Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Downloaded from https://iranpaper.

ir
https://www.tarjomano.com https://www.tarjomano.com

Dissipativity and EXAMINING

Optimal Control THE TURNPIKE


PHENOMENON

LARS GRÜNE

T
he close link between dissipativity and optimal with a peak of activities in the 1990s and monographs such as
control is already apparent in Willems’ first articles those by Freeman and Kokotovic’ [5] or by Sepulchre et al. [6].
on the subject. Particularly, [1] (which appeared In recent years, research on the link between dissipativity
one year before his famous dissipativity articles [2], and optimal control has been revived, with a particular
[3]) already contains much insight on this connec- focus on the turnpike phenomenon and applications in
tion for linear quadratic (LQ) optimal control problems (see model predictive control (MPC). This development was initi-
“Summary”). Around the same time, the role of a passivity- ated approximately 10 years ago in [7] by Diehl et al. in
like variant of dissipativity for inverse optimal control of which strong duality was used to construct a Lyapunov
nonlinear systems was described by Moylan and Anderson function for the closed-loop solution resulting from an eco-
[4], a result that triggered a significant amount of research, nomic MPC scheme. Soon after, it was realized in [8] that

Digital Object Identifier 10.1109/MCS.2021.3139724


Date of current version: 24 March 2022

74 IEEE CONTROL SYSTEMS » APRIL 2022 1066-033X/22©2022IEEE


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

strong duality is nothing other than strict dissipativity with


a linear storage function, and linearity of the storage func- Summary

T
tion is not really needed for this result (meaning that the his article describes the manifold relationships dissipativ-
same Lyapunov function can be constructed for general ity and optimal control, with a focus on recent develop-
nonlinear strict dissipativity). In both cases, the link between ments. In the early days of dissipativity theory, the questions
dissipativity and optimal control lies in the fact that the run- “how can optimal control be used to establish dissipativity?”
ning cost (or the stage cost in discrete time) serves as the and vice versa, “how can dissipativity can be used to estab-
supply rate in the dissipativity formulation. Although these lish (inverse) optimality?” were extensively investigated. Re-
results use strict dissipativity of the optimal control problem cently, the question “what does dissipativity tell us about the
within the MPC scheme in a formal way to obtain stability qualitative behavior of optimally controlled systems?” came
results via an appropriate Lyapunov function, the effect of into focus. This development was triggered by the observa-
strict dissipativity on optimal trajectories can also be ana- tion that dissipativity is extremely useful for the understand-
lyzed in a more geometrical fashion. More precisely, general- ing of (economic) model predictive control schemes, and
izing techniques that were already around in the 1990s (see, opened a new view on the relationship between dissipativity
for example, [9, Th. 4.2]), it was observed in [10, Th. 5.3 and and optimal control.
5.6] that strict dissipativity plus a suitable controllability
property is sufficient for the occurrence of the so-called
turnpike property. This property (first observed in mathe- which is briefly written as x + = f (x, u) . Here, f : R n # R m
matical economy in the works of Ramsey and von Neumann " R n is either the vector field in continuous time or the
in the 1920s and 1930s [11], [12], and first called this way by iteration map in discrete time, while , : R n # R m " R is
Dorfman et al. in the 1950s [13]) describes the fact that opti- called the running cost in continuous time and the stage cost in
mal trajectories stay close to an optimal equilibrium most of discrete time. To unify the notation, we use the symbol [a, b]
the time. Based on this observation, in [14]–[17], (see also [18, in both continuous and discrete time. In continuous time, it
Ch. 7] and [19]) the stability results from [8] could be extended denotes the usual closed interval {t ! R |a # t # b}, while in
to larger classes of MPC schemes and, in addition, nonaver- discrete time it denotes {t ! Z | a # t # b}, where Z is the
aged or transient approximate optimality of the MPC closed set of integers. For simplicity of exposition, this work is
loop could be established. Motivated by these new appli- limited to finite dimensional state space. However, note
cations, the relationship of strict dissipativity to classical that some of the results discussed in this article are also
notions of detectability for linear and nonlinear systems was available in infinite dimensional settings. Given an initial
also clarified [20]–[22]. This article surveys these recent value x 0 ! X, denote the set of control functions u ! U
developments and some of Willems’ early results. for which x (t) ! X and u (t) ! U holds for all t ! [0, T] by
U (x 0, T) . The optimal value function is then defined as
OPTIMAL CONTROL PROBLEMS AND
(STRICT) DISSIPATIVITY VT (x 0) : = inf J T (x 0, u),
u ! U(x 0, T)

