Anda di halaman 1dari 25

Automatica 36 (2000) 1401}1425

Survey Paper
Research on gain scheduling夽
Wilson J. Rugh *, Je! S. Shamma
Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA
Department of Mechanical and Aerospace Engineering, UCLA, Los Angeles, CA 90095, USA
Received 20 April 1998; received in revised form 6 December 1999

Current research on gain scheduling is clarifying customary practices as well as devising new
approaches and methods for the design of nonlinear control systems.

Abstract

Gain scheduling for nonlinear controller design is described in terms of general features of the approach and in terms of early
examples of applications in #ight control and automotive engine control. Then recent research is discussed, emphasizing work on
linearization-based scheduling and work on linear parameter-varying approaches.  2000 Elsevier Science Ltd. All rights reserved.

Keywords: Gain scheduling; Linear parameter-varying systems; Nonlinear control systems

1. Introduction tions papers. Examples in the 1970s include KaK llstroK m,


As stroK m, Thorell, Eriksson and Sten (1979) and Stein,
After decades of silence on the subject, the academic Hartmann and Hendrick (1977). And some papers from
research community in the 1990s addressed the topic of the 1980s on the subject of feedback linearization com-
gain scheduling with rapidly increasing interest. There is, ment on an `automatic gain schedulinga interpretation.
of course, a long history of gain scheduling in applica- This appears, for example, in the chapter on gain sched-
tions, an aspect we address brie#y in the sequel. But the uling in the "rst edition of As stroK m and Wittenmark
historically minded reader will discover few citations to (1995). However, the fact remains that theoretical treat-
the control research literature before 1990 in our dis- ments of gain scheduling as a worthy design methodo-
cussion. Indeed, gain scheduling has never been included logy for nonlinear control are rare until the 1990s (Rugh,
as a subject heading in the indexes of the IEEE Transac- 1991; Shamma & Athans, 1990).
tions on Automatic Control. `Gain controla does appear in Journals in various applications areas have more fre-
the 25th Anniversary Index (1956}1981), published in quently contained papers involving the design of gain
1981, though only one of the "ve listed papers is relevant scheduled controllers. A search of engineering index, Ei
to gain scheduling. To its credit, Automatica lists gain Compendex, covering both theoretical and applications
scheduling as a subject in its 1963}1995 cumulative in- papers, yields the data on growth of publications dealing
dex, published in 1995. Of the four citations given, the with gain scheduling and linear parameter-varying (LPV)
only one dated earlier than 1990 is a survey that includes systems shown in Table 1. It is gratifying to see a rapidly
a brief description of the subject. Also we note the ab- growing number of papers presenting innovative gain
sence of gain scheduling in the Automatica keyword list. scheduling approaches to various applications at con-
There are occasional, pre-1990 papers in these journals ferences on control, and we believe that new theoretical
that do mention gain scheduling, particularly applica- formulations and advances have played important roles
in generating this increased interest. However, in the

sequel, we discuss applications only in the context of
This paper was not presented at any IFAC meeting. This paper was historical examples.
recommended for publication in revised form by Editor M. Morari.
* Corresponding author. Tel.: #1-410-516-7004; fax: #1-410-516- Our goal is to survey the current status of recent
5566. research and some future possibilities. The coverage in-
E-mail address: rugh@jhu.edu (W.J. Rugh). cludes gain scheduling approaches that are classical, at

0005-1098/00/$ - see front matter  2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 0 0 ) 0 0 0 5 8 - 3
1402 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

Table 1 ly in the sequel. That is, we ignore smoothness assump-


Ei compendex data tions, existence/uniqueness questions, most speci"cation
of domains, ranges, dimensions, etc., and we rely on
1980s 1990}94 1995}98
context for reasonable interpretations of such matters.
Gain scheduling 25 101 185 At various points, certain time signals are treated as
LPV systems 0 4 54 constant parameters, and thus the argument notation
can become confusing. Typically, we will drop the argu-
ment (t) to unclutter displayed equations, though the
argument is often retained for emphasis when discussing
least in their application, as well as new approaches. particular variables.
These all share as a common core the use of linear
controller design techniques to address nonlinear prob-
lems. Though our treatment is not focused on particular 2. Gain scheduling
engineering areas, the research discussed is "nding its
most direct applications in the control of electro-mech- Many di!erent design notions can be viewed as gain
anical systems. For a survey focused on applications in scheduling. For example, precompensating a nonlinear
chemical process control, see Bequette and Doyle (2000). gain with the inverse gain function can be interpreted as
gain scheduling, as can switching gain values according
1.1. Organization to operating conditions or even according to preset
times. Various other ideas such as controller switching
The organization of the paper is as follows. In Sect- and controller blending "t a broad interpretation of gain
ion 2 we provide a general description of gain scheduling scheduling. Indeed gain scheduling seems mainly to be
approaches, both classical and modern. Also various a point-of-view taken in the design process. Certainly
features of gain scheduling, including advantages and examples of controllers designed by gain scheduling can
disadvantages, are listed. Then we give early examples of be found that, perhaps with a bit of diagram manipula-
gain scheduling applications, focusing on #ight control tion, would not be recognized as gain scheduled control-
and automotive engine control. lers. In this survey the focus is on gain scheduling in the
Gain scheduling is based on linear parameter-varying sense of continuously varying the controller coe$cients
plant models, and in Section 3 we review two ways in according to the current value of scheduling signals, also
which such models can arise: linearization about a mani- called scheduling variables, that may be either exogenous
fold of constant equilibria, and the so-called quasi-LPV signals or endogenous signals with respect to the plant.
approach. These are illustrated by a simple "rst-order However, other interpretations are not completely
PID control problem and by a pitch-axis missile model ignored.
* examples that are revisited in subsequent sections. In broad terms, the design of a gain scheduled control-
In Section 4 we review linearization-based gain sched- ler for a nonlinear plant can be described as a four-step
uling, focusing on recently developed analytical aspects. procedure, though di!ering technical avenues are avail-
Of particular interest are the notions of hidden coupling able in each step.
terms that arise as a consequence of the design approach, The "rst step is to compute a linear parameter-varying
and the extent to which closed-loop stability properties model for the plant. Historically, the most common ap-
result from the design process. proach is based on Jacobian linearization of the nonlin-
Linear parameter-varying and linear fractional trans- ear plant about a family of equilibrium points, also called
formation (LPV/LFT) approaches are treated in Sec- operating points or set points. This yields a parametrized
tion 5. We review how linear matrix-inequality- (LMI) family of linearized plants and forms the basis for what
based constructions of H optimal controllers can be we call linearization scheduling. Typically, the parametr-

adapted to compute gain scheduled controllers for LPV ization corresponds to "xed values of scheduling vari-
systems that guarantee stability and performance. ables that, in turn, are, or are functions of, internal plant
Section 6 presents an overview of analysis results re- variables and exogenous signals. Another approach is
garding the stability and performance of nonlinear gain quasi-LPV scheduling, in which the plant dynamics are
scheduled control designs. rewritten to disguise nonlinearities as time-varying
Finally, Section 7 gives a discussion of recent research parameters that are used as scheduling variables.
directions in gain scheduling. Depending on the speci"c plant structure, quasi-LPV
scheduling need not involve any Jacobian linearization
1.2. Notation of the plant dynamics.
The second step is to use linear design methods to
Because we cite sources for details on every topic, design linear controllers for the linear parameter-varying
mathematical notions are presented somewhat informal- plant model that arises in either linearization or quasi-
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1403

LPV scheduling. This design process may result directly E Gain scheduling design approaches are naturally com-
in a family of linear controllers corresponding to the patible with decompositions of the overall control
linear parameter-dependent plant, or there may be an problem. The decompositions are typically not hier-
interpolation process to arrive at a family of linear con- archical, and the interactions of sub-problems are cap-
trollers from a set of controller designs at isolated values tured by scheduling variables. Exogenous scheduling
of the scheduling variables. Traditionally, the designs are signals are typically state variables in a more complete
such that for each xxed value of the parameter, the linear model of the overall system.
closed-loop system exhibits desirable performance. More E Gain scheduling enables a controller to respond rap-
recent design approaches attempt to assure desirable idly to changing operating conditions. For this it is
performance for time-varying parameter trajectories in important that the selected scheduling variables re#ect
a speci"ed class. changes in plant dynamics as operating conditions
The actual gain scheduling, the third step, involves change.
implementing the family of linear controllers such that E The computational burden of linearization scheduling
the controller coe$cients (gains) are varied (scheduled) approaches is often much less than for other nonlinear
according to the current value of the scheduling vari- design approaches. On the other hand, quasi-LPV
ables. Various issues arise here, depending on whether we approaches are computationally intensive, but o!er
are using a quasi-LPV approach or a linearization sched- guaranteed stability and performance properties.
uling approach. E Gain scheduling often involves several ad hoc steps,
The fourth step is performance assessment. In the beginning with problem formulation. This can be suit-
best case, where analytical performance guarantees able in simple situations, but increasingly troublesome
are part of the design process, this may be relatively as more complicated controllers are designed.
simple. More typically, the local stability and perfor- E The linearization scheduling approach does not apply
mance properties of the gain scheduled controller might when little information is carried by plant lineariz-
be subject to analytical investigation, while the ations about constant equilibria. The quasi-LPV ap-
nonlocal performance evaluation is based on simulation proach, because of its conservative nature, may not
studies. yield controllers.
E Linearization gain scheduling depends on intuitive
2.1. Features of gain scheduling rules of thumb and extensive simulation for evaluation
of stability and performance. Typically, stability can be
A number of aspects of gain scheduling are listed assured only locally and in a `slow-variationa setting,
below, including aspects that have stimulated its wide- and typically there are no performance guarantees.
spread use in applications, have become well understood The quasi-LPV approach o!ers the potential of both
as limitations of the approach, and that have motivated stability and performance guarantees.
recent research.
2.2. Historical comments
E Gain scheduling employs powerful linear design tools
on di$cult nonlinear problems. This is true of both the As discussed in Section 1, very little was published on
classical linearization scheduling approaches and the gain scheduling in the archival control research literature
more recent quasi-LPV approaches. before 1990. One reason is that gain scheduling was
E Most performance speci"cations are phrased in linear considered an &applied' topic, and scheduling techniques,
terms, involving a mixture of time-domain and fre- particularly their hardware implementations, often were
quency-domain speci"cations, and these are directly proprietary or classi"ed as military secrets. We will com-
addressed, at least in a local sense, in a gain scheduling ment on two application areas * one commercial and
approach. one military * where gain scheduling has played an
E Gain scheduling does not require severe structural important role and where some documentation is avail-
assumptions on the plant model, and the approach can able. Both are considered in terms of developments only
be used in the absence of complete, analytical plant in the United States, and even with this restriction a care-
models (database models). In particular, linearization ful study of the history is beyond our scope. Thus we
scheduling can be used when plant information is con"ne remarks to some early examples of gain schedul-
limited to a few equilibria and the corresponding plant ing that are direct forerunners of current practice.
linearizations. Indeed there was great activity in these and other
E Design by gain scheduling preserves well-understood application areas in the period between the early work
linear intuition and is carried out using the physical we describe below and the present. We hope that our
variables in the plant model. (This is in contrast to brief comments will motivate others to attempt a more
nonlinear control approaches that involve coordinate complete story. In any case, the market has now spoken:
transformations.) gain scheduling is an e!ective and economical method
1404 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

