condition.
MOTIVATION Each individual controller is expected to achieve good per-
formance whenever the nonlinear plant is near the control-
A classical tradeoff in control design is model accuracy versus lers associated operating condition. As the plant varies from
model simplicity. While sophisticated models better represent one operating condition to another, the gains of the individual
a physical systems behavior, the resulting analysis and con- controllers are interpolated, or scheduled, to match the
trol design are more involved. changes in operating conditions. The nal result is a nonlin-
One manifestation of this tradeoff is the use of nonlinear ear controller which is constructed out of several local linear
versus linear models. Most physical systems exhibit nonlinear controllers.
behavior. Some common examples are saturations, rate limit- The implementation of a gain scheduled controller is de-
ers, hysteresis, and backlash. Predominantly nonlinear be- picted in Fig. 1. An auxiliary variable, usually called the
havior may be found in robotic manipulator dynamics, air- scheduling variable, is used to update the gains of the linear
craft or missile ight dynamics, undersea vehicle dynamics, controller. The scheduling variable should be a good indica-
jet engine combustion dynamics, and satellite attitude dy- tion of the current operating condition of the plant, and hence
namics. Although analysis and control design for nonlinear should be correlated with the plant nonlinearities. The sched-
systems remains an active topic of research, analysis of linear uling variable can be a combination of endogenous signals,
systems is signicantly less complicated, and there is an such as a plant measurements, or exogenous parameters
abundance of control design methodologies, ranging from which reect environmental conditions.
classical control to multivariable robust control. One example is missile autopilot design. Useful scheduling
One compromise is to linearize the system behavior, that variables are the angle-of-attack and dynamic pressure, both
is, approximate the behavior of a nonlinear system near a of which characterize the aerodynamic ight coefcients of
particular operating condition by a linear system. This simpli- the missile. The angle-of-attack is the angle between the mis-
cation allows one to draw upon analysis and design methods sile body and velocity vector and can be considered a state
for linear systems. However, this simplication comes at the variable, and hence endogenous to the missile dynamics. The
cost of certain limitations: dynamic pressure, which is a function of missle velocity and
atmospheric pressure, is indicative of the environmental con-
The nonlinear system must be conned to operating near ditions. Atmospheric pressure is clearly an exogenous signal.
the specied operating condition of the linearization. Since the dynamic pressure is also a function of missile veloc-
The linearization analysis may give misleading or incon- ity, it can be considered an endogenous variable. However,
clusive results. the velocity variations in a simplied model are decoupled
from the attitude dynamics, and hence, dynamic pressure
The linearization may ignore important nonlinear phe-
may be modeled as an exogenous signal. This sort of ambigu-
nomena which dominate the system behavior.
ity, namely that an exogenous signal is really an endoge-
nous signal in a more sophisticated model, is common.
Despite these limitations, linearization remains a widely used Gain scheduling has seen widespread industrial applica-
method for control system analysis and design. tion. It is perhaps the most prevalent nonlinear method for
In many cases, conning a nonlinear system to operating aircraft ight control and missle autopilot design. Other ap-
near a specied operating condition is too restrictive. One ex- plications include power systems, process control, and auto-
ample is ight control. An aircraft typically experiences sev- motive control. Despite its widespread usage, traditional gain
eral ight conditions, including take-off, cruising at various scheduling has been an ad hoc design approach accompanied
altitudes, specialized maneuvers, and landing. No single by heuristic guidelines.
linearization can adequately describe the aircraft dynamics at
all of these conditions. Another example is boiler-turbine con-
trol in power generation. Typical operating conditions include
power ramp up, steady power delivery at various levels, and Scheduling
power ramp down. Again, no linearization can adequately de- variables
scribe the dynamics at all operating conditions.
Environmental Operating
conditions conditions
Some immediate drawbacks of gain scheduling are the fol- Let r(x) denote the residual error of approximation; that is,
lowing:
r(x) = f (x) f (x0 ) + D f (x0 )(x x0 ) (4)
The design of several controllers at several linearization
points can be a tedious task.
