IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 8, NO.

5, SEPTEMBER 2000 749

Adaptive Nonlinear Control of Agile Antiair Missiles

Using Neural Networks
Michael B. McFarland, Member, IEEE, and Anthony J. Calise, Member, IEEE

Abstract—Research has shown that neural networks can be other concerns have prompted researchers to look beyond
used to improve upon approximate dynamic inversion controllers the classical methods that have historically dominated the
in the case of uncertain nonlinear systems. In one possible ar- field of missile autopilot design and to robust, nonlinear, and
chitecture, the neural network adaptively cancels linearization
errors through on-line learning. Learning may be accomplished “intelligent” control.
by a simple weight update rule derived from Lyapunov theory, Most nonlinear control techniques are based on linearizing
thus assuring the stability of the closed-loop system. In this paper, the equations of motion by the application of nonlinear feed-
the authors discuss the evolution of this methodology and its back. Known variously as feedback linearization or dynamic in-
application in a bank-to-turn autopilot design for an agile antiair version, this method relies heavily on knowledge of the plant
missile. First, a control scheme based on approximate inversion of
the vehicle dynamics is presented. This nonlinear control system is dynamics. An early application of this theory to the missile
then augmented by the addition of a feedforward neural network autopilot design problem is found in [1], while [2] presents a
with on-line learning. Finally, the resulting control law is demon- more sophisticated approach involving variable structure con-
strated in a nonlinear simulation and its performance is evaluated trol. More recently, neural networks have emerged as a means
relative to a conventional gain-scheduled linear autopilot. of explicitly accounting for uncertainties in the plant dynamics.
Index Terms—Nonlinear systems, adaptive control, Lyapunov Their on-line learning and functional approximation capabili-
methods, robustness, neural networks, missiles. ties make neural networks an excellent candidate for this appli-
cation. [3] is one example in which neural networks are investi-
I. INTRODUCTION gated for nonlinear control of missiles.
This paper concerns a neural-network-based approach to di-
A DVANCES in fighter aircraft technology continue to
create new challenges for designers of antiair weapon
systems. The introduction of low-observable aircraft has
rect adaptive control of nonlinear systems which originated in
[4] and was further developed in [5]. In the proposed architec-
ture, a simple dynamic inversion controller approximately lin-
increased demand for small lightweight missiles. This, in
earizes the vehicle dynamics. This controller is augmented by a
turn, has led to various control problems associated with
neural network which acts to improve the linearization by adap-
air-breathing propulsion, asymmetric airframes, and reduced
tively canceling inversion errors in real-time. The neural-net-
aerodynamic control surface area. Similarly, the development
work implementation features a stable on-line learning algo-
of supermaneuverable aircraft has motivated efforts to extend a
rithm derived from Lyapunov theory. While [4] and [5] dealt
missile’s flight envelope into the high angle of attack regime.
specifically with fighter aircraft and helicopter applications, re-
Accordingly, some proposed next-generation missiles employ
spectively, the work described in this paper adapts these pre-
nonaerodynamic control effectors similar to those associated
vious efforts for use in control of agile missiles. The result is an
with super-maneuverable aircraft. These so-called “agile”
autopilot that combines the best features of dynamic inversion,
missiles possess enhanced range because of their ability to
neural networks, and adaptive control to potentially increase the
execute rapid propulsive heading changes during the boost
effectiveness and versatility of tomorrow’s missile systems.
phase. Furthermore, these weapons can be deployed during
The preliminary study presented in [6] revealed that neural
high angle of attack maneuvers and may even engage targets in
networks are indeed capable of attaining sufficiently high
the rear hemisphere relative to the launch aircraft.
learning rates to make adaptation feasible during even the
The dynamics of an agile missile flying in bank-to-turn
most demanding aerial engagements. In fact, previous work
mode at a high angle of attack are inherently nonlinear and
documented in [7] shows that the methodology considered
may vary rapidly with time. Furthermore, these dynamics are
here compares favorably to traditional gain-scheduled linear
highly uncertain since aerodynamic data for vehicles operating
methods for this application.
under such conditions is difficult to obtain and may in fact
This paper begins with a review of the proposed controller
be a poor approximation to the actual dynamics. These and
architecture and its evolution to date. The methodology in ques-
tion is then applied to an agile antiair missile autopilot design
Manuscript received April 22, 1998; revised April 17, 1999. Recommended problem. First, the baseline control scheme consisting of an ap-
by Associate Editor, K. Wise. This work was sponsored by the U.S. Air Force
Research Laboratory Munitions Directorate under Contract F08630-95-1-0006. proximate inversion of the missile’s six-degree-of-freedom non-
M. B. McFarland is with Raytheon Missile Systems Company, Tucson, AZ linear dynamics is presented. Neural networks are then designed
85734 USA. to enhance this nonlinear controller. Simulation results com-
A. J. Calise is with the School of Aerospace Engineering, Georgia Institute
of Technology, Atlanta, GA 30332 USA. paring this technique to a traditional gain-scheduled implemen-
Publisher Item Identifier S 1063-6536(00)05738-9. tation are presented, and the effects of neural-network topology
1063–6536/00$10.00 © 2000 IEEE
750 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 8, NO. 5, SEPTEMBER 2000

