MODEL REFERENCE ADAPTIVE CONTROL uncertainties (see Case 2 in Figure 1 as an illustration of

the boundedness notion).

In this article, we discuss model reference adaptive
control in a basic setting (see Sections 2, 3, and 4). In
1. INTRODUCTION particular, for providing an introduction to the model
reference adaptive control design procedure, we follow
Every physical dynamical system, without an excep- a state feedback approach and focus on answering the
tion, is subject to exogenous disturbances and/or system fundamental question on how a control designer can
uncertainties. Specifically, the presence of exogenous make the state trajectories of an uncertain dynamical
disturbances arises from, for example, winds and turbu- system follow the state trajectories of a given reference
lences and the presence of system uncertainties arises model (i.e., state feedback design for state tracking).
from, for example, idealized assumptions, linearization, We also make connections to several other, relatively
and degraded modes of operation. Throughout this article, advanced model reference adaptive control methods
exogenous disturbances and system uncertainties are for interested readers (see Section 5). Finally, a fairly
often referred to as system anomalies since they generally standard notation is used in the following sections,
tend to negatively alter desired stability and performance where ℝ denotes the set of real numbers, ℝn denotes
characteristics of physical systems. the set of n × 1 real column vectors, ℝn×m denotes
When designing control laws for dynamical systems the set of n × m real matrices, ℝ+ (respectively, ℝ+ )
subject to anomalies, a fundamental problem is to achieve denotes the set of positive (respectively, nonnegative)
n×n
closed-loop system stability and performance. Similar real numbers, ℝn×n+ (respectively, ℝ+ ) denotes the set
in spirit to robust control methods, in particular, adap- of n × n positive-definite (respectively, nonnegative-
tive control theory presents effective system-theoretical definite) real matrices, IDn×n denotes the set of n × n
tools to address this fundamental problem. In contrast real matrices with diagonal scalar entries, ‖⋅‖2 denotes
to robust control methods, however, they have the capa- the Euclidian norm, ‖⋅‖F denotes the Frobenius matrix
bility to deal with system anomalies in a real-time (i.e., norm, 𝜆min (A) (respectively, 𝜆max (A)) denotes the mini-
online) manner. This implies that they are not tuned mum (respectively, maximum) eigenvalue of a real and
to a worst-case scenario as robust control methods and △
square matrix A ∈ ℝn×n , and “ =” denotes the equality by
they can continuously improve their performance in real
definition.
time. From a general point of view, adaptive control
methods are classified as either direct or indirect. This
article focuses on model reference adaptive control archi-
tectures, a well-known class of direct adaptive control 2. UNCERTAIN DYNAMICAL SYSTEM
methods.
While this is not intended to be a survey article (we In this article, we consider the dynamics of a physical sys-
refer to excellent books (1–9) and relatively recent surveys tem represented in the following state-space form:
(10–13) on this point), one should mention that the authors
of References 14 and 15 originally proposed the model ref- ẋ p (t) = Ap xp (t) + Bp Λu(t) + Bp 𝛿p (xp (t)),
erence adaptive control concept. Specifically, to control a
xp (0) = xp0 . (1)
given uncertain dynamical system (see Section 2), this con-
cept has two major components – a reference model (see Here, xp (t) ∈ ℝnp stands for a measurable (i.e., accessible)
Section 3) and a parameter adjustment mechanism (see state vector and u(t) ∈ ℝm stands for a control vector.
Section 4). In addition, Ap ∈ ℝnp ×np denotes a known system matrix
A desired closed-loop dynamical system response and Bp ∈ ℝnp ×m denotes a known control matrix. Further-
is captured by the reference model for which its state more, 𝛿p ∶ ℝnp → ℝm represents a time-invariant system
(respectively, output) is compared with the state (respec- uncertainty and Λ ∈ ℝm×m+ ∩ 𝔻m×m represents an unknown
tively, output) of the uncertain dynamical system. The control effectiveness matrix. For the well-posedness
system error signal resulting from this comparison drives of the model reference adaptive control problem, we
the parameter adjustment mechanism. This mechanism assume that the pair (Ap , Bp ) is controllable and the
then adjusts the controller parameters in real time for system uncertainty is composed of locally Lipschitz
driving the trajectories of the uncertain dynamical system functions.
to the trajectories of the reference model. Depending on A considerable set of physical systems such as aerial
the nature of the system uncertainties, the difference and ground robots are either explicitly (or approxi-
between these trajectories can asymptotically vanish mately) satisfy the state-space form in equation 1. To
for time-invariant structured system uncertainties (see elucidate this point, consider the mass, spring, and
Case 1 in Figure 1 as an illustration of the asymptotic damper system given in Figure 2. In addition, con-
convergence notion) or the distance of this difference can sider that all the physical model parameters 𝛼, 𝛽, and
remain bounded in time with often an user-adjustable △ △
̇
m are unknown. Letting xp1 (t) = p(t), xp2 (t) = p(t), and
bound for time-varying structured system uncertain- △
ties or time-invariant/time-varying unstructured system xp (t) = [xp1 (t), xp2 (t)]T , the equations of motion of this

J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright © 2019 John Wiley & Sons, Inc.
DOI: 10.1002/047134608X.W1022.pub2
2 Model Reference Adaptive Control

the system uncertainty 𝛿p (xp (t)), one can consider either

Reference Model Trajectory structured or unstructured parameterizations. Revisiting
0.4 Unc. Dyn. Sys. Trajectory (Case 1) the mass, spring, and damper problem in equation 2 as
0.3 Unc. Dyn. Sys. Trajectory (Case 2) an example, the system uncertainty physically appears in
0.2
0.1 the form that an unknown weight matrix [−𝛼∕m, −𝛽∕m]
0 multiplies a known basis function [xp1 (t), x3p2 (t)]T . In this
–0.1
–0.2 case, since the basis function is available to the model ref-
–0.3 erence adaptive control designer, we say that the nature
–0.4
–0.5 of the system uncertainty is structured (and otherwise we
1 say that it is unstructured). In general, there is a consid-
0.5 60 erable set of physical systems that has structured system
50
0 40
–0.5 20
30 uncertainties, specifically the ones that can be modeled
–1 10 t (s ) by the first principles. Motivated from this standpoint, we
0
here consider that the system uncertainty in equation 1
Figure 1. For a given reference model trajectory, this figure illus- satisfies the following structured parameterization (see,
trates two responses of a model reference adaptive control archi- for example, the books cited earlier)
tecture for a dynamical system with time-invariant structured
system uncertainty (Case 1) and time-varying structured sys- 𝛿p (xp ) = WpT 𝜎p (xp ), xp ∈ ℝnp , (3)
tem uncertainty (Case 2). Specifically, in Case 1, observe that the
uncertain dynamical system trajectory asymptotically converges as in the mass, spring, and damper example, where Wp ∈
to the reference model trajectory. Furthermore, in Case 2, the ℝs×m denotes an unknown weight matrix and 𝜎p ∶ ℝnp → ℝs
uncertain dynamical system trajectory converges to a close neigh-
denotes a known basis function of the form
borhood around the reference model trajectory (i.e., this conver-
gence is not exact as the one in Case 1).
𝜎p (xp ) = [𝜎p1 (xp ), 𝜎p2 (xp ), … , 𝜎ps (xp )]T . (4)

p(t) This parameterization also captures exogenous constant

α disturbances (bias) by setting, for example, the first entry
m u(t) of the basis function to one (i.e., 𝜎p1 (xp ) = 1).
Having said that the system uncertainties can be
β unstructured in certain applications, for example, con-
sider a state-space form of an aerial robot performing an
Figure 2. A mass, spring, and damper system with a linear aggressive maneuver at high angles of attack. In this case,
spring force (i.e., fs = −𝛼p(t), 𝛼 ∈ ℝ+ ) and a nonlinear damping the system uncertainty affecting this robot may not be
force (i.e., fd = −𝛽 ṗ 3 (t), 𝛽 ∈ ℝ+ ) in a frictionless surface (p(t) ∈ ℝ modeled exactly as in equation 3 and a structured basis
denotes the position of the mass with coefficient m ∈ ℝ+ and u(t) ∈
function may not be known to the designer. In this case
ℝ denotes the control input).
and other similar situations, it is of practice to utilize
universal function approximators (e.g., neural networks)
system, mp(t) ̈ = −𝛼p(t) − 𝛽 ṗ 3 (t) + u(t), can be rewritten as to parameterize such unstructured system uncertainties,
[ ] [ ] for example, in the form 𝛿p (xp ) = WpT 𝜎papp (xp ) + 𝜖 app (xp ),
0 1 0
ẋ p (t) = xp (t) + [1∕m] u(t) xp ∈ xp ⊂ ℝnp , where Wp ∈ ℝs×m stands for an unknown
0 0 1 ⏟⏟⏟
⏟⏟⏟ ⏟⏟⏟ Λ weight matrix, 𝜎papp ∶ xp → ℝs stands for a known basis
Ap Bp function (e.g., composed of radial basis functions), and
[ ] [ ]T [ ] [ ] 𝜖papp ∶ xp → ℝm stands for an unknown residual approx-
0 −𝛼∕m xp1 (t) p(0) imation error (see, for example, References 9, 21–31 and
+ , xp (0) = , (2)
1 −𝛽∕m x3p2 (t) ̇
p(0) references therein). Note that xp is a compact subset of
⏟⏟⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏟⏟
Bp 𝛿p (xp (t)) xp
ℝnp .
0
Considering the uncertain dynamical system given by
which is clearly in the form given by equation 1. equation 1 subject to the time-invariant structured system
Mathematically speaking, considering the way that uncertainty parameterization given by equation 3, we now
the system uncertainty appears in equation 1, it is often state the model reference adaptive control objective: In the
said that 𝛿p (xp (t)) is a matched system uncertainty (i.e., presence of system anomalies, design a real-time control
since Λ is nonsingular, the control vector can access to this algorithm in order to drive the trajectories of the uncertain
uncertainty). This state-space form is widely adopted in dynamical system to the trajectories of a reference model
the model reference adaptive control literature (see, for (i.e., to desired closed-loop dynamical system trajectories
example, the books cited earlier). There are also several determined by the designer). Based on the discussion
works that consider the cases when the nature of the sys- related to Figure 1, since we consider a time-invariant
tem uncertainty is unmatched (see, for example, relatively structured system uncertainty here, one expects the differ-
recent References 16–20 and references therein). ence between both trajectories to vanish asymptotically.
Next, for model reference adaptive control purposes, In order to start addressing the stated model reference
it is important to parameterize the system uncertainty adaptive control problem, the next section first discusses
in equation 1. To this end, depending on the nature of the reference model selection.
 Model Reference Adaptive Control 3