for nonlinear control design in practice. This is obvious To address failure tolerance, early autopilot designs
from discussions with practitioners, and from the rapidly often provided for reversion to a nominal constant gain
increasing, open publication of applications of gain in case of scheduling hardware failure. By the late 1950s
scheduling in conference proceedings and journals. With redundancy was in use. For example, separate airspeed
little exaggeration we can say, for at least the last few measurements were used in individually gain scheduled,
decades, `Machines that walk, swim, or #y are gain redundant controllers. Inaccuracies in the separate
scheduleda. measurements coupled with inaccuracies in the analog
Because of cost considerations and performance re- hardware became a dominant problem in failure-tolerant
quirements, it is not surprising that gain scheduling tech- designs, particularly as autopilot `authoritya increased.
niques were adopted in military applications relatively This provided an important motivation for rapid ad-
early, while commercial applications of gain scheduling option of digital autopilot systems.
began when digital computer control came on the scene. We brie#y describe below some interesting features of
In other words, gain scheduling is an old idea, but before early gain scheduled autopilots for a few well-known
digital implementation of controllers, it was expensive vehicles.
and di$cult to realize in hardware. This is re#ected in the To compensate for changing aero-surface e!ectiveness,
quarter-century di!erence in the adoption of gain sched- the B-52 autopilot (Sperry Gyroscope Co.), originally
uling in the two application areas discussed. developed around 1951, incorporated an airspeed-based
mechanism (bellows and linkage) to change the gain of
2.2.1. Autopilots and yight control the servo actuator's position feedback transducer. The
Gain scheduling in autopilots evolved naturally out of underlying design concept was model-following control
late World War II designs as dictated by stability di$cul- with an architecture that yielded di!erent transfer func-
ties with the new jet aircraft, the advent of the guided tions for the command and disturbance inputs. The loop
missile, and increased operating envelopes with more gain was scheduled to maintain appropriate closed-loop
demanding performance requirements. Major engineer- bandwidth in the face of changing surface e!ectiveness.
ing challenges involved the analog circuitry needed to This bandwidth viewpoint avoided much more complex
compute true airspeed from available measurements approaches that attempted to schedule several di!erent
while maintaining hardware simplicity. The most widely gains. Position feedback was eliminated later, ca. 1957,
known aspect is the fact that autopilot loop gain is highly by the use of so-called integrating servos, the analysis of
dependent on variations in control surface e!ectiveness which led to some interesting confusions as discussed in
due to altitude (dynamic pressure) and Mach number Osder (1996). Mechanisms for scheduling on dynamic
e!ects. Typically, the goal of scheduling was to maintain pressure were introduced to keep loop gains reasonably
constant loop-gain despite these e!ects. Many early im- constant.
plementations were based on gain switching when #ight The Nike Ajax missile (Bell Telephone Laboratories
conditions changed su$ciently. More sophisticated and Douglas Aircraft Co.), by 1951, incorporated hard-
approaches varied the gains in a continuous fashion, ware gain scheduling on dynamic pressure via a ram air
though the hardware complexity of the required sensors, tube extending from the nose that drove a multiple wind-
"lters, ampli"ers, and relays led in some cases to unac- ing potentiometer set. One potentiometer adjusted the
ceptably high failure rates. feedback gain in the "n control system and another
By the late 1950s, this situation was an important a!ected a gain in the roll control system. Designers of the
motivation for Air Force interest in adaptive control period report that the most di$cult task was hardware
methods that presumably could avoid the need for air design, due to the limited space and limited selection of
data. (Indeed gain scheduling was, and is, sometimes approved resistor and capacitance values Waymeyer
considered as a type of `open-loop adaptive controla.) (1997).
Detailed information about the state-of-the-art and In the Talos missile (Bendix Aviation, Johns Hopkins
views at the time is available in Gregory (1959). Applied Physics Laboratory) of the early 1950s, control-
ler gains were adjusted by a timer in the boost phase, and
then scheduled on altitude during the midcourse, beam-
riding phase. During the terminal homing phase, the
 Gain scheduling as we know it today, that is. Undoubtedly there autopilot compared the requested lateral acceleration
were earlier e!orts, particularly in view of the observation in Howard with the lateral acceleration produced by the control
(1973) that automatic #ight control is now at least 125 years old, surfaces and adjusted gains to maintain good response.
predating the Wright brothers #ight by some 30 years. See also, for This was called a `sensitivity feedback systema, and it
example, Draper (1955) and McRuer and Graham (1981).
 In the Introduction to this Proceedings, gain scheduling is de-
acted to compensate for changes in both altitude and
scribed as an open-loop adjustment of parameters by a process of speed (Paddison, 1982). Thus the #ight of a Talos exhib-
`omphaloskepsisa, a word that seems to be undeservedly little known ited a rudimentary form of gain scheduling at the outset,
today. followed by a scheduling approach that is now quite
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1405

standard, and "nally what might be called performance- closed-loop control of exhaust gas recirculation (EGR)
adaptive control. beginning in the early 1980s, and variable camshaft tim-
ing where gain scheduling e!orts are reported beginning
2.2.2. Automotive engine control in the mid 1980s. A continuing theme through reports of
Gain scheduling of automobile engine controllers be- these applications is that the economy and power of the
gan in conjunction with the use of microprocessor-based digital microprocessor enable more complex control
air/fuel mixture control and in response to the dual schemes.
imperatives of improved fuel economy and reduction of
exhaust emissions. Therefore, we focus on electronic fuel
control, particularly control of the air/fuel ratio, begin- 3. Linear parameter-varying plant descriptions
ning in the early 1970s.
Use of a catalytic converter requires precise control of Consider the nonlinear plant
the air/fuel ratio and necessitates feedback. This is ac-
x "a(x, u, w, v), z"c (x, u, w, v),
complished using a signal from an exhaust gas oxygen 
(EGO) sensor located in the exhaust pipe. The EGO y"c (x, w, v), (1)

provides a two-state signal indicating that the air/fuel
where x is the state, u is the input, z denotes an error
ratio is either lean or rich. Conventional (usually propor-
signal to be controlled, and y denotes a measured output
tional plus integral) control results in a limit cycle about
available to the controller and usually includes penalized
the desired value. By using asymmetric gains with respect
variables, tracking commands, and some of the state
to the sign of the error, the limit cycle and thus the
variables. Both w and v denote exogenous inputs to the
nominal air/fuel ratio can be shifted. Note that there is
plant. Informally, the signal w captures parametric de-
a substantial transport delay (load and engine-speed de-
pendence of the plant on exogenous variables, while the
pendent) from the time the fuel}air mixture is inducted
signal v represents external `input functionsa such as
until the signal appears at the EGO sensor (Cook,
reference commands, disturbances, and noises. The dis-
Grizzle & Sun, 1996).
tinction between the two is not well de"ned, but is more
An early paper, (Rivard, 1973), describing closed-
a matter of perspective. Ultimately, the practical distinc-
loop electronic fuel injection control refers to `circuits
tion will be in how they are treated in the design process.
[that] can be incorporated to control the system limit-
Namely, in linearization scheduling, we will linearize
cycle characteristica and one "gure shows a gain in-
Eqs. (1) with respect to v but will not linearize with respect
#uenced by `any sensed variable.a However the block
to w. Rather, the variable w is included as part of the
diagram of the system shows only constant-gain, integral
parametrization of the linearization family.
control.
As pointed out in Section 2, the "rst step in controller
In 1978 Ford Motor Company introduced its "rst
design by gain scheduling is to obtain a linear description
production, microprocessor-based feedback control sys-
or approximation for the nonlinear plant that involves
tem. That same year a gain-scheduled air/fuel control
a suitable set of scheduling variables. There are two
system was installed on about 30,000 Ford Pintos and
approaches to this: one is the classical linearization ap-
Mercury Bobcats sold in California. The carburetor used
proach and the other, more recent, is the quasi-LPV
a vacuum-actuated, variable ori"ce to meter fuel in re-
description.
sponse to the EGO sensor signal. Integral-only control
was used, and the gain (`ramp ratea) was dependent on
3.1. Jacobian linearization family
engine speed (Seiter & Clark, 1978). Also in 1978 Ford
described a concept vehicle incorporating fuel injection
A plant equilibrium point x , u , w , v is a point such
and air/fuel control by integral-only feedback with the    
that a(x , u , w , v )"0. The following de"nition de-
integrator gain scheduled on engine speed (Bailey,    
scribes a set of equilibrium points parametrized by
Cederquist, Florek, Hart & Meitzler, 1978). Papers
a scheduling variable denoted by p. We assume that
appearing in the same year from General Motors and
p includes the exogenous scheduling signal, w, and
Nissan describe carburetors with closed-loop air/fuel
p might also depend on the measured output, y. How-
control, but do not mention gain scheduling (Canale,
ever, since real-time values of p(t) will be used in the
Winegarden, Carlson & Miles, 1978; Masaki, Aono,
implementation of the scheduled controller, p must not
Akaeda & Minami, 1978). Developments were rapid in
depend on unmeasured internal variables or unmeasured
this period, however, and in 1982 the General Motors
exogenous signals.
system is described as employing fuel injection with gain
scheduling on engine speed and manifold pressure
De5nition 1. The functions x (p), u (p), w (p), and v (p)
(Gumbleton & Bowler, 1982).    
de"ne an equilibrium family for the plant (1) on the set S if
There are additional aspects of air/fuel regulation
where gain scheduling has been employed. These include a(x (p), u (p), w (p), v (p))"0, p3S. (2)
   
1406 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

Associated with this equilibrium family is the error The interpretation is that at each "xed p linearization (5)
equilibrium family, describes the local behavior of the nonlinear plant about
the corresponding equilibrium. As mentioned earlier, we
z (p)"c (x (p), u (p), w (p), v (p)), p3S (3)
      have not included linearization terms involving an asso-
and the measured output equilibrium family, ciated deviation variable w .
B
An important occurence is when p depends on state
y (p)"c (x (p), w (p), v (p)), p3S. (4)
     variables through the measured output (scheduling on
Often equilibrium points that are determined from a internal variables) in addition to w. Namely, these vari-
speci"ed z (p) are called trim points. For example, if the ables appear both as parameters (components of the
 vector p) and as dynamical variables (components of
objective is that z(t) track a constant reference input
r with zero steady-state error, then the equilibrium family y (t)) at di!erent places in the linearized plant. This has
B
should satisfy rami"cations in the theoretical analysis of gain schedul-
ing as a design method, a topic addressed in the sequel.
0"a(x (p), u (p), w (p), v (p)),
    In some cases, for example when a portion of the plant
0"z (p)"c (x (p), u (p), w (p), v (p)) description is a database, a family of plant linearizations
      might be obtained by interpolating coe$cients of plant
and p might be chosen to include r. In any case, we linearizations at isolated equilibria. Looking ahead to the
denote the dependence of the scheduling variable on second step of the gain scheduling process, sometimes
other variables by p"g(y, w). It is natural to assume that there is a choice: interpolate plant linearizations and
there exists a function w (p) such that w"w (g(y, w)). If design a family of linear controllers, or design controllers
 
p depends on a reference signal r, through the depend- for distinct linearizations and interpolate the controllers.
ence of y on r, we assume that there exists a function r (p)

such that r"r (g(y, w)). Usually, these functions simply Example 2. Consider the scalar nonlinear plant

strip o! the appropriate entries in the vector scheduling
variable. x "!x#u, y"tanh x,
Again we note that the scheduling variable p is treated
as a parameter throughout the design process, and then it z"r!y"r!tanh x. (8)
becomes a time-varying input signal to the gain sched-
uled controller implementation through the dependence We assume the objective is that y(t) track an exogenous
p(t)"g(y(t), w(t)). This is the main `tricka in the gain reference signal r(t) with acceptable transient response
scheduling approach. and zero steady-state error for constant r(t). In this
De"nition of the family of equilibrium points and the simple case the general notation is unwieldy, so we do not
selection of scheduling signals are problem dependent. include the reference signal r in y, but display it separate-
Furthermore, it can be di$cult to solve for the functions ly. Also, there is no exogenous scheduling signal present.
describing equilibrium families, though in many applica- However, there is an obvious equilibrium family that
tions physical insight is very helpful * the interpretation corresponds to zero tracking error, and this family can be
of `rest pointsa, also called `constant operating pointsa, parameterized by p3(!1,1) as
of physical systems is often highly intuitive.
Corresponding to a speci"ed equilibrium family there x (p)"u (p)"tanh\ p, r (p)"y (p)"p,
   
is a plant linearization family that can be written in the
linear parameter-varying form z (p)"0 (9)


   
x A(p) B (p) B (p) x and the deviation variables are
B   B
z " C (p) D (p) D (p) v , p3S, (5)
B    B x (t)"x(t)!tanh\ p, u (t)"u(t)!tanh\ p,
y C (p) D (p) 0 u B B
B   B
where, for example, y (t)"y(t)!p, r (t)"r(t)!p, z (t)"z(t). (10)
B B B
*a The corresponding linearized plant family is
A(p)" (x (p), u (p), w (p), v (p)),
*x    
(6)
x "!x #u , y "(1!p)x ,
*c B B B B B
D (p)"  (x (p), u (p), w (p), v (p))
 *u     z "r !y . (11)
B B B
and where the deviation variables are de"ned in the
In subsequent developments we take a classical view-
obvious fashion, for example,
point towards the linearized control problem, and con-
x (t)"x(t)!x (p), u (t)"u(t)!u (p). (7) sider (11) in terms of the parametrized transfer
B  B 
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1407

function The corresponding linearization family has the form

 
  
> (s) 1!p *c Md c
B " , p3(!1,1). (12) M L cos a! Mc ! L K sin a 1


; (s) s#1 a  *a L d a
B B " K B
q  *c q
B M K 0 B
Of course, a natural choice of scheduling variable in *a
a gain scheduling implementation is p"y. Another pos-
sibility is to choose p"r, although di!erences arise in
 