Then,
Gain scheduled designs typically assume a xed op-
erating condition, even though the operating condition is r(x)
varying. lim =0 (5)
xx o |(x x0 )|
Although the resulting controller is nonlinear, it is still
based on linearizations of plant dynamics, and hence
may neglect important nonlinear phenomena. where v denotes the Euclidean norm of v R n,
Two heuristic rules-of-thumb which guide successful gain |v| = (vT v)1/2 (6)
scheduled designs are the following:
Example 1. Let f: R 2 R 2 be dened as
The scheduling variable should vary slowly.
The scheduling variable should capture plant nonlinear- x1 |x1 | + x2
f (x) = (7)
ities. x21 x2
LINEARIZATION leads to
denote the approximation residual. Then as before, Definition 3. Let x0 be an equilibrium of Eq. (16).
r(x, u) x0 is stable if for any 0, there exists a 0 such that
lim
=0 (15)
xx 0
uu 0 |x x0 |2 + |u u0 |2
|x(0) x0 | < |x(t) x0 | (24)
Equations (5) and (15) indicate that the approximations Otherwise, x0 is unstable.
are accurate up to the rst order. x0 is asymptotically stable if in addition to being stable,
Linearizing about the upright equilibrium x0 0 leads to with the constant input u(t) u0 leads to the constant solu-
tion x(t) x0.
0 1 Proceeding as in the autonomous case, whenever f is con-
x = x (28)
g/ c tinuously differentiable, we may rewrite (33) as
It is also possible to use linearization methods to synthesize Example 10. Recall the simple pendulum example, but now
controllers for a nonlinear system. with a control torque input,
Consider the controlled system
d x1 x2
x = f (x, u) (33) = g (42)
dt x2 sin(x1 ) cx2 + u
Definition 7. The pair (x0, u0) is an equilibrium of Eq. (33) if
Linearizing about the upright equilibrium leads to
f (x0 , u0 ) = 0 (34)
. 0 1 0
The reasoning behind the term equilibrium is similar to x = x + u (43)
g/ c 1
the autonomous case. The initial condition x(0) x0 along
GAIN SCHEDULING 209
The feedback Then, the linearization about the equilibrium (x0, u0),
.
u = (k1 k2 )x (44) x = Ax + Bx
(53)
y = Cx
stabilizes the linearization for any k1 g/ and k2 c. From
Theorem 9, the feedback
where
u = u0 (k1 k2 )(x x0 )
(45) A = D1 f (x0 , u0 ), B = D2 f (x0 , u0 ), C = Dg(x0 ) (54)
= (k1 k2 )x
approximates the input-output behavior of Eq. (33) with mea-
stabilizes the upright equilibrium, where we used that x0 0 surement of Eq. (49). Here, x approximates x x0, y approxi-
and u0 0. mates y y0, and u exactly represents u u0.
In some cases, analysis of the linearization does not aid in Theorem 13. Let f in Eq. (33) and g in Eq. (49) be continu-
the construction of stabilizing feedback. ously differentiable, and let Eq. (53) be the linearization of
Eq. (33) about the equilibrium (x0, u0). Suppose the linear
Example 11. Consider the scalar nonlinear system feedback,
x = x + xu (46) z = Az + By
(55)
Linearizing about the equilibrium (x0, u0) (0, 0) leads to u = Cz
. stabilizes the linearization of Eq. (53). Then the equilibrium
x = x (47)
(x0, u0) of Eq. (33) is stabilized by the output feedback
which is not stabilizable. However, the constant feedback
u 2 leads to the closed loop equations z = Az + B(y g(x0 ))
(56)
u = u0 + Cz
x = x (48)
Example 14. Suppose we wish to control the simple pendu-
which is stable.