3. REFERENCE MODEL Now, since the selection of the pair (K1 , K2 ) yields to a
desired level of closed-loop system behavior in the absence
Since a reference model captures a desired closed-loop of system anomalies, we select the reference model as
dynamical system behavior, one needs to define such
behavior in order to construct a reference model. In prac- ẋ r (t) = Ar xr (t) + Br c(t), xr (0) = xr0 , (9)
tice, there generally exists a static (see Section 3.1) or where xr (t) ∈ ℝnp stands for the reference model state
dynamic (see Section 3.2) nominal control law, which vector. Based on this reference model selection subject
achieves a desired closed-loop system behavior in the to an existing static nominal control law, one can now
absence of system anomalies. To this end, it is of practice construct the system error between xp (t) and xr (t) as
to select the reference model based on the properties of
these nominal control laws. We below discuss the selection △
e(t) = xp (t) − xr (t), e(t) ∈ ℝnp , (10)
of the reference model under two subsections that are
aligned with the above discussion. and develop a parameter adjustment mechanism to drive
the trajectories of the uncertain dynamical system to the
3.1. Reference Model Selection in the Presence of a Static trajectories of this reference model in the presence of
Nominal Control Law system anomalies (see Section 4).
Consider the feedback control law given by
3.2. Reference Model Selection in the Presence of a
u(t) = un (t) + ua (t), (5) Dynamic Nominal Control Law
where un (t) ∈ ℝm and ua (t) ∈ ℝm , respectively, stand for the Similar to the preceding subsection, consider the feedback
nominal and adaptive control laws. Moreover, consider a control law in equation 5. Furthermore, consider a gen-
static nominal control law in the form given by eral dynamic nominal control law in the form, for example,
given by
un (t) = −K1 xp (t) + K2 c(t). (6)
ẋ c (t) = Ac xc (t) + Bc 𝜙p (t), xc (0) = xc0 , (11)
In equation 6, K1 ∈ ℝm×np stands for a feedback gain matrix
and K2 ∈ ℝm×nc stands for a feedforward gain matrix. In zc (t) = Cc xc (t) + Dc 𝜙p (t), (12)
addition, c(t) ∈ ℝnc represents a given uniformly continu-
ous bounded command. Since the static nominal control 𝜙p (t) = Ep xp (t) − c(t), (13)
law given by equation 6 is designed by a control user in
order to achieve a desired closed-loop system behavior in un (t) = −K1⋆ xp (t) − K2⋆ zc (t), (14)
terms of stability and performance in the absence of sys-
tem anomalies, we now select the reference model based where Ac ∈ ℝpa ×pa , Bc ∈ ℝpa ×nc , Cc ∈ ℝpb ×pa , and Dc ∈ ℝpb ×nc
on this insight. stand for nominal control design matrices, Ep ∈ ℝnc ×np
Motivated from the above discussion, we first ignore allows to choose a subset of xp (t) to be followed by
system anomalies as (Λ, Wp ) = (I, 0) in equation 1 and a given bounded command c(t), and K1⋆ ∈ ℝm×np and
set ua (t) ≡ 0 in equation 5. In this case, the following K2⋆ ∈ ℝm×pb denote feedback gain matrices. Moreover,
expression follows from equation 1 xc (t) ∈ ℝpa represents the nominal control state vector,
zc (t) ∈ ℝpb represents the nominal control output vector,
ẋ p (t) = Ap xp (t) + Bp un (t) and 𝜙p (t) ∈ ℝnc represents the command following error
= Ap xp (t) + Bp (−K1 xp (t) + K2 c(t)) between zp (t) = Ep xp (t) and c(t).
Since the dynamic control law given by equations
= (Ap − Bp K1 ) xp (t) + Bp K2 c(t) 11–14 is designed by a control user in order to achieve a
⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ ⏟⏟⏟
desired closed-loop system behavior in the absence of sys-
Ar ∈ℝnp × np Br ∈ℝnp × nc
tem anomalies, in what follows we next set (Λ, Wp ) = (I, 0)
= Ar xp (t) + Br c(t). (7) in equation 1 and ua (t) ≡ 0 in equation 5 similar to the
previous subsection. Specifically, defining
Specifically, since system stability is the first requirement
in any control system design, one must choose the nominal △
x(t) = [xTp (t), xTc (t)]T ∈ ℝnp + pa
, (15)
control law gain K1 to make Ar in equation 7 is Hurwitz
(i.e., the spectrum of Ar is composed of eigenvalues with one can write in this case
△ [ ]
negative real parts). Furthermore, we let zp (t) = Ep xp (t), A − Bp K1⋆ − Bp K2⋆ Dc Ep −Bp K2⋆ Cc
where Ep ∈ ℝnc × np allows to select a subset of xp (t) to be ̇ = p
x(t) x(t)
Bc Ep Ac
followed by c(t). To this end, it is of practice to select K2 ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
such that A⋆
r ∈ℝ
(np +pa ) × (np +pa )

−Ep A−1
r Br = I, (8) [ ]
Bp K2⋆ Dc
+ c(t)
holds. With this selection, one can readily show that −Bc
zp (t) → c(t) as t → ∞ for constant commands (we refer ⏟⏞⏞⏞⏟⏞⏞⏞⏟
B⋆
r ∈ℝ
(np +pa )×nc
to the last paragraph of Section 3.2 for a discussion
addressing time-varying commands). = A⋆r x(t) + B⋆r c(t). (16)
4 Model Reference Adaptive Control

Once again, A⋆r must be Hurwitz here by the nominal con- command following performance (Chapter 3 in Refer-
trol law design and equation 16 yields to a desired level ence 9, Chapter 1 in Reference 32), which is standard in
of closed-loop system behavior in the absence of system the literature. The former method is desired when the
anomalies. commands are not necessarily known before the imple-
As a consequence, one can now select the reference mentation of the control law (e.g., for cases when an
model in the form given by online path planning algorithm generates the command
trajectories). The latter method is desired otherwise when
ẋ r (t) = A⋆r xr (t) + B⋆r c(t), xr (0) = xr0 , (17) the commands are generated by a known exosystem.
where xr (t) ∈ ℝnp +pa stands for the reference model state
vector. Based on this reference model selection subject to 4. PARAMETER ADJUSTMENT MECHANISM
an existing dynamic nominal control law, one can also con-
struct the system error as
Based on two common reference model selections given in
△ Section 3.1 and Section 3.2, we now discuss how to deter-
e(t) = x(t) − xr (t), e(t) ∈ ℝ np +pa
, (18)
mine the adaptive control law in equation 5 and derive
where its role will be clear in Section 4. a parameter adjustment mechanism in order to asymp-
Based on the general structure of the dynamic nominal totically drive the trajectories of the uncertain dynamical
control law given by equations 11–14, we now illustrate system to the trajectories of a selected reference model.
how a control designer can select the corresponding We begin our discussion by introducing the system error
matrices to obtain, for example, a proportional–integral dynamics, which considers a reference model chosen first
nominal control law that is well adopted in model refer- based on a static nominal control law (see Section 4.1)
ence adaptive control architectures. For this purpose, we and then based on a dynamic nominal control law (see
set Ac = 0, Bc = I, Cc = I, and Dc = 0, where this selection Section 4.2).
also gives pa = nc and pb = nc . The special structure of the
proportional–integral nominal control law now follows 4.1. System Error Dynamics in the Presence of a Static
from the expressions 11–14 as Nominal Control Law