Md cos a
# L d ,
the gain scheduled controller, as will be discussed in due Md B
course. K

  
*c a
g " KM L 0 B #(KMd )d , (17)
Example 3. A simple model for the longitudinal dynam- B *a q L B
ics of a missile is B
where all partial derivatives are functions of the schedul-
a"MC (a, d, M) cos a#q, q "MC (a, d, M), (13) ing variables a and M, and the deviation variables have
L K
the obvious de"nitions. As discussed above, we do not
where the state variables are angle-of-attack a(t) and include deviation variables in M since it is a constant at
pitch-rate q(t). Mach number M(t) is an exogenous vari- equilibrium. However, the endogenous variable a enters
able in this model, and d(t) is the tail-"n de#ection. The (17) as both a constant parameter at equilibrium and as
aerodynamic coe$cients C (a, d, M) and C (a, d, M) typi- a dynamical variable through a .
L K B
cally are obtained empirically, though we assume that Since angle-of-attack can be di$cult to measure, per-
they have linear-in-tail-de#ection forms haps a preferable choice of scheduling variables is g(t)
and M(t). This involves a change of parametrization, and
C (a, d, M)"c (a, M)#d d, results in more complicated expressions for the equilib-
L L L
rium and linearization families. See, for example, the
C (a, d, M)"c (a, M)#d d. (14) formulation in Nichols, Reichert and Rugh (1993) where,
K K K
in the notation of (1), the measured output is
Also, we assume that the controlled output is normal
acceleration g(t), described by y(t)"[q(t) g(t) g (t)]2. (18)

g"KMC (a, d, M), (15)


L 3.2. Quasi-LPV description
where K is a constant. A typical control problem is to
design an autopilot that accepts as input a normal accel- This approach is based on the possibility of rewriting
eration command, g (t), from a guidance system, and the plant in a form where nonlinear terms can be hidden
produces the output d(t) with the objective that the plant with newly de"ned, time-varying parameters that are
output g(t) track g (t). That is, z(t)"g (t)!g(t). then included in the scheduling variable, in addition to w.
There are various possibilities in the formulation of Since nonlinear terms involve the state, this means that
linearization scheduling based on both endogenous some state variables must be relabeled as parameters in
and exogenous scheduling signals. One is to choose various parts of the model, while they remain dynamical
the exogenous variable M(t) and the endogenous variables elsewhere.
variable a(t) for scheduling, which requires that a(t) be Speci"cally, it is assumed that the state of plant (1) can
included in the measured output y(t). In this case, be divided into two partitions:
straightforward calculations give the input, state, and

 
output equilibrium family for (13) (parametrized by a x (t)
x(t)" ? , (19)
and M), x (t)
@
!c (a, M) where x comprises those state components that are
d (a, M)" K , ?
 d sometimes rewritten as parameters. With
K

   
d c (a, M) x (t)
q (a, M)"M L K !c (a, M) cos a, p(t)" ? , (20)
 d L w(t)
K

 
d c (a, M) the fundamental assumption of the approach is that
g (a, M)"KM c (a, M)! L K . (16)
 L d the plant description (1) can be written in the quasi-LPV
K
1408 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

form Now, the dynamics are in a quasi-LPV form with the


`parametera p"x .

   
x A (p) A (p) B (p) B (p) x 
?   ? ? ?
x A (p) A (p) B (p) B (p) x It is clear that quasi-LPV representations are not
@ "   @ @ @ , unique, and determining a suitable representation de-
z C (p) C (p) D (p) D (p) v
? @   pends on the intended gain scheduling implementation.
y C (p) C (p) D (p) D ((p) u( (p, u) One special class of systems that always admit a suitable
? @  
(21) representation are so-called `output nonlinear systemsa.
These systems are of the form
where p(t) is treated as a time-varying parameter that is

 
unknown a priori, but measurable for implementation d y y
purposes. In addition the function u( (p, u) is assumed to be "
(y)#AI (y) #BI u . (26)
dt x x
invertible with respect to u, that is, there exists a function P P
u( \ such that Note that state is partitioned into a measured output and
remaining states. Furthermore, the nonlinearities depend
u( \(p, u( (p, u))"u. (22) only on the measured output. Now assume there is an
equilibrium family parametrized by y such that
If the plant can be written in this form, with state-
dependent nonlinearities disguised in (21) as coe$cient
 
y
dependencies on p(t), then a linear control design for the
(y)#AI (y) #BI u (y)"0. (27)
x (y) 
control input u( yields a design for the actual control via P 
u"u( \(p, u( ).
Then using the state and control variables
Notice that in this formulation there is no issue of
plant linearization, and no explicit equilibrium point
 
y
family other than the equilibrium point at the origin for x" , u"u !u (y) (28)
x !x (y) 
the linear system (21). In other words, the quasi-LPV P P 
approach achieves linearity in plant description in a quite leads to the quasi-LPV dynamics
di!erent way than the linearization family approach.
Of course if a control design is satisfactory for all x "A(y)x#B(y)u, (29)
time-varying parameters p(t), then it is satisfactory when
the x component of p actually takes on the value of the where
?
dynamical variable x (t) in the original plant description.
?

AI (y)

In other words, description (21) introduces additional 0
A(y)"  ,
behaviors beyond the plant description (1), and this typi- 0 AI !/VP  (y)AI (y)
cally increases the di$culty of obtaining a suitable con-  /W 

 
troller. A particularly di$cult situation is that of rapidly BI
B(y)"  u (30)
varying p(t), in which case a controller must hedge BI !/VP  (y)BI
against behaviors of the quasi-LPV description that  /W 
might be well beyond the original plant's capabilities. and AI and BI are derived from an appropriate block
GH G
decomposition of AI and BI .
Example 4. Consider the nonlinear plant
Example 5. A quasi-LPV formulation for plant (8) based
x "sin x #x , x "x x #u. (23) on replacing x by p in a portion of the measured output
     
leads to the LPV plant
One quasi-LPV representation is
x "!x#u, y"(1/p)tanh p x, z"r!y. (31)

  
sin x /x 1 0
x "A(x)x#Bu"   x# u. (24)
x 0 1
 Example 6. A quasi-LPV formulation can be obtained
This representation is not suitable for gain scheduling for the missile dynamics in Example 3 as follows. Con-
unless the entire state is available for feedback, since the sider a change of variables
matrix A(x) depends on both state variables. Supposing
q( "q!q (a, M), dK "d!d (a, M),
that only x is available for feedback, a more suitable  

representation is g( "g!g (a, M). (32)


  
sin x /x 1 0
x "   x# u. (25) These variables represent the deviation from an equilib-
0 x 1 rium corresponding to the current a and M. The new

W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1409

state dynamics become The operation F (M, Q) denotes an `uppera LFT. We



can similarly de"ne the `lowera LFT
a"q( #Md cos a dK ,
L
Fl (M, Q)"M #M Q(I!M Q)\M . (38)
   
 
*q (a, M) *q (a, M)
q( "!  q( # Md !Md cos a  dK
*a K L *a A parameter-dependent plant, as in (5) or (21), admits an
LFT description if each of the parameter dependent
*q (a, M) matrices can be written in an LFT form F (M, *(p)) for
!  Q ,
M 
*M some constant matrix, M, and diagonal p-dependent
matrix, *(p), with the following structure. Suppose p has
g( "KMd dK . (33)
L p components, i.e.,
The term M Q can be viewed as a new disturbance input. p"[p p
  2 pN ]2. (39)
Then de"ning the parameters p "a and p "M, we
 
obtain a quasi-LPV representation of the form Then

 
a *(p)"diag(p I  ,p I  ,2,p I N ), (40)

 
a 0 1 0 b (p ,p )  L  L N L
?   q( where the n are suitable positive integers and I I denotes
q( " 0 a (p ,p ) b (p ,p ) b (p ,p ) . I L
   @   @   Q
M the n ;n identity matrix.
g( 0 0 0 d (p ) I I
  In general, an LFT description is capable of represent-
dK
ing any polynomial or rational matrix function of a
(34) scalar variable. The problem of constructing an LFT
description is essentially equivalent to that of construct-
3.3. Linear fractional transformation description ing a realization of a transfer function. Hence, there need
not be a unique representation, and some representations
Both the Jacobian linearization description, as in (5), may not be `minimala. For further discussion (see Zhou,
and the quasi-LPV description, as in (21), lead to a para- Doyle & Glover, 1996, Chapter 10).
meter-dependent family of linear systems. In some cases,
we can model or approximate the parameter dependence Example 7. Recall the linearized plant family (11). The
in the LPV system as a linear fractional transformation output matrix, C(p) can be written as
(LFT). The LFT description is a special case of an LPV
C(p)"1!p"F (M, *(p)) (41)
system. However, the particular structure will allow spe- 
cialized approaches to deriving gain scheduled control- with
lers (Section 5.6).

  
In the block diagram of Fig. 1, where M is partitioned 0 1 0
M " , M " ,
as  0 0  1

 
M M M "[!1 0], M "1 (42)
M"   , (35)  
M M
  and
a and b are related by

 
p 0
*" . (43)
b"(M #M Q(I!M Q)\M )a (36) 0 p
   
which we write as
b"F (M, Q)a. (37) 4. Linearization gain scheduling

The design of a linearization gain scheduled controller
entails designing a linear controller family corresponding
to the plant linearization family (5). This results in a lin-
ear controller family parametrized by p, with input signal
y (t) (including, if appropriate, r (t)) and output signal
B B
u (t). These are local deviation signals associated with the
B
linearized plant near the equilibrium parametrized by p.
As mentioned in Section 3.1, since p includes the
exogenous signal w, the linear controller design typically
will not involve explicit input of w (t) because w(t) is
B
Fig. 1. Upper linear fractional transformation F (M, Q).

viewed as the constant parameter w. We use the following
1410 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

notation to describe the linear controller family: available, and ignores the need to interpolate controllers
in the case of isolated point designs. However, a few ad
x 
   
F(p) G(p) x
B " B , p3S. (44) hoc approaches have been reported. These include
u H(p) E(p) y Nichols et al. (1993), where controller transfer functions
B B
are interpolated by linear interpolation of poles, zeros,
In the best case we might be able to directly design
and gains, (Reichert, 1992), where H controllers are
a family of controllers via some particular design 
interpolated by linearly interpolating the solutions of
method, for example the LPV or LFT methods discussed
Riccati equations, (Kellett, 1991) where state-space coef-
in Section 5. However, in many situations a set of so-
"cient matrices of balanced controller realizations are
called point designs are performed at selected equilibria,
linearly interpolated, (Hyde & Glover, 1993) where linear
with the restriction that all designs must be of the same
interpolation of state and observer gains is used, and
dimension. Then a family of linear controllers is com-
(Buschek, 1997; Kelly & Evers, 1997; Niemann & Stous-
puted by interpolating the point designs.
trup, 1999) where controller blending is used. Each of
these approaches has an intuitive appeal, and performed
4.1. Controller interpolation
satisfactorily in the application considered. On the other
hand, each approach can generate controllers that are
Corresponding to selected equilibria p ,2,p repre-
 ) not stabilizing at intermediate equilibria, a situation that
sentative of the equilibrium family p3S, suppose a linear
leads to instability in the closed-loop gain scheduled
controller is designed for each of the corresponding plant
system.
linearizations so that speci"cations are met. This results
In addition to these ad-hoc methods, theoretically
in an indexed collection of linear controllers,
justi"ed methods have been proposed. In Shahruz and
x 
   