lum under the output feedback
Now suppose that the state is not available for feedback. y = (1 0)x (57)
Rather, the control is restricted to measurements
Linearizing about the upright equilibrium leads to
y = g(x) (49)
. 0 1 0 def
By using a similar analysis, we can construct stabilizing out- x = x + u = Ax + Bu
put feedback for the nonlinear system based on stabilizing g/ c 1 (58)
output feedback for the linearization. def
y = (1 0)x = Cx
Definition 12. The equilibrium (x0, u0) of Eq. (33) is stabi-
The observer based controller
lized by the dynamic output feedback
z = (A BK HC)z + H y
z = F (z, y) (59)
(50) u = Kz
u = G(z)
stabilizes the linearization for appropriate gain matrices
if for some z0,
h1
(z0, g(x0)) is an equilibrium of Eq. (50) K = (k1 k2 ), H= (60)
h2
u0 G(z0)
(x0, z0) is an asymptotically stable equilibrium of the
Since x0, u0, y0 0, the same controller (with input y and
closed loop dynamics
output u) stabilizes the nonlinear pendulum.
x = f (x, G(z))
(51)
z = F (z, g(x)) BASIC GAIN SCHEDULING
Let (x0, u0) be an equilibrium of Eq. (33), and dene y0 Gain Scheduled Command Following
g(x0). In case g is continuously differentiable, we can approxi- We now outline the basic procedure of gain scheduling in the
mate Eq. (49) as context of command following. The nonlinear plant of interest
is Eq. (33). The objective is to make the measured output Eq.
y y0 + Dg(x0 )(x x0 ) (52) (49) approximately follow reference commands, r.
210 GAIN SCHEDULING
The primary motivation of gain scheduling is to address Example 17. Recall the controlled pendulum
local limitations associated with a control design based on a
single linearization. The main problem is that the perfor- d x1 x2
mance and even stability of the closed loop system can deteri- = g (67)
dt x2 sin(x1 ) cx2 + u
orate signicantly when the system is not operating in the
vicinity of the equilibrium.
An equilibrium family over the set S [, ] is
Example 15. Consider the scalar system
s g
xeq (s) = , ueq (s) = sin s (68)
x = x|x| + u 0
(61)
y=x
The associated linearization family is
By linearizing about the equilibrium (x0, u0) (0, 0), we ob-
tain the linear control u x r. For r 0, this control law 0 1 0
x = g x + u (69)
stabilizes the equilibrium (0, 0). The resulting closed loop sys- cos(s) c 1
tem is
x = x|x| x (62) Step 2: Fixed Operating Condition Designs. Let us select sev-
eral operating conditions,
For x(0) 1, the solution asymptotically approaches 0. How- {s1 , s2 , . . ., sN } S (70)
ever, for x(0) 1, the solution diverges to innity.
which characterize the variations within the operating en-
Step 1: Construction of Linearization Family. Gain scheduling velope.
attempts to overcome local limitations by considering a family At the ith equilibrium, si, we can linearize the plant dynam-
of linearizations, rather than a single linearization. ics about the equilibrium (xeq(si), ueq(si)) and design a stabiliz-
ing linear controller to achieve approximate command follow-
Definition 16. The functions (xeq( ), ueq( )) dene an equilib- ing using any suitable linear design methodology. The result
rium family for the nonlinear system Eq. (33) over the set is an indexed collection of controllers,
S if
z = Ai z + Bi y + Li r
(71)
f (xeq (s), ueq (s)) = 0 (63) u = Ci z
Switched Scheduling. The operating envelope is divided which are globally stable, as opposed to the local stability of
into disjoint regions, Ri, so that Example 15.
S R1 . . . RN (74)
Notice in this example that the linear control term (x
xeq(s)) has no effect on the closed loop equations. This is be-
and the controller matrices in (73) are scheduled ac- cause of the smooth scheduling implementation with s x as
cording to the scheduling variable. In this case, a desirable feedback
loop was eliminated in the scheduling implementation. It is
A1 , . s R 1 ;
.. also possible to introduce undesirable feedback during imple-
A(s) = , . . ., ueq (s) mentation.