ẋ c (t) = Ep xp − c(t), (19) Consider the uncertain dynamical system given by

equation 1 subject to the structured system uncertainty
un (t) = −Kx(t), (20) parameterization given by equation 3. In addition, con-
sider the feedback control law given by equation 5 subject
△ △
where K = [K1⋆ , K2⋆ ] ∈ ℝm×n and n = np + nc . In this case, to the static nominal control law given by equation 6.
the structures of A⋆r and B⋆r matrices used in the reference Then, one can write
model given by equation 17 have the form given by ẋ p (t) = Ap xp (t) + Bp Λ(un (t) + ua (t)) + Bp WpT 𝜎p (xp )
[ ]
A − Bp K1⋆ −Bp K2⋆ = Ap xp (t) + Bp un (t) + Bp Λ(I − Λ−1 )un (t)
A⋆r = p , (21)
Ep 0
+ Bp Λ(ua (t) + Λ−1 WpT 𝜎p (xp (t)))
[ ]
0 = Ar xp (t) + Br c(t) + Bp Λ(I − Λ−1 )un (t)
B⋆r = . (22)
−I
+ Bp Λ(ua (t) + Λ−1 WpT 𝜎p (xp (t)))
Furthermore, since A⋆r is considered to be a Hurwitz
matrix (under the rank condition = Ar xp (t) + Br c(t) + Bp Λ(ua (t) + W T 𝜎(⋅)), (25)
([ ]) where the aggregated system uncertainty matrix is
Ap Bp
rank = np + m, (23) denoted by
Ep 0
△
W = [Λ−1 WpT , (I − Λ−1 )]T , (26)
one can always make A⋆r
Hurwitz in the proportional–
integral nominal control law design since the pair (Ap , Bp ) and the aggregated basis function composed of known sig-
is assumed to be controllable (Chapter 13.4 in Reference nals is denoted by
9)), one can readily show that △
𝜎(xp (t), un (t)) = [𝜎pT (xp (t)), uTn (t)]T . (27)
−[Ep , 0]A⋆−1
r B⋆r =I (24)
In the above aggregated basis function, note that un (t) sat-
In other words, once the selection of the pair (K1⋆ , K2⋆ ) yields isfies equation 6. Based on equation 25 and the reference
to a Hurwitz A⋆r matrix, then zp (t) → c(t) as t → ∞ for con- model given by equation 9, the system error dynamics now
stant commands. follows from equation 10 as
To address time-varying commands, a control designer
̇ = ẋ p (t) − ẋ r (t)
e(t)
can pursue one of the following directions. First, the band-
width of the resulting reference model can be judiciously = Ar xp (t) + Br c(t) + Bp Λ(ua (t) + W T 𝜎(⋅))
adjusted to extract a desired level of time-varying com-
− Ar xr (t) − Br c(t)
mand following performance. Second, one can resort to
internal model principle to achieve a perfect time-varying = Ar e(t) + Bp Λ(ua (t) + W T 𝜎(⋅)). (28)
 Model Reference Adaptive Control 5

Now, since the goal of the adaptive control law is to sup- 4.2. System Error Dynamics in the Presence of a Dynamic
press system anomalies, it is of practice to choose it as Nominal Control Law
̂ T (t)𝜎(⋅),
ua (t) = −W (29) Consider the uncertain dynamical system given by
equation 1 subject to the structured system uncer-
where this selection is motivated by the term “ua (t) + tainty parameterization given by equation 3. In addition,
̂
W T 𝜎(⋅)” in equation 28. In equation 29, W(t) is an estimate consider the feedback control law given by equation 5
of the aggregated system uncertainty matrix satisfying subject to the dynamic nominal control law given by
a parameter adjustment mechanism to be defined below. equations 11–14. Then, following similar steps to the ones
As a consequence, the system error dynamics given by given in the preceding subsection, one can write
equation 28 can be equivalently written as
̇ = A⋆r x(t) + B⋆r c(t) + BΛ(ua (t) + W T 𝜎(⋅)),
x(t) (32)
̂
̇ = Ar e(t) − Bp Λ(W(t)
e(t)
T
− W) 𝜎(⋅)
△
⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟ where B = [BTp (t), 0]T . In equation 32, once again, the aggre-
̃
W(t)
gated system uncertainty matrix is denoted by equation 26
̃ T (t)𝜎(⋅).
= Ar e(t) − Bp ΛW (30) and the aggregated basis function composed of known sig-
nals is denoted by equation 27. In this aggregated basis
In practice, there can be situations that both Ap and function, note that un (t) satisfies equation 14. Based on
Bp in equation 1 are also unknown. If this is the case, equation 32 and the reference model given by equation 17,
then one may not readily construct a nominal control the system error dynamics now follows from equation 18
law since no prior information is available about the as
physical system; that is, un (t) ≡ 0 in equation 5. Yet, it is
still possible to construct the system error dynamics as ̇ = x(t)
e(t) ̇ − ẋ r (t)
in equation 30 when the following well-known matching
= A⋆r x(t) + B⋆r c(t) + BΛ(ua (t) + W T 𝜎(⋅))
conditions hold: There exist unknown matrices K1 and K2
such that Ar = Ap − Bp K1 and Br = Bp K2 hold. Note that − A⋆r xr (t) − B⋆r c(t)
these matching conditions only imply the knowledge of = A⋆r e(t) + BΛ(ua (t) + W T 𝜎(⋅)). (33)
the structure of Ap and Bp with some unknown parameter
entries in the sense that for a judiciously given Ar and Now, since the resulting system error dynamics is iden-
Br with Ar being Hurwitz and Br satisfying equation 8, tical in spirit to the one derived in the above subsection,
one can guarantee the existence of unknown matrices we choose the adaptive control law as equation 29, where
K1 and K2 to make both Ar = Ap − Bp K1 and Br = Bp K2 this allows one to equivalently write equation 33 as
satisfied. Now, under these matching conditions, one can
write e(t) ̂
̇ = A⋆r e(t) − BΛ(W(t)
T
− W) 𝜎(⋅)
⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
̇ = ẋ p (t) − ẋ r (t)
e(t) ̃
W(t)

= A⋆r e(t) ̃ T (t)𝜎(⋅).

− BΛW (34)
= Ar xp (t) + Br c(t) + Bp Λ(ua (t) + W T 𝜎(⋅))
− Ar xr (t) − Br c(t) Finally, following the discussion provided in the last para-
= Ar e(t) + (Ap − Ar )xp (t) + Bp Λ(u(t) graph of Section 4.1, we would like to highlight that it is
also possible here to write the system error dynamics as in
+ Λ−1 WpT 𝜎p (xp (t)) − Λ−1 K2 c(t)) equation 34 for the case when both Ap and Bp in equation 1
= Ar e(t) + Bp Λ(u(t) + W T 𝜎(⋅)), (31) are also unknown.

△
with W = [Λ−1 WpT , Λ−1 K1 , −Λ−1 K2 ]T being the aggregated 4.3. Derivation of the Parameter Adjustment Mechanism
system uncertainty that now also includes the unknown
△ In this subsection, we now derive a standard parameter
matrices K1 and K2 and 𝜎(xp (t), c(t)) = [𝜎pT (xp (t)), xTp (t), cT (t)]T
adjustment mechanism (see, for example, the books cited
being the corresponding aggregated basis function that ̂
earlier in this article) for W(t) used in the adaptive control
is again composed of known signals. Therefore, one can
̂ T (t)𝜎(⋅) in order to obtain the error signal given by equation 29 in order to asymptotically drive
choose u(t) = −W
the trajectories of the uncertain dynamical system to the
dynamics given by equation 30. For further details on
trajectories of a selected reference model. We begin by not-
the content provided in this paragraph, we refer to, for
ing that the system error dynamics given by equation 30
example, Section 9.5 of Reference 9. To summarize, one
based on a static nominal control law is identical in spirit
can write the system error dynamics given by equation 30
to the system error dynamics given by equation 34 based
whether Ap and Bp in equation 1 are known or they are
on a dynamic control law. Therefore, in what follows, we
unknown when the matching conditions hold for the latter
consider the system error dynamics given by
case. For this reason, we focus on the former case in this
article without loss of any generality and due to the fact e(t) ̃ T (t)𝜎(⋅),
̇ = Fe(t) − GΛW e(0) = e0 , (35)
that there generally exists a nominal control law designed
with some modeling knowledge in the feedback control of where the pair (F, G) = (Ar , Bp ) for the system error dynam-
physical systems subject to system anomalies. ics given by equation 30 and (F, G) = (A⋆r , B) for the system
6 Model Reference Adaptive Control

error dynamics given by equation 34. Note also that F is unless the closed-loop dynamical system is persistently
Hurwitz since either Ar or A⋆r is Hurwitz. In addition, F exciting (we refer to the books cited earlier on this
being Hurwitz implies that the Lyapunov equation given point). From a practical standpoint, equation 43 may
by not be necessary since equation 42 is sufficient to con-
0 = F T P + PF + I, (36) clude that an uncertain dynamical system of interest
can be asymptotically made to behave as a reference
holds for a unique positive-definite P (33, 34). model capturing a desired closed-loop dynamical system
Next, we utilize the Lyapunov function candidate in the performance.
form given by

(e, W) ̃ 12 )T (WΛ
̃ = eT Pe + 𝛾 −1 tr (WΛ ̃ 12 ). (37) 4.4. Role of the Learning Rate
[ ]
̃ 2 )T ⋅ (WΛ
Note that (0, 0) = eT Pe + 𝛾 −1 tr (WΛ
1
̃ 2)
1
= The positive scalar 𝛾 used in the parameter adjustment
̃
(e,W)=(0,0)
̃ > 0 for all (e, W)
̃ excluding (e, W) ̃ = (0, 0), and mechanism given by equation 39 is called the learning rate
0, (e, W)
̃ is radially unbounded. Moreover, 𝛾 ∈ ℝ used (adaptation gain) in the literature. Note that this scalar
(e, W) +
can be replaced by its positive-definite diagonal matrix ver-
in the last term of equation 37 is discussed in the next
sion provided that this matrix version is also properly used
subsection. The time derivative of this Lyapunov function
in the Lyapunov function candidate given by equation 37.
candidate is now calculated as
Specifically, the role of the learning rate in the model ref-
̇ = 2eT (t)Pe(t)
(⋅) ̇ + 2𝛾 −1 tr ΛW ̂̇
̃ T (t)W(t) erence adaptive control design is to control system error
transients.
̃ T (t)𝜎(⋅)
= 2eT (t)PFe(t) − 2eT (t)PGΛW To elucidate this point, note that
̂̇
̃ T (t)W(t)
+ 2𝛾 −1 tr ΛW
̃
(e(t), W(t)) ̃ )
≤ (e0 , W
̃ T (t)𝜎(⋅)
= eT (t)(F T P + PF)e(t) − 2eT (t)PGΛW
0