Optimal Control Problems and a control u ) ! U (x 0, T) with corresponding trajectory


Consider optimal control problems either in continuous time x)( ∙ ) is called optimal control for initial condition x0 and time
horizon T if
minimize J T (x 0, u) = #0 T , (x (t), u (t)) dt, (1) J T (x 0, u )) = VT (x 0) .

with respect to u ! U, u (t) ! U, and x (t) ! X for all t ! [0, T], The corresponding trajectory x)( ∙ ) is then called an opti-
T ! R 2 0, where U is an appropriate space of functions, mal trajectory.
X 1 R n and U 1 R m are the sets of admissible states and
admissible control inputs, respectively, and Dissipativity and Optimal Control
Dissipativity in the sense of Willems (as used in this article)
xo (t) = f (x (t), u (t)), x (0) = x 0, (2)
involves an abstract notion of energy that is stored in the
or in discrete time, system. For each admissible state x ! R n, denote the energy
T-1 in the system by m (x) . The function m : R n " R is called
minimize J T (x 0, u) = / , (x (t), u (t)), (3) the storage function, and it is usually assumed that m is
t=0
bounded from below to avoid an infinite amount of energy
with respect to u ! U, u (t) ! U and x (t) ! X for all t = 0, f, being extracted from the system. In continuous time, dissi-
T, T ! N, where U is an appropriate space of sequences, X pativity then demands that there exists another function,
and U are as mentioned previously, and the supply rate s : R n # R m " R, such that the inequality

x (t + 1) = f (x (t), u (t)), x (0) = x 0, (4) m (x (x)) # m (x 0) + #0 x


s (x (t), u (t)) dt (5)