are investigated. Finally, conclusions and future research direc-

tions are discussed.

II. CONTROL DESIGN METHODOLOGY

Consider a two-stage nonlinear system with the following Fig. 1. Open-loop nonlinear system.
structure:

(1)

where and remains nonsingular for all

. For reference, a block diagram of this system is included in
Fig. 1. What follows is an overview of the construction of an Fig. 2. Approximate dynamic inversion.
adaptive nonlinear control scheme with neural networks. Addi-
tional details and a complete stability proof are included in [8].
the definition of above provides a linearizing transformation
of the control. The inverse transformation is expressed by
A. Approximate Nonlinear Control
Suppose that the dynamics of the subsystem are well
(6)
known, but the nonlinearity is only known approximately.
Furthermore, suppose a stabilizing controller for the
subsystem with as input is also known. One example, and must be computed in real-time when the control is imple-
when it exists, is the dynamic inversion controller described by mented. The resulting closed-loop subsystem is described by

(2) (7)
where denotes a vector of external commands. Note, how-
ever, that is not necessarily an inverting controller. If, for in- where
stance, the plant dynamics include so-called “beneficial” non-
linearities, these terms need not be canceled by the nonlinear (8)
feedback. This is “lean” nonlinear control as described in [9]
with regard to integrator backstepping.
The mapping represents a nonlinear inver-
To construct a stabilizing control for the uncertain nonlinear
sion error. If is known and its inverse is computed accurately,
system in (1), first define the following error variables:
then and the system is ex-
actly linearized. However, since is uncertain, is generally
nonzero and only approximate linearization is possible. Fig. 2
(3) illustrates this approximate dynamic inversion process.
In the ideal case, integrator backstepping can be used to con-
We may use to close the loop on the subsystem, replacing
struct an expression for which stabilizes both errors and
in the in the equation of (1) with the expression for
. When there is nonzero inversion error, this expression may
from (2). Differentiating (3) and substituting then results in the
be augmented by an adaptive term to compensate for this im-
following error dynamics:
perfect inversion. The result is the following expression for the
pseudocontrol :

(4)
(9)
Note that no time-scale separation has been assumed, and
nowhere is it assumed that tracks perfectly. where is the adaptive control contribution to the pseu-
A technique derived from integrator backstepping will now docontrol. The first term in (9) is included to offset coupling
be used to construct the stabilizing controller for the composite between the and dynamics. It has become commonplace
system (1). We first turn our attention to the subsystem. In in applications of this type to reduce such coupling by enforcing
(1), the subsystem may be rewritten as time-scale separation between and . In that case, the first
term in (9) may be neglected. The second term is simply the
“command derivative” term in model-following control, and is
(5) necessary to achieve tracking of arbitrary continuous trajecto-
ries. For slowly varying commands, this term may also be ne-
where is a pseudocontrol input. If the mapping glected. Fig. 3 depicts the adaptive control architecture for the
is invertible and full-state feedback is available, subsystem.
MCFARLAND AND CALISE: ADAPTIVE NONLINEAR CONTROL OF AGILE ANTIAIR MISSILES USING NEURAL NETWORKS 751

Fig. 3. Adaptive control architecture.