̃ Λ 12 )T (W
= eT0 Pe0 + 𝛾 −1 tr (W ̃ Λ 12 ),
+ 2𝛾 −1 tr ΛW ̂̇
̃ T (t)W(t) 0 0

(44)
= − eT (t)e(t) − 2tr ΛW ̃ T (t)𝜎(⋅)eT (t)PG
̇
̃ T (t)W(t)
̂ which follows from equation 40. For the ease of the
+ 2𝛾 −1 tr ΛW ̃ = W,
following discussion, consider that e0 = 0 and W 0
̃ T (t)(𝛾𝜎(⋅)eT (t)PG
= − ||e(t)||22 − 2𝛾 −1 tr ΛW where this initialization of the model reference adaptive
̂̇
− W(t)). (38)
controller yields

̂̇ ̃
(e(t), W(t))
1 1
≤ 𝛾 −1 tr (WΛ 2 )T (WΛ 2 ) (45)
Motivated by the term “𝛾𝜎(⋅)eT (t)PG − W(t)” in equation 38,
select the parameter adjustment mechanism as ̃
Now, using (e(t), W(t)) ≥ eT (t)Pe(t) ≥ 𝜆min (P)||e(t)||22 in
̂̇
W(t) = 𝛾𝜎(⋅)eT (t)PG, ̂
W(0) ̂
=W (39) equation 45 gives
0
1 −1 1
Using equation 39 in equation 38, one can now write ||e(t)||2 ≤ 𝛾 − 2 𝜆min
2
(P)||WΛ 2 ||F , (46)
̇ =
(⋅) −||e(t)||22 ≤ 0, (40) which shows that the system error is inversely propor-
tional to the learning rate. As a consequence, it can be
which guarantees that the system error e(t) and the weight made small during the transient time subject to a large
̃
estimation error W(t) are Lyapunov stable. This result fur- learning rate.
̃
ther means that the pair (e(t), W(t)) are bounded for all On the expression given by equation 46, one should
t ∈ ℝ+ . Since 𝜎(⋅) is bounded for all t ∈ ℝ+ in this case, it also make the following two observations for interested
̇
follows from equation 35 that e(t) is bounded. As a conse- readers. In particular, the first one is that the upper
quence, (⋅) is bounded for all t ∈ ℝ+ . It now follows from bound on the system error given by equation 46 may
Barbalat’s lemma (9, 33, 34) that get overly conservative as the matrix dimension of the
1
̇
lim (e(t), ̃
W(t)) = 0. (41) product WΛ 2 increases. The second one is that this
t→∞ upper bound includes the unknown pair (Λ, W), where
Therefore, this implies that it may not be a priori verifiable at
lim e(t) = 0, (42) the design stage especially in the absence of exces-
t→∞
sive numerical (e.g., Monte Carlo) studies and vehicle
which shows that the parameter adjustment mechanism testing. Therefore, it is of practical interest to study
given by equation 39 asymptotically drives the trajecto- how model reference adaptive control architectures
ries of the uncertain dynamical system to the trajectories can achieve a priori, user-defined performance guar-
of the selected reference model. This conclusion addresses antees for safety-critical applications. To this end, the
the model reference adaptive control problem. authors of, for example, References 35–38 (see also ref-
Finally, the above result does not necessarily imply erences therein) use a system-error-dependent learning
rate 𝛾(e(t)) instead of the constant one to address this
̃
lim W(t) = 0, (43) problem.
t→∞
 Model Reference Adaptive Control 7

4.5. Time-Varying System Uncertainties ̃ ⋆ ̃ 1 2

F ≤ k ||W(t)||F +
𝜃 ,
2
2𝜃||W(t)|| (53)
k⋆
In the presence of time-varying system uncertainties, that
is, when 𝛿p (xp (t)) is replaced with 𝛿p (t, xp (t)) in equation 1, where the latter is an application of Young’s inequality
we consider the structured parameterization given by with an arbitrary scalar k⋆ ∈ ℝ+ . For mathematical con-
venience, we here set this scalar to k⋆ = 𝛼𝜆min (Λ). As a
𝛿p (t, xp ) = WpT (t)𝜎p (xp ), xp ∈ ℝnp , (47) consequence, the following expression now follows from
equation 50
instead of the one in equation 3, where it is generally
assumed that ||Wp (t)||F ≤ wp and ||Ẇ p (t)||F ≤ ẇ p with (⋅) ̃
̇ ≤ −||e(t)||2 − 𝛼𝜆 (Λ)||W(t)|| 2 𝜃2
. (54)
F +
wp ∈ ℝ+ and ẇ p ∈ ℝ+ (it is possible to remove the latter
2 min
𝛼𝜆min (Λ)
assumption for applications involving sudden change ̇ ≤ 0 outside the com-
Note that equation 54 implies (⋅)
in dynamics; see, for example, References 28, 39). This pact set
parameterization now captures exogenous (time-varying)
disturbances by setting, for example, the first entry of ⎧ ⎫
the basis function to one (i.e., 𝜎p1 (xp ) = 1). In this case, ⎪ ⎪
△⎪ 𝜃 ⎪
̃
 = ⎨(e(t), W(t)) ∶ ||e(t)||2 ≤
only the definition of the aggregated system uncertainty 1 ⎬ ∩ {(e(t),
matrix in equation 26 changes to ⎪ (𝛼𝜆min (Λ)) 2 ⎪
⎪ ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟⎪
△ ⎩ 𝜁1 ⎭
W(t) = [Λ−1 WpT (t), (I − Λ−1 )]T , (48)
⎫
and the system error dynamics given by equation 35 ⎪
remains the same. Note here that ||Wp (t)||F ≤ wp and ̃ ̃ 𝜃 ⎪
W(t)) ∶ ||W(t)||F ≤ . (55)
̇
||Ẇ p (t)||F ≤ ẇ p imply ||W(t)||F ≤ w and ||W(t)|| ̇ with
F ≤ w
𝛼𝜆min (Λ) ⎬
⎪
w ∈ ℝ+ and ẇ ∈ ℝ+ . ⏟⏞⏞⏟⏞⏞⏟⎪
𝜁2 ⎭
For ensuring robustness of model reference adaptive
control designs subject to time-varying system uncertain- Therefore, the evolution of (⋅) is upper bounded by
ties, it is of practice to modify the parameter adjustment
̂ (⋅) ≤ max (⋅)
mechanism in equation 39 with a leakage term −𝛼 W(t), ̃
(e(t),W(t))∈
𝛼 ∈ ℝ+ (called 𝜎-modification term (40)), with a sys-
̂ = 𝜆max (P)𝜁12 + 𝛾 −1 𝜆max (Λ)𝜁22
tem error-dependent leakage term −𝛽||e(t)||2 W(t), 𝛽 ∈ ℝ+ ( )
(called e-modification term (41)), or by resorting to a pro- 𝜆max (P) 𝛾 −1 𝜆max (Λ)
=𝜃 2
+ 2 2 (56)
jection operator (42). For example, with the leakage term 𝛼𝜆min (Λ) 𝛼 𝜆min (Λ)
̂
−𝛼 W(t) added to equation 39, the modified parameter since (⋅) cannot grow outside . This establishes the
adjustment mechanism becomes ̃
boundedness of the pair (e(t), W(t)).
̂̇
W(t) ̂
= 𝛾(𝜎(⋅)eT (t)PG − 𝛼 W(t)). (49) Using 𝜆min (P)||e(t)||22 ≤ (⋅) in equation 56, one can
obtain the bound of the system error for t ≥ T as
Note that all these modification terms are also effective 1

tools for ensuring robustness of model reference adaptive (max(e(t),W(t))∈

̃ (⋅)) 2
||e(t)||2 ≤ 1
control designs subject to not only time-varying structured (𝜆min (P)) 2
system uncertainties but also time-invariant/time-varying ( )1
unstructured system uncertainties. 𝜃 𝜆max (P) 𝛾 −1 𝜆max (Λ) 2
= + 2 2
(𝜆min (P)) 2 𝛼𝜆min (Λ) 𝛼 𝜆min (Λ)
1
Based on the discussion related to Figure 1, one now
expects the boundedness of the system error dynamics −1 ̇
(
given by equation 35 subject to the modified parameter ||Λ||F (𝛼w + 𝛾 w) 𝜆max (P)
=
adjustment mechanism given by equation 49 in the case of (𝜆 (P)) 2
1
𝛼𝜆min (Λ)
min
time-varying system uncertainties. To elucidate this point, )1
the time derivative of the Lyapunov function candidate in 𝛾 −1 𝜆 (Λ) 2
+ 2 2max = 𝜁 ⋆, (57)
equation 37 can be calculated as 𝛼 𝜆min (Λ)
̇ = − ||e(t)||2 − 2𝛼 trΛW
(⋅) ̃ T (t)W(t)
̂ where this bound 𝜁 ⋆ can be made small by the model refer-
2
̃ T (t)W(t)
̇ ence adaptive control designer through judiciously tuning
− 2𝛾 −1 trΛW
the parameter pair (𝛾, 𝛼) (e.g., see Figure 3). The above
̃ T (t)(W(t)
= − ||e(t)||22 − 2𝛼 trΛW ̃ + W(t)) analysis can be also used for the system-error-dependent
̃ T (t)W(t)
̇ leakage term and the projection operator with appropriate
− 2𝛾 −1 trΛW
changes.
̃
≤ − ||e(t)||22 − 2𝛼||W(t)Λ
1
̃
2 || + 2𝜃||W(t)||
F,
2
F (50) Finally, we note that the above modification terms are
also popular system-theoretical tools to achieve closed-
where loop system stability when the uncertain dynamical
△
̇
𝜃 = ||Λ||F (𝛼w + 𝛾 −1 w). (51) system given by equation 1 has added perturbation terms.
To elucidate this point in a simple setting, consider, for
In equation 50, note that
example, the dynamics of a physical system represented
̃ 1
2 || ≤ −2𝛼𝜆 ̃ in the following state-space form
min (Λ)||W(t)||F ,
2 2
−2𝛼||W(t)Λ F (52)
8 Model Reference Adaptive Control

