You are on page 1of 6

Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 52-29 (2019) 49–54
Transient
Transient Performance
Performance Improvement
Improvement in
in
Transient Performance
Reduced-Order Model Improvement
Reference in
Transient Performance
Reduced-Order Model Improvement
Reference in
Reduced-Order Model Reference 

Adaptive
Reduced-Order
Adaptive Control
Model Systems
Reference
Adaptive Control
Control Systems
Systems 
Adaptive Control Systems
Stefan Ristevski ∗∗ K. Merve Dogan ∗∗ ∗∗∗
∗∗ Tansel Yucelen ∗∗∗
Stefan Ristevski ∗∗ K. Merve Dogan ∗∗ ∗∗∗∗
∗∗ Tansel Yucelen ∗∗∗
Stefan Ristevski ∗Jonathan K. Merve
Jonathan A.
A. Muse
Dogan
Muse ∗∗∗∗
∗∗
Tansel Yucelen ∗∗∗
Stefan Ristevski Jonathan K. Merve Dogan
A. Muse ∗∗∗∗ ∗∗∗∗ Tansel Yucelen ∗∗∗
∗ Jonathan A. Muse
∗ Department of Mechanical Engineering, University of South Florida,
∗ Department of Mechanical Engineering, University of South Florida,
Department
Tampa, of Mechanical
Florida 33620, Engineering,
USA (e-mail: University of South Florida,
stefanr@mail.usf.edu)
∗ Tampa, Florida 33620, Engineering,
USA (e-mail:Universitystefanr@mail.usf.edu)
∗∗Department of Mechanical of South Florida,
∗∗
Department
Tampa, Floridaof
of Mechanical
33620, USA Engineering,
(e-mail: University of
of South
stefanr@mail.usf.edu) South Florida,
∗∗ Department
Tampa,
Tampa,
Mechanical
Florida
Florida 33620,
33620,
Engineering,
USA
USA (e-mail:
(e-mail:
University
stefanr@mail.usf.edu)
dogan@mail.usf.edu)
Florida,
Department of Mechanical Engineering,
∗∗ ∗∗∗Tampa, Florida 33620, USA (e-mail: dogan@mail.usf.edu)
University of South Florida,
Department
Department of Mechanical
of 33620, Engineering, (e-mail:University of SouthSouth Florida,
∗∗∗Tampa,
Department
∗∗∗Tampa,
Florida
Floridaof Mechanical
Mechanical
33620,
USAEngineering,
USA Engineering,
(e-mail:
University
University of
dogan@mail.usf.edu)
dogan@mail.usf.edu) of South
Department
Florida, Tampa, of Mechanical Engineering, University of South
∗∗∗∗Florida,
∗∗∗ Tampa,ofFlorida
Department Florida
Mechanical
33620, USA
USA (e-mail:
33620,Engineering, (e-mail: yucelen@usf.edu)
yucelen@usf.edu)
University of South
Air
∗∗∗∗Florida,
Air Force
Force Research
Tampa,
Research Laboratory,
Florida 33620, USA
Laboratory, Wright-Patterson
(e-mail:
Wright-Patterson Air
Air Force
Force Base,
yucelen@usf.edu) Base,
Florida,
∗∗∗∗Dayton,
Air Force Tampa,
OhioResearchFlorida
45433, USA 33620,
(e-mail:
Laboratory, USA (e-mail: yucelen@usf.edu)
jonathan.muse.2@us.af.mil)
Wright-Patterson Air Force Base,
∗∗∗∗Dayton, Ohio 45433, USA (e-mail: jonathan.muse.2@us.af.mil)
Air Force
Dayton, OhioResearch
45433, Laboratory,
USA (e-mail: Wright-Patterson Air Force Base,
jonathan.muse.2@us.af.mil)
Dayton, Ohio 45433, USA (e-mail: jonathan.muse.2@us.af.mil)
Abstract: The objective of model reference adaptive control systems is to drive the trajectories
Abstract:
Abstract: The
The objective
objective of model
model reference
of system reference adaptive
adaptive control
control systems
systems is
is to
to drive
drive the trajectories
the capturing
trajectories
of an uncertain dynamical to the trajectories of a given reference model a
of an
Abstract: uncertain dynamical
The objective system to the trajectories of a given reference model capturing a
of an uncertain
desired closed-loop systemofperformance.
dynamical model reference
system to the adaptive
To trajectories
this end, mostcontrol systems
of adaptive
a given is to drive
reference
control model
signals the trajectories
capturing
take the forma
desired
of an = closed-loop
uncertain system performance.
dynamical system to the To this end, most
n trajectories of adaptive
a given control signals
reference modeltake the forma
capturing
∈R
desired T
closed-loop systemwhere performance. To this end,the most adaptive control signals take the form
u (t) −
−Ŵ Ŵ T (t)σ(x(t)), x(t) ∈
x(t) n denotes
RTo state vector of
of anan uncertain dynamical
uaa (t) =closed-loop
desired (t)σ(x(t)),
system where
performance. denotes
this end, the
moststate vectorcontrol
adaptive uncertain
signals take dynamical
the form
u (t)
system,
a = −
σ :: R
Ŵ T
n → Rs denotes a known basis
n(t)σ(x(t)),
s where x(t) ∈ R n
denotes
n function,
the
and state
Ŵ vector
(t) ∈ R s×mof an uncertain
s×m denotes an estimation of dynamical
system,
u (t) = σ
− ŴR T → R denotes
(t)σ(x(t)), where a known
x(t) ∈basis
R function,
denotes and
the Ŵ
state (t) ∈
vectorR of denotes
an an
uncertain estimation
dynamical of
a unknown
system,
the σ : R weight
n
→ R matrix
s
denotesW ∈ Rs×m
a known basis function,
satisfying smand ∈ R (heredenotes
Ŵ (t)laws
update s×m
m an estimation
denotes the number of
nweights matrix W ∈ Rs×m satisfying sm update laws s×m
s×m
the unknown
system, R weight
→ R Inmatrix
σ : inputs). denotes a known basis function, and Ŵreduced-order,
(t)laws (here
∈ R (heredenotes m denotes the
an estimation number of
of
the control
unknown
of control inputs). this
Inmatrix paper,
W
this paper, ∈ Rwe focus on
satisfying a class
sm of
update m computationally
denotes the numberless
the
of unknown
expensive,
control weight
model
inputs). reference
In this W ∈ Rwe
adaptive
paper,
s×mfocus on a class of reduced-order, computationally less
we satisfying
control
focus systems
on a sm
class update
that are
of lawspredicated
only (here mcomputationally
reduced-order, denotes
on a the number
scalar update
less
expensive,
of control model
inputs).reference this adaptive
In contribution
paper, we control
focus systems
on aa classthatofare only predicated
reduced-order, on a scalar update
computationally less
law.
law. Specifically,
expensive, model
Specifically, our
reference
our adaptive
contribution is
is to
control
to utilize
systems
utilize a command
that are
command governor
only architecture
predicated
governor on a scalar
architecture in
in order
update
order to
to
expensive,
law. Specifically,
improve model reference
transient our
performance adaptive
