero error, and do not need a cut o ff filter making F( z) = 1. In
the difference equation in (5.3), the right hand side is then zero, m aking the
sensitivity transfer function zero. The desired trajectory Yd(z) and the disturbance
W(z) repeat every repetition. Every repetition contains p time steps limiting the
number o f frequencies that can be observed to only those that are periodic with period
p, and there is not way to have any other observable frequency in the system. With
the forcing function zero in difference equation (5.3), a sufficient condition for the
error to go to zero is (5.4). For (5.4) to be satisfied, the GBI Theorem m ust produce a
negative value. Suppose that the learning law was the simple integral control based
ILC having L(z) = (frz. Examining the transfer function for (ii) above we see that the
coefficient o f the highest power includes the learning gain <f>, requiring the use o f the
generalized Bode integral theorem, and allowing the integral to be negative because
N ow consider using a zero phase low pass filter FL(z) and a learning control
law containing a linear phase lead, L(z) = <pzy w ith any integer y > 1 corresponding
to a phase lead that can help keep equation (5.4) satisfied up to a higher frequency
before the low pass filter has to cut o ff the learning. N ow the right hand side o f the
difference equation (5.3) is non zero and there is a sensitivity transfer function (5.3)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
88
to consider. The standard Bode integral theorem does not apply for various reasons.
Increasing y above one makes more poles than zeros in the sensitivity transfer
function. It also allows one to have a non unity value for the gain K r in equations
(4.7.4. 8 ). And in addition, since we are using a noncausal zero phase filter, there will
be poles outside the unit circle. Hence, ILC can bypass the waterbed effect in the
Again, for stability we would like to satisfy (5.4), and this means that we m ust
be able to bypass the standard waterbed effect again for the transfer function o f (ii)
above. This is accomplished in several ways. The poles outside the unit circle in the
zero-phase low-pass filter help decrease the integral according to (4.7). Also,
increasing y above one makes the coefficient o f the highest power o f r in the
num erator contain the learning gain <f> as a multiplicative factor making the first term
in (4.7) help.
Figure 5.1 plots the sensitivity function for an example system. We use the
same exam ple system for essentially all numerical results presented. It is a m odel o f
the closed loop transfer function for each link o f a Robotics Research Corporation
robot. In continuous form the transfer function is o f third order as given in (3.16)
where d = 8.8 rad/sec, co0 = 3 7 rad/sec, and £ = 0.5. A sample rate o f 100 Hz is
considered, and the low pass filter is a 5th order Butterworth with a cutoff at 100
rad/sec equivalent to 15.9 Hz. Using the filter in zero phase form gives a 10th order
attenuation. The learning gain tf>= \ and the phase lead parameter y = 5. It is clear
from Fig. 5.1 that the learning control law has very effectively bypassed the waterbed
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
89
result. The plot goes to Nyquist frequency, and although it is very slightly above unity
for frequencies above the cutoff, there was no significant penalty paid for the good
performance below the cutoff. N ote that the sensitivity function and the GBI Theorem
do not have any information about the finite num ber o f time steps p in the ILC
problem. They talk about steady state frequency response, and Fig. 5.1 gives results
for all frequencies. Which points on the plot refer to the problem at hand are
determ ined by the frequencies that can be seen in p time steps o f data. For large p
there will be many points, and for small p there are a small number o f frequencies. O f
course the plot is for steady state response, and therefore only applies for sections o f
the p step trajectory after initial condition influences have become negligible.
(5.5) for real time repetitive control. This sensitivity function differs from the
previous S i in that the Ts are replaced by z p . This fact prevents a phase lead in the
learning law from influencing the highest power o f z in the denominator. It would
require that y be larger than p , and this means we would need a future error. Also,
the zero-phase filter FL(z) m ust be replaced by the causal filter FR(z). This precludes
having poles outside the unit circle. The conclusion is that when using real time
repetitive control one is necessarily subject to the standard waterbed effect of (4 .5)
N ow consider the transfer function of (ii) above. It does not really apply as a
transfer function from one period to the next, except when the learning is sufficiently
slow that one can make a quasistatic assumption. Nevertheless, we are still interested
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
90
in satisfying equation (5.4) for stability, and this requires that the transfer function
involved bypasses the standard waterbed result. Both causal and zero phase filters can
producing the gain term on the right in the GBI Theorem (4.7). In the case o f a real
time zero phase filter, there are no poles outside the unit circle, but the H x (z~l)
Long Term Batch Update Repetitive Control: So far the discussions o f repetitive
control in this chapter have only considered the implementation with real time control
controller. Suppose we use the learning control laws as discussed above in a long
term batch update mode. The repetitive control problem can have a constant
command and a periodic disturbance o f period p time steps, or it can have only a
periodic command, or a periodic command and disturbance both o f the same period.