4 5.5
γ=1 γ=1
γ = 2.5 γ = 2.5
3.5 γ=5 5
γ=5
γ = 10 γ = 10
γ = 50 4.5 γ = 50
3 γ = 1000 γ = 1000

2.5 4

3.5
ζ⋆

⋆
ζd
2

1.5 3

1 2.5

0.5 2

0 1.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
α α

Figure 3. An illustration of the system error bound 𝜁 ⋆ given Figure 4. An illustration of the system error bound 𝜁d⋆ given by
by equation 57 as a function of the parameter pair (𝛾, 𝛼) with equation 64 as a function of the parameter pair (𝛾, 𝛼) with d⋆ = 1,
w = 1, ẇ = 1, 𝜆min (P) = 1, 𝜆max (P) = 1, ||Λ||F = 1, 𝜆min (Λ) = 1, and w = 1, ẇ = 1, 𝜆min (P) = 1, 𝜆max (P) = 1, ||Λ||F = 1, 𝜆min (Λ) = 1, and
𝜆max (Λ) = 1. 𝜆max (Λ) = 1.

̇ ≤ 0 outside the compact set

where (⋅)
ẋ p (t) = Ap xp (t) + Bp Λu(t) + Bp 𝛿p (t, xp (t)) + d(t), △
̃
 ′ = {(e(t), W(t)) ̃
∶ ||e(t)||2 ≤ 𝜁1′ } ∩ {(e(t), W(t)) ∶
xp (0) = xp0 , (58)
̃
||W(t)|| F ≤ 𝜁2 },
′
(63)
where the perturbation term d(t) ∈ ℝ is added under the np
△ 1 △
assumption ||d(t)||2 ≤ d⋆ , d⋆ ∈ ℝ+ . Here, it is considered
2 2
with 𝜁1′ = ( 𝛼𝜆 2𝜃 (Λ) + 4𝜆2max (P)(d⋆ )2 ) 2 and 𝜁2′ =( 𝛼𝜆 𝜃 + 2𝜆2max
min (Λ)
that this perturbation term is unmatched (i.e., BTp d(t) = 0). 1
min
1

Otherwise, it can be included in 𝛿p (t, xp (t)) based on the (P)(d⋆ )2 ) ∕(𝛼𝜆min (Λ)) . Following the previous analysis
2 2

above discussion. In this case, one now has the system steps identically, one can now establish the boundedness
̃
of the pair (e(t), W(t)), where the bound of the system error
error dynamics given by
for t ≥ T is given by
e(t) ̃ T (t)𝜎(⋅) + d (t),
̇ = Fe(t) − GΛW e(0) = e0 , (59)
u ( )1
𝜆max (P) ′ 2 𝜆 (Λ) ′ 2 2

where (F, G, du (t)) = (Ar , Bp , d(t)) in the presence of the ||e(t)||2 ≤ (𝜁 ) + 𝛾 −1 max (𝜁 ) = 𝜁d⋆ .
𝜆min (P) 1 𝜆min (P) 2
static nominal control law and (F, G, du (t)) = (Ar , Bp , Dd(t))
(64)
in the presence of the dynamic nominal control law with
△
D = [I, 0]T . For both cases, we note that ||du (t)||2 ≤ d⋆ , Unlike the previous bound 𝜁 ⋆ given by equation 57,
which follows from ||d(t)||2 ≤ d⋆ . which is derived in the absence of the considered pertur-
Based on the modified parameter adjustment mecha- bation term d(t), the new bound 𝜁d⋆ given by equation 64
nism given by equation 49, the expression in equation 54 may not be made small by the model reference adaptive
now becomes control designer (e.g., see Figure 4) since the term d(t)
𝜃2 is considered to be unmatched. To address this issue
(⋅) ̃
̇ ≤ − ||e(t)||2 − 𝛼𝜆 (Λ)||W(t)|| 2
especially when d⋆ is not sufficiently small, we refer to,
F +
2 min
𝛼𝜆min (Λ)
for example, Reference 20. Once again, the above analysis
+ 2eT (t)Pdu (t). (60) guaranteeing closed-loop system stability can be also used
for the system-error-dependent leakage term and the
For the last term on the right-hand side of equation 60, one
projection operator with appropriate changes.
can now obtain an upper bound given by

2eT (t)Pdu (t) ≤ 2𝜆max (P)d⋆ ||e(t)||2

5. DISCUSSIONS
≤ 5||e(t)||22 + 2𝜆2max (P)(d⋆ )2 , (61)

where Young’s inequality is used. Using equation 61 in In this section, we first present a model reference adap-
equation 60, we can then write tive control design example on the mass, spring, and
damper system shown in Figure 2 (Section 5.1) and then
̃
̇ ≤ − 5||e(t)||2 − 𝛼𝜆 (Λ)||W(t)|| 𝜃2
(⋅) 2
F + make connections to several other available model ref-
2 min
𝛼𝜆min (Λ) erence adaptive control methods for interested readers
+ 2𝜆2max (P)(d⋆ )2 , (62) (Section 5.2).
 Model Reference Adaptive Control 9

5.1. Illustrative Numerical Example adaptive control design for three learning rates. Specifi-
cally, when 𝛾 = 0, a desirable command following perfor-
Consider the mass, spring, and damper system given
mance could not be obtained since this case corresponds to
by equation 2 with a linear spring force fs = −𝛼p(t) and
ua (t) ≡ 0 here. When we set 𝛾 = 0.5, a desirable command
a nonlinear damping force fd = −𝛽 ṗ 3 (t) in a frictionless
following performance is achieved around t = 70 s. When
surface; see Figure 2. In addition, consider that all the
this learning rate is increased to 𝛾 = 5.0, a desirable com-
physical model parameters 𝛼, 𝛽, and m are unknown
mand following performance is achieved around t = 25 s.
(we set these unknown parameters to 𝛼 = 1, 𝛽 = 1, and
For cases when 𝛾 ≠ 0, the designed parameter adjustment
m = 10 in this numerical study). Based on the known pair
mechanism asymptotically drives the trajectories of the
(Ap , Bp ) of this physical system, we design a static nominal
considered uncertain dynamical system to the trajectories
control law in the form given by equation 6, where we set
of the selected reference model. While we consider here a
K1 = [0.250, 1.000] and K2 = [0.250]. Note that this selec-
static nominal-control-law-based adaptive control design,
tion of the static nominal control law gains yields to a ref-
the same conclusion can be easily illustrated with an adap-
erence model response captured by equation 9 in the form
[ ] [ ] tive control design using a dynamic nominal control law.
0 1 0
ẋ r (t) = xr (t) + c(t), (65)
−0.250 −1.000 0.250
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏟⏞⏟ 5.2. Further Reading
Ar Br
We begin by noting that several topics presented here
which has a natural frequency of 0.5 rad s−1 and a damping in detail partially build on the model reference adaptive
ratio of 1.0. Furthermore, with Ep = [1, 0] in equation 65, control overview sections of the author documented in, for
this reference model design guarantees equation 8. We example, References 20, 28, 37, 38, 43–51. In addition, we
also choose a filtered square-wave reference command c(t). would like to refer to the recent book (9) as well as other
For executing the parameter adjustment mechanism books (1–8) for reading related to the topics covered in this
given by equation 39 for the adaptive control law in article on model reference adaptive control theory. For the
equation 29, we solved the Lyapunov equation given by nonlinear stability analysis tools and methods adopted
equation 36 as in this article, we also refer to the excellent books (33,
[ ] 34).
2.625 2.000 While the goal of this article is to provide an intro-
P= . (66)
2.000 2.500 duction to the model reference adaptive control design
Moreover, in this numerical study, all initial conditions procedure in a basic setting, we also would like to make
are set to zero. Figure 5 presents the above model reference connections to several other, relatively advanced methods
for interested readers. Once again, this is not intended to
be a survey article, and the following list only presents a
1.5
1
xr1(t) fraction of available related literature from the viewpoint
xp1(t), γ = 0
of the author (see, for example, relatively recent surveys
xr1(t), xp1(t)

0.5 xp1(t), γ = 0.5

0 xp1(t), γ = 5.0 (10–13) for a detailed literature review as well as the
–0.5 aforementioned books). In particular, for the purpose of
–1 achieving stringent transient performance characteristics,
–1.5 a model reference adaptive control designer may resort
0 10 20 30 40 50 60 70 80 90 100
to high learning rates as noted in Section 4.4. However,
t (s) as it is known, parameter adjustment mechanisms with
0.8
xr2(t)
0.6 xp2(t), γ = 0
high learning rates may yield to signals with high-
xr2(t), xp2(t)