contribution of iscontrol
this to utilize
class systems
of adaptive that
a command are only
control predicated
governor
systems. on athe
architecture
We prove scalar
in update
order
stability to
of
improve
law. transient
Specifically, performance
our contribution of this
is class of
to utilize adaptive
a command control systems. We prove the stability of
the overall
improve closed-loop
transient system using
performance of Lyapunov
this class of stability
adaptive theorygovernor
control and we also
systems. architecture
We present
prove anin
the order to
illustrative
stability of
the overall
improve closed-loop
transient system using
performance of Lyapunov
this class of stability control
adaptive theory and we also
systems. We present
prove an illustrative
the stability of
numerical
the overall
numerical example
closed-loop for demonstrating
system using the
Lyapunov efficacy of
stability the proposed
theory and architecture.
we also present an illustrative
the overall example
numerical closed-loop
example
for system
demonstrating
for demonstrating
the efficacy
using Lyapunov
the efficacy
of the
stability
of the
proposed
theory
proposedand we architecture.
also present an illustrative
architecture.
© 2019, IFAC
numerical (International
example Federation of the
for demonstrating Automatic
efficacyControl)
of the Hosting
proposed by Elsevier Ltd. All rights reserved.
architecture.
Keywords:
Keywords: UncertainUncertain dynamical
dynamical systems;
systems; reduced-order
reduced-order model model reference
reference adaptive
adaptive control;
control;
Keywords:performance
transient Uncertain dynamical
improvement; systems; reduced-order
stability analysis. model reference adaptive control;
transient performance
Keywords:performance improvement;
Uncertain dynamical stability
systems; analysis.
reduced-order
transient improvement; stability analysis. model reference adaptive control;
transient performance improvement; stability analysis.
1. INTRODUCTION
∈R
s×m
1.
1. INTRODUCTION
INTRODUCTION
function,
function, and and Ŵ (t) ∈
Ŵ (t) Rs×m denotes an
s×m denotess×m an
estimation
estimation of of the
the
Model reference adaptive control systems, which are orig- function,
unknown and
weight Ŵ (t) ∈
matrix R W ∈ R
denotes an estimation
s×m satisfying sm of the
update
Model reference 1. INTRODUCTION
adaptive control systems, which are orig- unknown
function, weight
and matrix
Ŵ (t) ∈ the W
R number
s×m ∈ R satisfying sm update
Model reference adaptive
inally proposed in Whitaker controlet systems,
al. (1958) which
and are Osburn laws
laws (here
orig- unknown m
m denotes
(hereweight matrix
denotes theW ∈denotes
number Rs×m
s×m an estimation
of control
satisfyinginputs). ofThat
sm update the
inally
Model proposed
inally
(1961), reference
proposed
have three
in Whitaker
adaptive
inmajor
Whitaker control
elements
et al.
al. a(1958)
et systems,
— (1958)
reference
and
which Osburn
are
andmodel,
Osburn an unknown
orig- is, the(here
laws weight
number m of matrix
denotesupdate the W
laws
number R of
∈ increases
of
control
satisfying
as the
control
inputs).
product
inputs).
That
sm update be-
That
(1961), have three inmajor elements —
al. a reference model, an is, the(here
number of update the laws increases as the product be-
inally proposed
(1961),
adaptive have threesignal,
control Whitaker
major anetupdate
elements
and — a(1958)
law. Inand
reference Osburn
model,
particular, an laws tween
is,
tweenthe the
number
the
m denotes
number
number of of
update
of basis
basis
number
function
laws ofentries
increases
function
control
entriesas inputs).
and
the
and the
product
the
That
number
numberbe-
adaptive
(1961),
adaptive control
have three
control signal,
major
signal, and an
elements
and an update law.
— a reference
update In particular,
model,
law. In particular, an is, tweenthe the
of control numbernumber
inputs of increases.
update
of basislaws As increases
function entries
it is now as and
the product
well-known,the number be-
model
the reference model captures a desired closed-loop system of control inputs increases. As it is now well-known, model
the
the reference
adaptive
reference model
control
model captures
signal, and
captures a
an
a desired
update
desired closed-loop
law. In
closed-loop system
particular,
system tween
reference
of the
control number
adaptive
inputs of basis
control
increases. function
systems
As it is entries
can
now and
provide
well-known,the
poor number
tran-
model
behavior for the uncertain dynamical system to (asymp- reference adaptive control systems can provide poor tran-
behavior for
the reference the
model uncertain
capturesdynamical
a desired system
closed-loopto
to (asymp-
system of control
sient inputs increases.
performance — Astheit is nowcan well-known, poormodel
behavior
totically
totically
forapproximately)
or
or
the uncertain
approximately)
dynamical
follow.
follow. For
system
this
this purpose,
Forsystem
(asymp-
purpose, the
the sient
reference
reference
adaptive
performance
adaptive — during
control
during
control the estimation
systems
systems
provide
estimation
can provide
(learning)
(learning)
poor
tran-
tran-
of
of
behavior
totically
trajectoriesfor the uncertain
or approximately)
of the uncertainfollow. dynamical
dynamical For this to (asymp-
and the sient
purpose,
system the unknown
performance system — parameters.
during the To this
estimation end, the research
(learning) of
trajectories of the uncertain dynamical system and the the unknown
sientunknown
performance system parameters.
—parameters.
during the ToTo this
estimation end, the research
toticallyerror,
trajectories
system or approximately)
of which denotesfollow.
the uncertain dynamical
the For this
difference purpose,
system
betweenand the the results
the in, system
for example, Duarte and end,(learning)
thisNarendra the research
(1989); of
system error, which denotes the difference between the results
the unknownin, for example,
