Proceedings of the 19th World Congress

The International Federation of Automatic Control

Cape Town, South Africa. August 24-29, 2014

A proportional integral extremum-seeking

control approach
Martin Guay ∗ , Denis Dochain ∗∗
∗
Department of Chemical Engineering, Queen’s University, Kingston,
ON, Canada E-mail address: [email protected]
∗∗
Université Catholique de Louvain, Centre for Systems Engineering
and Applied Mechanics (CESAME), Louvain-la-Neuve, Belgium
[email protected]

Abstract: This paper proposes an alternative extremum seeking control design technique for
the solution of real-time optimization control problems. The technique considers a proportional-
integral approach that avoids the need for a time-scale separation in the formulation of the
ESC. It is assumed that the equations describing the dynamics of the nonlinear system and
the cost function to be minimized are unknown and that the objective function is measured.
The dynamics are assumed to be asymptotically stable and relative order one with respect to
the objective function. The extremum-seeking problem is solved using a time-varying parameter
estimation technique.

Keywords: Extremum-seeking control, Real-time optimization, Time-varying systems

1. INTRODUCTION objective function is parameterized in a known fashion. A

three-time scale approach is proposed to establish the com-
Extremum-seeking control (ESC) has been the subject bined adaptive estimation and extremum seeking control
of considerable research effort over the last decade. This algorithms. Recent work reported in Ghaffari et al. [2012]
approach, which dates back to the 1920s Leblanc [1922], is and Moase et al. [2010] have proposed a Newton-based
an ingenious mechanism by which a system can be driven extremum-seeking technique that provides an estimate of
to the optimum of a measured variable of interest Tan the inverse of the Hessian of the cost function. This tech-
et al. [2010]. The revived interest in the field was primarily nique can effectively alleviate the convergence problems
sparked by Krstic and co-workers who provided an elegant associated with the increase of the gain of the Newton
proof of the convergence of a standard perturbation based update. Other alternative techniques such as proposed
extremum seeking scheme for a general class of nonlinear Zhang and Ordóñez [2009] and Zhang and Ordóñez [2012]
systems. The main drawback of ESC is the lack of transient make use of sampled gradient measurements to improve
performance guarantees. As highlighted in the proof of the convergence properties of ESC techniques that imple-
Krstic and Wang Krstic and Wang [2000], the stability ment numerical optimization techniques. A sliding-mode
analysis relies on two components: 1) an averaging analysis approach is presented in Fu and Özgüner [2011].
of the persistently perturbed ESC loop and 2) a time-
Although the limitations associated with the tuning of
scale separation of ESC closed-loop dynamics between the
ESC is generally well understood, the limitations associ-
fast transients of the system dynamics and the slow quasi
ated with the two time-scale approach to ESC remains
steady-state extremum-seeking task.
problematic. Under the two time-scale assumption, the
Over the last few years, many researchers have considered optimization operates at a quasi steady-state, or slow,
various approaches to overcome the limitations of ESC. In time-scale such that the search for optimal operating con-
Krstic [2000], the performance limitations associated with ditions does not affect the process dynamics. To over-
ESC were considered in detail. The non-local properties on come the time-scale separation, one must incorporate some
ESC was studied in Tan et al. [2006]. This work extends knowledge of the transient behaviour of the process dy-
the work in Krstic and Wang [2000] by considering the namics. In the case where a model is available, one can
case where the fast dynamics can be assumed to be uni- use adaptive extremum seeking technique as proposed in
formly global asymptotically stable along the equilibrium Guay and Zhang [2003] to stabilize a nonlinear system to
manifold. In Adetola and Guay [2007], Guay et al. [2004] the unknown optimum of a known but unmeasured cost
and Cougnon et al. [2011], an alternative ESC algorithm function. If a model is not available but similar systems are
is considered where an adaptive control and estimation available, the use of multi-unit extremum seeking control
approach is used. The key aspect of this approach is that techniques Srinivasan [2007] can be used to steer both
the equilibrium map is parameterized and the parame- systems in a neighbourhood of the unknown optimum.
ters are estimated with the help of a tailored adaptive Both classes of techniques can solve the steady-state op-
estimation technique. The results in Nesic et al. [2010] timization ESC problem without the need for time-scale
unify the approaches based on singular perturbation and separation. In Scheinker and Krstic [2013], Lie bracket
parameter estimation by considering the case where the averaging techniques are considered to stabilize unknown