FI GI x Behtash (1992) an algorithm for linear interpolation of
B " B , k"1,2, K. (45) eigenvalue-placement state feedback gains is presented.
u HI EI y
B B In Stilwell and Rugh (1999, 2000), approaches to linear
Here we use the same notation for the variables in each dynamic controller interpolation are presented that sat-
controller, assuming that the dimensions are the same for isfy a linear stabilization requirement, namely, that the
every k, and leave it understood that each variable is interpolated controller stabilize the corresponding lin-
a deviation variable with respect to its indexed equilib- earized plant for every p3S. Then it is shown that
rium. That is, for example, y (t)"y(t)!y (p ). a bound on p(t) can be computed that guarantees local
B  I
A linear parameter-varying controller (44) can be stability properties for the gain scheduled nonlinear
based on collection (45) of point designs in many di!erent system.
ways. For example, the equilibrium family can be divided We should note that LPV approaches o!er another
into disjoint regions, S"S 626S and the controller way to interpolate controllers (Packard, 1994; Apkarian
 )
coe$cients varied by switching according to the obvious & Gahinet, 1995). A particular structure for the control-
rule: p3S implies use of the corresponding coe$cients ler is pre-determined, and then the controller state-space
I
FI, GI, HI, EI. This has the advantage of simplicity in matrices for every p3S are speci"ed by certain matrix
de"ning the controller family, but the disadvantage of parameters that are computed during the design.
discontinuities (jumps) in the controller coe$cients, pos-
sible chattering behavior, and so on. 4.2. Gain scheduling
Usually, the indexed controllers are interpolated with
respect to the parameter p in a smoother, at least con- From a broad perspective, the objective is to compute,
tinuous, way. Interpolating state-space controller coe$- based on the linear controller family (44), a controller
cients is one option. For example, a coe$cient matrix that has the general form
function F(p) is computed to have certain smoothness
x "f (x, y, w), u"h(x, y, w) (46)
properties for p3S and satisfy F(p )"FI, k"1,2, K.
I
Another possibility is to interpolate features of the with the obvious input and output signals corresponding
input}output representations of the indexed set of to the original nonlinear plant description (1). The notion
controllers, for example transfer-function coe$cient of linearization gain scheduling arrives at this controller
interpolation. via replacement of the parameter p in (44) by g(y(t), w(t)),
Much of the literature on linearization scheduling as- though additional considerations arise.
sumes that a continuously interpolated controller is It is easy to show that if a nonlinear plant and nonlin-
ear controller both have equilibria at the origin, then the
corresponding closed-loop system has an equilibrium at
 A theoretician tends to think of point designs at a few selected the origin. Furthermore, the linearization of the closed-
equilibria, but there are #ight control systems with several thousand loop system about the origin can be computed as the
design points! closed-loop system for the linearized plant and linearized
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1411

controller. (See Khalil, 1996, Section 11.2.1) for an ex- Example 8. Returning to Example 2, and the linearized
plicit discussion with linear state feedback.) This same plant transfer function (12), consider unity feedback with
result holds at nonzero equilibria if the controller equilib- a PI controller acting on the tracking error. Such a linear
rium `matchesa the plant equilibrium. (See Rugh (1983) controller has a transfer function of the form
for various aspects related to nonzero equilibria of non-
; (s) k (p)
linear systems.) Thus, the objective in linearization sched- B "k (p)# G . (51)
uling is that the equilibrium family of the controller (46) Z (s) N s
B
match the plant equilibrium family, so that the closed-
In this case we can select the PI gains (as functions of the
loop system maintains suitable trim values, and, second,
scheduling variable) so that the family of closed-loop
the linearization family of the controller is the designed
linear systems (12), (51) has a transfer function that is
family of linear controllers.
independent of p. For example, choosing
For the equilibrium conditions of plant (1) and con-
troller (46) to match, there must exist a function x (p) 1 1
 k (p)" , k (p)" (52)
such that N 3(1!p) G (1!p)
0"f (x (p), y (p), w (p)),
   yields the closed-loop transfer function
u (p)"h(x (p), y (p), w (p)), p3S. (47)
    > (s) ()s#1
B "  . (53)
In words, (47) requires that at equilibrium the controller R (s) s#()s#1
B 
supply the correct constant input signal to the plant to
maintain the trim condition on which the plant's equilib- Assuming that the characteristics of this transfer function
rium family is based. Then the second requirement is that are suitable, we have completed the linear design process,
the controller linearization, about each such controller and it remains to implement the gain scheduled control-
equilibrium, is exactly the linear controller (44) designed ler.
for that equilibrium. Thus we require, for example, A standard realization of the parametrized PI control-
ler (51) can be written in the form (44), again with the
*f reference signal explicitly displayed, as
(x (p), y (p), w (p))"F(p),
*x   


x
x  B
  
*h 0 !1 1
(x (p), y (p), w (p))"E(p), p3S. (48) B " y , p3S. (54)
*y    u k (p) !k (p) k (p) B
B G N N r
B
The most direct way to accomplish this is to specify, in
The corresponding controller equilibrium family x (p)
terms of the coe$cients in (44), the controller family 
should be such that (47) is satis"ed with
x "F(p)[x!x (p)]#G(p)[y!y (p)],
  u (p)"tanh\ p, y (p)"r (p)"p (55)
  
u"H(p)[x!x (p)]#E(p)[y!y (p)]#u (p). (49)
   and this gives x (p)"(1/k (p)) tanh\ p. Then (49) gives
At any p3S, it is easy to see that this controller satis"es  G
a linear controller family that simpli"es to
the equilibrium condition (47). Furthermore, a simple
exercise in di!erentiation * continuing to view p as x "r!y, u"k (p)x#k (p)[r!y]. (56)
G N
a parameter * shows that the requisite linearization
Upon replacing the parameter p by the current value of
conditions exempli"ed in (48) are satis"ed. Of course, it is
the measured output, y(t)"tanh x(t), we obtain the
possible that there are many other controllers of the form
gain-scheduled, nonlinear controller
(46), in addition to (49), that satisfy the required condi-
tions. However (49) is a customary choice in applications. x "r!y,
Implementation of the gain scheduled controller based
on (49) is the conceptually simple matter of replacing the u"k (y(t))x#k (y(t))[r!y]
G N
parameter p by its dependence on other variables accord-
3x#r!y
ing to p(t)"g(y(t), w(t)). Thus (49) leads to the gain sched- " . (57)
uled controller 3(1!y)

x "F(g(y, w))[x!x (g(y, w))] Simulating the response of the nonlinear closed-loop
 system (8), (57) with zero initial conditions, we encounter
#G(g(y, w))[y!y (g(y, w))], instability for constant inputs r larger than about 0.8.

u"H(g(y, w))[x!x (g(y, w))] This might be viewed as a disappointing result, for the
 linear closed-loop family (11), (56) is stable for every
#E(g(y, w))[y!y (g(y, w))]#u (g(y, w)). (50) p3(!1,1).
 
1412 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

4.3. Hidden coupling *k (p)


u "k (p)x # G x (p)y #k (p)[r !y ]
B G B *p  B N B B
We have seen that (49) satis"es the basic relations
guaranteeing that at equilibria corresponding to con- 1 1 2p tanh\p
" x # [r !y ]# y .
stant values of the parameter p the linearized controller 1!p B 3(1!p) B B 1!p B
matches the designed linear controller family (44). How- (59)
ever, the implemented gain scheduled controller is
described by (50). To analyze the behavior of the gain- We note that this exhibits an extra term when compared to
scheduled, closed-loop system (1), (50), it is a natural "rst (56). Using the relationship y "(1!p)x , the correspond-
step to linearize about closed-loop equilibria to assess B B
ing linearized closed-loop system (11), (59) can be written as
local behavior. This can be accomplished by separately

   
linearizing the plant and controller. However lineariz- x !4/3#2p tanh\p 1/(1!p) x
B " B
ation of (50) is a di!erent matter than linearization of (49) x  !(1!p) 0 x
since we no longer have the arti"cial parameter p ex- B B
plicitly appearing in the controller. (Although p con-
 
1/(3(1!p))
tinues to parametrize the equilibrium family of the # r
1 B
controller.) This di!erence can lead to terms in the lin-
earization of (50) that are not present in (44). with output given by
Linearization of (50) about an equilibrium speci"ed by
p gives, after careful di!erentiation and evaluation,

x
y "[(1!p) 0] B .
B x
x  "F(p)x #G(p)y B
B B B
It is straightforward to check that this linear system is

 
*x (p) *y (p)
! F(p)  #G(p)  unstable for "p"'0.725. Thus we see that when the
*p *p hidden coupling term is taken into account the instability
encountered in simulation, as mentioned in Example 8, is

 
*g *g to be expected.
; (y (p), w (p))y # (y (p), w (p))w ,
*y   B *w   B
As discussed from various perspectives in Rugh (1991),
u "H(p)x #E(p)y Nichols et al. (1993) and Lawrence and Rugh (1995),
B B B
there is a potential remedy for the appearance of hidden

 
*u (p) *x (p) *y (p)
#  !H(p)  !E(p)  coupling terms. To illustrate these results in the present
*p *p *p context, we state the following.

 
*g *g Theorem 10. Given the linear controller family (44), there
; (y (p), w (p))y # (y (p), w (p))w . (58)
*y   B *w   B exists a gain scheduled controller (50) with no hidden
coupling terms if there exists a controller equilibrium func-
Comparing this to (44) exhibits the additional terms in tion x (p) satisfying
the actual linearized controller when compared to the 
designed linear controller family. The additional terms *x (p) *y (p)
F(p)  #G(p)  "0,
act as additional feedback terms in the case of compo- *p *p
nents of y (t) involving plant state variables, and as
B
exogenous disturbances in the case of components in- *u (p) *x (p) *y (p)
 !H(p)  !E(p)  "0, for p3S. (60)
volving w (t) or r (t). These additional terms are called,
B B *p *p *p
appropriately, hidden coupling terms.
The conditions in (60) form a set of "rst-order partial
Example 9. Linearizing the gain scheduled controller di!erential equations with the various coe$cients speci-
(57) of Example 8 about an equilibrium p"y gives "ed by the linear controller family coe$cients and the
plant equilibrium functions. Existence conditions for
x  "r !y , a solution x (p) are known, but in general are quite
B B B 
restrictive. However, in simple cases the satisfaction of
(60) might be apparent after suitably modifying the real-
ization of the designed linear controller family.
 For many years various practitioners have been well aware of the
phenomenon we call hidden coupling, and have used terms such as
`"ctitious gainsa and `hidden loopsa. However the issues seems not to Example 11. Returning again to Example 8, consider, in
have been analyzed in any general context until relatively recently. place of (54), a di!erent realization of the parametrized PI
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1413

controller transfer function in (51): interesting implementation for gain scheduled controllers
that avoids hidden coupling terms and has a simpler


x
x  B form than (50).
  
0 !k (p) k (p)
B " G G y , p3S. (61) To illustrate, consider a nonlinear plant (1) with error
u 1 !k (p) k (p) B
B N N r signal z(t)"y(t)!r(t), where r is a reference input (separ-
B ately displayed for convenience), and y(t) is a measured
output with the same dimension as u(t). Suppose the
Writing out (60), remembering that in this example we objective is to have zero steady-state error for constant
are separately displaying the r-component of the general reference signals. That is, y (p)"r (p) for p3S. Suppose
measured output y, yields the simple condition  
also that the designed linear controller family for the
plant linearization family
*u (p) *x (p)
 !  "0, p3S. (62)

   
*p *p x A(p) B(p) x
B " B , p3S (65)
y C(p) 0 u
Obviously, x (p)"u (p) is a solution, and thus the gain B B
 
scheduled controller (50) becomes has the following form, a special case of (44), with slightly
di!erent notation,
x "k (y(t))[r!y], u"x#k (y(t))[r!y]. (63)
G N

 
x?


x ? F (p) 0 G (p) 0 B
It is easy to verify that indeed there are no hidden B ? ? x@
coupling terms in this controller. Computing the lin- x @ " 0 0 I !I B , p3S. (66)
B y
earized closed-loop system gives u H (p) H (p) 0 0 B
B ? @ r
B

     
x !4/3 1 x (1/3)/(1!p)
B " B # r , Clearly, this is a general type of integral-error controller,
x  !1 0 x 1/(1!p) B
B B and for this controller structure it is reasonable to as-
sume that both F (p) and H (p) are square and invertible
? @

x
y "[(1!p) 0] B . (64) for every p3S. We further assume that the time deriva-
B x tive of the plant measured output signal is available or is
B
computed. Then, rather than (50), consider a gain sched-
This linear system is stable regardless of the value of uled controller that uses y (t), but does not involve plant
p3(!1,1), and simulation of the nonlinear closed-loop equilibrium functions:
system (8), (63) indicates stability for constant reference
inputs over the entire operating range r3(!1,1). x ? "F (g(y, w))x? #G (g(y, w))y ,
? ?
Although in many cases the conditions (60) are restrict- x @ "H (g(y, w))x? #H (g(y, w))[y!r], u"x@ . (67)
? @
ive, there often is freedom to select whether hidden coup-
ling terms are due to an exogenous signal or an internal This controller has the equilibrium family
variable. Typically, the consequences of hidden coupling
x? (p)"0, y (p)"r (p), u (p)"x@ (p), (68)
terms in internal variables are more severe, since these     
represent feedback terms not accounted for in the linear since y (t) is zero at any equilibrium, and the correspond-
design process. (Lawrence & Rugh, 1995). ing controller linearization family is given by

 
x?