system DuarteTo
parameters. andthis Narendra
end, the (1989);
research
trajectories
system error,of the
which uncertain
denotes dynamical
the difference system
betweenand the Lavretsky
results in, (2009);
for
trajectories of the uncertain dynamical system and the ref- Lavretsky (2009); Volyanskyy et al. (2009); Hovakimyan Volyanskyy
example, Duarte et al.
and (2009); Hovakimyan
Narendra (1989);
trajectories
systemmodel,
trajectories of
error, the
which
of are
the uncertain
denotesdynamical
uncertain the difference
dynamical system
system and
between
and the
the ref-
the results
ref- and Caoin,
Lavretsky for example,
(2010);
(2009); Yucelen
Volyanskyy Duarte
and Johnson
et al. and(2009);
Narendra
(2013b); (1989);
Gruenwald
Hovakimyan
erence
erence model, are fed
fed into
into thethe update
update law. law. The update
The and
update law
law and Cao (2010); Yucelen and Johnson (2013b); Gruenwald
trajectories
erence of
model, are
then provides the uncertain
fed into
(online) dynamical
the update
estimates for the system
law.unknown
The update the
systemref-
law et Lavretsky
and Cao (2009);
(2010); Volyanskyy
Yucelen and et
Johnson al.
al. (2015); Yucelen and Haddad (2012); Yucelen et al. (2009);
(2013b); Hovakimyan
Gruenwald
then
erenceprovides
model, (online)
are fed into estimates
the update for
for thelaw.unknown
Thetheupdate systemlaw et et al.
and
(2013);
(2015);
al.Cao (2010);
(2015);
Yucelen
Yucelen
Yucelen
Yucelen
and
and
andand Haddad
Johnson
Haddad (2012);
(2013b);
(2012); Yucelen et
et al.
Gruenwald
Yucelen al.
then provides
parameters
parameters
and
and
(online)
delivers
delivers
estimates
these
these
estimates
estimates
the unknown
to
to the
system
adaptive
adaptive (2013);
et al. Yucelen
(2015); and Johnson
Yucelen Johnson
and
(2013a); Gibson
(2013a);(2012);
Haddad
et
et al.
GibsonYucelen al. (2013);
(2013);
et al.
then provides
parameters and(online)
delivers estimates
these for
estimatesthe unknown
to the system
adaptive Yucelen
(2013); et
Yucelenal. (2014);
control signal, where this signal drives the trajectories of Yucelen et al. (2014); Gruenwald et al. (2017) (see alsoand Gruenwald
Johnson (2013a); et al.
Gibson(2017)
et (see
al. also
(2013);
control
parameterssignal,
and where this
delivers signal
these drives
signalestimates the trajectories
thetotrajectories
the adaptive of (2013);
references
of Yucelen Yucelen
ettherein) and
al. (2014); Johnson
have (2013a);
presented Gibson
several et al.
important (2013);
ap-
control
the
the
signal,
uncertain
uncertain
where
dynamical
dynamical
thissystem
system
drives
to
to the
the trajectories
trajectories of
of the
the references
Yucelen et therein)
al. (2014); haveGruenwald
presented et
Gruenwald
al. (2017)
several
et al.
(see also
important
(2017) (see ap-
also
control
the signal,
uncertain
reference model. where thissystem
dynamical signal todrives
the the trajectories
trajectories of the of references
proaches to therein)
address have
the presented
poor transient several important
performance ap-
aspect
reference model. proaches
references to address
therein) the
have poor transient
presented performance
several important aspect
ap-
the uncertain
reference model. dynamical system to the trajectories of the of model
proaches toreference
address adaptive
the poor control
transient systems,
performance which are
aspect
In of model toreference adaptive control systems, which are
In the
the literature,
reference model. most
literature, most adaptive
adaptive control
control signals
signals take
take the
the proaches
predicated
of model
predicated
address
on
reference
on the
the
the common poorform
adaptive
common
transient
control
form
performance
discussed above.
systems,
discussed above. whichaspect
are
In theualiterature,
form (t) = − Ŵ Tmost
T adaptive
(t)σ(x(t)) (see, control
for signalsLavretsky
example, take the of model reference
predicated on the adaptive
common control
form systems,
discussed above. which are
form
In theu (t) = − Ŵ (t)σ(x(t)) (see, for example, Lavretsky
form
and uaaliterature,
Wise = −Ŵ T
(t)(2012); most
Yucelen adaptive
(t)σ(x(t)) (see,control
(2019) and signals therein),
references
for example, take the In this
this paper,
predicated on thewewecommonfocus
focus on formaa discussed
class
class of reduced-order,
of above.
and
form Wise
u (t)(2012);
= − Ŵ TYucelen
(t)σ(x(t)) (2019)
(see, and
for referencesLavretsky
example, therein), In
Lavretsky In this
paper,
computationally
paper, less
we focus
on
expensive,
on a model
class of
reduced-order,
reference adaptive
reduced-order,
where
and
where x(t)
Wise
a ∈ R
(2012); n
n denotes
Yucelen
∈ Rn denotes
x(t)(2012); the
the state
(2019)
state andvector of
references
vector an uncertain
therein),
of an uncertain computationally less expensive, model reference adaptive
In this systems
control paper, we that focus
are on
only a model
class reference
predicatedof reduced-order,
on a
and
whereWise
x(t)
dynamical
dynamical
∈ R
system, Yucelen
denotes
σ : R
σ : Rnthe
(2019)
states and references
n → Rs denotes a known basis
nthe vector
Rs denotes
→state
of an therein),
uncertain computationally
control systems less
that expensive,
are only predicated on a scalar
adaptive
scalar
where x(t) system,
dynamical ∈ Rn denotes
system, σ : R → R vector ofa
denotes a
known
an
known
basis control
uncertain
basis
computationally
update
update
systems
law
law (see,
(see,
less
that
for
for
expensive,
example,
model
are only Gruenwald
example, Gruenwald
reference
predicated onal.adaptive
et
et
a (2018)
al.
scalar
(2018)
 This research was supportednby the Air control systems that are only predicated onal.aThatscalar
dynamical
 system, σ : R → R s Force Research
denotes a
This research was supported by the Air Force Research Laboratory knownLaboratory
basis and
update
and Section
law
Section 2
(see,of
2 of for this
for paper
example,
thisexample, for an
Gruenwald
paper forGruenwald overview).
an overview). et (2018)is,