978-3-902823-62-5/2014 © IFAC 377

19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

dynamical systems using ESC. The approach does not Note that, in contrast to standard ESC, convexity of the
explicitly rely on the need for time-scale separation but cost function h(x) is required. We also require the following
it requires a known CLF of the unknown control system. properties for the dynamics:
ESC problems cannot be currently solved in the absence Assumption 2. The dynamics (1) are such that:
of time-scale separations if explicit process models or mul- (1) the cost function h(x) decreases in the direction of
tiple identical units are not available. This paper attempts f (x):
to bridge this gap in the application of ESC. It proposes ∂h ∂h
a proportional-integral ESC design technique. This tech- f (x) + g(x)u ≤ −αkx − π(u)k2 , ∀x ∈ D(u),
nique can be interpreted as a generalization of the stan- ∂x ∂x
dard approach where the integral action corresponds to (2) the matrix valued function g(x) is full rank ∀x ∈
the standard ESC control task used to identify the steady- D(u),
state optimum. The proportional control action is designed ∀u ∈ U.
to ensure that the measured cost function is optimized
instantaneously. The approach considers an alternative Assumption 2 states that h is non-decreasing in along the
parameterization of the ESC problem in which the rate of vector field f (x) + g(x)u over some neighbourhood of the
change of the output is parameterized directly without the steady-state manifold x = π(u) at a fixed value of the
need to invoke a time-scale separation argument. Under input u. It also states that the cost function is relative
suitable assumption on the dynamics of the system, this order 1 in a neighbourhood of the origin.
action can be shown to minimize the cost over short times
while reaching the optimum steady-state conditions. Finally, we will require the following additional assumption
concerning the steady-state cost function `(u).
The paper is organized as follows. A brief description of Assumption 3. The equilibrium steady-state map `(u) is
the ESC problem is given in section 2. In section 3, the pro- such that
posed ESC formulation is presented for a known cost func-
tion and process dynamics. The proposed proportional- ∇u `(u)(u − u∗ ) ≥ αu ku − u∗ k2
integral ESC controller is described in section 4. A simula- for some positive constant αu ∀u ∈ U.
tion example is presented in 5 followed by brief conclusions
in 6. 3. EXTREMUM SEEKING CONTROLLER WITH
FULL INFORMATION
2. PROBLEM DESCRIPTION
In this section, we propose the extremum-seeking control
Consider a nonlinear system approach that will form the basis of the development
in later sections. Let us first consider the cost function
y = h(x) and compute its time derivation:
ẋ = f (x) + g(x)u (1)
y = h(x) (2) ẏ = Lf h + Lg hu (4)
n where Lf h and Lg h are the Lie derivatives of h(x) with
where x ∈ R is the vector of state variables, u is the vector
of input variables taking values in U ⊂ Rp and y ∈ R is respect to f (x) and g(x), respectively. The Lie derivative
the variable to be minimized. It is assumed that f (x) and is the directional derivative of the function h(x) given by:
g(x) a smooth vector valued functions of x and that h(x) ∂h ∂h
is a smooth function of x. Lf h = f, Lg h = g.
∂x ∂x
The objective is to steer the system to the equilibrium By the relative order assumption it follows that Lg h 6= 0
x∗ and u∗ that achieves the minimum value of y(= h(x∗ )). in a neighbourhood of the unknown optimum x∗ .
The equilibrium (or steady-state) map is the n dimensional We propose the following controller:
vector π(u) which is such that:
u = −kLg h + û (5)
f (π(u)) + g(π(u))u = 0. where û is a steady-state bias term to be estimated. Let
the optimal steady-state input be given by u∗ . The error
The equilibrium cost function is given by: in the deviation bias is denoted by ũ = u∗ − û. Pose the
function
y = h(π(u)) = `(u) (3) 1
V = y + ũT ũ
Thus, at equilibrium, the problem is reduced to finding the 2
Its time derivative is given by:
minimizer u∗ of y = `(u∗ ). Let D(u) be a neighbourhood
of the steady-state x = π(u). ˙
V̇ = Lf h − kkLf gk2 + Lg hû − ũû.
Some additional assumptions are required concerning the Let û˙ = −Lg h. Upon substitution of ũ = u∗ − û, one
cost function h(x). obtains:
V̇ = Lf h − kkLf gk2 + Lg hu∗
Assumption 1. The cost h(x) is such that
∂h(x∗ ) By assumption, it follows that:
(1) =0
∂x
∂ 2 h(x) V̇ ≤ −αkx − π(u∗ )k2 − kkLg hk2
(2) > αI, ∀x ∈ Rn
∂x∂xT Since g(x) is everywhere full rank and x∗ is the unique
where α is a strictly positive constant. point where ∇x h(x∗ ) = 0. Thus the system reaches the