APRIL 2022 « IEEE CONTROL SYSTEMS 75


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

holds for all x $ 0, all control functions u ! U (x 0, x), and all connection was first made in [1], which remarkably appeared
initial conditions x 0 ! X. As in the optimal control problem the year before Willems’ seminal articles [2], [3], which
(1), x(t) is supposed to satisfy (2). In discrete time, the introduced and discussed dissipativity in a comprehen-
demanded inequality analogously reads sive way. More precisely, in [1], continuous-time, linear,
x-1
controllable dynamics f (x, u) = Ax + Bu and quadratic run-
m (x (x)) # m (x 0) + / s (x (t), u (t)), (6) ning cost , (x, u) = x T Qx + 2u T Cx + u T Ru without any defi-
t=0
niteness assumption on Q and R, are studied. The article
with x(t) satisfying (4). It appears that this discrete-time vari- provides necessary and sufficient conditions for the exis-
ant of (5) was first used by Byrnes and Lin in [23]. The inter- tence of a storage function m in terms of the finiteness of
pretation of these inequalities is as follows: they demand optimal value functions of certain related optimal control
that the energy m (x (x)) in the system after a certain time x is problems. These characterizations led to the concepts of
not larger than the initial energy m (x 0) plus the integral or available storage and required supply, which are described in
sum over the supplied energy, expressed at each time instant “Available Storage and Required Supply.” What is impor-
by s (x (t), u (t)) . In this interpretation, if s (x (t), u (t)) is nega- tant here is that these characterizations involve constraints
tive, then negative energy is supplied, which means that on the asymptotic behavior of the optimal control problems
energy is extracted from the system. In continuous time, if m under consideration. This already indicates the fundamen-
is continuously differentiable then using the chain rule, tal connection between dissipativity and the long-term
behavior of optimal trajectories, an inspiring research topic
d m (x (t)) = Dm (x (t)) xo (t) = Dm (x (t)) f (x (t), u (t)) until today (as recent publications such as [24] show).
dt
can equivalently be rewritten in infinitesimal form, Dissipativity and Inverse Optimal Control
Passivity is a particular case of dissipativity with supply rate
Dm (x) f (x, u) # s (x, u) . (7)
s (x, u) = u T y = u T h (x) or variants thereof, where y = h (x) is
In discrete time, equivalent to (6), the one-step form the output of the system. As stated in the introduction of [2],
Willems introduced dissipativity as a “generalization of
m (x ) # m (x) + s (x, u) (8)
+
the concept of passivity.” The link between passivity and
can be used, where the brief notation x + = f (x, u) is used. inverse optimal control goes back to [4]. It is described here
To establish these equivalences, the inequalities (7) and (8) following the presentation in [6, Sec. 3.3]. To this end, con-
need to hold for all pairs (x, u) = (x (t), u (t)) that appear in sider the continuous-time functional (1) on the infinite
(5) or (6), respectively. This dissipativity concept can be time horizon, that is, with T = 3 and a running cost , of
2
related to the optimal control problem from the previous the form , (x, u) = ,t (x) + u . Consider, moreover, dynam-
section by setting s (x, u) := , (x, u) for the running or stage ics (2) with f of the form f (x, u) = f1 (x) + f2 (x) u. Then, there
cost , from (1) and (3), respectively. It appears that this exists a cost function ,t $ 0 such that the feedback control

Available Storage and Required Supply


O ptimal control can be used to compute storage functions for
dissipative systems, provided the supply rate s and, in case
$ #0 t- s (x (x), ut (x)) + a^
+ sup
x (x) - x e hdx

# T-- ts (x (x), u (x)) + a^ x (x) - x e h dx


of strict dissipativity, the K 3 -function a are known. More pre- T $ 0, u ! U (x (t), T - t) t

#0 t- s (x (x), ut (x)) + a^ x (x) - x e h dx + V (x (t)),


cisely, the system is strictly dissipative if and only if the optimal
=
value function

V (x 0) := sup #0 T- s (x (x), u (x)) + a^ x (x) - x e h dx which implies (9) for m = V. If each x ! R n can be reached from
T $ 0, u ! U (x 0, T) the equilibrium xe, then another optimal control characterization
of a storage function is given by the required supply. In this
is finite for all initial values x0. In this case, m = V is a storage
case, define
function called available storage. An analogous construction
works without a in the case of nonstrict dissipativity and with V (x) := inf # T s (x (x), u (x)) - a^ x (x) - x e h dx.
T $ 0, u ! U (x 0, T): 0
a sum instead of the integral in the case of discrete-time sys- x (0) = x e, x (T) = x
tems. That V is indeed a storage function follows for any t > 0
and u t ! U (x 0, t) from the inequalities Then, strict dissipativity holds if and only if V is bounded from
below and, again, m = V is a storage function. Here, the storage
V (x 0) = sup # T- s (x (x), u (x)) + a^ x (x) - x e h dx function property follows from the fact that steering from xe to
T $ 0, u ! U (x 0, T) 0
x(t) via x0 cannot be cheaper than steering from xe to x(t) in the
$ sup # T- s (x (x), u (x)) + a^ x (x) - x e h dx
T $ t, u ! U (x 0, T) 0 optimal way. For details on both constructions, refer to [2].