under
Thisthe
under the
Universal
research was Technology
Universal
supported by
Technology
Corporation
the Air Force
Corporation
Grant
Research
Grant
Laboratory
17-S8401-02-C1.
17-S8401-02-C1.
update
and law (see,
Section 2 of this paper for an et al.That
overview). (2018)
That
is,
is,
 This research was supported by the Air Force Research Laboratory
under the Universal Technology Corporation Grant 17-S8401-02-C1. and Section 2 of this paper for an overview). That is,
under the Universal
2405-8963 Technology
© 2019, IFAC Corporation
(International Grantof17-S8401-02-C1.
Federation Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2019.12.620
50 Stefan Ristevski et al. / IFAC PapersOnLine 52-29 (2019) 49–54

this scalar update law does not change as the number of the form σ(x(t)) = [σ1 (x(t)), σ2 (x(t)), . . . , σs (x(t))]T such
basis function entries and the number of control inputs that σ (x(t))2 ≤ k1 + k2 x(t)2 holds, where k1 ∈ R+
change, unlike the most model reference adaptive control and k2 ∈ R+ .
systems discussed above. Specifically, our contribution
is to utilize a command governor architecture inspired Next, consider the reference model, which captures a
by the research results in Yucelen and Johnson (2013a) desired closed-loop system performance, given by
in order to improve transient performance of this class ẋr (t) = Ar xr (t) + B r c(t), xr (0) = xr0 , (2)
of adaptive control systems. In contrast to Yucelen and where xr (t) ∈ R is the reference model state vector,
n
Johnson (2013a), however, the resulting architecture of c(t) ∈ Rm is a desired bounded command, Ar  A−BK1 ∈
this paper is computationally less expensive since it is Rn×n is the reference model matrix that is considered
predicated on the reduced-order model reference adaptive to be Hurwitz, and B r  BK2 ∈ Rn×m is the reference
control formulation. We prove the stability of the overall model input matrix. Since the pair (A, B) is known and
closed-loop system using Lyapunov stability theory and controllable, here K1 ∈ Rm×n and K2 ∈ Rm×m represent
we also present an illustrative numerical example for the matrices utilized in the nominal feedback control law
demonstrating the efficacy of the proposed architecture. (details below, where we also refer to Yucelen (2019) for
We use a fairly standard notation in this paper. In par- further reading). As standard in most model reference
ticular, we use R to denote the set of real numbers, Rn adaptive control system formulations, the objective here
to denote the set of n × 1 real column vectors, Rn×m to is to synthesize a control signal such that x(t) of the
denote the set of n×m real matrices, R+ (respectively, R+ ) uncertain dynamical system practically follows xr (t) of the
to denote the set of positive (respectively, non-negative- reference model.
definite) real numbers, Rn×n
+ to denote the set of n × n Motivated by this standpoint, consider the feedback con-
positive-definite (respectively, non-negative-definite) real trol signal given by
matrices, Sn×n to denote the set of n × n symmetric real
u(t) = un (t) + ua (t), (3)
matrices, (·)T to denote the transpose, (·)−1 to denote
the inverse, tr(·) to denote the trace,  · 2 to denote the where un (t) ∈ Rm and ua (t) ∈ Rm are respectively the
Euclidean norm, ‘’ to denote the equality by definition, nominal and adaptive control laws. In particular, let the
λ (A) (respectively, λ (A)) to denote the minimum (respec- nominal control law have the form given by
tively, the maximum) eigenvalue of a matrix A ∈ Rn×n , un (t) = −K1 x(t) + K2 c(t). (4)
and tanh(x) = [tanh(x1 ), . . . , tanh(xn )]T , x ∈ Rn , to Before we state the adaptive control law, note that
denote the vector tangent hyperbolic function. 
ẋ(t) = Ax(t) + B − K1 x(t) + K2 c(t) + ua (t)
We finally provide a definition of the (rectangular) pro- 
+W T σ (x(t))
jection operator that is necessary for the main result of
this paper (see, for example, Exercise 11.3 in Lavretsky = (A − BK1 ) x(t) + BK2 c(t)
and Wise (2012)). Specifically, consider convex hypercube  
+B ua (t) + W T σ (x(t))
Ω = {θ ∈ Rn : (θimin ≤ θi ≤ θimax )i=1,··· ,n }, where  
Ω ∈ Rn and θimin and θimax respectively represent the = Ar x(t) + B r c(t) + B ua (t) + W T σ (x(t)) , (5)
minimum and maximum bounds for the ith component follows by using (3) and (4) in (1). Now, let the reduced-
of the n-dimensional parameter vector θ. For a sufficiently order adaptive control law be given by
small positive constant 0 , in addition, consider another  
ua (t) = −v̂(t)σ (x(t)) 2 tanh B T P e(t)µ−1 , (6)
hypercube Ω0 = {θ ∈ Rn : (θimin + 0 ≤ θi ≤ θimax −
0 )i=1,··· ,n }, where Ω0 ⊂ Ω. The projection operator where µ ∈ R+ is a design parameter and v̂(t) ∈ R+ is
Proj : Rn × Rn → Rn can then be component-wise defined an estimate of v, which satisfies the scalar update law
as Proj(θ, y) = (θimax − θi )yi /0 when θi > θimax − 0 and predicated on the projection operator given by
 