First, apply the desired output as the command to the feedback controller, and wait
until the response is considered to be in steady state. Take a substantial data set o f this
steady state response, and compute the zero-phase filtered value and the learning
update. Then apply the result, wait for steady state, make the next update in the
command, and continue the process. Since all responses used are considered to have
reached steady state, the frequency response information in the sensitivity transfer
learning control case. The possible repetitive control laws that can be used when
implemented in this batch update mode, can be the sam e as the range o f learning
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
91
control laws. A nd with sufficiently long waiting between each update, one should be
able to get arbitrarily close to the steady state response indicated by the learning
control sensitivity transfer function, e.g. Fig. 5.1. We will dem onstrate that this is the
However, the above thinking only applies to frequencies that are periodic with
period p tim e steps. Unlike the learning control problem w here one only has p time
steps o f data and hence there are no frequencies in the problem that are not
expressible w ith period p, in the repetitive control problem w e can have deterministic
disturbances that are periodic with som e unrelated frequency. The long term updates
must be tim ed to coincide with the phase o f the disturbance/comm and o f period p
time steps. Any unrelated determ inistic frequency will produce a learned response
whose phase could be at worst 180 degrees o ff in the next application o f the learned
corrective signal. And this can produce errors that can double the original error size.
This will be demonstrated later (Figs. 5.13-5.16), and the underlying mechanism is
similar to the discussion of Figs. 5.5 and 5.6 below. Although one can get
amplification o f errors in this batch update regime, the Bode Integral Theorem type o f
thinking does not apply to the problem. As formulated below, one needs a time
history o f the output over one or m ore periods for batch update vectors, and the
frequency response is seen as the frequency observed going down the entries in the
column vector. In addition, there are start and end effects in the colum n vectors, so
that the output is not a single frequency coming out for a single frequency going in. A
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
92
From the point o f view o f creating guaranteed stability using equation (5.4).
there is not distinction between using a causal low pass filter o f a zero phase low pass
filter F(z). However, the performance o f the system according to the sensitivity
transfer function (5.5) can be very different. Figure 5.2 plots the steady state
am plitude frequency response o f this sensitivity function when a 10th order causal
Butterworth filter is used with the third order feedback control system and linear
phase lead learning law discussed above. Figure 5.3 is the corresponding plot using a
real time zero phase 5th order Butterworth filter, which is effectively 10th order
because o f the combination o f forward and backward filtering. These plots go only up
to coT = 1.5 , but the remainder o f the plots up to Nyquist at n hold no surprises.
These plots look very similar, and both show the possibility o f amplifying by a factor
o f more than 5 at some frequencies different from those periodic w ith period p.
However, looking at where the curve approaches the horizontal axis near the point
coT = 0.5 one sees that the minimum points are different. In Fig. 5.2 the frequencies
that the system is trying to eliminate having period p are indicated by circles, and it is
clear that the causal filter does a poor jo b o f attenuating at these desired frequencies.
Figure 5.4 plots ju st these frequencies for both cases, and it is clear that even for the
fundamental o f period p the causal filter does not attenuate as much as 50% o f the
error, and that for the first harmonic it has actually made the error worse by a factor
2.3. This is easily understood by looking at the frequency transfer function for the
small co the M{co) is close to unity, and this function goes through a m inim um near
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
93
zero when the phase angles match, i.e. when a> = \ 2 x /(Tp)]l + 0{co) /(Tp) where i is
any integer. For the zero phase filter the 9(co) is zero, and this means the minima are
at the desired frequencies o f period p , but when 8{a>) is not zero for the causal zero
phase filter, the minima occur at the wrong frequencies. Since the minima are narrow,
be attenuated.
The above comments apply to the situation when the low pass filter is used as
in (5.4) to produce stabilization. In [13] a causal filter was used on the error only, and
stabilization was accomplished instead by quantization. The linear phase lead was
used to compensate for both the system phase and the filter phase, and good
performance was obtained. A batch zero phase low pass filter combined with this can
periodically eliminate any transient accumulation effect in the final error levels.
accomplished, and at the same tim e show some limitations o f the Bode integral
thinking. Again examine the numerator o f the sensitivity transfer function (5.5), but
for simplicity supposes that we are able to satisfy (5.4) without needing low pass
term is given by \[e'“>Tp - l j = [2 - 2 cos( 6>7/ 7)]‘' 2. For the frequencies periodic with
period p, this give the desired zero value. For frequencies half way between looking
back p time steps is looking back an odd number o f h alf periods o f the oscillation, so
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
94
that the terms adds the signals rather than subtracting th em . Thus the plot shown in
Fig. 5.5 (shown up to tt! 5 but the behavior remains th e same to Nyquist), and the
plot goes from the m inim a that cancel the frequencies o f interest to m axim a that
amplify by a factor o f 2 the frequencies one h alf w a y between the harm onics o f
period p. N ow exam ine what the average is over all frequencies up to Nyquist:
(1 /tt) f | txp(i9p) \dQ = 1.2732. This means that the te rm amplifies the disturbances
JO
when averaged over all frequencies. However, the Generalized Bode Integral
Theorem applies to this transfer function, and it gives th e average o f the logarithm as
zero. From this one m ight presume that there was no overall amplification on the