4.4. Velocity implementation x ? F (p) 0 0 0 B
B ? x@
x @ " H (p) 0 H (p) !H (p) B
The requirement that the linearization family of the B ? @ @ y
gain scheduled controller match the designed linear con- u 0 I 0 0 B
B r
troller family (no hidden coupling) can be relaxed, lead- B

 
ing to a linearization property comprising two conditions.
G (p)
These are that at each equilibrium the linearized closed- ?
loop, gain-scheduled system should have the same eigen- # 0 y , p3S. (69)
values and same transfer function as do the linear 0
closed-loop system composed of the linearized plant and
the designed linear controller (Kaminer, Pascoal, Khar- It is not di$cult to verify that the family of linear closed-
gonekar & Coleman, 1995). For a typical type of tracking loop systems (65), (66) and the linearized closed-loop
problem and integral-error controller, this leads to an system family (65), (69) are related by the coordinate
1414 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

transformation

 
I 0 0
P(p)" !F\(p)G (p)C(p) F\(p) 0 (70)
? ? ?
H\(p)H (p)F\(p)G (p)C(p) !H\(p)H (p)F\(p) H\(p)
@ ? ? ? @ ? ? @

and thus the linearization property is satis"ed. system satisxes the following property. If the exogenous
More general situations are treated in Kaminer et al. signals satisfy ""v (t)"", ""w (t)""(k, for t50, and if for some
(1995) and various additional features of this implemen- p3S,
tation are discussed. In addition this gain scheduling idea

   
has been used recently in the case where a nonlinear x(0) x (p)
!  (c, (71)
transformation of the plant achieves the property that x(0) x (p)

the plant linearization about an equilibrium (or trim-
ming) trajectory * a trajectory along which the dynamics then
are in equilibrium * is constant parameter (Kaminer,

   
x(t) x (g(y(t), w(t))
Pascoal, Hallberg & Silvestre, 1998). A well-known !  (o (72)
example of such a situation, where a nonlinear trans- x(t) x (g(y(t), w(t))

formation is not required, is the linearization about a cir-
cular orbit of a satellite in an inverse-square-law for t50.
gravitational "eld.
This o!ers some guidance on the stabilization proper-
ties of a gain scheduled controller that has no hidden
4.5. Stability properties
coupling terms, but the need for simulation studies to
assess stability and performance throughout the operat-
Based on standard results relating stability of a nonlin-
ing range remains.
ear system to stability of its linearization about an equi-
librium, we can state that if the linearized closed-loop
system at any "xed equilibrium p3S (implying that
exogenous signals are constant) is asymptotically stable, 5. LPV design methods
then the closed-loop nonlinear system is asymptotically
stable in a neighborhood of x (p), x (p). Indeed, if there The "rst step in a gain scheduled design is to derive
  a linear parameter-dependent description for the nonlin-
are no hidden coupling terms involving y , then design of
B ear plant (Section 2). Whether this representation is
a stabilizing linear controller family can be assumed to
guarantee asymptotic stability of the linearized closed- derived by Jacobian linearization or the quasi-LPV
loop system in a neighborhood of every p3S. Unfortu- approach, the "nal product is a family of linear systems
nately, this type of result provides rather small comfort parametrized by the scheduling variable, p. This moti-
since the essence of the gain scheduling approach is to vates the concept of a linear parameter-varying (LPV)
accommodate system operation across a range of operat- system, i.e., a linear system whose dynamics depend on
ing conditions, not in a small neighborhood of any one time-varying exogenous parameters whose trajectories
equilibrium. are unknown a priori but can be measured on-line by the
However, there is another stability result that can be controller.
used to address this situation. Essentially, the closed-loop We will now consider the LPV representation as our
system is not restricted to remain in a neighborhood of starting point, neglecting that it was originally derived
any single equilibrium, but is assumed to be slowly vary- from a nonlinear system. However, the interpretation of
ing and to have initial state su$ciently close to some the results of our analyses certainly depends on whether
equilibrium in S. Then the conclusion is that the closed- the Jacobian linearization or quasi-LPV approaches
loop system remains in a neighborhood of the equilib- underly the LPV representation.
rium manifold (Kelemen, 1986; Khalil & Kokotovic,
1991; Lawrence & Rugh, 1990). 5.1. Set-up

Theorem 12. For plant (1), suppose the gain-scheduled Consider an LPV plant of the form
controller (50) is such that there are no hidden coupling

   
x A(p(t))
B (p(t)) B (p(t)) x
terms and the eigenvalues of the linearized closed-  
loop system satisfy Re[j]4!e(0 for every p3S. z " C (p(t)) D (p(t)) D (p(t)) v . (73)
  
Then given o'0 there exist positive constants k and c y C (p(t)) D (p(t)) 0 u
such that the response of the nonlinear closed-loop  
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1415

The dynamics depend on the time-varying parameter to exhibit exponential stability for all admissible time-
vector, varying parameter trajectories satisfying (75)}(76), it
clearly must exhibit exponential stability for all constant
p(t)"[p (t) p (t) 2 p (t)]2 (74) parameter trajectories. In other words, for any xxed
  N
p 3S, the time-invariant system,
whose time variations are constrained by 

 
x x
p(t)3SLRN, (75) "A (p , 0)
 
(81)
x  x
p(t)3RLRN. (76) must be exponentially stable. However, stability of the
family of frozen parameter systems (81) need not imply
Eqs. (75) and (76) impose magnitude bounds and rate
stability of the time-varying systems (80).
bounds, respectively, on the admissible parameter vari-
ations. Motivation for introducing consideration of the
Example 13. Consider the `time-varying oscillatora,
parameter rate is provided by the stability analysis in
Section 5.2. Typical constraints are of the form
 
0 1
x " x, (82)
S"+p3RN : "p (t)"4p ,, R"+o3RN : "p (t)"4o , !(1#p(t)/2) !0.2
G G G G
(77) where !14p(t)41. This LPV system can be viewed as
a damped mass}spring system where the spring has
for pre-speci"ed bounds p and o , i"1,2, p.
G G a time-varying sti!ness coe$cient. Clearly for all con-
Our objective is to design an LPV controller that
stant p, (82) exhibits exponentially stable behavior. How-
measures y, p, and possibly p, to produce u so that the
ever, for the time-varying parameter p(t)"cos 2t, the
e!ect of the exogenous signal, v, on the error (penalized
dynamics become unstable. An intuitive explanation is
output) z, is minimized in an appropriate sense (to be
that if the spring is sti! while contracting and soft while
de"ned later). It is important to note that the LPV
expanding, then the mass oscillations will steadily grow.
dynamics (73) often describe a `generalizeda plant de-
scription that includes the actual system to be controlled
Example 13 demonstrates the well-known fact that the
as well as auxiliary dynamic weighting "lters that re#ect
location of frozen parameter eigenvalues do not necessar-
closed-loop performance criteria.
ily provide an indication of time-varying stability.
Combining an LPV controller of the form
A classical result from di!erential equation analysis is
x  that the time-varying system (80) is stable if the rates of

   
F(p(t), p(t)) G(p(t), p(t)) x
" (78) time variation in (76) are `su$ciently slowa. This su$-
u H(p(t), p(t)) E(p(t), p(t)) y cient condition is captured in the following theorem.
with the LPV plant (73) produces the LPV closed-loop
system dynamics Theorem 14. Suppose there exists X(p)"X2(p)'0 such
that

     
x x
A (p(t), p(t)) B (p(t), p(t)) N *X(p)
x    x (79)
C (p(t), p(t)) D (p(t), p(t)) X(p)A (p, o)#A2 (p, o)X(p)# o (0 (83)
z     G *p
v G G
for appropriately de"ned parameter-dependent matrices for all p3S and o3R. Then the unforced closed-loop system
A (p, p), B (p, p), C (p, p), and D (p, p). (80) is exponentially stable for all parameter trajectories
   
Note that we have exhibited controller dependence on satisfying (75)}(76).
parameter rates of variation as well as instantaneous
parameter values. Although it is often desirable to con- Under mild conditions, exponential stability implies
strain the controller to be independent of parameter deriv- bounded input/bounded output (BIBO) stability of the
atives, stability issues render this a di$cult objective. LPV closed-loop system (79).
Theorem 14 assures that
5.2. Stability analysis <(x, p)"x2X(p)x (84)
For the unforced closed-loop system is a Lyapunov function for the closed-loop system with
admissible time-varying parameter trajectories. Typi-

 
x x
"A (p(t), p(t)) (80) cally, determining whether there exists an X(p) satisfying
x   x (76) must be done numerically.
1416 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

5.3. LPV state-feedback stabilization The advantage over (88) is that (91) has an a$ne depend-
ence on the search variables, >(p) and EI (p). An important
In this section, we will use a simple example of state- implication is that the set of >(p) and EI (p) which satisfy
feedback stabilization to illustrate the main ideas used in (91) is convex. Eq. (91) is a particular case of a linear
LPV controller synthesis. matrix inequality (LMI). Computing solutions to LMIs
Consider an LPV plant of the special structure is a convex optimization problem, and e$cient software
is becoming increasingly available (Boyd, Ghaoui, Feron
x "(A #pA )x#Bu, (85)
  & Balakrishnan, 1994).
where the scalar parameter satis"es Because of the dependence on p, computing solutions
to (91) requires satisfying an in"nite family of LMIs.
!14p(t)41, !o4p(t)4o. (86) Therefore, a computational implementation will require
approximation and/or simpli"cation. Two approaches to
Important simplifying features are that the dynamics deriving a "nite set of LMIs are (1) discretizing the
matrix, A #pA , has an a$ne dependence on p, and
  parameter dependence and (2) imposing "xed structures
that the control matrix, B, is constant. on the search variables.
The objective is to "nd a state-feedback law u"E(p)x
so that the closed-loop LPV system
5.3.1. Parameter discretization
x "(A #pA #BE(p))x (87) By forming a grid of parameter values and using
 
a "nite di!erence scheme to approximate partial deriva-
is exponentially stable for all parameter trajectories sat-
tives, it is possible to approximately convert (91) to
isfying (86).
a "nite collection of LMIs. For this example, we can
Theorem 14 states that the closed-loop system is expo-
divide the interval [!1,1] into N intervals of width h.
nentially stable if there exists a X(p)"X2(p)'0 such
With a forward di!erence approximation of the deriva-
that
tive, (91) becomes the "nite collection of LMIs, indexed
X(p)(A #pA #BE(p)) by k"0,2, N!1, given by
 
*X(p) (A #(kh!1)A )>(kh!1)#BEI (kh!1)
#(A #pA #BE(p))2X(p)#o (0 (88)  
  *p
# >(kh!1)(A #(kh!1)A )2#EI (kh!1)2B2
for all "p"41 and "o"4o.  
This provides a basis for constructing the feedback $ o(>(kh!1#h)!>(kh!1))/h(0. (92)
gain E(p). The main idea is to numerically search for an
X(p) and E(p) which satisfy (88). One di$culty is that (88) Note that since the parameter rate of variation, o, enters
contains cross terms involving X(p) and E(p) which make linearly into (91), it is su$cient to evaluate o at its
a direct numerical search of a nonconvex optimization extreme values in order to satisfy (91) for all o.
problem. Once the collection of >(kh) and EI (kh) is found (if
It is possible to transform (88) to a more computation- possible), then a continuous >(p) and EI (p) can be formed
ally attractive form. First, note that (88) is equivalent to by interpolation. The left-hand side of (92) being su$-
ciently negative de"nite assures that (91) will be satis"ed
(A #pA #BE(p))X\(p)
  by the interpolated matrix function.
#X\(p)(A #pA #BE(p))2
 
*X\(p) 5.3.2. Fixed structure search
!o (0. (89) By imposing a "xed structure on >(p) and EI (p), it is
*p
possible to exactly convert (91) to a "nite collection of
Now perform the change of variables, LMIs. Let us impose

>(p)"X\(p), EI (p)"E(p)X\(p), (90) >(p)"> #p> , EI (p)"EI #pEI . (93)


   
so that the existence of an X(p) and E(p) satisfying (88) is
Then (91) becomes
equivalent to the existence of a >(p)">2(p)'0 and
EI (p) such that
(A #pA )(> #p> )#B(EI #pEI )
     
(A #pA )>(p)#BEI (p)#>(p)(A #pA )2
    #(> #p> )(A #pA )2
*>(p)    
#EI (p)2B2!o (0. (91)
*p #(EI #pEI )2B2!o> (0. (94)
  
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1417

Because the left-hand side is quadratic in p, constructing denote the closed-loop input/output relationship de"ned
solutions to (94) requires discretization over !14 by (79) with zero initial conditions. The induced norm is
p41. The term involving p is A > #> A2 , so sup- de"ned as
   
pose that we further impose the de"niteness condition
(z2(t)z(t) dt)
""¹ "" " sup sup  . (99)
A > #> A2 50. (95) XT G  (v2(t)v(t) dt)
    N     T 
Then the quadratic function of p is convex, and it su$ces The induced norm (99) generalizes to LPV systems the
to check that (94) is satis"ed at the extreme values of p. H norm of linear time-invariant systems in that it
The "nal set of LMIs is then (95) along with measures the largest possible ampli"cation of signal en-
ergy over all parameter trajectories satisfying (75)}(76).
(A #A )(> #> )#B(EI #EI )
      We will say that a controller achieves a performance
#(> #> )(A #A )2#(EI #EI )2B2 of c if the unforced closed-loop system (80) is exponen-
      tially stable for all admissible parameter trajectories
$o> (0, and if the induced norm of the closed-loop system (79)

(A !A )(> !> )#B(EI !EI ) satis"es
     
#(> !> )(A !A )2#(EI !EI )2B2$o> (0 ""¹ "" (c. (100)
       XT G 
(96) As in stability analysis, a su$cient condition for achiev-
ing closed-loop performance can be derived in terms of
and the de"niteness conditions the existence of a certain parameter-dependent positive-
> $> '0. (97) de"nite matrix.
 