B. Neural-Network-Based Adaptation
Fig. 4. Sigma-Pi neural network.
In [4], the adaptive component of (9) is assigned the form
A. Approximate Dynamic Inversion
The equations of motion of a symmetric missile about its ve-
locity vector are as follows:
(10)

for with scalar weights and the -di-

mensional set of basis functions . Since the adaptive (12)
control term is intended to approximate the nonlinear inversion
error, the formulation in (10) is equivalent to parameterizing
with a neural network that is linear in its adjustable parameters. (13)
Such simple networks are often referred to as “single-layer” net-
works because they have one output layer and no hidden layers. where and denote airspeed, angle of attack, and sideslip
An example of this class of neural networks is represented by angle. Here, , and are the body-axis components of
the Sigma-Pi architecture depicted in Fig. 4. acceleration (including gravitational effects) while , and
It can be shown using Lyapunov stability theory, as in [4]–[6], are the body-axis angular rates. The moment equations have the
that the adaptive control achieves global asymptotic stability form
when the weight update rule is chosen as follows:

(11)

where is an adaptation gain, or learning rate, and (14)

denotes the magnitude of a deadzone introduced as an element
of the stability proof as in [4]. Without loss of generality, the where and represent aerodynamic moments about the
weights may be assigned zero initial values. body axes. Finally, , and are roll, pitch, and yaw
Note from Fig. 3 that the network input depends on the moments of inertia, respectively, and are approximated as con-
current network output. Thus, a critical assumption in the sta- stants. Also, for the symmetric missile under investigation note
bility proof involves the existence of a fixed-point solution for that .
the output . Such a condition is guaranteed to be satisfied In order to simplify roll control in bank-to-turn flight, we now
when bounded basis functions are used. Furthermore, when introduce an aerodynamic bank angle about the velocity
a stable fixed-point exists, a simple iterative scheme can be em- vector. The bank angle dynamics are described by (15) below.
ployed to compute fixed-point solution for the network output.
The deadzone is used to account for the fact that the network
(15)
may be incapable of exactly representing the inversion error
using a finite set of basis functions.
The following development assumes the presence of a guidance
III. MISSILE AUTOPILOT APPLICATION law that commands angle of attack, sideslip angle, and bank
angle. There are several possible alternatives for the roll com-
In this section, the neural-network-based adaptive control mand, including body roll angle , body roll rate , and sta-
methodology described above is used to design a nonlinear bility-axis roll rate .
autopilot for a next-generation antiair missile depicted in Introducing the more compact notation
Fig. 5. This vehicle is described in detail in [10]. In the current
scenario, the objective is to design a bank-to-turn autopilot that
tracks guidance commands in angle of attack and bank angle (16)
while holding sideslip near zero.
752 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 8, NO. 5, SEPTEMBER 2000

Fig. 5. Agile missile configuration.

the last two equations in (13) may be rewritten along with (15) where the function refers to the right-hand-side of (14).
as follows: The missile’s body-axis angular rate dynamics must be approx-
imately inverted to determine the required control input. For an
(17) approximate inversion, the body-axis moments are represented
linearly
Here, and are given by (18) and (19) below.

(18)
sec sec (23)

The body angular rate dynamics of (14) are then rewritten as

(19) follows:

(24)
The autopilot design methodology presented above will now
be applied to the system described by (17) and (14). First,
where is approximated linearly by introducing
body-axis angular rate commands are computed from (2) as
follows:

(20)
and (25)
The gain matrix may be chosen as diagonal with positive
elements. An integral term may also be included in (20), but has
been omitted for simplicity. No adaptive control is necessary at
this stage of the design, since the nonlinear dynamics are suffi- with
ciently well known. Note that (20) makes use of available accel-
eration and velocity information, as well as estimates of angle
of attack and sideslip angle constructed from inertial data. The (26)
accelerometer measurements are filtered and biased appropri-
ately to account for gravitational effects. Note that may be (27)
inverted, since it is nonsingular except when , which
should not occur in bank-to-turn flight. (28)
Next, a first-order filter is applied to the guidance commands
, and . Its outputs are the filtered commands and their
Performing the approximate dynamic inversion, the control
rates.
input is then given by (29)

(21) (29)

The next step in the control design is the computation of the

At this point in the control design procedure, we turn our
pseudocontrol input . Here, the commands generated by (16)
attention to the body-axis angular rates. First, a pseudocontrol
are filtered to obtain
input is defined by