average, which we know is incorrect. The source o f the difficulty is that the integrand
in the GBI Theorem contains a logarithm o f a function that repeatedly goes to zero,
Figure 5.6 gives the corresponding plot when a re a l time zero phase low pass
filter is used (employing a 5th order Butterworth w ith a c u to ff at 100 rad/sec). Now in
place o f moving back and forth between zero and two, th e amplitude o f this decreases
as the value o f M(co) decays from unity. The average o f z p —F{z) over all
frequencies up to N yquist is 1.080 for the zero-phase filter case, and 1.087 for the
causal low pass filter case (10th order). The zero-phase filter gives the better value,
but the difference is not large. The big difference is th a t the causal case put the
minima in the wrong place. The corresponding Bode integral values are 1.006 x 10“ 13
for the zero-phase filter case, and 2.509 x 10-14 for ther causal filter. This time the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
95
better value is for the causal filter, but this difference might be at the numerical noise
function becomes larger than 2. for the peaks in Figs. 5.2 and 5.3. A t low frequencies
the factor [1 - G(z)L(z)] w ill be small, and can be near zero if the DC gain o f G is
unity and the learning gain is unity. Then one would observe am plification by a factor
o f approximately 2, as observed in Figs. 5.2 and 5.3 for the first peak. At high
frequencies the F in the num erator and the denominator goes to zero, making the
sensitivity transfer function approach unity at all frequencies in this range. For
intermediate frequencies before F has become small, the square bracket term in the
denom inator will introduce a phase change. For G having a phase lag as the
frequency goes up, the [1 —G(z)Z(z)] will exhibit a corresponding phase lead. The
denom inator will go through a minimum around where the phase o f z p matches that
o f the bracket term (for F zero phase), and this shifts the peaks o f Fig. 5.3 toward the
early side o f the intervals between harmonics o f period p. A nd o f course, the extra
am plification above a factor o f 2 at these points occurs because the denominator can
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
96
establishing this for L(z) = (fc.Y. The characteristic polynomial for (5.5) is the
num erator o f the square bracket term on the left o f this equation. Suppose the
denom inator has only roots inside the unit circle (G and F are asym ptotically stable).
counterclockwise around the unit circle, out to infinity along a branch cut, around at
infinite radius clockwise, and back along the branch cut. Then if the phase angle
comes back to its original value at the end, there are no roots outside the unit circle.
Instead o f using the term in square brackets on the left o f (5.5), we could use
z~pF(z)\\ —G(z)L(z)] and see if the phase angle measured from +1 comes back to
its original value to determ ine stability. This new expression is strictly proper
provided y —1 < p when G has one more zero than pole which is generically the case,
and y < p when the number o f zeros equals the number o f poles in G. Note that these
inequalities correspond to not asking for a future error in the learning control law.
U nder these conditions, the phase angle condition cited above defines the true
Note that for z on the unit circle, the magnitude [z^j = 1, and then satisfying (5.4) will
automatically satisfy the necessary and sufficient condition developed here. The
learning control, and examination o f the distinction between this sufficient condition
for stability and the true stability boundary are discussed in detail in [19,31],
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
97
Testing f o r the Stability Boundary Without a Filter: The normal Nyquist plot for
F W C K xW r)
z'-fW
When there is no low pass filter, F(z) = 1 , and the denom inator o f (5.18) contains p
roots on the unit circle. The contour m ust be adjusted to go around each o f these roots
when making the Nyquist stability analysis. This difficulty can be avoided by using
The equation (2.4), Z = (- W / 180°) + Pq + n ' I 2 , determines stability. Pick any point
(p on the positive real axis and track the change in phase angle from this point to the
points Q{e'°) as 9 goes from zero to 180 degrees, and this is W. The Pq is the
number o f poles of O outside the unit circle, i.e. for this application, the number of
zeros o f G outside. The n' is the num ber o f poles of O on the unit circle, which is p in
this application. Then Z is the num ber o f zeros o f 1 + 0 outside the unit circle, i.e. the
Introducing a Filter F: When a Butterworth low pass filter F is introduced the term
z p - F(z) no longer has roots on the unit circle except at z = 1, since |/r(e "9)| < 1 for
all 0 < 9 < 180°. But this makes the usual Nyquist plot singular at zero frequency
However, the Butterworth filter has a factor (z + 1)'" for its numerator, where the m is
the order o f the filter (or two times the order o f the filter in the batch zero phase case
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
98
with two applications o f the filter). There for by using O, the singularity has been
switched from zero to N yquist frequency, and again it becom es hard to plot and count
encirclem ents.
Assum e that the unit circle condition (5.4) is satisfied by for repetitive control
gains in the range 0 <<f> < $ max . T hen consider plotting O. One has trouble finding W,
but one knows Pq and n' , and this implies that we know W for the above range of <fi.
The range o f stabilizing (f> will continue past the value (pxmax, which is a sufficient
condition for stability, until it hits the actual stability boundary <f>2max. And this new
gain is obvious for the beginning o f the plot of Q, giving the interval on <p up to the
first loop o f the curve. Hence, this forms a simple m ethod to answ er the following
question. Pick a repetitive controller that satisfies the sufficient condition for stability
(5.4) for sufficiently small gain (f>, and then ask how far beyond the maximum
allowed gain satisfying (5.4) can one go before actually reaching the true stability
boundary? This approach allows one to adjust the gain to reach the stability boundary,
but it does not allow one to adjust the cutoff frequency upward violating (5.4). Figure
5.7 shows the 0 Ijnaxand f ° r the third order system (3.16) using a normal causal
10th order Butterworth filter w ith a 100 rad/sec cutoff and a sam ple rate o f 100 Hz.