Eq. (91) has been exactly converted to a "nite set of Theorem 15. A controller achieves a performance level of
LMIs, however, the conversion comes at the cost of c if there exists X(p)"X2(p)'0 such that

N *X(p)

 
X(p)A (p, o)#A2 (p, o)X(p)# o X(p)B (p, o) c\C2 (p, o)
  *p G  
B2 (p, o)X(p) G G !I c\D2 (p, o) (0 (101)
 
c\C (p, o) c\D (p, o) !I
 

increased conservatism. Imposing (95) and (96) reduces for all p3S and o3R.
the possibilities for "nding >(p) and EI (p) that satisfy (91).
This reduction further increases the potential conserva- Theorem 15 is an LPV generalization of the well-
tism of (91) which is only a su$cient condition for stabil- known bounded real lemma (e.g., Boyd et al., 1994) for
ity in the "rst place. linear time-invariant systems, which states that the norm
bound (100) is equivalent to the existence of a particular
quadratic storage function.
5.4. Performance analysis In the time-invariant case, condition (101) is both ne-
cessary and su$cient for achieving a performance level of
A useful measure of closed-loop performance is the c. The main sources of conservatism in the LPV version
induced norm of the mapping from the exogenous signals that reduce (101) to a su$cient condition are that the
to the error signals. The induced norm allows us to storage function which characterizes the norm bound
quantify closed-loop performance objectives such as (100) need not be quadratic, and that parameters
tracking error reduction, disturbance and noise rejection, and their time derivatives are treated as independent
control authority limitation, and robust stability. Certain quantities.
important criteria, in particular those dealing with transi-
ent phenomena, are not directly addressed with this kind 5.5. LPV control design
of performance measure. However, such criteria can be
indirectly addressed by the inclusion of weighting func- The analysis result of Theorem 15 is also the basis for
tions in the plant description (73). See Doyle, Wall and synthesizing controllers. It imposes a combined con-
Stein (1982) for further motivation and discussion of straint on the unknown controller matrices F, G, H, and
induced norms as performance measures. E and an unknown X (all parameter dependent). By
Let eliminating explicit dependence on controller matrices, it
is possible to convert (101) to an LMI form that will
z"¹Nv (98) allow a numerical search for a feasible X(p). The basic
XT
1418 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

procedure for controller construction is as follows: and the term B (p(t))v in the state dynamics is replaced
 
by B (p(t))M\v . The general case will follow from
(1) Transformation of (101) to convex optimization  
similar methods but with signi"cantly more complicated
problem. formulas.
(2) Numerical search.
(3) Construction of controller. 5.5.1. Transformation to convex optimization
Several of the main ideas in this procedure parallel The unknown quantities in (101) are the controller
those in Section 5.3 for constructing stabilizing state matrices and the de"nite matrix, X(p). It is possible to
feedback. The methods described here have been de- isolate the controller matrices in (101) by rewriting the
veloped by several researchers, including Apkarian and inequality as

   
Adams (1998), Apkarian and Gahinet (1995), Apkarian et E H E H 2
al. (1995), Becker and Packard (1994), Packard (1994), M(X)#¸(X) R2(X)#R(X) ¸2(X)(0
G F G F
Scherer, Gahinet and Chilali (1997), Wu, Yang, Packard (103)
and Becker (1996) and Yu and Sideris (1997). Our treat-
ment follows that of Balas et al. (1997). for appropriately de"ned M(X), ¸(X), and R(X), where
In order to simplify matters, we will assume that the explicit dependence on p and o is omitted. Let ¸ (X) and
,
LPV plant (73) has the form R (X) be X-dependent matrices whose columns are the
,
orthogonal complements of the columns of ¸(X) and

   
x A(p(t))
B (p(t)) B (p(t)) B (p(t)) x R(X), respectively. It turns out that (103) is equivalent to
   [Boyd et al., 1994, Section 2.6.2]
z C (p(t)) 0 0 0 v
 "   .
¸ (X)2M(X)¸ (X)(0, R (X)2M(X)R (X)(0.
z

C (p(t))

0 0 I v
 , , , ,
y C (p(t)) 0 I 0 u (104)

(102) These ideas lead to the following alternative character-
ization of (101).
This form assumes no direct feedthrough from exogenous
signals to penalized outputs and assumes that the control Theorem 16. There exists an X(p)"X2(p)'0 that satisxes
(101) if and only if there exist X (p) and > (p) such that
penalty and measurement noise have been normalized.  


N *X (p)


AI 2(p)X (p)#X (p)AI (p)#  o !cC2 C X (p)B (p) C2 (p)
  *p G     
G G
B2 (p)X (p) !cI 0 (0, (105)
 
C (p) 0 !cI



N *> (p)


AK 2(p)> (p)#> (p)AK (p)!  o !B (p)B2 (p) > (p)C2 (p) B (p)
  *p G     
G G
C (p)> (p) !cI !0 (0, (106)
 
C (p) 0 !cI


 
X (p) c\I
 '0 (107)
For example, suppose the plant measurement is c\I > (p)

y"C (p(t))x#Mv . for all p3S and o3R, where
 
AI (p)"A(p)!B (p)C (p),
Assuming that M is square and invertible, we can de"ne  
the new variable AK (p)"A(p)!B (p)C (p). (108)
 
The connection between X (p) and > (p) in The-
v "Mv  
  orem 16 and X(p) in (101) is that X(p) and X\(p) can be
partitioned as
in which case the measurement equation becomes

   
X (p) * > (p) *
X(p)"  , X\(p)"  . (109)
y"C (p(t))x#v * * * *
 
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1419

Eqs. (105)}(107) now represent LMI constraints on un- parameter, it is possible to perform a p-dependent change
known quantities X (p) and > (p). of variables to remove p dependence as follows (Balas et
 
al., 1997). In the scalar parameter case, the controller
5.5.2. Numerical search dynamics from Theorem 17 take the form
Because of the underlying LMI structure, the set of
x "F (p)x#pF (p)x#G(p)u. (112)
matrix functions X (p) and > (p) which satisfy  
 
(105)}(107) is convex. However, constructing solutions to Now consider the change of variables
(105)}(107) is still an in"nite-dimensional optimization
problem because a solution is required for all p3S and x "¹(p)x,

(113)
o3R.
where ¹(p) is invertible. It is easy to show that if ¹(p)
As in the state-feedback stabilization problem, it is
satis"es the linear matrix di!erential equation
possible to replace (105) and (106) with a "nite collection
of LMIs through either parameter discretization or d¹(p)
imposing "xed structures on the search variables. These "!¹(p)F (p), (114)
dp 
issues are addressed in Apkarian and Adams (1998) and
Zhou et al. (1996). then the resulting controller dynamics in the new co-
ordinates, x , will be independent of p. The multi-

5.5.3. Controller construction parameter case seems to require partial di!erential equa-
It is possible to construct a controller directly from tions whose analytical solvability is questionable.
X (p) and > (p) that satis"es the conditions of The- An alternative is to simply restrict the choice of > to
  
orem 16. The following result is from Balas et al. (1997). be independent of p. However, this approach limits the
possible solutions of (105)}(107), and may be conserva-
Theorem 17. Let X (p) and > (p) satisfy the conditions tive.
 
of Theorem 16. Then a controller that achieves a perfor- Finally, the issue of controller discretization for imple-
mance level of c is given by mentation is considered in Apkarian (1997).

F(p, p)"A(p)#c\(Q\(p)X (p)M(p)B2 (p)


  5.6. LFT design methods
#B (p)B2 (p))>\(p)#B (p)<(p)
    In the special case of LFT parameter dependence, it is
#Q\(p)X (p)M(p)C (p)#Q\(p)=(p, p), possible to construct a gain scheduled controller by solv-
 
ing LMI feasibility problems that are less complex than
G(p)"!Q\(p)X(p)M(p), H(p)"<(p), the general LPV case. The required computational bur-
E(p)"0, (110) den is comparable to the standard linear time-invariant
case. This simpli"cation comes at the cost of several
where drawbacks:
Q(p)"X (p)!>\(p),
  E The LFT representation assumption can be restrictive.
E Parameter rate-of-variation bounds are not exploited.
<(p)"!cB2 (p)>\(p)!C (p), E Robustness to parameter variations is achieved by
  
M(p)"!cX\(p)C2 (p)!B (p), grossly overestimating the e!ects of admissible para-
   meter trajectories.
A (p)"A(p)#B (p)#<(p),
4  The implications of these drawbacks is that the resulting

 
C (p) gain scheduled controller can be conservative in that it
C (p)"  ,
4 C (p)#<(p) seeks to achieve the desired performance not only for
 system dynamics with admissible parameter trajectories
=(p, p)"A2 (p)>\(p)#>\(p)A (p) satisfying (77), but also for system dynamics that can
4   4
exhibit a much broader range of behaviors.
#c\C2 (p)C (p)
4 4 An LPV plant with LFT parameter dependence can be
# c\>2 (p)B (p)B2 (p)>2 (p) written in the state-space form
   

   
N *>\(p) x A B B B x
#  p. (111) N  
*p z C D D D v
G N " N NN N N N . (115)
z C D D D v
Note that the controller dynamics matrix, F(p, p), de-  N  
pends explicitly on p. In case of a single scheduling y C D D 0 u
 N 
1420 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

Here z and v are `arti"ciala signals related by


N N
v "*(p)z , (116)
N N
where * takes the diagonal form of (40).
We will impose a "xed structure on the controller
which mimics the LFT parameter dependence of the
plant, namely,

   
x  F G G x
 N
u " H E E y (117)
 N
u H E E yN
N N N NN
where u and y are related by
N N
u "*(p)y . (118)
N N
It is important to note that in the above plant and Fig. 3. Augmented LTI plant for control design.
controller descriptions, all p dependence is contained in
*(p) * all other matrices are constant.
The design procedure is to temporarily view the para-
meter variations as `unknown perturbationsa on an Fig. 3 shows the modi"ed block diagram for control
otherwise time-invariant system and use time-invariant design. The block P represents the augmented plant

optimization methods along with `structureda small- used for control design. The twice-repeated block * rep-
gain analysis to construct the time-invariant portion of resents parameter time variations which will be viewed as
the compensator. After the design, the parameter vari- structured unknown perturbations. We will assume that
ations are once again incorporated to produce an LPV the parameter magnitude bounds have been normalized
controller. These ideas have been developed in Apkarian to
and Gahinet (1995), Packard (1994) and Packard, Zhou,
Pandey, Leonhardson and Balas (1992). The treatment "p "41. (120)
G
here follows Apkarian and Gahinet (1995).
Fig. 2 shows the original block diagram. The blocks Our time-invariant control objective is as follows:
P and K represent the time-invariant portion of the
*2' *2'
LPV plant, P , and controller, K , respectively, and Construct a time-invariant controller, K , for
*.4 *.4 *2'
the LFT relationships are the time-invariant plant, P , such that for any

perturbation, *(p), with p( ) ) satisfying (120), the
P "F (P , *), K "Fl (K , *). (119) closed-loop system satis"es the induced-norm condi-
*.4  *2' *.4 *2'
tion (100).