A ,
N
s R N .
ueq (s1 ), Theoretical Foundations
s R1 ;
.. The gain scheduled controller is designed so that stability and
= . (75)
performance are achieved whenever the plant is in the vicin-
u (s ), s R N
eq N ity one of the design operating conditions. Since the plant ac-
tually varies throughout the entire operating regime, an im-
Continuous Scheduling. Any interpolation algorithm is portant question is to what degree the local properties of the
used to construct continuous matrices which interpolate individual operating point designs carry over to the global
the design conditions so that system.
The overall closed loop equations for a gain scheduled sys-
A(si ) = Ai , . . ., ueq (si ) = ueq (si ) (76) tem are
The xed operating point designs assume a constant y = g(x) yeq (s)
scheduling variable which is actually time varying. r = r yeq (s)
Implementing the gain scheduled controller introduces
feedback loops which are not present in the xed op- In general, the scheduling variable, s, can be written as
erating point designs.
s = (x, r) (84)
Example 18. Recall the system of Example 15. The equilib-
rium family for an appropriate function, . Clearly, the overall system is
nonlinear and hence, requires nonlinear methods for analysis.
xeq (s) = s, ueq (s) = s|s| (77)
An analysis of these equations (see Bibliography for
sources) leads to the conclusion that the overall gain sched-
leads to the linearization family uled system will exhibit similar stability and performance as
. the local designs whenever (1) the scheduling variable, s,
x = 2|s|x + u (78) changes sufciently slowly, and (2) the plant dynamics are
predominantly nonlinear in the scheduling variable.
Because of the simplicity of this system, we are able to design The following sections provide some insight into these re-
controls for all s, rather than selected s. A suitable linear de- strictions.
sign for command following is
u = 3|s|x + (r x)
(79) LPV Systems. It is convenient to consider slow variation re-
striction in the context of linear parameter varying (LPV) sys-
tems. LPV systems are dened to be linear systems whose
Implementing this design using smooth scheduling leads to
the gain scheduled control dynamics depend on exogenous time varying parameters
which are unknown a priori, but can be measured upon opera-
u = ueq (s) + u tion of the control system.
(80) An LPV system can be represented in state space form as
= ueq (s) 3|s|(x xeq (s)) + ((r xeq (s)) (x xeq s))
x = A( )x + B( )u
For the scheduling variable s x, the control becomes (85)
y = C( )x
u = x|x| + (r x) (81)
where is a time varying parameter. Typical assumptions on
This feedback leads to the closed loop dynamics are magnitude bounds; for example,
and rate bounds; for example, By subtracting equation (91) from equation (89), we obtain
| | max (87) d s 0
= M(s) + B(u ueq (s)) (92)
dt v v veq (s)
LPV systems form the underlying basis of gain scheduling.
It is convenient to associate the parameter with the sched-
uling variable and the LPV structure with the linearization If veq(s) is differentiable,
family, although this is not always the case as will be seen.
The following is a classical result from differential equa- d
veq (s) = Dveq (s)s
tions stated in an LPV context. dt
= Dveq (s)M12 (s)(v veq (s)) + Dveq (s)B1 (u ueq (s))
Theorem 2. If the equilibrium xo 0 of Eq. (85) is asymptoti- (93)
cally stable for all constant , then it is asymptotically stable
for all time varying provided that max is sufciently small. where the matrices M12(s) and B1 are appropriate sub-matri-
ces of M(s) and B.
The relevance of Theorem 19 to gain scheduling is as fol- Combining these equations leads to the alternate form of
lows. A closed loop LPV system is such that good stability and Eq. (89),
performance is expected for xed values of the parameter/
scheduling variable. However, performance and even stability
d s (94)
can deteriorate in the presence of parameter time variations.