The control law as before uses a phase lead parameter o f y = 5. This time the
parameter con is allowed to vary along the horizontal axis. The sufficient condition
(5.4) produces maximum values o f the gain that are very close to the true stability
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
99
discussion address each type o f problem, we com m ent on the stability o f batch update
repetitive control as discussed in the next section. W hen updates are made every p
time steps, stability is determ ined by seeing whether all eigenvalues o f the matrix A
o f equation (5.20) below have magnitude less than one, and for the case o f updates
every 5p one uses the A o f equation (5.30) below. Using p = 100 and the third order
system o f (3.16), the m axim um eigenvalue in the former case is 0.9680 and in the
The developments o f the previous sections say that if the repetitive control
law does an update in real time, it will be subject to the limitations o f the waterbed
effect. On the other hand, iterative learning control with zero phase filtering is not
subject to this effect. By m aking batch updates in repetitive control, and by waiting
until reaching steady state before making an update, one expects that one can obtain
the same performance in repetitive control as in learning control, at least for tracking
errors o f period p. In this section we investigate what the final error levels should
look like for this mode o f operation, for the frequencies periodic with p steps, and
factor o f 2 is again possible, but by doing batch updates, amplifications above this
value are no longer likely to occur in the steady state response. It takes some
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
100
mathematical development to create the equivalent for batch update o f the frequency
response o f the sensitivity transfer function, and this is the purpose the section.
update at the end o f every period. This is a limiting situation that m ight require some
fast com putation in the last time step for the zero phase filter. In the next section we
will relax the computation requirement. Using the underbar notation for the history o f
variables in a period as introduced for equations (5.1) and (5.2), we can write the
relevant equations as
£j = + A mxj(0)
y j = C X j + w; xM (0) = CXj (5.19)
u J+ 1 = F ( U j + L e j )
For sim plicity both the forward and the reverse time initial conditions for the filter F
are set to zero for this development. Matrices P r,Am are P r,Am with the output
matrix C v removed, and the matrix C is a block diagonal matrix with C v along the
diagonal. For this repetitive control problem the first entry in x } is the initial
condition, and according to the third equation o f (5.19) the initial condition for period
j + 1 is equal to the final state for period j . A minimal state vector to use for the
periodic updates must contain the p step history o f the inputs as contained in and
the initial condition for each period x y( 0 ) . From this state xj the first o f equations
(5.19) gives the complete state history for period j as an output equation.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
101
Developing the update equations for the two parts o f th is new state vector and
x=y+1 = A= x= / + = y - w—)'
B (_«/ (5.20)
r c
——a zo
xo C
--- Pr
x 1 ' 0 '
;A = :B =
-F L C A xn F (I —LCPX)_ - F L .
Suppose that y - w is constant so that the command a n d the disturbances are both
periodic w ith period p tim e steps. Then when steady s ta te response is reached,
!„ = { /-C k „ -K ) (5.21)
analog o f the frequency transfer function for such a batch: update process. W hat we
want to know is the frequency content o f the history o f tfie error that one observes
going down the column matrix e ss. We can substitute a sin u so id for either the desired
trajectory or the output disturbance, and obtain the response. This process works for
all sinusoids that produce constant values for y or w Tor all repetitions, i.e. for
frequencies that have p time steps as their period, a fundam ental with that period and
all harmonics up to Nyquist. Since the desired trajectory is supposed to have this
period, this equation is sufficient for studying the tracking ezrror to commands.
behaves when confronted with a disturbance that does not h ap p en to be at one o f the
frequencies targeted for elimination by the repetitive controller. It is not obvious how
to do this for all frequencies, but we can generate the frequency response information
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
102
illustrated for tripling the num ber o f frequencies addressed. To plot two points
betw een each o f the ones obtained above, consider frequency com ponents in the
output disturbance that are periodic with period 3p. This can be represented w ithin the
period j + 2, and repeat for J = 1,4,7,10,... This produces the dynamic equations from
x j^ . = A x j + B y j - w 2 ;A = AJ ' , B = A ZB AB s] (5.22)
To develop the associated steady state error, first find the steady state value o f the
state vector for these j , by setting x j = x ^ , = x v[, and then propagate this state to the
Then the steady state error history is periodic with 3p, and is given by e for
Consider the third order system again, and consider a period p o f 100 time
steps. To study the final error level associated with the 5th harmonic at 5 Hz. set the
command (5.21) to equal the sampled version o f sin(lOzzr). The resulting steady state
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
103
error is plotted in Fig. 5.8. The corresponding plot for the 14 H z harmonic is given in
Fig. 5.9. A linear phase lead param eter y o f five steps was used. In matrix form for
the batch update, this means that there is no learning in the last four steps o f the data
set used, and this can be the cause o f some o f the end effects observed. The more
important consideration is the zero phase filter initial conditions at the start and the
end o f this time period. It is clear that one needs to pad the ends to allow time for
decay o f these initial condition effects, so that the choice o f their values has
negligible influence.
intersperse the batch update computation among the real tim e computations was
presented in [13]. This com putes one time step o f filtering update o f the batch zero
phase filter during each time step o f the real time process. To allow substantial time
for the effect o f filter initial conditions to decay, the forward filtering is started one
allow substantial time for the filter initial conditions to decay for the reverse path, an
extra period is added on the end as well. Therefore, the batch update has the
following features: the filtering is started at the beginning o f some period f. , and it
continues filtering the appropriate function as the data arrives through three periods,
including periods, I , t + 1 and t + 2. Starting with the 4th period, the results o f the
forward filtering are sent through the filter in reverse order, one time step o f the
forward filtered result is processed for each time step in real time. Once two more
periods have elapsed, the desired zero phase filter result is available for the period H
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
104
continue the filtering through period £. This creates a batch zero-phase filter update
every 5 periods.