378
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

largest invariant on the set Lg h = 0 which occurs at the η̃˙ = −K η̃ − cT θ̇, η̃(0) = 0. (12)
point x∗ with corresponding input u∗ .
We can now define the proposed parameter estimation
As a result, the ESC closed-loop system converges to the update. Let Σ ∈ Rp×p be the solution to the following
steady-state optimum of the cost h(x). The main feature matrix differential equation
of the proposed ESC is the combination of a proportional
component u − û = −kLg h with integral action given by
Σ̇ = ccT − kT Σ + δI (13)
û˙ = −Lg h.
with initial conditions Σ(0) = α1 I  0, where α1 , δ and kT
are strictly positive constants to be assigned. The inverse
4. EXTREMUM-SEEKING CONTROLLER of Σ is then given as the solution to the matrix differential
equation:
In this section, the minimization of y is performed in real-
time. Since the dynamics of the system are unknown, one
Σ̇−1 = − Σ−1 ccT Σ−1 + kT Σ−1 − δΣ−2 (14)
must consider an adaptive control approach to implement
the ESC (5). In the following, we parameterize the un- with initial condition Σ−1 (0) = α1 I. Based on (7),(8) and
known dynamics (4). Defining θ0 = Lf h and θ1 = Lg h, we (11), one considers the following parameter update law
propose the following parameterization: proposed in Adetola and Guay [2009]:

ẏ = θ0 + θ1 u = φT θ (6) ˙
θ̂ =Proj(Σ−1 (c(e − η̂) − σ θ̂), θ̂), θ̂(0) = θ0 ∈ Θ0 , (15)
T T
where φ = [1, u ] and θ = [θ0 , θ1T ]T .
The design of the
extremum seeking routine is based on the dynamics (6). where σ is a positive constant. Proj{τ, θ̂} denotes a Lips-
The first step consists in the estimation of the time-varying chitz projection operator Krstic et al. [1995] such that
parameters θ0 and θ1 . In the second step, we define a
controller, based on the controller (5), that achieves the −Proj{τ, θ̂}T θ̃ ≤ −τ T θ̃, (16)
extremum-seeking task.

θ̂(0) ∈ Θ0 =⇒ θ̂ ∈ Θ, ∀t ≥ 0 (17)
4.1 Parameter estimation
where Θ , B(θ̂, zθ ), where B(θ̂, zθ ) is the ball centered
The first element of the proposed time-varying parameter at θ̂ with radius zθ . Following standard arguments from
estimation scheme is the following output prediction model adaptive control, the filter parameter, c(t), must satisfy
for (6). Let ŷ represent the predicted output for a given the following assumption.
value of the parameter estimates θ̂ = [θ̂0 , θ̂1T ]T . The output Assumption 4. There exists constants α2 > 0 and T > 0
prediction error is denoted by e = y − ŷ and the parameter such that
estimation error is given by θ̃ = θ − θ̂. We consider the Z t+T
following prediction dynamics: c(τ )c(τ )T dτ ≥ α2 I (18)
t
˙ ∀t > 0.
ŷ˙ = φT θ̂ + Ke + cT θ̂, (7)
where K is a positive constant to be assigned and where 4.2 Controller design
the time varying parameter c ∈ Rp is the solution of the
differential equation: The input space U is defined as U = {u | kuk ≤ zu } where
zu is a positive constant that identifies the upper limit on
ċT = −KcT + φT (8) the size of the norm the control input u. Let P (u) = kuk2 −
with initial conditions c(0) = 0. The prediction error zu2 and define τ = −k θ̂+d(t) where d(t) is a bounded dither
dynamics are given by: signal with kd(t)k ≤ D and k > 0.
The extremum-seeking controller considered is given by:
˙ (9)
ė =u̇T θ̃ − Ke − cT θ̂
with initial conditions e(0) = y(0) − ŷ(0). Following u = −kg θ̂1 + û + d(t) (19)
standard arguments, we define the auxiliary variable η = 1
û˙ = θ̂1 (20)
e − cT θ̃. The dynamics of η are as follows: τI
where d(t) is a dither signal to be determined, kg and τI
η̇ = −Kη − cT θ̇, η(0) = e(0) (10) are tuning parameters taken as positive constants.
A filtered estimate, η̂, of η is also defined. The η̂ dynamics Theorem 1. Let Assumptions 1 to 4 hold. Consider the
are given by: extremum-seeking controller (19) and the parameter esti-
mation algorithm (14) and (15). Then there exists tuning
parameters kg , kT , K and τI∗ such that for all τI > τI∗ . the
η̂˙ = −K η̂. (11)
system converges exponentially to an O(D/τI ) neighbour-
As a result, the dynamics of the estimation error η̃ = η − η̂ hood of the minimizer x∗ of the measured cost function
are y.