dt v veq (s)
Theorem 19 provides a sufcient condition for the xed pa-
rameter properties to carry over to the varying parameter
0 M12 (s) s
setting. =
0 M22 (s) Dveq (s)M12 (s) v veq (s)
(95)
Example 20. A classical example of instability from xed pa- B1
rameter stability is the time-varying oscillator, + (u ueq (s))
B2 Dveq (s)B1
0 1
=
x(t) x(t) (88) which can be written as
(1 + (t)/2) 0.2
These equations can be viewed as a mass-spring-damper sys- d s s
= Anew (s) + Bnew (s)u (96)
tem with time-varying spring stiffness. For xed parameter dt v v
values, (t) o, the equilibrium xo 0 is asymptotically sta-
ble. However, for the parameter trajectory (t) cos(2t), it where
becomes unstable. An intuitive explanation is that the stiff-
ness variations are timed to pump energy into the oscilla- = v(t) veq (s(t)),
v(t)
u(t) = u(t) ueq (s(t)) (97)
tions.
The original equations now take a quasi-LPV form, where
Quasi-LPV Representation. It is also convenient to consider
the parameter is actually an endogenous variable. Note that
the relationship between the scheduling variable and plant
no linearization approximations were made to bring Eq. (89)
nonlinearities in an LPV setting.
to the form Eq. (96).
The relationship between LPV systems and gain schedul-
This transformation shows that an underlying LPV struc-
ing is not limited to linearization families. Consider the fol-
ture exists, even without linearizations, in the extreme case
lowing special nonlinear plant in which the scheduling vari-
that the plant dynamics are nonlinear only in the scheduling
able is a subset of the state,
variable. Any additional nonlinearities not captured by the
scheduling variable enter as high order perturbations in Eq.
d s s
= (s) + M(s) + Bu (89) (96). This transformation then reveals the importance of the
dt v v scheduling variable to capture the plant nonlinearities.
These equations represent the extreme case where the nonlin- Example 21. Consider the nonlinear system
earities are entirely captured in the scheduling variable, s.
Let (xeq(s), ueq(s)) be an equilibrium family, with
x1 |x1 | + x2
x = (98)
s x21 x2 + u
xeq (s) = (90)
veq (s)
and let s x1 be the scheduling variable. These equations
so that take the form of Eq. (89). The resulting equilibrium family is
s s
0 = (s) + M(s) + Bueq (s) (91) xeq (s) = , ueq (s) = s3 |s| (99)
veq (s) s|s|
GAIN SCHEDULING 213
Performing the transformations described above leads to the In order to convexify the problem, consider the change in
quasi-LPV form variables
d s 0 1 s Q = P1 , Yi = Ki P1 (106)
=
dt x2 (s|s|) 0 s2 2|s| x2 (s|s|)
(100) Given Q 0 and Yi, one can solve for the original variables P
0 and Ki. With these variables, condition Eq. (105) is equiva-
+ (u (s3 |s|) lent to
1
One invariance objective is for the closed loop linearization H. K. Khalil and P. K. Kokotovic, On stability properties of nonlinear
family to be constant; for example, systems with slowly varying inputs, IEEE Trans. Autom. Control,
36: 229, 1991.
.
x = Adesx + Bdes r (113) M. Vidyasagar, Nonlinear Systems Analysis, Nonlinear Systems Analy-
sis, Englewood Cliffs, NJ: Prentice-Hall, Inc., 1993.
One can state appropriate conditions involving partial differ- A. Bacciotti, Local Stabilizability of Nonlinear Control Systems, Singa-
ential constraint equations under which there exists any G pore: World Scientic Publishing Co., 1992.
which achieves this objective. H. K. Khalil, Nonlinear Systems, 2nd Ed., New York: Macmillan,
The intention is that improved closed loop performance is 1996.
possible if the closed loop linearizations have some sort of in-
variance property. Such improved performance, if any, is dif- Overview of Gain-Scheduling and Its Theoretical Foundations
cult to quantify, but simulation studies indicate the poten-
J. Wang and W. J. Rugh, On parameterized linear systems and lin-
tial benets. earization families for nonlinear systems, IEEE Trans. Circuits
Syst., 34: 650657, 1987.