The subscript j is now reserved to count the batch updates, so that it increases
one unit every 5 periods. The Xj (0) is the system state at the start o f these 5 periods.
The vectors giving the history o f the states for each o f these 5 periods are
same p step history for all 5 periods. Following the method o f equation (5.29) the
As before the minimal state vector needed to know what happens every repetition is
that given in equation (5.20) containing an update of the initial system state at the
start o f the 5 periods, xy(0), and the repetitive control action, u] . Cascading the
equations in (5.24) produces the update for the first o f these two
Xj . l (0) = A uxJ(0) + An uj
(5.25)
du = ( Q A,ay ^ d n = U + C A „ + (£ A a,)* + (C T t„)3 + ( C 4 „ ) 4]
Now we need to develop the update equation for uJ+l. The zero phase filtering is
done on three periods o f the error, so we define e} to contain the error history for the
first three periods o f the five periods for update j, and it can be written as
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
105
Li *7.
- c (5.26)
*7 = L, -W j 2 *72
L, 7^ _
Here we have allow ed the disturbance for each o f the periods to be different so that
the final equation can be used to analyze disturbance o f any period. O f course, when
the disturbance contains only period p , then all o f the w are the same and can be
replaced by vv. During the first three periods, a causal zero phase filter is being
applied forward in tim e. Denote this filter by its matrix o f M arkov parameters P~ (we
let the initial condition be zero, since we are allowing a full period to the initial
condition effects to decay). Then the result o f the forward filtering o f the three
“j
P3{ “ j (5.27)
y.j
Note that the learning law L is being applied for all three periods to the error history.
M ultiplying this by P.; produces the backwards filtering o f all three periods. The
product P.r P3 is the zero phase filter F. Then it is the center period that is o f interest,
producing
Hm = L F i Hj +Le j } (5.28)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
106
Com bining equations (5.24), (5.26), and (5.28) and packaging the result w ith equation
(5.25) produces the state equation for the repetitive control batch update
0
x
— j-r I
= ,4x + Ba y■Ld —-wJ - \B
---------- j — ’ =0 = (5.29)
v —— w j>~_ J « FL.
JLd
CAm ~i 1
CAxo(CAxo) ^ 2 2 = I » F 1 - K .F L C I + AxnC
CAxo(CAxu)2 1 J + Axoc + (A xoc ) \
and the A u ,Ar are given in equation (5.25), where these are the four partitions of
From these equations one can find the steady state error for disturbances of
period p , and also for disturbances o f period 5p, immediately. Consider the second
case containing the disturbance sequence w,,vv2,w3,vv4,u ’5 that occurs for the 5
x = { I - A Y xBa (5.31)
This can be used in equations (5.24) to produce the steady state values for the five
periods x vvl ,x u2,x „3 ^,,5 >and then the steady state errors for the five periods are
= y d - C x „ , - w, ; / = 1,23,4,5 (5.32)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
107
Setting all w, to w gives the equations for determining the final error level in
Figure 5.10 shows the steady state error for the fundam ental frequency o f I
Hz. The error is quite small, but it does contain not only the fundamental but also
some harmonics. Figure 5.11 does a DFT on the 500 points and shows the harmonic
content. Figure 5.12 gives the result for a 2 Hz input, and the presence o f harmonics
is no longer obvious.
It is o f interest to understand the relationship between the final error level and
the learning gain. Reference [27] looks at the steady state covariance equation and
suggests that in the presence o f w hite noise the final error level will get better as the
learning or repetitive control gain gets smaller. The action taken by the learning law
as a result o f the noise will be smaller. Simulation here gave the opposite result for
the amplitude is about 6 x 10-13 using a gain o f <p = 1, and when the gain is decreased
the learning control sensitivity transfer function, St = (1 —F) /(I —F + FGL) , for low
frequencies. In this range F is near unity and this means that the denominator is
dominated by the term FGL, whose value is approximately equal to 0, and hence
decreasing this gain increases the final error level in this range. This suggests that
final error levels when using a cutoff filter has a tradeoff in the choice o f gain for
final error level with respect to the performance for terms periodic with p and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
108
N ow consider what happens to the final error level w hen there are
determ inistic disturbances at frequencies other than those o f period p. One can o f
course apply the method outlined in equation (5.23) to the 5p approach in this section
have the ability to examine four evenly spaced frequencies within the one Hertz
intervals between harmonics o f period p. Making use of these results. Fig. 5.13 shows
the steady state error for a sine wave input at 1/5 Hz, and also for a cosine input o f the
same frequency. Note that the error during the second one second period is nicely
cancelled, but the algorithms assumes that this same error will appear for all 5 o f the
one second periods, and therefore does inappropriate things for the other 4 seconds.