The control objective can be restated as "nding


a time-invariant K such that
*2'

  
* 0
F Fl (P , K ), (c. (121)
  *2' 0 *
G 
A su$cient condition for the control objective can be
obtained by applying structured small gain analysis
(Zhou et al., 1996, Chapter 11). De"ne

    
* 0 * 0
L" ¸'0 : ¸ " ¸ . (122)
0 * 0 *

The set L comprises symmetric, positive-de"nite ma-


trices that commute with the structured unknown per-
turbations. This class can be used to derive the following
Fig. 2. LFT plant and controller. su$cient condition for the control objective.
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1421

Theorem 18. The controller K achieves (121) if there This plant can be put in LFT form as follows. Recall
*2'
exists an ¸3L such that from Example 7 that

   
M M p 0

   
¸ 0 ¸\ 0 1!p"F   , (125)
Fl (P ,K ) (1. (123) M M 0 p
0 c\  *2' 0 I  
G 
for appropriate M , M , M , and M . Then (124) can be
We are now in a position to employ LMI methods    
rewritten in the form of (115) as
for control design. The procedure follows that taken in
x "!x#u, e "!k e#k v!k (M v #M x),
Section 5.5:     N 
z "M v #M x, z "k e, z "u,
N  N    
(1) Express (123) as an LMI involving the state-space y"v!(M v #M x), (126)
matrices of K , ¸, ¸\, and some unknown de"-  N 
*2' where v and z are related by
nite matrix, X. N N
(2) Manipulate the resulting LMI to remove explicit
 
p 0
dependence on the controller matrices. v " z . (127)
N 0 p N
(3) Compute a feasible X, ¸, and ¸\.
(4) Construct a controller. Now to construct an LFT controller, we introduce the
2;1 vector signals u and y as in Fig. 3. The new time-
The small-gain condition of (123) involves scaling by the N N
invariant control problem corresponds to the generalized
multipliers ¸ and ¸\. Generally speaking, LMIs asso- plant

   
  
x 0
   
!1 0 0 0 1 0
e x
!k M !k 0 !k M k
      0 0

      
e
u 0 0 0


N 0 0 0 I y
z 0 M 0 N
N M 0  0 0
"  v
N . (128)
z 0 k 0 0 0 0 0
  v
z 0 0 0 0 0 1 0


    
0 !M 1 u

  
y !M 0  0 0
 u
y 0 0 I 0 0 N
N 0 0

ciated with multipliers involve terms containing both Note that the controller, K , will have three inputs
*2'
¸ and ¸\, and hence the resulting feasibility search is and outputs. Imposing the LFT relationship
not convex. However, the special structure of this prob-

 
p 0
lem is such that the resulting feasibility search is convex. y " u (129)
The details are omitted here, but may be found in Apkar- N 0 p N
ian and Gahinet (1995). produces the desired single-input/single-output LPV
Example 19. Consider the synthesis problem associated controller (Fig. 2).
with the LPV plant
6. Stability and performance
x "!x#u, e "!k e#k (v!(1!p)x),
 
z "k e, z "u, y"v!(1!p)x. (124) By the nature of the design process, a gain scheduled
   controller is constructed to exhibit desirable stability and
In this set-up, v is a reference command input, (1!p)x performance properties provided that the system is oper-
is the tracking output; e is a "ltered version of the ating in the vicinity of a design operating condition.
tracking error, z and z penalize the tracking error and However, a gain scheduled design must also perform well
 
control input, and y, the tracking error, is also the mea- for other operating conditions as well as during
sured output. The parameters k and k determine the transitions among operating conditions.
 
"lter bandwidth and relative weighting for the tracking The stability and performance properties of a gain
error, respectively. scheduled design are typically inferred from extensive
1422 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

simulations. It is important to note that any such proper- 1986; Khalil & Kokotovic, 1991; Lawrence & Rugh,
ties are not necessarily assured by the stability and perfor- 1990). Though scheduling variable variations are re-
mance properties of the "xed operating condition quired to be slow and initial conditions are restricted to
designs. A similar phenomenon was seen in the time- be close to the equilibrium surface, large variations in the
varying oscillator, Example 13, where a system which is scheduling variable are possible so long as deviations
stable for constant parameter trajectories becomes un- from the equilibrium surface are not large. Once again
stable for time-varying parameter trajectories. the ideas behind the heuristic guidelines can be readily
Two heuristic guidelines for a gain scheduled design interpreted in the context of Theorem 12.
are that the scheduling variable should `vary slowlya and Finally, it is important to note that the LPV and LFT
`capture plant nonlinearitiesa. It is possible to interpret approaches in Section 5 represent a departure from tradi-
these guidelines in terms of the gain scheduling design tional gain scheduling in that they address parameter
process itself. Namely, by considering only linearizations time variations during the design process itself, as op-
about "xed operating conditions, a traditional gain posed to doing analysis after the fact. By doing so, these
scheduled design neglects future transitions among oper- methods can potentially alleviate the restriction to slow
ating conditions and neglects certain second-order non- variations.
linearities. The heuristic guidelines simply state that as
long as the e!ects of these neglected items are not signi"-
cant, then one should expect stability and performance of 7. Current directions
the overall gain scheduled design.
There have been relatively few research papers on the As mentioned in Section 6, a major limitation of lin-
stability and performance of gain scheduled systems. earization gain scheduling is that stability properties of
A complicating factor is that although gain scheduled the closed-loop system can be guaranteed only in
designs are based on linearizations, the overall system is a neighborhood of the equilibrium manifold and under
still nonlinear. This essentially prohibits the possibility an assumption of slow variation of signals. Recently,
of nonconservative stability and performance analysis a notion of `extended local linear equivalencea has been
results. proposed to address the problem (Leith & Leithead,
Kamen and Khargonekar (1984) and Shahruz and 1996). However, the theoretical basis of claims regarding
Behtash (1992) consider LPV systems with the focus on this notion are somewhat controversial as indicated by
constant parameter trajectories. Kamen and Khar- the exchange in Leith, Leithead, Lawrence and Rugh
gonekar (1984) addresses achievable continuity and (1998).
smoothness properties of LPV controllers. Shahruz and Because the performance of linearization scheduled
Behtash (1992) derives methods to assure constant para- system typically deteriorates as exogenous reference and
meter stability at a continuum of parameter values based scheduling signals vary more rapidly, attempts have been
on a "nite collection of designs at discrete parameter made to use information about these signals in addition
values. to their instantaneous values. Early attempts based on
As for LPV systems with varying parameter trajecto- using the history of the scheduling variable (Shamma
ries, the fact that "xed parameter stability implies stabil- & Cloutier, 1992), or the time derivative of the scheduling
ity for slow parameter variations (which can inferred variable (Sureshbabu & Rugh, 1995), met with limited
from Theorem 14) is a classical di!erential equations success. Of course, the LMI approach described in Sec-
result (Desoer, 1969). These ideas were extended in tion 5 directly addresses the issue by permitting control-
Shamma and Athans (1991) to incorporate robust perfor- ler dependence on the derivative of the scheduling signal.
mance in the presence of unmodeled dynamics. It is interesting to speculate on the possibility of new
Shamma and Athans (1990) considers quasi-LPV sys- approaches that use additional information.
tems and derives (conservative) su$cient conditions for An important question is whether linearization sched-
the robust stability and performance of gain scheduled uling can be made to better address the nonlinear control
designs. These conditions also require that the scheduling context in which it is used. For example, can the various
variable varies su$ciently slowly. In case the system linear controller point designs be formulated so that they
cannot be brought to a quasi-LPV form, Shamma and contribute to an overall nonlinear objective? As matters
Athans (1990) shows that stability and performance of stand, the point designs are not related to each other in
the gain scheduled design still holds as long as nonlin- any speci"c way and might be, in some sense, con#icting.
earities in the state variables other than the scheduling An approach to this issue in the context of linear-quad-
variables are small. This result provides some quantitat- ratic point designs and a nonlinear overtaking optimal
ive interpretation of the guideline of `capturing plant control objective can be found in Tan and Rugh (1998).
nonlinearitiesa. This work shows that the linear-quadratic point designs
As previously discussed, Theorem 12 provides a stabil- should be related in a certain way so that at each equilib-
ity condition for linearization gain scheduling (Kelemen, rium the gain scheduled control approximates a so-called
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1423

weak overtaking optimal control. Much remains to be lems other than gain scheduling, e.g., controller satura-
done in assessing the e!ectiveness of such an approach. tion (Wu & Grigoriadis, 1999).
Along these lines of coordinated gain scheduled con- Taking a brief look at the future, we note again that the
trol designs is the idea of `stability-based switchinga. increasing number of papers reporting successful imple-
Recent work (Leonessa, Haddad & Chellaboina, 1988; mentations documents the utility of gain scheduling in
McConley, Appleby, Dahleh & Feron, 1997) considers practice. Reasons for this utility include the basis in
scheduling the controller transitions based on stability well-developed linear control theory, room for incorpo-
properties of the linear point designs. The idea is to select ration of physical intuition in the gain scheduling meth-
operating conditions and to perform linear point designs odology, and the availability of software for the requisite
so that each operating condition is in the `capture re- computations. While future research directions are di$-
giona of a neighboring operating condition. In this man- cult to predict, the breadth of contemporary applications
ner, one can assure stable transitions among operating of gain scheduling will provide a strong driving force for
conditions by appropriately switching controllers. theoretical inquiry and a ready test bed for new ideas.
The topic of state-dependent Riccati equations (SDRE)
has connections with LPV approaches to gain scheduling
(Cloutier, D'Souza & Mracek, 1996a,b). Suppose plant Acknowledgements
(1) includes no exogenous signals and can be written in
the form The authors are grateful to David J. Yost, Walter K.
Waymeyer, Stephen S. Osder, and H. Max Davis for
x "A(x)x#B(x)u, z"C (x)x#D (x)u, providing information on early practices in #ight control,
 
y"x. (130) and to Je!rey A. Cook for historical information on
engine control. Also the authors thank Andrew Packard
Here the coe$cient matrices are functions of the plant for providing the workshop notes (Balas et al., 1997) and
state, and typically they are highly nonunique. With Douglas A. Lawrence for suggesting the PID example.
p"y this is an LPV representation of the form (5). The
focus is on optimality with respect to minimization of the
performance index References


 Apkarian, P. (1997). On the discretization of LMI-synthesized linear
J" x2Q(x)x#u2R(x)u dt. (131) parameter-varying controllers. Automatica, 33(4), 655}661.
 Apkarian, P., & Adams, R. (1998). Advanced gain-scheduling tech-
This leads to consideration of the state-dependent Riccati niques for uncertain systems. IEEE Transactions on Control Systems
equation Technology, 6(1), 21}32.
Apkarian, P., & Gahinet, P. (1995). A convex characterization of gain-
A2(x)P(x)#P(x)A(x)!P(x)B(x)R\(x)B2(x)P(x) scheduled H controllers. IEEE Transactions on Automatic Control,

AC-40(5), 853}864.
#Q(x)"0, (132) Apkarian, P., Gahinet, P., & Becker, G. (1995). Self-scheduled
H-in"nity control of linear parameter-varying systems: A design
the appropriate solution of which gives the control as example. Automatica, 31(9), 1251}1261.
As stroK m, K. J., & Wittenmark, B. (1995). Adaptive control (2nd ed.).
u"!R\(x)B2(x)P(x)x. (133) Reading, MA: Addison-Wesley.
Bailey, R. L., Cederquist, A. L., Florek, J. J., Hart, D. L., & Meitzler,
In a sense, this control can be viewed as being gain A. H. (1978). An IIEC-2 low-emission concept car. SAE Paper
scheduled on the plant state. The approach appears to No. 780206, Society of Automotive Engineers.
have worked well in some examples (Cloutier et al., Balas, G., Fialho, I., Lee, L., Nalbantoglu, V., Packard, A., Tan, W.,
1996a,b), but the theoretical aspects are complex. It is Wemho!, E., Wolodkin, G., & Wu, F. (1997). Theory and applica-
tion of linear parameter varying control techniques. Workshop
shown in Huang and Lu (1996) that nonlocal optimality notes: American Control Conference, June.
is achieved only for a particular choice of state- Bamieh, B., & GiarreH , L. (1999). Identi"cation of linear parameter
dependent plant representation, and this choice requires varying models. Proceedings of the 38th IEEE conference on decision
knowledge of the optimal value function (solution of and control, Phoenix, AZ, December.
a Hamilton}Jacobi equation). Becker, G., & Packard, A. (1994). Robust performance of linear
parametrically varying systems using parametrically-dependent
Research in LPV control design has been focused linear feedback. Systems & Control Letters, 23, 205}215.
primarily on computational approaches as in Section 5 Bequette, W., & Doyle, F. J. (2000). A review of gain scheduled control
to construct quadratic parameter-dependent Lyapunov in chemical process systems, submitted for publication.
functions. The computational construction of nonquad- Boyd, S., El Ghaoui, L., Feron, E., & Balakrishnan, V. (1994). Linear
ratic Lyapunov functions is considered in Shamma and matrix inequalities in system and control theory. Philadelphia, PA:
SIAM.
Xiong (1999) and Tu and Shamma (1998). Other emerg- Buschek, H. (1997). Robust autopilot design for future missile systems.
ing directions are identi"cation of LPV systems (Bamieh Proceedings of the AIAA guidance, navigation, and control conference,
& GiarreH , 1999) and utilization of LPV tools for prob- New Orleans, LA. AIAA Paper No. A97-37172.
1424 W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425