76 IEEE CONTROL SYSTEMS » APRIL 2022


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

The term inverse in inverse optimality stems from the fact that the cost
function ,t is not known but, instead, only implicitly defined by f and k.

u = - k (x) is optimal and stabilizing if and only if the and xe is an equilibrium of the control system, that is, there
closed-loop system is a control value ue with f (x e, u e) = 0 in continuous time
xo (t) = f1 (x (t)) + f2 (x (t)) u (t) and f (x e, u e) = x e in discrete time. Written explicitly, the cor-
responding dissipation inequalities read
y (t) = k (x)

#0 s (x (t), u (t)) - a ^ x (t) - x e h dt (9)


x
is zero-state detectable and dissipative with supply rate m (x (x)) # m (x 0) +
s (x, u) = u T k (x) + (1/2) k (x) T k (x) and continuously differen-
tiable storage function m. Here zero-state detectability means and
that every solution with u / 0 and y / 0 behaves like a solu-
/ s (x (t), u (t)) - a^ x (t) - x e h . (10)
x-1

tion of an asymptotically stable system. The term inverse in m (x (x)) # m (x 0) +


t=0
inverse optimality stems from the fact that the cost function
,t is not known but, instead, only implicitly defined by f and The infinitesimal or one-step versions (7) and (8) of
k (for details, refer to the references cited). Rather than speci- these inequalities then become
fying ,t and deriving k, one specifies or designs k and derives
Dm (x) f (x, u) # s (x, u) - a ^ x - x e h (11)
,t. One may then ask why inverse optimality is an important
property, given that one has only limited control about what and
m (x ) # m (x) + s (x, u) - a ^ x - x h, 
is actually optimized by the feedback law k. The answer is + e
(12)
that optimal feedback laws have a range of beneficial prop-
erties, above all related to robustness (refer to [5] and [6] and respectively. For some applications, it is necessary to use
a ( (x, u) - (x , u ) ) in place of a ^ x - x h . In this case, one
e e e
the references therein for details). The fact that k is a stabiliz-
ing feedback law is crucial for this result. Hence, again, there speaks about strict (x, u)-dissipativity. It is easily shown that
is a connection between dissipativity and the long-term (9) and (10) are more demanding than (5) and (6), respec-
behavior of optimal trajectories. In summary, the two classi- tively, as they not only demand that the energy difference
cal lines of research m (x (t)) - m (x 0) is bounded by the integral or sum over the
»» revealed that optimal control problems can indicate supplied energy s, but that some of the energy is actually
whether dissipativity holds (for example, via the avail- dissipated if the system is not in the equilibrium xe and the
able storage or the required supply) amount of dissipated energy increases the further away the
»» showed that dissipativity conditions on stabilizing state x (x) is from xe. At first glance, it seems that the differ-
feedback laws imply that they solve an optimal con- ence between strict and nonstrict dissipativity is only quan-
trol problem for a suitable cost function. titative. Indeed, it follows readily from the definition that if
The recent renaissance of the investigation of the link the system is dissipative with supply s(x, u), then it is strictly
between dissipativity and optimal control, which is described dissipative with supply s (x, u) + a ^ x h . However, if the
in the remainder of this article, investigates a third aspect. supply function s is not merely a parameter that is played
Namely, its goal is to with but a function that results from some modeling proce-
»» use dissipativity to make statements about the long- dure, then there is both a quantitative and qualitative differ-
time behavior of optimal and near-optimal trajectories. ence between strict and nonstrict dissipativity. This applies
Before turning to the description of this recent devel- for the setting in which s is derived from the running or
opment, a stricter variant of the dissipativity property stage cost , of an optimal control problem, that is, when
is introduced. s = ,. When using strict dissipativity for the analysis of the
control system’s behavior, it is often necessary to demand
Strict Dissipativity s (x e, u e) = 0, as shown in the next section. In general, this is
Strict dissipativity is simply dissipativity with storage function a restrictive condition. However, it is not restrictive in the
s (x, u) - a ^ x - x e h
case where s is derived from the optimal control problem.
This is because the optimal trajectories for the cost , (x, u)
in place of s(x, u), where a ! K 3 with are the same as for the cost , (x, u) - , (x e, u e) . It is thus conve-
nient and not restrictive to use the supply rate
K 3 := {a : [0, 3) " [0, 3)|a c ontinous, strictly increasing,
unbounded, and a (0) = 0} s (x, u) = , (x, u) - , (x e, u e) . (13)