yi > 0, Proj(θ, y) = (θi − θimin )yi /0 when θi < θimin + 0 ˙
v̂(t) = γProj v̂(t), σ (x(t)) 2 B T P e(t)2
and yi < 0, and Proj(θ, y) = yi otherwise, y ∈ Rn . In the
light of this definition, note that (θ − θ∗ )T (Proj(θ, y) − v̂(0) = v̂0 , v̂0 ∈ R+ , (7)
y) ≤ 0 holds, where θ∗ ∈ Ω0 , θ ∈ Ω, and y ∈ Rn . where γ > 0 is the learning rate, e(t)  xr (t) − x(t) ∈ Rn
2. REDUCED-ORDER MODEL REFERENCE is the system error, and P ∈ Rn×n
+ is the solution of the
ADAPTIVE CONTROL FORMULATION Lyapunov equation given by
In this section, we provide an overview of the reduced- 0 = Ar T P + P Ar + R, (8)
order model reference adaptive control formulation, which with R ∈ Rn×n+ begin a design matrix. Here, the minimum
can be viewed as a case of the results in, for example, Gru- value for the projection operator is set to zero and its
enwald et al. (2018). Mathematically speaking, consider maximum value is set to a finite constant (larger than
the uncertain dynamical system given by v). In this case, note that v̂(t) ≥ 0 trivially holds by
 
ẋ(t) = Ax(t) + B u(t) + W T σ (x(t)) , x(0) = x0 , (1) definition. Since (7) is a monotonically increasing function,
where x(t) ∈ Rn is the system state vector, u(t) ∈ here one can adopt a leakage term in the update law (see,
Rm is the control signal, A ∈ Rn×n and B ∈ Rn×m for example, Dogan et al. (2018)), where the addition of
are respectively the known system and input matrices such a leakage term does not change the main results of
such that the pair (A, B) is controllable, W ∈ Rs×m is this paper.
unknown weight matrix with the upper bound W 2 ≤ Next, using (2), (5), and (6), the system error dynamics
v, and σ(x(t)) : Rn → Rs is known basis function in can be written as
Stefan Ristevski et al. / IFAC PapersOnLine 52-29 (2019) 49–54 51

ė(t) = ẋ(t) − ẋr (t) in (16) as well as the projection operator-based scalar
  update law, existence of a compact set is assured such
= Ar e(t) − v̂(t)σ (x(t)) 2 B tanh B T P e(t)µ−1
that V̇ (e(t), ṽ(t)) ≤ 0 outside of this set. Hence, uniform
+BW T σ (x(t)) , e(0) = e0 . (9) ultimate boundedness of the solution (e(t), ṽ(t)) is now
We are now ready to state the following theorem, which immediate Khalil (2002); Lavretsky and Wise (2012). 
can be viewed as a case of the results in, for example, Note that both p1 and p2 appearing in (16) depends on
Gruenwald et al. (2018). µ = µκ, where µ is a design parameter utilized in the
Theorem 1. Consider the uncertain dynamical system adaptive control law given by (6). Therefore, small values
given by (1), the nominal control signal given by (4), the of µ can be utilized to make both p1 and p2 small, and
adaptive control signal given by (6), the update law given hence, the corresponding ultimate bound small. While this
by (7), and the reference model given by (2). The solution is desired from a practical stability point of view, it should
(e(t), ṽ(t)) is then uniformly ultimately bounded, where be also noted that tanh(·) function approaches to sign(·)
ṽ(t)  v̂(t) − v. function for sufficiently small values of µ that can cause the
chattering phenomenon as is well-known. Hence, µ should
Proof. To show the boundedness of the solution (e(t), v̂(t)), be selected as small judiciously.
consider the Lyapunov-like function given by
V (e, ṽ) = eT P e + γ −1 ṽ 2 , (10) 3. TRANSIENT PERFORMANCE IMPROVEMENT
where V (0, 0) = 0 and V (e, ṽ) > 0 for all (e, ṽ) = In this section, we utilize a command governor architecture
(0, 0). Differentiating (10) along the closed-loop system inspired by the research results in Yucelen and Johnson
trajectories gives (2013a) in order to improve transient performance of the
˙ adaptive control system overviewed in Section 2. Note
V̇ (e(t), ṽ(t)) = 2ėT (t)P e(t) + 2γ −1 ṽ(t)v̂(t). (11) that the research results in Yucelen and Johnson (2013a)
Now, using (9) in (11) yields develop this architecture for the common model reference
V̇ (e(t), ṽ(t)) adaptive control system formulation that requires sm up-
  date laws. Hence, in contrast to Yucelen and Johnson
= eT (t) Ar T P + P Ar e(t) + 2σ T (x(t)) W B T P e(t) (2013a), we present a new theory in what follows, which
 T is computationally less expensive owing to the fact that it
−v̂(t)σ (x(t)) 2 tanh eT (t)P Bµ−1 B T P e(t)
is predicated on the reduced-order model reference adap-
˙
+2γ −1 ṽ(t)v̂(t) tive control system formulation of the previous section.
= −eT (t)Re(t) + 2σ T (x(t)) W B T P e(t) Mathematically speaking, consider the uncertain dynam-
  ical system given by (1) and the adaptive control signal
−2v̂(t)σ (x(t)) 2 tanh eT (t)P Bµ−1 B T P e(t) given by (6). For the results of this section, in addition,
˙
+2γ −1 ṽ(t)v̂(t). (12) consider the reference model given by (2) and the nominal
control signal given by (4) with the command given by
Next, we state an intrinsic property of tanh (·) function
from Polycarpou and Ioannou (1993) as c(t) = cd (t) + Gg(t). (17)
  Here, cd (t) ∈ R is a desired bounded command (same as
m
tanh eT (t)P Bµ−1 B T P e(t) ≥ B T P e(t)2 − µ, (13)
c(t) used alone in the previous section) and Gg(t) ∈ Rm
where µ  µκ ∈ R+ and κ = 0.2785. Using (13) in (12) is the command governor signal with G  K2−1 B L =
further yields K2−1 (B T B)−1 B T and det(B T B) = 0 (Yucelen and John-
V̇ (e(t), ṽ(t)) son (2013a)), which corrects the feedback control and
  reference model trajectories for transient performance im-
≤ −eT (t)Re(t) − 2v̂(t)σ (x(t)) 2 B T P e(t)2 − µ provement (details below). Furthermore, g(t) ∈ Rn is the
˙
+2vσ (x(t)) 2 B T P e(t)2 + 2γ −1 ṽ(t)v̂(t) output of the command governor dynamics given by
= −eT (t)Re(t) − 2ṽ(t)σ (x(t)) 2 B T P e(t)2 f˙(t) = −λf (t) + λe(t), f (0) = 0, (18)
˙
+2µv̂(t)σ (x(t)) 2 + 2γ −1 ṽ(t)v̂(t) g(t) = λf (t) + (Ar − λI)e(t). (19)
T
= −e (t)Re(t) Here f (t) ∈ R is the command governor state and λ ∈ R+
n
  is the scalar command governor gain.
˙ − σ (x(t)) 2 B T P e(t)2
+2ṽ(t) γ −1 v̂(t)
In this case, note that the system error dynamics given by
+2µv̂(t)σ (x(t)) 2 . (14) (9) remain the same. Note also that one can equivalently
At this point, note that the upper bound of the basis rewrite the system error dynamics as
function can be written as  
B L (Ar e(t) − ė(t)) = − ua (t) + W T σ (x(t)) . (20)
σ (x(t)) 2 ≤ k 1 + k2 e(t)2 . (15)
To elucidate the motivation behind the command governor
Here, k 1  k1 + k2 xr ∗ ∈ R+ , where xr (t)2 ≤ xr ∗ ∈ R+ . dynamics design, consider the uncertain dynamical system
Now, using (7) and (15) in (14) gives ẋ(t) = Ar x(t) + B r cd (t)
V̇ (e(t), ṽ(t)) ≤ −p0 e(t)22 + p1 e(t)2 + p2 , (16)   
+B B L g(t) − −ua − W T σ (x(t)) , (21)
where p0  λ (R), p1  2µv̂ ∗ k2 , and p2  2µv̂ ∗ k 1 (note
that v̂(t) ≤ v̂ ∗ ∈ R+ holds owing to the projection which follows by using (17) in (5). Using (20), in addition,
operator utilized in (7)). Finally, considering the structure (21) can be rewritten as
52 Stefan Ristevski et al. / IFAC PapersOnLine 52-29 (2019) 49–54