379
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

Proof: We consider the Lyapunov function: We let the tuning constants of the ESC be such that:
1 1 K = kη1 I + kη2 cT c, kT γ1 = kθ̃ ,
W = η̃ T η̃ + θ̃T Σθ̃.
2 2 where kη1 > 1/2, kη2 > k/2, σ = δ and kθ̃ > k22γ2 . Let
Taking the derivative of W yields:
1 k1 kT γ1 k2 γ2 γ2 1
ka = K − I − ccT , kb = − , kc = + ,
kT T 2 2 2 2 2k2 2k1
Ẇ ≤ − η̃ T K η̃ + η̃ T cT θ̇ + θ̃T Σθ̇ − θ̃ Σθ̃ and rewrite the last inequality as follows:
2 (21)
δ 1 σ
+ θ̃T θ̃ − θ̃T c(e − η̂) + θ̃T ccT θ̃ + σ θ̃T θ̂ Ẇ ≤ −ka kη̃k2 − kb kθ̃k2 + kc kθ̇k2 + kθk2 .
2 2 2
where the property of the property (16) of the projection Next we consider the proposed ESC at constant û using
algorithm is invoked. Substituting for the Lyapunov function candidate: V = W + y. Using the
last inequality, its time derivative is given by:
cT θ̃ = e + η = e − η̂ + η̃ σ
and completing the squares yields the following inequality: V̇ ≤ − ka kη̃k2 − kb kθ̃k2 + kc kθ̇k2 + kθk2
2
kT T + θ0 + θ1T u + θ1T d.
Ẇ ≤ − η̃ T K η̃ + η̃ T cT θ̇ + θ̃T Σθ̇ − θ̃ Σθ̃
2 Substituting the proposed ESC:
δ 1 1 σ
+ θ̃T θ̃ − (e − η̂)T (e − η̂) + σ θ̃T θ̂ + η̃ T η̃ V̇ ≤ − ka kη̃k2 − kb kθ̃k2 + kc kθ̇k2 + kθk2
2 2 2 2
Noting that θ = θ̂ + θ̃, one can rewrite the above inequality + θ0 − kg θ1T θ̂1 + θ1T û + θ1T d
as follows:   or,
1 kT T σ
Ẇ ≤ − η̃ T K − I η̃ + η̃ T cT θ̇ + θ̃T Σθ̇ − θ̃ Σθ̃ V̇ ≤ − ka kη̃k2 − kb kθ̃k2 + kc kθ̇k2 + kθk2
2 2 2
+ θ0 + θ1T û − kg θ1T θ1 + kg θ1T θ̃1 + θ1T d.
 