APRIL 2022 « IEEE CONTROL SYSTEMS 77


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

THE TURNPIKE PROPERTY control problem has the turnpike property if there is an equi-
The turnpike property demands that optimal trajectories librium xe such that for any f > 0 and K > 0, there is a C > 0
(and in some variants, also near-optimal trajectories) stay such that for all time horizons T > 0 and all optimal trajec-
in the vicinity of an equilibrium xe most of the time. Of tories x)( ∙ ) with x ) (0) ! B K (x e) + X, the set of times
course, this statement must be made mathematically pre-
! B f (x e)}
Q := {t ! [0, T] |x ) (t) Y
cise to analyze it mathematically. The meaning of “most of
the time” is that the amount of time the trajectory spends satisfies Q # C. Here, Q is the Lebesgue measure of Q in
outside a given neighborhood of the equilibrium xe is continuous time or the number of elements contained in Q
bounded, and the bound is independent of the optimiza- in discrete time, and B K (x e) and B f (x e) denote the balls
tion horizon and of the initial condition (at least for the ini- around xe with radii K and f, respectively. Figure 1 depicts
tial conditions contained in a bounded set). This article the sketch of a trajectory exhibiting the turnpike phenome-
focuses on the turnpike property for finite-horizon optimal non. The set Q contains seven time instants, which are
control problems. Refer to [25] for an infinite-horizon ver- marked with short, red vertical lines on the t-axis. The equi-
sion and the relationship between the finite- and infinite- librium xe is represented by the blue dashed line. Figure 2
horizon cases and to [24] for the impact of dissipativity shows optimal trajectories for varying time horizons
on infinite-horizon optimal control. Formally, an optimal T = 5, 7, f, 19 for two simple, 1D, discrete-time, optimal
control problems. The first problem is given by
5.5 x + = u, , (x, u) = - log (5x 0.34 - u),
5 X = [0, 10], U = [0.01, 10], (14)
4.5
and the second by
4
3.5 x + = 2x + u, , (x, u) = u 2, X = [- 2, 2], U = [- 3, 3] . (15)
3
x ∗(t )

The model (14) is an optimal investment problem from


2.5
Bε (x e )

[26], in which x denotes the investment in a company, and


2
5x 0.34 is the return from this investment (including the
1.5
investment itself) after one time period. As x + = u is the
1
investment in the next time period, 5x 0.34 - u is the amount
0.5
of money that can be used for consumption in the current
0
0 5 10 15 20 25 time period, and the optimal control problem thus models
t the maximization of the sum of the logarithmic utility func-
tion log (5x 0.34 - u) over the time periods. Clearly, it is opti-
FIGURE 1 An illustration of the turnpike property in discrete time.
mal to spend all the available money until the end of the
The set Q contains seven time instants, which are marked with
short, red, vertical lines on the t-axis. The equilibrium xe is repre- time horizon T, which is why all optimal trajectories end at
sented by the blue dashed line. x = 0. However, in between, the trajectories spend most of

5.5
5 1.8
4.5 1.6
4 1.4
3.5 1.2
3 1
x ∗(t )

x ∗(t )

2.5 0.8
2 0.6
1.5 0.4
1 0.2
0.5 0
0 –0.2
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
t t
(a) (b)

FIGURE 2 (a) The optimal trajectories for example (14) and (b) the optimal trajectories for example (15) for time horizon length T = 5, 7, f, 19.