ẋ(t) = Ar x(t) + B r cd (t) V̇ (e(t), ṽ(t), f (t))


   
+BB L g(t) − [Ar e(t) − ė(t)] . (22) ≤ −eT (t)Re(t) − 2v̂(t)σ (x(t)) 2 B T P e(t)2 − µ
˙
+2vσ (x(t)) 2 B T P e(t)2 + 2γ −1 ṽ(t)v̂(t)
In (22), “Ar x(t) + B r cd (t)” represents the ideal system
response. As a consequence, one should then aim in sup- −2γ2 λf T (t)f (t) + 2γ2 λf T (t)e(t)
pressing the effect of “[Ar e(t) − ė(t)]” such that the uncer- = −eT (t)Re(t) − 2ṽ(t)σ (x(t)) 2 B T P e(t)2
tain dynamical system trajectories can closely follow the ˙
given ideal reference model given by +2µv̂(t)σ (x(t)) 2 + 2γ −1 ṽ(t)v̂(t)
−2γ2 λf T (t)f (t) + 2γ2 λf T (t)e(t)
ẋri (t) = Ar xri (t) + B r cd (t). (23)
= −eT (t)Re(t)
The command governor dynamics given by (18) and (19)  
can indeed achieve this objective by judiciously selecting ˙ − σ (x(t)) 2 B T P e(t)2
+2ṽ(t) γ −1 v̂(t)
the scalar command governor gain λ.
+2µv̂(t)σ (x(t)) 2
We now further explain the last sentence of the above para-
graph by resorting to the results in Yucelen and Johnson −2γ2 λf T (t)f (t) + 2γ2 λf T (t)e(t). (28)
(2013a). Specifically, consider the transfer function from Furthermore, using (7) in (28) yields
e(s) → g(s) in the Laplace domain given by V̇ (e(t), ṽ(t), f (t)) ≤ −eT (t)Re(t) + 2µv̂(t)σ (x(t)) 2
 −1
Ge→g (s) = Ar − s λ−1 s + 1 In , (24) −2γ2 λf T (t)f (t) + 2γ2 λf T (t)e(t). (29)
which is derived taking the Laplace transform of the In the continuation of the proof, we assume that R in (8)
command governor dynamics given by (18) and (19). As it can be chosen such that it takes the form R = R0 + γ2 λI,
is clear, (24) approximates [Ar e(t) − ė(t)] as λ increases; where R0 ∈ Rn×n+ Yucelen and Johnson (2013a). This does
that is, the closed-loop system trajectories recovers the not in any way place any restrictions on the choice of R
ideal reference model trajectories. The next theorem now owing to the fact that γ2 is a free parameter and can be
presents the main result of this paper. chosen sufficiently small. Now, applying this in (29) yields
Theorem 2. Consider the uncertain dynamical system V̇ (e(t), ṽ(t), f (t))
given by (1), the nominal control signal given by (4) with ≤ −eT (t)R0 e(t) − γ2 λeT (t)e(t) + 2µv̂(t)σ (x(t)) 2
(17), the adaptive control signal given by (6), the update
−2γ2 λf T (t)f (t) + 2γ2 λf T (t)e(t)
law given by (7), the reference model given by (2) with
2 2
(17), and the command governor dynamics given by (18) ≤ −λ (R0 )e(t)2 − γ2 λf (t)2
and (19). The solution (e(t), ṽ(t), f (t)) is then uniformly 2
−γ2 λ (e(t)2 − f (t)2 ) + 2µv̂(t)σ (x(t)) 2
ultimately bounded.
2 2
≤ −λ (R0 )e(t)2 − γ2 λf (t)2
Proof. To show the boundedness of the solution (e(t), ṽ(t),
f (t)), consider the Lyapunov-like function given by +2µv̂(t)σ (x(t)) 2 . (30)
In Section 2, we considered that σ (x(t))2 ≤ k1 +
V (e, ṽ, f ) = eT P e + γ1−1 ṽ 2 + γ2 f T f, (25) k2 x(t)2 , which can be expressed in an equivalent form
where γ1 ∈ R+ , γ2 ∈ R+ , V (0, 0, 0) = 0, and V (e, ṽ, f ) > 0 σ (x(t))2 ≤ k1 + k2 e(t)2 + k2 xr (t)2 . In particular,
for all (e, ṽ, f ) = (0, 0, 0). Differentiating (25) along the xr (t) was bounded for the results of Section 2. In this
closed-loop system trajectories gives case, however, we do not have a-priori knowledge on the
boundedness of c(t) given by (17). Hence, we now address
˙
V̇ (e(t), ṽ(t), f (t)) = 2ėT (t)P e(t) + 2γ −1 ṽ(t)v̂(t) this issue in the remainder of this proof.
+2γ2 f T (t)f˙(t). (26) Using the command signal defined in this section by (17),
the reference model given by (2) can be rewritten as
Now, using (9) and (18) in (26) yields ẋr = Ar xr (t) + B r cd (t) + Gg(t)
 