δ
− σ− θ̃T θ̃ + σ θ̃T θ
2
By Assumption 2, we obtain:
By completing the squares, we remove the indefinite terms σ
as follows:  V̇ ≤ − ka kη̃k2 − kb kθ̃k2 + kc kθ̇k2 + kθk2
 2
1 k1 T 1 k2
Ẇ ≤ − η̃ T
K − I − cc η̃ + kθ̇k2 + θ̃T Σθ̃ − αkx − π(û)k2 − kg θ1T θ1 + kg θ1T θ̃1 + θ1T d.
2 2 2k1 2
  Completing the squares using constants k3 and k4 , we
1 T kT T δ remove the indefiniteterms andcollect terms to obtain:
+ θ̇ Σθ̇ − θ̃ Σθ̃ − σ − kθ̃k2
2k2 2 2 kg
σ σ V̇ ≤ − ka kη̃k2 − kb − kθ̃k2 − αkx − π(û)k2
+ kθ̃k2 + kθk2 2k3
2 2  
kg k3 k4 σ
where k1 and k2 are strictly positive constants. − kg − − − kθ1 k2
2 2 2
The boundedness of the matrix Σ can be shown as follows. σ 1
By integration, one gets: + kc kθ̇k2 + θ02 + kdk2 .
Z t 2 2k4
 σ  As a result, we see that the proposed ESC in the absence
Σ =e−kT t Σ(0) + e−kT (t−τ ) c(τ )c(τ )T + I dτ
0 2 of integral action can approach a neighbourhood of the
Z t  σ  unknown optimum whose depends on the distance of û
≥ e−kT (t−τ ) c(τ )c(τ )T + I dτ from the true unknown steady-state optimum u∗ . The next
t−T 2
σ step of the proof focusses on proving that this distance can
−kT T be minimized for some large enough τI .
≥e (α2 + )I = γ1 I
2
where Assumption 4 is invoked. By the boundedness of c, To do this, we consider the equilibrium response, given by
one can also write, x = π(û), of the system at a specific û. The equilibrium
map is such that:
σ  t −kT (t−τ )
 Z
Σ ≤Σ(0) + β2 + e dτ I f (π(û)) + g(π(û))û = 0.
2 0
 σ  The output trajectory along the equilibrium manifold, also
≤ α1 + β2 + I = γ2 I. called the quasi steady-state response, can be described by
2
the differential equation:
As a result, we get that: γ1 I ≤ Σ ≤ γ2 I and γ2−1 I ≤ Σ−1 ≤
dy ˙ T ∂π T ∂ 2 h(π(û))
γ1−1 I. By the boundedness of Σ and Γ, one can write: =û f (π(û))
  dτ ∂ û ∂x∂xT
1 k1 ∂π T ∂ 2 h(π(û))
Ẇ ≤ − K − I − ccT kη̃k2 + û˙ T g(π(û))û
2 2 ∂ û ∂x∂xT 
 
kT γ1 σ k2 γ2 δ T

− + − − kθ̃k2 ˙ T ∂π ∂h(π(û)) ∂f (π(û)) ∂g(π(û))
2 2 2 2 + û + û
  ∂ û ∂x ∂x ∂x
γ2 1 σ ∂h(π(û)) dû
+ + kθ̇k2 + kθk2 + g(π(û)) .
2k2 2k1 2 ∂x dτ

380
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

where dτ = τI dt represents the quasi steady-state time-

 
kg
scale. Along the steady-state manifold, it follows that: Ẇ ≤ − ka kη̃k2 − kb − kθ̃k2 − αkx − π(û)k2
2k3
 
dy ∂h(π(û)) dû dû kg k3 k4 σ
= g(π(û)) = θ1s (û) − kg − − − kθ1 k2
dτ ∂x dτ dτ 2 2 2
σ 1
where θ1s (û) is the equilibrium value of θ1 . + kc kθ̇k2 + θ02 + kdk2
2 2k 4
The steady-state output map is h(π(u)) = `(u). Differen-
 