0.4 xp2(t), γ = 0.5 frequency content (e.g., compare the oscillations between
xp2(t), γ = 5.0
0.2 𝛾 = 0.5 and 𝛾 = 5.0 cases in Figure 5), which can result
0 in system instability due to, for example, the presence
–0.2 of unmodeled system dynamics (46, 52, 53). To this end,
–0.4
0 10 20 30 40 50 60 70 80 90 100
the authors of References 5, 43, 44, 46, 54–64 (also see
references therein) make notable contributions to the field
t (s)
4
u(t), γ = 0
of model reference adaptive control theory, where their
2
u(t), γ = 0.5 algorithms have the ability to suppress high-frequency
u(t), γ = 5.0
oscillations.
u(t)

0 From the robustness point of view, important practical

–2 developments in model reference adaptive control theory
is the consideration of the presence of actuator (amplitude
–4
0 10 20 30 40 50 60 70 80 90 100 saturation, rate saturation, and/or dynamics) limitations
t (s) and unmodeled dynamics in feedback loops. We refer to
the contributions documented in References 48, 65–82 on
Figure 5. Command following performance with the model ref- the former topic and References 7, 39, 50, 52, 83–97 on
erence adaptive control design in Section 5.1 for the mass, spring, the latter topic (also see references therein). Furthermore,
and damper system given by equation 2, where it is considered for applications where parameter convergence equation
that all the physical model parameters 𝛼, 𝛽, and m are unknown. 43 is desired under a relaxed persistency of excitation
10 Model Reference Adaptive Control

condition, we refer to relatively recent contributions in 5. N. Hovakimyan and C. Cao. 1 Adaptive Control Theory:
References 30, 98–101 and references therein. For adap- Guaranteed Robustness with Fast Adaptation. SIAM, 2010.
tive control architectures that employ a gain-scheduled, 6. S. Sastry and M. Bodson. Adaptive Control: Stability, Con-
time-varying, and/or nonlinear reference models, we vergence and Robustness. Courier Corporation, 2011.
also refer to, for example, notable research results in 7. K. S. Narendra and A. M. Annaswamy. Stable Adaptive Sys-
References 102–112. tems. Courier Corporation, 2012.
As noted earlier, Sections 2, 3, and 4 of this article 8. K. J. Åström and B. Wittenmark. Adaptive Control. Courier
focus on a state feedback model reference adaptive con- Corporation, 2013.
trol design for state tracking. For important research 9. E. Lavretsky and K. A. Wise. Robust and Adaptive Control
contributions that focus on state feedback model ref- with Aerospace Applications. Springer: London, 2013.
erence adaptive control design for output tracking and 10. K. A. Wise, E. Lavretsky, and N. Hovakimyan. Adaptive
output feedback model reference adaptive control design Control of Flight: Theory, Applications, and Open Problems,
for output tracking, we refer to not only to the books in American Control Conference, 2006; IEEE: Minneapolis,
cited above but also, for example, the relatively recent Minnesota, USA, 2006, p 6.
contributions in References 9, 17, 63, 113–121. Finally, 11. B. D. Anderson and A. Dehghani. Ann. Rev. Control 2008,
with regard to emerging developments in model reference 32(2), pp 123–135.
adaptive control theory, we refer to References 122–124 12. W. S. Black, P. Haghi, and K. B. Ariyur, Adaptive Systems:
for secure cyber-physical systems applications, References History, Techniques, Problems, and Perspectives, Systems
125–132 for control applications over wireless networks 2014, 2(4), pp 606–660.
with reduced communications, References 51, 133 for 13. G. Tao. Automatica 2014, 50(11), pp 2737–2764.
applications involving human-control interactions, and 14. H. P. Whitaker, J. Yamron, and A. Kezer. Design of Model
References 134–144 for cooperative control of multiagent Reference Control Systems for Aircraft. Instrumenta-
systems (also see references therein). tion Laboratory, Massachusetts Institute of Technology:
Cambridge, MA, 1958.
15. P. V. Osburn, H. P. Whitaker, and A. Kezer. New Develop-
ACKNOWLEDGMENT ments in the Design of Adaptive Control Systems. Institute
of Aeronautical Sciences, 1961.
16. A. J. Koshkouei and A. S. Zinober. IFAC Proc. Vol. 2000,
The author would like to thank his related research collab-
33(14), pp 167–172.
orators (Sivasubramanya Balakrishnan, Eric A. Butcher,
17. T. Yucelen and W. M. Haddad. Int. J. Control 2012, 85(6),
Anthony J. Calise, Animesh Chakravarthy, Girish Chowd-
706–721.
hary, Gerardo De La Torre, Eric Feron, Eric N. Johnson,
Mario L. Fravolini, Frank Fresconi, Wassim M. Haddad, 18. N. T. Nguyen. Multi-objective Optimal Control Modification
Adaptive Control Method for Systems with Input and
Ali T. Kutay, Kilsoo Kim, Frank Lewis, Jonathan A. Muse,
Unmatched Uncertainties, in Proc. of AIAA Guidance,
Morad Nazari, Nhan Nguyen, Mehrdad Pakmehr, Farzad
Navigation, and Control Conference, 2014; pp 1–15.
Pourboghrat, Simon Schatz, Jeff Shamma, Rifat Sipahi,
19. C. D. Heise and F. Holzapfel. Uniform Ultimate Bounded-
James Steck, Bong-Jun Yang, and Yildiray Yildiz) and
ness of a Model Reference Adaptive Controller in the Pres-
graduate students (Ali Albattat, Ehsan Arabi, K. Merve
ence of Unmatched Parametric Uncertainties, in Automa-
Dogan, Benjamin C. Gruenwald, S. Burak Sarsilmaz, tion, Robotics and Applications (ICARA), 2015 6th Interna-
Daniel Wagner, and Kevin Wilcher) for many fruitful tional Conference on; IEEE, 2015; pp 149–154.
discussions on model reference adaptive control theory as 20. E. Arabi, T. Yucelen, and B. C. Gruenwald. In Robotics and
well as its applications. Mechatronics for Agriculture; Zhang, D.; Wei, B., Eds.; CRC
Press, 2017.
21. K. S. Narendra and K. Parthasarathy. IEEE Trans. Neural
RELATED ARTICLES Networks 1990, 1(1), pp 4–27.
22. F.-C. Chen and H. K. Khalil. Int. J. Control 1992, 55(6), pp
Adaptive Control; Intelligent Control; Nonlinear Control 1299–1317.
Systems, Analytical Methods; Nonlinear Control Systems, 23. R. M. Sanner and J.-J. E. Slotine. IEEE Trans. Neural Net-
Design Methods; Stability Theory, Nonlinear; Robust Con- works 1992, 3(6), pp 837–863.
trol. 24. A. Yeşildirek and F. L. Lewis. Automatica 1995, 31(11), pp
1659–1664.

BIBLIOGRAPHY 25. B. S. Kim and A. J. Calise. J. Guid. Control Dynam. 1997,

20(1), pp 26–33.
1. P. V. Kokotović. Foundations of Adaptive Control. Springer- 26. A. J. Calise and R. T. Rysdyk. IEEE Control Syst. Mag. 1998,
Verlag: New York 1991. 18(6), pp 14–25.
2. M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic. Nonlin- 27. M. L. Fravolini and G. Campa. IEEE Trans. Neural Net-
ear and Adaptive Control Design. Wiley: New York, 1995. works 2011, 22(4), pp 627–638.
3. G. Tao. Adaptive Control Design and Analysis, vol. 37. John 28. T. Yucelen and A. J. Calise. J. Guid. Control Dynam. 2011,
Wiley & Sons, 2003. 34(4), p 933.
4. P. Ioannou and B. Fidan. Adaptive Control Tutorial. Tech. 29. A. J. Calise and T. Yucelen. J. Guid. Control Dynam. 2012,
Rep., 2006. 35(3), pp 807–815.
 Model Reference Adaptive Control 11