78 IEEE CONTROL SYSTEMS » APRIL 2022


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

the time near the equilibrium x e . 2.2344, that is, it exhibits for all t ! [0, T] with constants C, v 2 0. It is easily shown
the turnpike property. Why this occurs is explained later. that this property implies the aforementioned turnpike
The task modeled in the second example is to keep the property (as for each f 2 0, x 2 0 with C ^e - v 1 t + e - v 2 (T - t)h 1 f
state of the system in X with as little quadratic control effort for all t ! [x, T - x] exists, regardless of how large T is).
, (x, u) = u 2 as possible. In the long run, it is beneficial to stay Another variant demands the turnpike behavior not only
near x e = 0 because the control effort for staying in X is very for the optimal trajectories x)( ∙ ), but also for all trajectories
small near this equilibrium. Thus, it makes sense for the corresponding to control functions satisfying J (x 0, u) #
optimal trajectories to stay near x e = 0 for most of the time. V (x 0) + d, that is, for all near-optimal trajectories. For even
At the end of the horizon, however, it is beneficial to turn off more variants and a historic discussion of the turnpike
the control completely because this entirely reduces the con- property, refer to the recent survey in [35].
trol effort to zero and does not violate the state constraint
(provided that the control is turned off sufficiently late). Strict Dissipativity and the Turnpike Property
In Figure 2(a) and (b), it is clearly visible that the number of Under a boundedness assumption on the optimal value
states outside a neighborhood of the respective equilibrium functions, it is fairly easy to prove that strict dissipativity
(indicated by the blue dashed line) remains constant with implies the turnpike property. The boundedness assumption
increasing horizon length, which is exactly what the turnpike demands that there are a K 3 -function c and a constant C > 0,
property demands. These examples show that the turnpike such that the optimal value and the storage functions satisfy
property appears in very simple problems. It is, however,
known that it also occurs in much more complicated prob- VT (x) - VT (x e) # c ^ x - x e h + C and
lems, including optimal control problems governed by par- m (x) # c ^ x - x h + C (16)
e

tial differential equations; [27]–[34] is just a small selection of


recent references where this is investigated. Interestingly, the for all x ! X and all T $ 0. The first condition can be ensured
two examples exhibit a particular form of the turnpike prop- by a reachability condition: if the equilibrium xe can be
erty, in which the exceptional points (that is, the points cor- reached from every x ! X with costs that are bounded in
responding to the times in Q) lie only at the beginning and at bounded subsets of X, then the inequality for VT (x) - VT (x e)
the end of the time interval. One speaks of the approaching arc can be established for all T > 0; for details, see [10, Sec. 6]. The
at the beginning and the leaving arc at the end of the time derivation of the turnpike property is illustrated under these
horizon. Although the general definition of the turnpike boundedness assumptions in the discrete-time setting; refer
property allows for excursions from xe also in the middle of to [36, Sec. 2] for a continuous-time version of this result.
the time interval (as indicated in the sketch in Figure 1), these Using (10) and (13) and choosing D > 0 such that m (x) $ - D
do not appear in the two examples. The reason for this will be for all x ! X yields
explained in the next section.
T- 1
The optimal trajectories in Figure 2(a) and (b) nicely J (x 0, u) = / , (x (t), u (t))
illustrate the source of the name turnpike property, which t= 0

/ a^ x (t) - x e h + T, (x e, u e) - m (x 0) + m (x (T))
T- 1
was coined in [13]: The behavior of the optimal trajectory is $
similar to a car driving from the initial to the endpoint, t= 0