(30)
(22)
MCFARLAND AND CALISE: ADAPTIVE NONLINEAR CONTROL OF AGILE ANTIAIR MISSILES USING NEURAL NETWORKS 753

Fig. 6. Step response comparison. Fig. 7. Intercept trajectories.

which are used in (9) to give

(31)

where is the adaptive control component, which is included

to cancel nonlinear inversion errors. As in the outer loop, is
a diagonal matrix of positive gains and the error variable
has been introduced. Note that we have neglected the
coupling term from (9). We will instead choose
and to impose time-scale separation upon the closed-loop
error dynamics.
In practice, the stability and control derivatives in (24) could
be scheduled as a function of flight condition, along with trim
values of the control inputs. This, however, presumes that an
accurate full-envelope vehicle model is available. In the cur-
rent implementation, the moment coefficients are treated as con-
stants, shifting the burden of gain scheduling to the neural net-
work. The dynamic inversion control law may be simplified fur- Fig. 8. Angle of attack response.
ther by introducing additional approximations as simulation re-
sults warrant. The neural network would then be required to ad- in each channel is therefore described by (10) presented previ-
ditionally compensate for those effects that are neglected by the ously. Moreover, a stable learning rule is again given by (11).
dynamic inversion. Various neural network topologies and choices of inputs have
In the actual implementation, the inputs computed by (29) are been considered in this study. The results presented in the fol-
distributed among the missile’s aerodynamic control surfaces lowing section use a second-order Sigma-Pi network with inputs
and RCS thrusters using any suitable control allocation algo- chosen as normalized values of , and as well as a constant
rithm. bias term.
B. Neural-Network Architecture
C. Simulation Results
Substituting (20) into (17) and (29) into (24), the closed-loop
Results were obtained using the neural-adaptive nonlinear au-
error dynamics can be written as follows:
topilot described above in a nonlinear, six-degree-of-freedom
(6-DOF) simulation of a proposed agile antiair missile. In this
particular simulation, RCS thrusters and BTT guidance are only
(32) employed during high-flight. The autopilot results presented
below therefore do not reflect the terminal stage of the intercept,
Any simple linearly parameterized feedforward neural network during which control is returned to a gain-scheduled autopilot.
which is capable of approximately reconstructing the nonlinear Fig. 6 compares this autopilot with a more conventional gain-
inversion error may be used to compute the adaptive contri- scheduled autopilot for a step command of 90 in angle of at-
bution to the pseudocontrol. The adaptive pseudocontrol input tack. This plot is representative of 6-DOF simulation results for
754 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 8, NO. 5, SEPTEMBER 2000