V̇ (e(t), ṽ(t), f (t)) = Ar xr (t) + B r cd (t) + G λf (t) + (Ar − λI) e(t)
 
= eT (t) Ar T P + P Ar e(t) + 2σ T (x(t)) W B T P e(t) = Ar xr (t) + B r cd (t) + Gdξ(t), (31)
 T
−v̂(t)σ (x(t)) 2 tanh eT (t)P Bµ−1 B T P e(t) where d  [λ, (Ar −λI)] and ξ(t)  [f (t), e (t)] . Since,
T T T

˙ + 2γ2 f T (t) (−λf (t) + λe(t)) Ar is Hurwitz there exist b1 ∈ R+ and b2 ∈ R+ such that
+2γ −1 ṽ(t)v̂(t)
(Section 13.5, Lavretsky and Wise (2013))
= −eT (t)Re(t) + 2σ T (x(t)) W B T P e(t)
  xr (t)2 ≤ b1 + b2 ξ(t)2
−2v̂(t)σ (x(t)) 2 tanh eT (t)P Bµ−1 B T P e(t)
≤ b1 + b2 f (t)2 + b2 e(t)2 . (32)
˙ − 2γ2 λf T (t)f (t)
+2γ −1 ṽ(t)v̂(t) Then, we can rewrite bound of σ (x(t)) given in (15) as
+2γ2 λf T (t)e(t). (27) σ (x(t))2 ≤ a1 + a2 f (t)2 + a3 e(t)2 , (33)
Taking advantage of the intrinsic property of tanh (·) given where a1  k1 + k2 b1 , a2  k2 b2 , and a3  k2 (1 + b2 ).
in (13) once again, (27) further yields Using (33) in (30) further yields
Stefan Ristevski et al. / IFAC PapersOnLine 52-29 (2019) 49–54 53

V̇ (e(t), ṽ(t), f (t))


2 2
≤ −λ (R0 )e(t)2 − γ2 λf (t)2
+2µv̂ ∗ (a1 + a2 f (t)2 + a3 e(t)2 )
 
= − λ (R0 )e(t)2 − 2µv̂ ∗ a3 e(t)2
 
− γ2 λf (t)2 − 2µv̂ ∗ a2 f (t)2 + 2µv̂ ∗ a1 . (34)

where v̂(t) ≤ v̂ ∗ owing it to the projection operator


used in the update law given by (7). Finally, considering
the structure in (34) as well as the projection operator-
based scalar update law, existence of a compact set is
assured such that V̇ (e(t), ṽ(t), f (t)) ≤ 0 outside of this
set. Hence, uniform ultimate boundedness of the solution
(e(t), ṽ(t), f (t)) is now immediate Khalil (2002); Lavretsky
and Wise (2012). 

4. ILLUSTRATIVE NUMERICAL EXAMPLE


Consider the second-order uncertain dynamical system
   
0 1 0
ẋ(t) = x(t) + [u(t) + W T x(t)], (35)
10 0 0.1
Fig. 1. Closed-loop response of the reduced-order model
where x(t) = [x1 (t), x2 (t)]T , W T = [−5, −25], and reference adaptive control architecture proposed in
x(0) = [0, 0]T . For this study, we choose the second-order Theorem 1 with µ = 1. Arrow indicates the direction
reference model     in which closed-loop response moves as learning rate
0 1 0 γ increases (i.e., light gray lines indicate lower values
ẋr (t) = 2 xr (t) + 2 u(t), (36) of γ ∈ {1, 10, 25, 50, 100} than dark gray lines).
−ωn −2ζωn ωn
with a natural frequency of ωn = 0.2 rad/s and a damping
ratio of ζ = 0.7. As a consequence, the nominal controller
gains become K1 = [0.16, 0.64], K2 = 0.16. We also choose
R = I2 for the Lyapunov equation in (8). Here, we are
interested to follow the desired command cd (t) = ±1.
In Figures 1 and 2, the closed-loop responses of the
reduced-order model reference adaptive control architec-
ture in Theorem 1 are shown for learning rates γ ∈
{1, 10, 25, 50, 100} (we use µ = 1 in Figure 1 and
µ = 1/15 in Figure 2). Recall that xri (t) = xr (t) in these
figures. Two observations follow from these figures: First,
the closed loop system response improves as learning rate γ
increases; however, still there is a considerable gap between
the trajectories of the ideal reference model and the tra-
jectories of the uncertain dynamical system. Second, while
the latter choice of µ provides improved closed-loop system
response in x1 (t), it causes chattering phenomenon as
highlighted in the last paragraph of Section 2 (a sampling
rate of 100 Hz is utilized in our all numerical results).
In Figure 3, the closed-loop responses of the proposed
reduced-order model reference adaptive control architec-
Fig. 2. Closed-loop response of the reduced-order model
ture augmented with command governor in Theorem 2 is
reference adaptive control architecture proposed in
shown. Specifically, the scalar command governor gain is
Theorem 1 with µ = 1/15. Arrow indicates the direc-
selected as λ ∈ {1, 5, 10, 15, 25}, and we use γ = 10
tion in which closed-loop response moves as learning
and µ = 1. Clearly, the gap between the trajectories
rate γ ∈ {1, 10, 25, 50, 100} increases (i.e., light gray
of the ideal reference model and the trajectories of the
lines indicate lower values of γ than dark gray lines).
uncertain dynamical system get sufficiently small as λ
increases, without further need to reduce µ that can demonstrates the efficacy of the proposed architecture in
induce chattering phenomenon as discussed above. The Theorem 2.
last subfigure of Figure 3 also shows how the command 5. CONCLUSION
governor mechanism alters the given desired command to As it is well-known, model reference adaptive control sys-
allow for improved transient performance. To summarize, tems can provide poor transient performance (that is,
the closed-loop system response observed in Figure 3 is during the estimation of the unknown system parameters).
superior than the ones in Figures 1 and 2, where this While there are research results addressing this problem,
54 Stefan Ristevski et al. / IFAC PapersOnLine 52-29 (2019) 49–54

mance improvement in model reference adaptive control.