30. G. Chowdhary, H. A. Kingravi, J. P. How, and P. A. Vela. American Control Conference; IEEE: Baltimore, Maryland,
IEEE Trans. Neural. Netw. Learn. Syst. 2015, 26(3), pp USA, 2010, pp 93–98.
537–550. 56. V. Stepanyan and K. Krishnakumar. M-mrac for Nonlinear
31. E. Arabi, B. Gruenwald, T. Yucelen, M. Fravolini, and N. Systems with Bounded Disturbances, in IEEE Conference
Nguyen. Model Reference Neuroadaptive Control Revisited: on Decision and Control; IEEE: Orlando, Florida, USA,
How to Keep the System Trajectories on a Given Compact 2011, pp 5419–5424.
Set, in AIAA Guidance, Navigation, and Control Conference 57. E. Lavretsky. Reference Dynamics Modification in Adaptive
and Exhibit, 2017. Controllers for Improved Transient Performance, in AIAA
32. J. Huang. Nonlinear Output Regulation: Theory and Appli- Guidance, Navigation, and Control Conference, 2011; pp
cations, vol. 8; SIAM, 2004. 1–13.
33. H. K. Khalil. Nonlinear Systems. Prentice Hall: Upper Sad- 58. V. Stepanyan and K. Krishnakumar. J. Guid. Control
dle River, New Jersey, 2002. Dynam. 2012, 35(4), pp 1370–1374.
34. W. M. Haddad and V. Chellaboina. Nonlinear Dynamical 59. T. E. Gibson, A. M. Annaswamy, and E. Lavretsky. Adap-
Systems and Control: A Lyapunov-based Approach. Prince- tive Systems with Closed-loop Reference Models: Stability,
ton University Press, 2008. Robustness and Transient Performance, arXiv preprint
arXiv:1201.4897, 2012.
35. J. A. Muse. A Method for Enforcing State Constraints in
Adaptive Control, in AIAA Guidance, Navigation and Con- 60. N. Harl, K. Rajagopal, and S. Balakrishnan. J. Guid. Control
trol Conference, 2011. Dynam. 2013, 36(4), pp 1194–1209.
36. T. Yucelen and J. S. Shamma. Adaptive Architectures for 61. Y. Yang, S. Balakrishnan, L. Tang, and R. G. Landers.
Distributed Control of Modular Systems, in American Con- IEEE/ASME Transactions On Mechatronics 2013, 18(3),
trol Conference, 2014; pp 1328–1333. pp 867–877.
37. E. Arabi, B. C. Gruenwald, T. Yucelen, and N. T. Nguyen. 62. V. Pappu, J. E. Steck, K. Rajagopal, and S. Balakrishnan.
Int. J. Control 2018, 91(5), pp 1195–1208. Modified State Observer Based Adaptation of a General Avi-
ation Aircraft: Simulation and Flight Test, in Proceedings
38. E. Arabi and T. Yucelen. A Generalization to Set-Theoretic
of the AIAA Guidance, Navigation and Control Conference,
Model Reference Adaptive Control Architecture for Enforc-
2014; pp 13–17.
ing User-defined Time-varying Performance Bounds, in
American Control Conference, 2017; pp 5077–5082. 63. T. Rajpurohit, W. M. Haddad, and T. Yucelen. J. Guid. Con-
trol Dynam. 2015, 39(1), pp 16–31.
39. T. Yucelen and A. J. Calise, J. Guid. Control Dynam. 2014,
37(5), pp 1583–1594. 64. T. Yucelen and S. Balakrishnan. Adaptive Set-Theoretic
Emulator Reference Architecture (ASTERA): Control of
40. P. A. Ioannou and P. V. Kokotovic. Automatica 1984, 20(5),
Uncertain Dynamical Systems with Performance Guaran-
pp 583–594.
tees and Smooth Transients, in Proceedings of the AIAA
41. K. Narendra and A. Annaswamy. IEEE Trans. Autom. Con- Guidance, Navigation and Control Conference, 2018.
trol 1987, 32(2), pp 134–145. 65. S. P. Karason and A. M. Annaswamy. Adaptive Control in
42. J.-B. Pomet and L. Praly. IEEE Trans. Autom. Control 1992, the Presence of Input Constraints, in American Control Con-
37(6), pp 729–740. ference, 1993. IEEE: San Francisco, California, USA, 1993;
43. T. Yucelen and W. M. Haddad. Int. J. Adapt Control Signal pp 1370–1374.
Process. 2012, 26(11), pp 1024–1055. 66. A. M. Annaswamy and S. Karason. Automatica 1995, 31(10),
44. T. Yucelen and W. M. Haddad. IEEE Trans. Autom. Control pp 1421–1431.
2013, 58(4), pp 1080–1085. 67. C.-Y. Su and Y. Stepanenko. Int. J. Adapt Control Signal
45. T. Yucelen and E. Johnson. Int. J. Adapt Control Signal Pro- Process. 1997, 11(2), pp 141–153.
cess. 2013, 27(12), pp 1065–1085. 68. M. R. Akella, J. L. Junkins, and R. D. Robinett. Structured
46. T. Yucelen, G. De La Torre, and E. N. Johnson. Int. J. Control Model Reference Adaptive Control with Actuator Satura-
2014, 87(11), pp 2383–2397. tion Limits, in Proc. AIAA/AAS Astrodynam. Specialist
47. B. Gruenwald and T. Yucelen. Int. J. Control 2015, 88(11), Conf. Exhibit, Collection of Tech. Papers (A98-37348), 1998;
pp 2305–2315. pp 10–13.
69. E. N. Johnson. Limited Authority Adaptive Flight Control.
48. B. C. Gruenwald, D. Wagner, T. Yucelen, and J. A. Muse, Int.
PhD dissertation, Georgia Institute of Technology, 2000.
J. Control 2016, 89(12), pp 2434–2452.
70. E. N. Johnson and A. J. Calise. Neural Network Adaptive
49. B. C. Gruenwald, T. Yucelen, and J. A. Muse, Machines 2017,
Control of Systems with Input Saturation, in American
5(1), p 9.
Control Conference. Proceedings of the 2001, vol. 5; IEEE:
50. K. M. Dogan, B. C. Gruenwald, T. Yucelen, and J. A. Muse. Arlington, Virginia, USA, 2001; pp 3527–3532.
Int. J. Control 2018, 91(8), pp 1774–1784.
71. E. N. Johnson. J. Guid. Control Dynam. 2003, 26(6), pp
51. T. Yucelen, Y. Yildiz, R. Sipahi, E. Yousefi, and N. Nguyen, 906–913.
Int. J. Control 2018, 91(10), pp 2314–2331 (just-accepted).
72. M. Schwager, A. M. Annaswamy, and E. Lavretsky.
52. C. Rohrs, L. Valavani, M. Athans, and G. Stein. IEEE Trans. Adaptation-based Reconfiguration in the Presence of
Autom. Control 1985, 30(9), pp 881–889. Actuator Failures and Saturation, in American Control
53. Z. T. Dydek, A. M. Annaswamy and E. Lavretsky. IEEE Con- Conference, 2005; pp 2640–2645.
trol Systems 2010, 30(3), pp 32–48. 73. C. Cao and N. Hovakimyan. 1 Adaptive Controller for Sys-
54. J. A. Muse. An ∞ Norm Minimization Approach. PhD dis- tems in the Presence of Unmodelled Actuator Dynamics, in
sertation, Georgia Institute of Technology, 2010. IEEE Conference on Decision and Control, 2007.
55. V. Stepanyan and K. Krishnakumar. Mrac Revisited: Guar- 74. E. Lavretsky and N. Hovakimyan. Systems & control letters
anteed Performance with Reference Model Modification, in 2007, 56(11), pp 722–729.
12 Model Reference Adaptive Control