W hen the reaction is exactly wrong one expect to reach a m axim um error that
version o f the mechanism discussed for Figs. 5.5 and 5.6. Figure 5.14 gives the
corresponding plots for sine wave inputs o f frequency 1/5 and 4/5 Hz. At 8 Hz the
coT = 0.502 which corresponds to starting up the knee o f the curve in Fig. 5.1. The
performance for this frequency that is periodic with period p is still relatively good
with an amplitude o f 5 x 10- 3 . For a frequency o f 8.2 Hz the steady state error level
to a sine wave input is shown in Fig. 5.15. The frequency 8.8 Hz produces a very
sim ilar plot, while 8.4 and 8.6 Hz produce plots o f the same nature but this time the
amplitudes o f oscillation in the first and third second are larger than the amplitudes in
the fourth and fifth seconds. Figure 5.16 gives the corresponding result above the
cutoff where the learning process is much attenuated, at the frequency o f 20.2 Hz
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
109
which is associated with coT = 1.268 . The response looks like the input except that it
is superim posed on a slow small am plitude oscillation. Only the tim e from 1 to 2
seconds is shown, but the remaining 4 seconds appear like a simple continuation o f
the same patter. For 20.6 Hz the behavior is sim ilar except that instead o f one small
amplitude oscillation in the one second shown, there are 3 such oscillations.
N ow exam ine the comparison between the performance o f ILC as seen in the
sensitivity transfer function o f Fig. 5.1, and the performance obtainable in long term
batch updates in repetitive control. To w hatever extent the long term batch updates
correspond to steady state frequency response updates associated w ith the transfer
function from update to update used in (5.4), the behavior should be the same as that
in Fig. 5.1. Actually, it should be possible to have Fig. 5.1 be m ore precisely
applicable to the batch update repetitive control than it is to learning control. In the
approach steady state frequency response behavior, while the learning control
problem is always limited to p time steps. O f course this perform ance is only
Equation (5.21) could be described as the sensitivity matrix for Ip updates analogous
to the sensitivity transfer function SL , and using equations (5.31) and (5.32) gives the
5p analog. Using this second analog, taking the discrete Fourier transform o f the 500
points in the error, and computing the difference with Fig. 5.1 produces the 5p curve
in Fig. 5.17. Agreement is quite good, with the peak o f the difference being
4.5 x 10 We can improve this agreement by allowing more time to reach steady
state response between updates, by following the same algorithm but using 2p in
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
110
place o f p. Figure 5.17 also displays the result for this case, but one cannot see the
actual size o f the error. Figure 5.18 plots the result w ith a different scale, and we see
that the amplitude the difference in performance betw een the learning control result in
Fig. 5.1 and this longer term batch update is roughly 6 x 10~15.
5.10 Conclusions
Repetitive control aims for zero error for the period associated with an
im portant disturbance source, at least up to some cutoff frequency for the learning.
This means that the frequency transfer function from disturbance to error m ust go to
near zero at the fundamental frequency and at integer multiples up to the cutoff. If the
occur between these harmonics, or above the cutoff. It is shown here that real time
repetitive control is subject to this waterbed effect, and that the response can be quite
sensitive, going to zero for a chosen frequency and being amplified by a factor o f 5 at
a very nearby frequency. It is shown that batch update repetitive control can bypass
the waterbed effect for frequencies o f period p , approaching the same sensitivity
transfer function behavior observed for iterative learning control. For other
frequencies between those o f period p, there can be amplification, and the mechanism
2. This is in contrast to the real time repetitive control situation where the
peaks that amplify by a factor o f over 5. Thus, long term batch updates have better
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Ill
steady state performance than real time repetitive control. This improved performance
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
112
55 0.8
0.6
0.4
CO
0.2
0.5
coT
Fig. 5.2 Amplitude o f frequency response o f SR(z) using a causal low pass filter.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
113
CO
coT
Fig. 5.3 Amplitude o f frequency response o f SR(z) using a zero phase low pass Filter.
2.5
>
o
<o
*o
3
■<— *
*5
on
CO
0.5
2.5
coT
Fig. 5.4 Comparison o f Figs. 2 and 3 at frequencies periodic with period p steps.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
114
0.5
0.5
0.5 2.5
coT
Fig. 5.6 Magnitude plot o f \^-F(z)\ w ith real-time zero-phase filter
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
115
2 . 5,
N y q u ist sta b ility boundary for p = 1 0 0
0.5
0 i
0.32 0.34 0.36 0.38 0.4 0.42 0.44
C0o T
Fig. 5.7 Comparison o f true stability boundary and monotonic decay condition.
0.5
-0.5
0.2 0.4
0.6 0.8
Time
Fig. 5.8 5 Hz command y j and steady state error with batch zero phase update every p
steps.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
116
0.5
-0.5
o
fc
W
GO
GO
Time
Fig. 5.10 Steady state error for 1 Hz command, zero phase batch update every 5p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
0 *
*1 " T S " "T " 1 ' ill>
Frequency
Fig. 5.11 Frequency content o f Fig. 5.10.
x 10 -1 0
Time
Fig. 5.12 Steady state error for 2 H z command, batch update every 5p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
118
cos
sin
Time
Fig. 5.13 Steady state error to 1/5 Hz sine and cosine input, batch update every 5p.