other maneuvers, which indicate that the two controllers achieve

similar performance.
An important engagement geometry for an agile antiair mis-
sile is the so-called “merge” scenario, in which the launch air-
craft and the target are passing one another as the missile is
fired. In this case, the missile initially executes an agile turn at
very high angles of attack, and intercept occurs in the rear hemi-
sphere relative to the launch aircraft. Fig. 7 illustrates this sce-
nario, which takes place primarily in the local horizontal plane.
Autopilot angle of attack and sideslip angle tracking re-
sponses are depicted in Figs. 8 and 9. Sideslip angle is held to
less than 4 throughout the maneuver, a level similar to that
of the gain-scheduled design. The shape of the angle of attack
transient response is a result of both the nonlinear aerody-
namics of the missile and the neural network’s initial attempts
to compensate for these nonlinearities. This will subsequently
be discussed in relation to the neural network’s effectiveness
for error reconstruction. In Fig. 10, the missile’s roll angle Fig. 9. Sideslip response.
response is presented. Implementation concerns which are
beyond the scope of this paper dictated that the missile achieve
roll control using a redefined roll angle, neither nor . This
is merely a detail that does not significantly affect the overall
control design methodology.
Figs. 11–13 compare the RCS thrust commands for the
gain-scheduled and neural network-based autopilots. Sat-
uration limits are activated by stressful maneuvering, a
phenomenon which has not been explicitly accounted for in
the control design process. In the pitch case, such saturation is
unavoidable. In yaw, however, high control activity is a result
of the neural network canceling aerodynamic nonlinearities. In
fact, Figs. 11 and 12 illustrate that although command levels are
similar for both autopilots throughout the majority of the flight,
the neural-network control exhibits oscillatory behavior associ-
ated with adaptive learning. Reducing the network’s adaptation
gain would reduce the frequency of these oscillations but
degrade tracking performance. It has been shown in [12] that Fig. 10. Roll angle response.
selecting a different neural-network structure or set of basis
functions to provide a more effective parameterization of the
linearization error reduces these undesirable oscillations. Other
techniques for improving transient response include adding
damping to the weight update rule, an idea addressed in [12], or
incorporating tuning functions in the backstepping design [9].
Figs. 14–16 illustrate the effectiveness of neural networks in
inversion error reconstruction. In these figures, the solid lines
represent the actual value of the inversion error , which may
be computed exactly in the simulation. The dashed lines repre-
sent the adaptive pseudocontrol component. When the neural
network perfectly reconstructs linearization errors, these two
quantities have equal values. Fig. 14 illustrates the network’s
ability to effectively compensate for inversion errors in the pitch
dynamics. The yaw-channel network performs similarly to the
pitch-channel network, as illustrated by Fig. 15.
In Fig. 16, however, the neural network does not respond to
transient errors in the roll dynamics. Increasing will cause the
network to track these transients more aggressively, but such Fig. 11. Pitch RCS thrust.
high gains lead to excessive oscillations in the presence of even
small errors. In the results presented here, adaptation gains were performance was achieved. As with the control oscillations dis-
first assigned a small initial value and increased until the desired cussed above, choosing a different network topology or set of
MCFARLAND AND CALISE: ADAPTIVE NONLINEAR CONTROL OF AGILE ANTIAIR MISSILES USING NEURAL NETWORKS 755

Fig. 12. Yaw RCS thrust. Fig. 15. Yaw inversion error comparison.

Fig. 13. Roll RCS thrust. Fig. 16. Roll inversion error comparison.

architecture in [12]. Finally, Fig. 14 also illustrates the lag asso-

ciated with learning in the neural network, which is unavoidable
given available measurements.

IV. CONCLUSIONS
A nonlinear bank-to-turn missile autopilot based on approx-
imate dynamic inversion has been proposed. With the aid of
an adaptive neural network, this autopilot tracks guidance com-
mands in angle of attack and bank angle while holding sideslip
angle near zero. A similar development could be carried out for
the skid-to-turn case by replacing the bank angle state with
roll angle and maintaining .
Nonlinear 6-DOF simulation results indicate that the perfor-
mance of the neural-adaptive nonlinear autopilot is comparable
to that of an existing gain-scheduled autopilot. One advantage
of the neural-network-based approach is that it eliminates the
Fig. 14. Pitch inversion error comparison. time-consuming process of designing a different autopilot at
each of numerous flight conditions. Additionally, the use of
basis functions is one way to improve the performance of the neural networks enables the nonlinear controller to effectively
roll network. This was demonstrated using a single hidden-layer adapt on-line to uncertain nonlinear aerodynamic phenomena,
756 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 8, NO. 5, SEPTEMBER 2000