Canale, R. P., Winegarden, S. R., Carlson, C. R., & Miles, D. L. (1978). Lawrence, D. A., & Rugh, W. J. (1995). Gain scheduling dynamic linear
General motors phase II catalyst system. SAE Paper No. 780205, controllers for a nonlinear plant. Automatica, 31(3), 381}390.
Society of Automotive Engineers. Leith, D. J., & Leithead, W. E. (1996). Appropriate realization of
Cloutier, J. R., D'Souza, C. N., & Mracek, C. P. (1996a). Nonlinear gain-scheduled controllers with application to wind turbine regula-
regulation and nonlinear H-in"nity control via the state-dependent tion. International of Control, 65(2), 223}248.
Riccati equation technique, Part 1, theory. Proceedings of the xrst Leith, D. J., Leithead, W. E., Lawrence, D. A., & Rugh, W. J. (1998).
international conference on nonlinear problems in aviation and aero- Comments on Gain scheduling dynamic linear controllers for
space. Daytona Beach, FL (pp. 117}130). a nonlinear plant. Automatica, 34(8), 1041}1043.
Cloutier, J. R., D'Souza, C. N., & Mracek, C. P. (1996b). Nonlinear Leonessa, A., Haddad, W. M., & Chellaboina, V. (1988). Nonlinear
regulation and nonlinear H-in"nity control via the state-dependent system stabilization via stability-based switching. Proceedings of
riccati equation technique, Part 2, examples. Proceedings of the xrst the 37th IEEE conference on decision and control, Tampa, FL
international conference on nonlinear problems in aviation and aero- (pp. 2983}2996).
space, Daytona Beach, FL (pp. 130}147). Masaki, K., Aono, S., Akaeda, M., & Minami, H. (1978). Development
Cook, J. A., Grizzle, J. W., & Sun, J. (1996). Engine control. In W.S. of the Nissan electronically controlled carburetor system. SAE
Levine, The control handbook (pp. 1261}1274) New York: IEEE Paper No. 780204, Society of Automotive Engineers.
Press (Chapter 74.1). McConley, M. W., Appleby, B. D., Dahleh, M. A., & Feron, E. (1997).
Desoer, C. A. (1969). Slowly varying system x "A(t)x. IEEE Trans- Control Lyapunov function approach to robust stabilization of
actions on Automatic Control, AC-14(6), 780}781. nonlinear systems. Proceedings of the American control conference,
Doyle, J. C., Wall, J. E., & Stein, G. (1982). Performance and robustness Albuquerque, NM, June (pp. 416}419).
analysis for structured uncertainty. Proceedings of the 21st IEEE McRuer, D., & Graham, D. (1981). Eighty years of #ight control:
conference on decision and control, December (pp. 629}636). triumphs and pitfalls of the systems approach. Journal of Guidance
Draper, C. S. (1955). Flight control. Journal of the Royal Aeronautical and Control, 4(4), 353}362.
Society, 59, 451}477. Nichols, R. A., Reichert, R. T., & Rugh, W. J. (1993). Gain scheduling
Gregory, P. C. (Ed.) (1959). Proceedings of the self adaptive yight control for H-in"nity controllers: A #ight control example. IEEE Transac-
systems symposium, Wright-Patterson AFB. OH. WADC Technical tions on Control Systems Technology, 1(2), 69}79.
Report 59-49. Niemann, H., & Stoustrup, J. (1999). An architecture for implementa-
Gumbleton, J. J., & Bowler, L. L. (1982). General motors computer tion of multivariable controllers. Proceedings of the American control
command control system development. SAE Paper No. 82091, conference, San Diego, CA, June (pp. 4029}4032).
Society of Automotive Engineers. Osder, S. S. (1996). Evolution of automatic #ight controls. Material from
Howard, R. W. (1973). Automatic #ight controls in "xed wing aircraft: a short course at Honeywell Air Transport Division.
The "rst 100 years. Aeronautical Journal, 77, 533}562. Packard, A. (1994). Gain-scheduling via linear fractional transforma-
Huang, Y., & Lu, W.-M. (1996). Nonlinear optimal control: Alterna- tions. Systems and Control Letters, 22, 79}92.
tives to Hamilton}Jacobi equation. Proceedings of the 35th Confer- Packard, A., Zhou, K., Pandey, P., Leonhardson, J., & Balas, G. (1992).
ence on Decision and Control, Kobe, Japan (pp. 3942}3947). Optimal, constant I/O similarity scaling for full-information and
Hyde, R. A., & Glover, K. (1993). The application of scheduled state-feedback control problems. Systems & Control Letters, 19,
H-in"nity controllers to a vstol aircraft. IEEE Transactions on Auto- 217}280.
matic Control, 38(7), 1021}1039. Paddison, F. C. (1982). The Talos control system. Johns Hopkins APL
KaK llstroK m, C. G., As stroK m, K. J., Thorell, N. E., Eriksson, J., & Sten, L. Technical Digest, 3(2), 154}156.
(1979). Adaptive autopilots for tankers. Automatica, 15, 241}254. Reichert, R. (1992). Dynamic scheduling of modern-robust-control au-
Kamen, E. W., & Khargonekar, P. P. (1984). On the control of linear topilot designs for missiles. IEEE Control Systems Magazine, 12(5),
systems whose coe$cients are functions of parameters. IEEE Trans- 35}42.
actions on Automatic Control, AC-29(1), 25}33. Rivard, J. G. (1973). Closed-loop electronic fuel injection control for
Kaminer, I., Pascoal, A., Hallberg, E., & Silvestre, C. (1998). Trajectory the internal combustion engine. SAE Paper No. 73005, Society of
tracking for autonomous vehicles: An integrated approach to guid- Automotive Engineers.
ance and control. Journal of Guidance, Control, and Dynamics, 21(1), Rugh, W. J. (1983). Linearization about constant operating points: An
29}38. input}output viewpoint. Proceedings of the IEEE conference on
Kaminer, I., Pascoal, A. M., Khargonekar, P. P., & Coleman, E. (1995). decision and control, San Antonio, TX (pp. 1165}1169).
A velocity algorithm for the implementation of gain-scheduled Rugh, W. J. (1991). Analytical framework for gain-scheduling. IEEE
controllers. Automatica, 31(8), 1185}1191. Control Systems Magazine, 11(1), 79}84.
Kelemen, M. (1986). A stability property. IEEE Transactions on Auto- Scherer, C., Gahinet, P., & Chilali, M. (1997). Multiobjective output-
matic Control, 31(8), 766}768. feedback control via LMI optimization. IEEE Transactions on
Kellett, M. G. (1991). Continuous scheduling of H-in"nity controllers Automatic Control, AC-42(7), 896}911.
for a ms760 paris aircraft. In P.H. Hammond, Robust control system Seiter, R. E., & Clark, R. J. (1978). Ford three-way catalyst and feedback
design using H-inxnity and related methods (pp. 197}223). London, fuel control system. SAE Paper No. 780203, Society of Automotive
England: Institute of Measurement and Control. Engineers.
Kelly, J. H., & Evers, J. H. (1997). An interpolation strategy for schedul- Shahruz, S. M., & Behtash, S. (1992). Design of controllers for linear
ing dynamic compensators. Proceedings of the AIAA guidance, navi- parameter-varying systems by the gain scheduling technique. Jour-
gation, and control conference, New Orleans, LA. AIAA Paper No. nal of Mathematical Analysis and Applications, 168(1), 195}217.
A97-37173. Shamma, J. S., & Athans, M. (1990). Analysis of nonlinear gain
Khalil, H. K. (1996). Nonlinear systems (2nd ed.). Upper Saddle River, scheduled control systems. IEEE Transactions on Automatic
NJ: Prentice-Hall. Control, AC-35(8), 898}907.
Khalil, H. K., & Kokotovic, P. V. (1991). On stability properties of Shamma, J. S., & Athans, M. (1991). Guaranteed properties of gain
nonlinear systems with slowly varying inputs. IEEE Transactions on scheduled control of linear parameter-varying plants. Automatica,
Automatic Control, 36(2), 229. 27(3), 898}907.
Lawrence, D. A., & Rugh, W. J. (1990). On a stability theorem for Shamma, J. S., & Cloutier, J. R. (1992). Trajectory scheduled missile
nonlinear systems with slowly varying inputs. IEEE Transactions on autopilot design. Proceedings of the xrst IEEE conference on control
Automatic Control, 35(7), 860}864. applications, Dayton, OH, September.
W.J. Rugh, J.S. Shamma / Automatica 36 (2000) 1401}1425 1425

Shamma, J. S., & Xiong, D. (1999). Set-valued methods for linear Wilson J. Rugh received the B.S. degree in
parameter varying systems. Automatica, 35(6), 1081}1089. electrical engineering from Penn State
University, University Park, PA, in 1965,
Stein, G., Hartmann, G. L., & Hendrick, R. C. (1977). Adaptive control
and the M.S. and Ph.D. degrees in electri-
laws for F-8 #ight test. IEEE Transactions on Automatic Control, cal engineering from Northwestern Uni-
22(5), 758}767. versity, Evanston, IL, in 1967 and 1969,
Stilwell, D. J., & Rugh, W. J. (1999). Interpolation of observer state respectively. Since 1969 he has been on the
feedback controllers for gain scheduling. IEEE Transactions on faculty of Johns Hopkins University, cur-
Automatic Control, 44(6), 1225}1229. rently as the Edward J. Schaefer Professor in
Stilwell, D. J., & Rugh, W. J. (2000). Stability preserving interpolation the Department of Electrical and Computer
methods for the synthesis of gain scheduled controllers. Automatica, Engineering. He is the author of numerous
36(5), 665}671. research articles and three books * most
recently `Linear System Theorya, 2nd ed, Prentice Hall, 1996.
Sureshbabu, N., & Rugh, W. J. (1995). Output regulation with deriva-
Dr. Rugh has been an associate editor of the IEEE Transactions on
tive information. IEEE Transactions on Automatic Control, 40(10), Automatic Control, and currently is an associate editor of Mathematics
1755}1766. of Control, Signals, and Systems, and Systems & Control Letters. In
Tan, H., & Rugh, W. J. (1998). Nonlinear overtaking optimal control: 2000 he serves as President-Elect of the IEEE Control Systems Society.
Su$ciency, stability, and approximation. IEEE Transactions on Au-
tomatic Control, 43(12), 1703}1718.
Tu, K.-H., & Shamma, J. S. (1998). Nonlinear gain-scheduled control Je4 S. Shamma was born in New York,
design using set-valued methods. Proceedings of the American con- NY, in 1963, and raised in Pensacola, FL.
trol conference, Philadelphia, PA. He received the Ph.D. degree in 1988 from
the Massachusetts Institute of Technology,
Waymeyer, W. K. (1997). History information on gain scheduling. Department of Mechanical Engineering.
Personal communication. He has held faculty positions at the Uni-
Wu, F., & Grigoriadis, K. M. (1999). LPV-based control of systems with versity of Minnesota, Minneapolis, and the
amplitude and rate actuator saturation constraints. Proceedings of University of Texas, Austin. Since 1999, he
the American control conference, San Diego, CA, June (pp. has been with the University of California,
3191}3195). Los Angeles, where he is currently a
Wu, F., Yang, X. H., Packard, A., & Becker, G. (1996). Induced Professor of Mechanical and Aerospace
L -norm control for LPV systems with bounded parameter Engineering. His main research interest is
 nonlinear control theory and design with applications to mechanical,
variation rates. International Journal of Control, 6(9/10),
aerospace, and manufacturing systems. He has served on the editorial
983}998. boards of the IEEE Transactions on Automatic Control and Systems
Yu, J., & Sideris, A. (1997). H control with parametric Lyapunov & Control Letters. He is a recipient of a 1992 NSF Young Investigator

functions. Systems & Control Letters, 30, 57}69. Award and the 1996 Donald P. Eckman Award of the American
Zhou, K., Doyle, J. C., & Glover, K. (1996). Robust and optimal control. Automatic Control Council, and was a Plenary Speaker at the 1998
Upper Saddle River, NJ: Prentice-Hall. American Control Conference.

Anda mungkin juga menyukai