2/5 Hz
4/5 Hz
Time
Fig. 5.14 Steady state error to 2/5 Hz and 4/5 Hz sine inputs, batch update every 5p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
119
s_
O
H
M
C/2
C /2
U !
1 2 3 4 3
Time
Fig. 5.15 Steady state error to 8.2 Hz sine input, batch update every 5p.
\
0.5
LU 0
C/2
C /2
-0.5
-1
1 1-2 1.4 1.6 1.8
Time
Fig. 5.16 Steady state error to 20.2 Hz sine input, batch update every 5p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
120
x 10 ’ 12
10p
2.5
coT
Fig. 5.17 Difference in frequency response amplitudes o f jS^I and \Sr\ for 5 and 10p, for
frequencies periodic with period p.
Co
1
2.5
coT
Fig. 5.18 Enlarge scale graph o f Fig. 5.17 for 10p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
121
C hapter 6
C onclusions
design requirement, and stability conditions are discussed in chapter 2. Results show
convergence to zero tracking error in both ILC and RC. In RC the advantages o f this
condition over the stability condition defining the actual boundary o f stability are the
ease o f com putation and the fact that the condition implies monotonic decay o f error.
In chapter 3 methods are developed to predict final error level o f ILC and RC
when using linear phase lead com pensator as well as zero-phase filter. These predict
the steady state error levels that will be reached for each choices made by the control
system designer. These methods can be used to optimize the choice o f the learning
process.
to learning control problems using the sensitivity function generated in chapter 3. The
establishes the waterbed effect. The waterbed effect states that the transfer function
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
122
from command to the resulting error, or from output disturbance to resulting error,
has the property that attenuation in some frequency range m ust be paid for by
Integral Theorem that shows how the waterbed effect is bypassed by learning control
law designs. This includes introducing more zeros than poles, modifying the gain in
the sensitivity transfer function, and introducing poles outside the unit circle with
non-causal filtering.
control are studied using filtering for a frequency cut o ff the learning for the purposes
easier to satisfy the stability condition. It is shown that when using filtering in a real
time repetitive control mode, the frequency response stability condition is only
slightly more restrictive than the Nyquist stability condition defining the stability
engineering practice, one prefers to use the simple frequency response based stability
condition because o f the ease o f computation. It is shown that for real time repetitive
control, one is subject to the waterbed effect. On the other hand, it is possible to
bypass this effect for those frequencies having the period o f interest, by making batch
updates o f the repetitive control signal. Sufficiently slow batch updates in repetitive
control will result in the same situation that applies to iterative learning control for
these frequencies, where it is possible for certain classes o f learning control laws to
bypass the waterbed effect. Methods are developed to predict the disturbance to error
frequency response characteristics for the range o f possible batch updates rates.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
123
R eferrences
1. Arimoto, S., Kawamura, S., and M iyazaki, F., "Better operation o f robots by
learning." Journal o f Robotic System. 1984, V o l.l, p p .123-140
2. Bode, H.W., Netw ork Analysis and Feedback Am plifier Design. Princeton. NJ:
Van Nostrand, 1945.
4. Casalino, G. and Bartolini, G., "A learning procedure for the control movements
o f robotic m anipulators," IASTED Sym posium on Robotics and Automation.
Amsterdam, The Netherlands, 1984, pp. 108-111.
6. Elci, H., Longman, R.W ., Juang, J.-N., and Ugoletti, R., "Discrete frequency
based learning control for precision m otion control," Proceedings o f the 1994
IEEE International Conference on Systems, Man, and Cybernetics, San Antonio.
TX, USA, 1994, pp. 2767-2773.
7. Elci, H., Longman, R.W., Juang, J.-N., and Ugoletti, R., "Experiments in the use
o f learning control for maximum precision robot trajectory tracking," Proceedings
o f the 1994 Conference on Information Science and Systems, D epartm ent o f
Electrical Engineering , Princeton University, Princeton, NJ, USA, 1994, pp. 951 -
958.
8. Elci, H., Longman, R. W., Phan M. Q., Juang, J.-N. and Ugoletti, R . , "Automated
Learning Control throught Model Updating for Precision M otion Control."
Adaptive Structures and Composite M aterials: Analysis and Applications, AD-
Vol. 45/M D-Vol. 54, AS M E 1994, pp. 299-314.
9. Francis, B., and Zames, G., "On H-infinity optim al sensitivity theory for SISO
feedback system," IEEE Transactions on Autom atic Control, 1984, Vol. 29, no. 1.
pp. 9-16.
10. Hara, S., Omata, T., and Nakano, M., "Stability o f repetitive control systems."
Proceedings o f th 24th IEEE Conference on Decision and Control, Fort
Lauderdale, FL., 1985, pp. 326-327.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
124
11. Hara, S., and Yamamoto, Y., "Synthesis o f repetitive control system s and its
application, " Proceedings o f th 24th IEEE Conference on Decision and Control.