which are difficult to model for purposes of design and simula- [10] K. A. Wise and D. J. Broy, “Agile missile dynamics and control,” in
tion. Proc. AIAA Guidance Navigation Contr. Conf., San Diego, CA, 1996,
paper AIAA-96-3912.
Neural-network-based adaptive control of nonlinear systems [11] M. B. McFarland and A. J. Calise, “Robust adaptive control of non-
is a maturing technology area which promises to be applicable linear systems using neural networks,” in Proc. Amer. Contr. Conf., Al-
to a wide range of systems. The favorable results described here buquerque, NM, 1997, pp. 1996–2000.
[12] , “Multilayer neural networks and adaptive nonlinear control of
have already led to more advanced research, including modi- agile anti-air missiles,” in Proc. AIAA Guidance Navigation Contr.
fications for robustness to input unmodeled dynamics, which Conf., New Orleans, LA, 1997.
are the subject of [11]. The approach has also been extended [13] R. T. Rysdyk, F. Nardi, and A. J. Calise, “Robust adaptive nonlinear
flight control applications using neural networks,” in Proc. Amer. Contr.
to accommodate an important class of neural networks with Conf., CA, 1999, pp. 2595–2599.
one hidden layer as described in [12]. These multilayer feed-
forward networks feature improved approximation capabilities
over their single-layer counterparts, but at the price of nonlin-
Michael B. McFarland (S’94–A’97–M’99) re-
earity in their adjustable parameters. Additional detail regarding ceived the B.S.E. degree in aerospace engineering
robust redesign and nonlinear network parameterizations may (with high honors) from the University of Florida,
also be found in [8]. Moreover, the problem of simultaneous Gainesville, in 1991, and the M.S. degree in hybrid
algorithms for optimal aeroassisted orbit transfer
nonlinear parameterization and robustness to input unmodeled guidance and the Ph.D. degree in adaptive nonlinear
dynamics has been treated in [13]. Each of these advancements control of missiles using neural networks, both
represents another step toward the creation of a design technique from the School of Aerospace Engineering, Georgia
Institute of Technology, Atlanta, in 1992 and 1997,
powerful enough for tomorrow’s demanding flight control ap- respectively.
plications. From 1991 to 1999, he was a Research Aerospace
Engineer with the Air Force Research Laboratory Munitions Directorate
(AFRL/MN), Eglin Air Force Base. He is a Senior Systems Engineer with
REFERENCES Raytheon Missile Systems Company, Tucson, AZ. His research interests
[1] M. Tahk, M. Briggs, and P. K. A. Menon, “Application of plant inver- include applications of nonlinear and adaptive control, as well as neural
sion via state feedback to missile autopilot design,” in Proc. IEEE Conf. networks.
Decision Contr., Austin, TX, 1986, pp. 730–735.
[2] M. Innocenti and A. Thukral, “Simultaneous reaction jet and aerody-
namic control of missile systems,” in Proc. AIAA Guidance Navigation
Contr. Conf., Monterey, CA, 1993, pp. 347–354. Anthony J. Calise (M’72) received the B.S. degree
[3] D. M. McDowell, G. W. Irwin, and G. McConnell, “Online neural con- from Villanova University, Philadelphia, PA, in 1964,
trol applied to a bank-to-turn missile autopilot,” in Proc. AIAA Guidance and the M.S. and Ph.D. degrees from the University
Navigation Contr. Conf., Baltimore, MD, 1995, pp. 1286–1294. of Pennsylvania, Philadelphia, in 1966 and 1968, re-
[4] B. S. Kim and A. J. Calise, “Nonlinear flight control using neural net- spectively, all in electrical engineering.
works,” AIAA J., vol. 20, pp. 26–33, Jan.–Feb. 1997. He was a Professor of Mechanical Engineering at
[5] J. Leitner, A. J. Calise, and J. V. R. Prasad, “Analysis of adaptive neural Drexel University, Philadelphia, for eight years. He
networks for helicopter flight controls,” AIAA J., vol. 20, pp. 972–979, also worked in industry for ten years at Raytheon
Sept.–Oct. 1997. Missile Systems Company, Tucson, AZ, where he
[6] M. B. McFarland and A. J. Calise, “Neural networks for stable adap- was involved with analysis and design of inertial
tive control of air-to-air missiles,” in Proc. AIAA Guidance, Navigation navigation systems, optimal missile guidance, and
Contr. Conf., Baltimore, MD, 1995, pp. 1280–1285. aircraft flight path optimization. He is currently with the School of Aerospace
[7] , “Neural-adaptive nonlinear autopilot design for an agile anti-air Engineering, Georgia Institute of Technology, Atlanta. His research intrests
missile,” presented at the AIAA Guidance Navigation Contr. Conf., San include optimal control theory, aircraft flight control, optimal guidance of
Diego, CA, 1996, paper AIAA-96-3914. aerospace vehicles, adaptive control using neural networks, robust linear
[8] M. B. McFarland, “Adaptive nonlinear control of missiles,” Ph.D. dis- control, and control of flexible structures.
sertation, Georgia Inst. Technol., Atlanta, GA, 1997. Dr. Calise received the USAF Systems Command Technical Achievement
[9] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adap- Award and the AIAA Mechanics and Control of Flight Award. He is a Fellow
tive Control Design. New York: Wiley, 1995. of the AIAA.