Example 23. Consider the nonlinear system
J. S. Shamma and M. Athans, Analysis of nonlinear gain-scheduled
control systems, IEEE Trans. Autom. Control, 35: 898907, 1990.
x1 x2
= (114) W. J. Rugh, Analytical framework for gain-scheduling, IEEE Control
x2 x1 x2 + x21 + u Syst. Magazine, 11: 7984, 1991.
J. S. Shamma and M. Athans, Gain scheduling: Potential hazards
The equilibrium family suggested by x1 is and possible remedies, IEEE Control Syst. Magazine, 12: 101
107, 1992.
s
xeq (s) = , ueq (s) = s2 (115)
0 LPV System Analysis and Control
J. S. Shamma and M. Athans, Guaranteed Properties of Gain Sched-
Suppose we want the closed loop linearization family to have uled Control of Linear Parameter Varying Plants, Automatica, 27:
dynamics matrices 559564, 1991.
S. M. Shahruz and S. Behtash, Design of controllers for linear param-
0 1 0 eter-varying systems by the gain scheduling technique, J. Math.
Ades = , Bdes = (116)
1 2 1 Anal. Appl., 168: 125217, 1992.
A. Packard, Gain scheduling via linear fractional transformations,
An obvious selection for G(x, r) is Syst. Control Lett., 22: 7992, 1994.
P. Apkarian and P. Gahinet, A convex characterization of gain-sched-
G(x, r) = x1 x2 x21 x1 2x2 + r (117) uled H controllers, IEEE Trans. Autom. Control, 40: 853864,
1995.
Then setting u G(x, r) leads to the linear closed loop dynam- P. Apkarian and R. J. Adams, Advanced gain-scheduling techniques
ics for uncertain systems, IEEE Trans. Control Syst. Technol., 60: 21
32, 1998.
0 1 0
x = x+ r (118)
1 2 1 Extended and Pseudo-Linearization
C. Reboulet and C. Champetier, A new method for linearizing nonlin-
The above choice for G achieves what is called feedback linear- ear systems: the pseudolinearization, Int. J. Control, 40: 631
ization, since the resulting closed loop dynamics are linear. 638, 1984.
It is not necessary that feedback linearization is achieved. W. T. Baumann and W. J. Rugh, Feedback control of nonlinear sys-
For example, consider the modied dynamics tems by extended linearization, IEEE Trans. Autom. Control, 31:
4046, 1986.
x1 x2 + x22 u J. Wang and W. J. Rugh, Linearized model matching for single-input
= (119)
x2 x1 x2 + x21 + u nonlinear systems, IEEE Trans. Autom. Control, 33: 793796,
1988.
The resulting equilibrium families and linearization families J. Huang and W. J. Rugh, On a nonlinear servomechanism problem,
Automatica, 26: 963972, 1990.
are the same. Furthermore, the same choice of G achieves the
desired invariance objective. J. Huang and W. J. Rugh, Approximate noninteracting control with
stability for nonlinear systems, IEEE Trans. Autom. Control, 36:
295304, 1991.
FOR FURTHER REFERENCE D. A. Lawrence and W. J. Rugh, Input-output pseudolinearization for
nonlinear systems, IEEE Trans. Autom. Control, 39: 22072218,
Nonlinear Systems Analysis 1994.
M. Kelemen, A stability property, IEEE Trans. Autom. Control, 31: I. Kaminer, A. M. Pascoal, P. P. Khargonekar, and E. Coleman, A
766768, 1986. velocity algorithm for the implementation of gain-scheduled con-
D. A. Lawrence and W. J. Rugh, On a stability theorem for nonlinear trollers, Automatica, 31: 11851192, 1995.
systems with slowly varying inputs, IEEE Trans. Autom. Control, D. A. Lawrence and W. J. Rugh, Gain scheduling dynamic linear con-
35: 860864, 1990. trollers for a nonlinear plant, Automatica, 31: 381390, 1995.
GALLIUM ARSENIDE FIELD EFFECT TRANSISTOR LOGIC CIRCUITS 215
JEFF S. SHAMMA
The University of Texas at Austin