Fort Lauderdale, FL., 1985, pp. 1387-1392.
12. Hsin, Y.-P., Longman, R. W., Solcz, E. J., and de Jong, J., "Experimental
Com parisons o f Four Repetitive Control Algorithms," Proceedings o f th 31st
Annual Conference on Information Sciences and Systems. John Hopkins
University, Department of Electrical and Computer Engineering, Baltimore,
M aryland, 1997, pp.854-860.
13. Hsin, Y.-P., Longman, R. W., Solcz, E. J., and de Jong, J., "Experiments bridging
learning and repetitive control," Advances in the Astronautical Sciences, 1997,
Vol. 95, pp. 671-690.
14. Hsin, Y.-P., Longman, R. W., Solcz, E. J., and de Jong, J., "Stabilization due to
Finite W ord Length in Repetitive and Learning Control," Advances in the
Astronautical Sciences, Vol. 97, 1997, pp. 817-836.
15. Huang, Y.-C. and Longman, R. W ., "The Source o f the Often Observed Property
o f Initial Convergence Followed by Divergence in Learning and Repetitive
Control," Advances in the Astronautical Sciences, Vol. 90, 1996, pp. 555-572.
16. Inoue, T., Nakano, M., and Iwai,S., "High accuracy control o f a protron
cynchrotron m agnet power supply," Proceedings o f the 8th World Congress o f
IFAC, XX, 1981, pp. 216-221.
17. Lang, S., Complex Analysis, 3rd edition, Springer-Verlag, New York, 1993.
19. Longman, R. W., "Iterative learning control and repetitive control for engineering
practice," International Journal o f Control, 2000, Vol. 73, no. 10, pp. 930-954.
20. Longman, R. W., "Real time HR zero-phase low-pass filtering for robust
repetitive control," Proceedings o f the 6th International Conference on Control,
Automation, Robotics and Vision, Singapore, December 2000, to appear.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
125
24. Nakano, M., and Hara, S., "M icroprocessor-based repetitive control,"
Microprocessor-Based Control Systems, 1986, pp. 279-296.
25. Oh, S. J., Longman, R. W. and Phan, M. Q., "Use o f Decoupling Basis Functions
in Learning Control for Local Learning and Improved Transients," Advances in
the Astronautical Sciences, AAS/AIAA Spaceflight Mechanics Meeting, 1997,
pp. 651-670.
26. Omata, T., Nakano, M., and Inoue, T., "Application o f repetitive control method
to multivariable systems," Transactions o f SICE, 1984, Vol. 20, pp.795-800.
27. Phan, M. Q. and Longman, R.W., "A mathematical theory of learning control for
linear discrete multivariable systems," Proceedings o f the AIAA/AAS
Astrodynamics Conference, Minneapolis, Minnesota, USA, 1988, pp.740-746.
28. Pierre, D. A., "Reformulated Nyquist criterion for discrete-time systems," IEEE
Transactions on Education, 1989, Vol. 3, n o .l, pp. 59-61.
29. Plotnik, A. and Longman, R. W., "Subtleties in the use of zero-phase low-pass
filtering and cliff filtering in learning control," Proceedings o f the 1999
AIAA/AAS Astrodynamics Conference, Girdwood, Alaska, Advances in
Astronautical Science , Vol. 103, 1999, pp.673-692.
30. Seron, M. S., Braslavsky, J.H., and Goodwin, G. C., Fundamental Limitations in
Filtering and Control, Springer Verlag, 1997.
31. Songchon. T. and Longman, R.W., "Comparison o f the stability boundary and the
frequency response stability condition in learning and repetitive control," Second
International Workshop on Multidimensional (nD) Systems, Czocha Castle,
Lower Silesia, Poland, 2000, pp. 121-126.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
126
32. Songchon, T. and Longman, R. W., "Iterative learning control and w aterbed
effect," Proceedings o f the 2000 AIAA/AAS Astrodynamics Conference, Denver,
Cororado, 2000, pp. 444-453.
34. Tomizuka, M., Tsao, T.-C., and Chew, K.-K.," Analysis and synthesis o f discrete
time repetitive controllers," Journal o f Dynamic Systems, M easurement, and
Control, 1989, Vol. 111, pp. 353-358.
35.U chiyam a, M., "Form ulation of high-speed m otion pattern o f a m echanical arm
by trial (in Japanese)," Transactions o f the Society fo r Instrumentation and
Control Engineer, 1978, Vol. 14, pp. 706-712.
36. Wang, Y. and Longm an, R. W., "Use o f Non-Causal Digital Signal Processing in
Learning and Repetitive Control," Advances in the Astronautical Sciences. Vol.
90, 1996, pp. 649-668.
37. Wen, H. P., Phan, M. Q. and Longman, R. W., "Bridging Learning Control and
Repetitive Control using Basis Functions," Advances in the Astronautical
Sciences, 1998 AAS/AIAA Spaceflight M echanics Meeting, Vol. 99. 1998,
pp.335-354.
39. Wirkander, S.-L. and Longman, R. W., "Lim it Cycles for Improved Perform ance
in Self-Tuning Learning Control." Advances in the Astronautical Sciences. Vol.
102, 1999, pp. 763-781.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.