75. A. Leonessa, W. M. Haddad, T. Hayakawa, and Y. Morel. Int. Switchings and Unmodeled Dynamics, IEEE Transactions
J. Adapt Control Signal Process. 2009, 23(1), pp 73–96. on Cybernetics, 2016.
76. C. Wen, J. Zhou, Z. Liu, and H. Su. IEEE Trans. Autom. Con- 96. K. M. Dogan, T. Yucelen, B. C. Gruenwald, and J. A.
trol 2011, 56(7), pp 1672–1678. Muse. On Model Reference Adaptive Control for Uncertain
77. G. De La Torre, T. Yucelen, and E. Johnson. Reference Con- Dynamical Systems with Unmodeled Dynamics, in IEEE
trol Architecture in the Presence of Measurement Noise and 55th Conference on Decision and Control; IEEE: Las Vegas,
Actuator Dynamics, in American Control Conference, 2014. USA, 2016; pp 377–382.
78. J. A. Muse. Frequency Limited Adaptive Control Using a 97. K. M. Dogan, B. C. Gruenwald, T. Yucelen, and J. A. Muse.
Quadratic Stability Framework: Guaranteed Stability Lim- On the Stability of Adaptive Control Systems in the Pres-
its, in AIAA Guidance, Navigation, and Control Conference, ence of Control and State Dependent Unmodeled Dynam-
2014. ics, in AIAA Guidance, Navigation, and Control Conference,
79. B. C. Gruenwald, T. Yucelen, J. A. Muse, and D. Wagner. 2017.
Computing Stability Limits of Adaptive Controllers in the 98. G. Chowdhary and E. Johnson. Concurrent Learning for
Presence of High-order Actuator Dynamics, in IEEE Con- Convergence in Adaptive Control without Persistency of
ference on Decision and Control, 2016. Excitation, in IEEE Conference on Decision and Control,
80. B. C. Gruenwald, J. A. Muse, and T. Yucelen. Am. Control 2010; pp 3674–3679.
Conf. 2017, 4430–4435. 99. G. Chowdhary, T. Yucelen, M. Mühlegg, and E. N. John-
81. B. C. Gruenwald, K. M. Dogan, T. Yucelen, and J. A. son. Int. J. Adapt Control Signal Process. 2013, 27(4), pp
Muse. Model Reference Adaptive Control Framework for 280–301.
Uncertain Dynamical Systems with High-order Actuator 100. G. Chowdhary, M. Mühlegg, and E. Johnson. Int. J. Control
Dynamics and Unknown Actuator Outputs, in ASME 2014, 87(8), pp 1583–1603.
Dynamic Systems and Control Conference, 2017. 101. S. B. Roy, S. Bhasin, and I. N. Kar. Combined Mrac for
82. B. C. Gruenwald, J. A. Muse, D. Wagner, and T. Yucelen. In Unknown Mimo Lti Systems with Parameter Convergence,
Advances in Intelligent and Autonomous Aerospace Systems; IEEE Trans. Autom. Control, 2017.
Valasek, J., Eds.; AIAA Press, 2018. 102. Y. Kawaguchi, H. Eguchi, T. Fukao, and K. Osuka. Passivity-
83. B. Riedle and P. V. Kokotovic. IEEE Trans. Autom. Control based Adaptive Nonlinear Control for Active Steering, in
1985, 30, pp 1027–1030. Control Applications, 2007. CCA 2007. IEEE International
84. S. M. Naik, P. R. Kumar, and B. E. Ydstie. IEEE Trans. Conference on; IEEE, 2007; pp 214–219.
Autom. Control 1992, 37, pp 182–197. 103. A. M. Annaswamy, J. Jang, and E. Lavretsky. Adaptive
85. A. J. Calise, N. Hovakimyan, and M. Idan. Automatica 2001, Gain-scheduled Controller in the Presence of Actuator
37(8), pp 1201–1211. Anomalies, in AIAA Guidance, Navigation, and Control
86. N. Hovakimyan, F. Nardi, A. Calise, and N. Kim. IEEE Conference and Exhibit, 2008; pp 18–21.
Trans. Neural Networks 2002, 13(6), pp 1420–1431. 104. S. K. Scarritt. Nonlinear Model Reference Adaptive Control
87. M. Matsutani, A. M. Annaswamy, T. E. Gibson, and E. for Satellite Attitude Tracking, in Proceedings of the AIAA
Lavretsky. Trustable Autonomous Systems Using Adaptive Guidance, Navigatoin and Control Conference and Exhibit,
Control, in IEEE Conference on Decision and Control, 2011; 2008.
pp 6760–6764. 105. J. Jang. Adaptive Control Design with Guaranteed Margins
88. M. Matsutani, A. M. Annaswamy, and E. Lavretsky. for Nonlinear Plants. PhD dissertation, Massachusetts
Guaranteed Delay Margins for Adaptive Control of Scalar Institute of Technology, 2009.
Plants, in IEEE Conference on Decision and Control, 2012; 106. J. Wang, N. Hovakimyan, and C. Cao. 1 Adaptive Augmen-
pp 7297–7307. tation of Gain-scheduled Controller for Racetrack Maneu-
89. H. Hussain, M. Matsutani, and A. M. Annaswamy. Robust ver in Aerial Refueling, in AIAA Guidance, Navigation, and
Adaptive Control in the Presence of Unmodeled Dynamics: Control Conference, vol. 5739, 2009.
A Counter to Rohrs’s Counterexample, in AIAA Guidance 107. X. Wang and N. Hovakimyan. Systems & Control Letters
Navigation and Control Conference, 2013. 2012, 61(4), pp 455–463.
90. H. Hussain, M. Matsutani, A. M. Annaswamy, and E. 108. F. Peter, M. Leitão, and F. Holzapfel. Adaptive Aug-
Lavretsky. Adaptive Control of Scalar Plants in the Pres- mentation of a New Baseline Control Architecture for
ence of Unmodeled Dynamics, in IFAC International Tail-controlled Missiles Using a Nonlinear Reference Model,
Workshop, 2013. in AIAA Guidance, Navigation, and Control Conference,
91. M. Matsutani. Robust Adaptive Flight Control Systems 2012; p 5037.
in the Presence of Time Delay. PhD dissertation, Mas- 109. M. Pakmehr. Towards Verifiable Adaptive Control of Gas
sachusetts Institute of Technology, 2013. Turbine Engines. PhD dissertation, Georgia Institute of
92. S. Tong, T. Wang, Y. Li, and H. Zhang. IEEE transactions on Technology, 2013.
cybernetics 2014, 44(6), pp 910–921. 110. M. Pakmehr and T. Yucelen. Model Reference Adaptive
93. S. Tong, T. Wang, and Y. Li. IEEE Transactions on Fuzzy Control of Systems with Gain Scheduled Reference Models,
Systems 2014, 22(3), pp 563–574. arXiv preprint arXiv:1403.3738, 2014.
94. M. L. Fravolini, T. Yucelen, B. Gruenwald, M. Dogan, and M. 111. M. Pakmehr and T. Yucelen. Adaptive Control of Uncertain
R. Napolitano. Comparison of Robust and Probabilistic Lmi- Systems with Gain Scheduled Reference Models and Con-
based Design of Adaptive Flight Controllers with Uncertain strained Control Inputs, in American Control Conference,
Input Dynamics, in AIAA Guidance, Navigation, and Con- 2014; pp 691–696.
trol Conference, 2016. 112. T. Yucelen, B. Gruenwald, J. A. Muse, and G. De La Torre.
95. Y. Li, S. Sui, and S. Tong. Adaptive Fuzzy Control Design Adaptive Control with Nonlinear Reference Systems, in
for Stochastic Nonlinear Switched Systems with Arbitrary American Control Conference, 2015; pp 3986–3991.
 Model Reference Adaptive Control 13

113. J. Guo, Y. Liu, and G. Tao. Multivariable mrac with State 131. A. Albattat, B. Gruenwald, and T. Yucelen. J. Dyn. Syst.
Feedback for Output Tracking, in American Control Confer- Meas. Contr. 2017, 139(7), pp 074501.
ence; IEEE, 2009; pp 592–597. 132. D. Zhang and B. Wei. Adaptive Control for Robotic Manipu-
114. K. Y. Volyanskyy, W. M. Haddad, and A. J. Calise. IEEE lators. CRC Press, 2017.
Trans. Neural Networks 2009, 20(11), pp 1707–1723. 133. T. Yucelen, Y. Yildiz, R. Sipahi, E. Yousefi, and N. Nguyen.
115. M. A. Santillo and D. S. Bernstein. J. Guid. Control Dynam. Stability Analysis of Human-Adaptive Controller Interac-
2010, 33(2), p 289. tions, in AIAA Guidance, Navigation, and Control Confer-
116. J. B. Hoagg and D. S. Bernstein. J. Guid. Control Dynam. ence, 2017.
2012, 35(6), pp 1767–1786. 134. T. Yucelen and M. Egerstedt. Control of Multiagent Systems
117. G. Tao, S. Chen, X. Tang, and S. M. Joshi. Adaptive Control under Persistent Disturbances, in American Control Confer-
of Systems with Actuator Failures. Springer, 2013. ence, 2012; pp 5264–5269.
118. D. P. Wiese, A. M. Annaswamy, J. A. Muse, M. A. Bolender, 135. Y. Liu and Y. Jia. IET Control Theory Appl. 2012, 6(13), pp
and E. Lavretsky. J. Guid. Control Dynam. 2015, 38(12), pp 2002–2008.
2429–2440. 136. T. Yucelen and E. N. Johnson. Int. J. Control 2013, 86(9), pp
119. T. E. Gibson, Z. Qu, A. M. Annaswamy, and E. Lavretsky. 1540–1553.
IEEE Trans. Autom. Control 2015, 60(10), pp 2728–2733. 137. Z. Peng, D. Wang, H. Zhang, G. Sun, and H. Wang. IET Con-
120. K. Kim, A. J. Calise, and T. Yucelen. Int. J. Adapt Control trol Theory Appl. 2013, 7(8), pp 1079–1087.
Signal Process. 2018, 31(11), pp 1608–1622. 138. G. De La Torre, T. Yucelen, and J. D. Peterson. Resilient Net-
121. E. Lavretsky. Robust and Adaptive Output Feedback Con- worked Multiagent Systems: A Distributed Adaptive Con-
trol for Non-minimum Phase Systems with Arbitrary Rel- trol Approach, in IEEE Conference on Decision and Control;
ative Degree, in AIAA Guidance, Navigation, and Control IEEE, 2014; pp 5367–5372.
Conference, 2017. 139. Z. Li, Z. Duan, and F. L. Lewis. Automatica 2014, 50(3), pp
122. T. Yucelen, W. M. Haddad, and E. M. Feron. Cyber-Phys. Sys. 883–889.
2016, 2(1-4), pp 24–52. 140. W. Wang, J. Huang, C. Wen, and H. Fan. Automatica 2014,
123. X. Jin, W. M. Haddad, and T. Yucelen. An Adaptive Control 50(4), pp 1254–1263.
Architecture for Mitigating Sensor and Actuator Attacks 141. T. Yucelen, J. D. Peterson, and K. L. Moore. Control of Net-
in Cyber-physical systems, in IEEE Trans. Autom. Control, worked Multiagent Systems with Uncertain Graph Topolo-
2017. gies, arXiv preprint arXiv:1501.03439, 2015.
124. L. An and G.-H. Yang. Inf Sci 2018, 423, pp 145–156. 142. K. M. Dogan,B. C. Gruenwald, T. Yucelen, J. A. Muse, and
125. X. Wang and N. Hovakimyan. 1 Adaptive control of event- E. A. Butcher. Distributed Adaptive Control of Networked
triggered networked systems, in American Control Confer- Multiagent Systems with Heterogeneous Actuator Dynam-
ence (ACC), 2010; IEEE, 2010; pp 2458–2463. ics, in American Control Conference, 2017; pp 5605–5610.
126. A. Sahoo, H. Xu, and S. Jagannathan. Neural Network- 143. G. De La Torre and T. Yucelen. Int. J. Control 2018, 91(3),
based Adaptive Event-Triggered Control of Nonlinear pp 495–507.
Continuous-Time Systems, in Intelligent Control (ISIC), 144. B. Sarsilmaz, T. Yucelen, and T. Oswald. A Distributed
2013 IEEE International Symposium On; IEEE, 2013; pp Adaptive Control Approach for Heterogeneous Uncertain
35–40. Multiagent Systems, in AIAA Guidance, Navigation, and
127. X. Wang, E. Kharisov, and N. Hovakimyan. IEEE Trans. Control Conference, 2018.
Autom. Control 2015, 60(9), pp 2500–2505.
128. A. Sahoo, H. Xu, and S. Jagannathan. IEEE Trans. Neural. TANSEL YUCELEN
Netw. Learn. Syst. 2016, 27(3), pp 497–509. University of South Florida, Tampa, FL, US
129. A. Sahoo, H. Xu, and S. Jagannathan. IEEE Trans. Neural.
Netw. Learn. Syst. 2016, 27(1), pp 151–164.
130. A. Albattat, B. C. Gruenwald, and T. Yucelen. Sensors 2016,
16(8), p 1297.