1 1
tiating with respect to τ along the steady-state yields: − αu kũk2 + Lg Lh kũkkx − π(û)k
τI τI
dy ∂` dû 1
= . − ũT θ̃1 .
dτ ∂u dτ τI
Finally, we bound the last term for some positive constant
Thus following development above, one can always identify k5
the gradient of the steady-state map `(u) as follows:  
kg k5
Ẇ ≤ − ka kη̃k2 − kb − − kθ̃k2 − αkx − π(û)k2
∂` ∂h(π(û)) 2k3 2
= g(π(û)).  
∂u ∂x kg k3 k4 σ
− kg − − − kθ1 k2
Remark 1. This observation suggests a unification of the 2 2 2
proposed approach with the standard ESC which considers σ 1
+ kc kθ̇k2 + θ02 + kdk2
only the gradient of the steady-state objective function.  2 2k
4
1 1 1
Consider the Lyapunov function candidate, W = V + − αu − 2 kũk2 + Lg Lh kũkkx − π(û)k.
 T τI 2τI k5 τI
2 ũ ũ. Differentiating, one obtains:
  As a result, it follows that there exists a strictly positive
kg τI such that the last inequality can be written as:
Ẇ = − ka kη̃k2 − kb − kθ̃k2 − αkx − π(û)k2
2k3  
2 kg k5
Ẇ ≤ − ka kη̃k − kb − − kθ̃k2
 
kg k3 k4 σ
− kg − − − kθ1 k2 2k3 2
2 2 2  
kg k3 k4 σ
σ 1 − kg − − − kθ1 k2
+ kc kθ̇k2 + θ02 + kdk2 − ũT û˙ 2 2 2
2 2k4
σ 1
+ kc kθ̇k2 + θ02 + kdk2
2 2k4
Upon substitution of û˙ = − τ1I θ̂1 , the following inequality − αβ1 (τI∗ )kx − π(û)k2 − αu β2 (τI∗ )kũk2 .
results:
  from corresponding positive constants β1 (τI∗ ) and β2 (τI∗ ).
2 kg As a result,iIt follows that, for every τI > τI∗ , η̃, θ̃, ũ and
Ẇ ≤ − ka kη̃k − kb − kθ̃k2 − αkx − π(û)k2
2k3 θ1 converge to an O(D/τI ) neighbourhood of the origin.

kg k3 k4 σ
 As u approaches a neighbourhood of u∗ , the state x enters
− kg − − − kθ1 k2 a neighbourhood of the steady-state optimum π(u∗ ). This
2 2 2
is achieved by using estimation gains K and kT that are
σ 1  larger than the optimization gain kg to ensure that all
+ kc kθ̇k2 + θ02 + kdk2 + ũT θ̂1
2 2k4 τI constants multiply the corresponding norms (kη̃k2 , kθ̃k2 ,
The term ũT θ1 can be rewritten as: kθ1 k2 , kũk2 and kx − π(û)k2 ) are negative.

∂h(π(û)) 5. SIMULATION EXAMPLE

ũT θ1 = g(π(û))ũ
∂x  
∂h(x) ∂h(π(û)) In this section, we consider the following dynamical sys-
+ g(x) − g(π(û)) ũ.
∂x ∂x tem:
ẋ1 = − x2 + u
By local convexity of the steady-state objective function,
it follows that: ẋ2 =x1 − x2
  The cost function to be minimized is given by: y = x41 +
T 2 ∂h(x) ∂h(π(û)) 6(x1 − 1). The optimum occurs at u∗ = −1.145, x∗1 =
ũ θ1 ≤ −αu kũk − g(x) − g(π(û)) ũ.
∂x ∂x −1.145, x∗2 = −1.145 where y ∗ = −11.15. The tuning
parameters are chosen as: kT = 100, K = 100I, kg = 0.1
Given that h(x) and g(x) are smooth, it follows that there and τI = 1. The dither signal is d(t) = sin(1000t). The
exists Lipschitz constants Lg and Lh such that: simulation results are shown in Figures 1 and 2. Figure
1 shows the objective function and the corresponding
ũT θ1 ≤ −αu kũk2 + Lg Lh kũkkx − π(û)k. input trajectories for the ESC. Figure 2 shows the state
trajectories. The results demonstrate the effectiveness of
As a result, one can write the following inequality for W: the PI-ESC to locate the unknown optimum.