Machines, 5(1), 9.
1
Gruenwald, B.C., Arabi, E., Yucelen, T., Chakravarthy,
0 A., and McNeely, D. (2018). Decentralised adaptive
-1
architectures for control of large-scale active–passive
0 5 10 15 20 25 30 35 40 45 50 modular systems with stability and performance guar-
antees. International Journal of Control, 1–15.
0.3
0.2
Gruenwald, B.C., Yucelen, T., and Fravolini, M. (2015).
0.1
0
Performance oriented adaptive architectures with guar-
-0.1
anteed bounds. In AIAA Infotech@ Aerospace, 0120.
0 5 10 15 20 25 30 35 40 45 50
Hovakimyan, N. and Cao, C. (2010). L1 Adaptive Control
100 Theory: Guaranteed Robustness with Fast Adaptation.
50 SIAM.
0
-50
Khalil, H.K. (2002). Nonlinear systems, volume 3. Prentice
0 5 10 15 20 25 30 35 40 45 50
hall Upper Saddle River, NJ.
Lavretsky, E. and Wise, K. (2012). Robust and Adap-
tive Control: With Aerospace Applications. Advanced
40
Textbooks in Control and Signal Processing. Springer
20
London.
0
0 5 10 15 20 25 30 35 40 45 50 Lavretsky, E. (2009). Combined/composite model refer-
ence adaptive control. IEEE Transactions on Automatic
5 Control, 54(11), 2692–2697.
0
Lavretsky, E. and Wise, K.A. (2013). Robust adaptive con-
-5

-10
trol. In Robust and adaptive control, 317–353. Springer.
0 5 10 15 20 25 30 35 40 45 50 Osburn, P. (1961). New developments in the design of
adaptive control systems. February 1961 Instituteof
Fig. 3. Closed-loop response of the reduced-order model Aeronautical Sciences, Paper–No.
reference adaptive control architecture augmented Polycarpou, M.M. and Ioannou, P.A. (1993). A robust
with command governor proposed in Theorem 2 with adaptive nonlinear control design. In 1993 American
γ = 10 and µ = 1. Arrow indicates the direction in Control Conference, 1365–1369. IEEE.
which closed-loop response moves as command gov- Volyanskyy, K.Y., Haddad, W.M., and Calise, A.J. (2009).
ernor gain λ ∈ {1, 5, 10, 15, 25} increases (light gray A new neuroadaptive control architecture for nonlinear
lines indicate lower values of λ than dark gray lines). uncertain dynamical systems: Beyond σ- and e- modifi-
cations. IEEE Transactions on Neural Networks, 20(11),
they are predicated on the common model reference adap- 1707–1723.
tive control system formulation that requires sm update Whitaker, H., Yamron, J., and Kezer, A. (1958). Design of
laws with s being the number of elements on the basis model reference control systems for aircraft. Cambridge,
function and m being the number of control inputs. Moti- MA: Instrumentation Laboratory, Massachusetts Insti-
vated by this standpoint, this paper focused on a class of tute of Technology.
reduced-order, computationally less expensive, model ref- Yucelen, T. (2019). Model reference adaptive control.
erence adaptive control systems that only require a scalar Wiley Encyclopedia of Electrical and Electronics Engi-
update law. Specifically, our contribution was to utilize a neering, 1–13.
command governor architecture in order to improve tran- Yucelen, T., De La Torre, G., and Johnson, E.N. (2013).
sient performance of this class of adaptive control systems. Frequency-limited adaptive control architecture for
In addition to proving the stability of the overall closed- transient response improvement. In 2013 American
loop system using Lyapunov stability theory, we presented Control Conference, 6631–6636. IEEE.
an illustrative numerical example for demonstrating the Yucelen, T., De La Torre, G., and Johnson, E.N. (2014).
efficacy of the proposed architecture. Improving transient performance of adaptive control
architectures using frequency-limited system error dy-
REFERENCES namics. International Journal of Control, 87(11), 2383–
2397.
Dogan, K.M., Gruenwald, B.C., Yucelen, T., and Muse, Yucelen, T. and Haddad, W.M. (2012). Low-frequency
J.A. (2018). Relaxing the stability limit of adaptive learning and fast adaptation in model reference adaptive
control systems in the presence of unmodelled dynamics. control. IEEE Transactions on Automatic Control,
International Journal of Control, 91(8), 1774–1784. 58(4), 1080–1085.
Duarte, M.A. and Narendra, K.S. (1989). Combined Yucelen, T. and Johnson, E. (2013a). A new command
direct and indirect approach to adaptive control. IEEE governor architecture for transient response shaping.
Transactions on Automatic Control, 34(10), 1071–1075. International Journal of Adaptive Control and Signal
Gibson, T.E., Annaswamy, A.M., and Lavretsky, E. Processing, 27(12), 1065–1085.
(2013). On adaptive control with closed-loop reference Yucelen, T. and Johnson, E.N. (2013b). Artificial basis
models: Transients, oscillations, and peaking. IEEE functions in adaptive control for transient performance
Access, 1, 703–717. improvement. In AIAA Guidance, Navigation, and
Gruenwald, B., Yucelen, T., and Muse, J. (2017). Direct Control Conference, 4754.
uncertainty minimization framework for system perfor-

You might also like