381
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

6. CONCLUSION known control directions. to appear in IEEE Trans. Aut.

Contr., 2013.
In this paper, an alternative proportional-integral ESC B Srinivasan. Real-time optimization of dynamic systems
technique was proposed. The technique is based on the using multiple units. International Journal of Robust
time-varying parameter estimation that allows one to ex- and Nonlinear Control, 17(13):1183–1193, 2007.
ploit a simple parameterization of the unknown dynamics Y. Tan, D. Nesic, and I. Mareels. On non-local sta-
of the cost function. The ESC algorithm is shown to bility properties of extremum seeking control. Auto-
provide local asymptotic convergence of the closed-loop matica, 42(6):889 – 903, 2006. ISSN 0005-1098. doi:
system to the unknown optimum. In contrast to existing 10.1016/j.automatica.2006.01.014.
ESC techniques, no time-scale separation is required to Y. Tan, W.H. Moase, C. Manzie, D. Nesic and, and I.M.Y.
achieve the steady-state optimum. Future work will be Mareels. Extremum seeking from 1922 to 2010. In 29th
focussed on the generalization of the technique to a large Chinese Control Conference (CCC), pages 14 –26, july
class of systems. In particular, the proposed ESC will be 2010.
considered for the solution of steady-state optimization C. Zhang and R. Ordóñez. Robust and adaptive design
problems in unstable systems of numerical optimization-based extremum seeking con-
trol. Automatica, 45(3):634–646, 2009.
REFERENCES C. Zhang and R. Ordóñez. Extremum-Seeking Control and
Applications. Springer, 2012.
V. Adetola and M. Guay. Parameter convergence
in adaptive extremum-seeking control. Automat-
100
ica, 43(1):105 – 110, 2007. ISSN 0005-1098. doi: E SC
Opti mum
10.1016/j.automatica.2006.07.021. 50
V. Adetola and M. Guay. Robust adaptive mpc for systems
y
with exogeneous disturbances. In American Control 0
Conference, St Louis, MO, 2009.
P. Cougnon, D. Dochain, M. Guay, and M. Perrier. On- −50
0 1 2 3 4 5 6 7 8 9 10
line optimization of fedbatch bioreactors by adaptive
extremum seeking control. Journal of Process Control, 5

21(10):1526 – 1532, 2011. ISSN 0959-1524. 0

L. Fu and Ü. Özgüner. Extremum seeking with sliding −5

u

mode gradient estimation and asymptotic regulation for −10

a class of nonlinear systems. Automatica, 47(12):2595– −15

2603, 2011. −20

0 1 2 3 4 5 6 7 8 9 10
A. Ghaffari, M. Krstic, and D. Nesic. Multivariable t
newton-based extremum seeking. Automatica, 48(8):
1759–1767, 2012. Fig. 1. Plot of the cost function and the corresponding
M Guay and T. Zhang. Adaptive extremum seeking control as a function of time.
control of nonlinear dynamic systems with parametric
uncertainties. Automatica, 39:1283–1293, 2003.
3
M. Guay, D. Dochain, and M. Perrier. Adaptive extremum x1
x2
seeking control of continuous stirred tank bioreactors Opti mum

with unknown growth kinetics. Automatica, 40:881–888, 2

2004.
M Krstic. Performance improvement and limitation in 1

extremum seeking control. Systems and Control Letters,

x1, x2

39(5):313–326, 2000. 0

M Krstic and H.H. Wang. Stability of extremum seeking

feedback for general dynamic systems. Automatica, 36 −1

(4):595–601, 2000.
M. Krstic, I. Kanellakopoulos, and P. Kokotovic. Nonlin- −2
ear and Adaptive Control Design. Wiley: Toronto, 1995.
M Leblanc. Sur l’électrification des chemins de fer au −3
0 1 2 3 4 5 6 7 8 9 10
moyen de courants alternatifs de fréquence élevée. Re- t
vue Générale de l’Electricité, 1922.
William H Moase, Chris Manzie, and Michael J Brear.
Fig. 2. State trajectories for the PI-ESC closed-loop pro-
Newton-like extremum-seeking for the control of ther-
cess.
moacoustic instability. IEEE Trans. Autom. Contr., 55
(9):2094–2105, 2010.
D. Nesic, A. Mohammadi, and C. Manzie. A systematic
approach to extremum seeking based on parameter
estimation. In 2010 49th IEEE CDC, pages 3902 –3907,
dec. 2010. doi: 10.1109/CDC.2010.5716937.
A. Scheinker and M. Krstic. Minimum-seeking for clfs:
Universal semiglobally stabilizing feedback under un-

382