/ a^ x (t) - x e h + T, (x e, u e) - m (x 0) - D.
T- 1
where the equilibrium—that is, the blue dashed line—
$
plays the role of a highway (or turnpike as highways are t= 0
called in parts of the United States). If the time is suffi-
ciently large, it is beneficial to first go to the highway (even Moreover, using the constant control u / u e yields
if this is associated with some additional cost), stay there T- 1
for most of the time, and then leave the highway at the end. VT (x e) # J T (x e, u) = / , (x e, u e) = T, (x e, u e) .
t= 0
There may be different reasons for the occurrence of the
leaving arc. A constraint x (N) = xt may be imposed for the Together, this implies for x = x ) with x ) (0) = x 0,
terminal state, which forces the trajectory to leave the turn-
c ^ x 0 - x h + C $ VT (x 0) - VT (x )
e e
pike, or it may simply be beneficial to leave the turnpike
because this reduces the cost of the overall trajectory. As no
/ a^ x ) (t) - x e h - m (x 0) - D.
T- 1
$
terminal constraints are imposed, the latter must be the k= 0
case in these examples, and it is actually easy to check that
this is the case. Besides the definition given previously, Given an arbitrary f 2 0, if x ) spends too much time
there are a number of alternative definitions of the turn- outside B f (x e), then this inequality is violated. From this
pike property. A widely used variant is the exponential turn- fact, the existence of the constant C in the turnpike prop-
pike property, which demands an inequality of the type erty can be concluded. The inequalities in (16) are needed to
ensure that the constant C in the turnpike property depends
x ) (t) - x e # C ^e - vt + e - v (T - t)h only on the size K of the ball B K (x e) containing x0 and not

APRIL 2022 « IEEE CONTROL SYSTEMS 79


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

on the individual initial value x0. It is easily confirmed that a similar inequality as mentioned previously shows that an
both examples (14) and (15) are strictly dissipative, the excursion from the equilibrium xe followed by a return to xe
former with the storage function m (x) = a (x - x e) /x e and results in costs that are significantly higher than staying in
the latter with the storage function m (x) = - x 2 /2. This xe. Under the assumption that the optimal values are con-
explains why the turnpike behavior is observed in these tinuous near xe (in the sense that the optimal value function
examples. Note that in both examples, the boundedness of near xe has roughly the same values as in xe —a property
X is important to ensure that m is bounded from below on that can again be rigorously established for locally control-
X. With a more involved proof, strict dissipativity with lable systems), this implies that such excursions will not
suitable bounds on the problem data can also be used to happen in (near-)optimal trajectories. In contrast, an excur-
obtain an exponential turnpike property [37]. Alternatively, sion from the equilibrium xe can be cheaper than staying in
the exponential turnpike property can be established using xe if the solution does not return to xe after the excursion.
necessary optimality conditions (refer to [27], [31], [34], and This is precisely the effect that creates the leaving arc of the
[38] for a selection of articles in this direction). However, optimal trajectories. As the gain that can be obtained from
the strict dissipativity-based approach has the advantage the leaving arc is bounded by the value of the storage func-
that it works for arbitrary nonlinearities, while the optimal- tion m (x (T)) at the terminal state of the trajectory, it is
ity condition-based approach is typically limited to linear important that m is bounded from below. Indeed, (15) would
systems or to small nonlinearities via linearization argu- cease to be strictly dissipative if X = [- 2, 2] was changed to
ments. The approach based on optimality conditions, in X = R. In this case, it is easily shown that the optimal trajec-
turn, has the advantage that it also provides a turnpike tories would tend to !3 (with optimal control u ) / 0)
property for the adjoint states. For the dissipativity-based instead of staying near x e = 0 most of the time, that is, the
approach, such an implication is, thus far, known only for turnpike property also ceases to exist.
the particular case of a so-called interval turnpike prop-
erty, in which the solutions are required to stay near the Related Properties
turnpike during a sufficiently large interval of time [39]. Strict dissipativity generalizes a lot of well-known properties
The fact that strict dissipativity implies the turnpike for optimal control problems, ensuring a certain asymptotic
property immediately raises the question of how much behavior for the optimal trajectories. First, it is easily seen that
stronger strict dissipativity is than the turnpike property. strict dissipativity holds with storage function m / 0 for any
The answer lies in the observation that the inequality cost function , satisfying , (x e, u e) = 0 and , (x, u) $ a ^ x - x e h
chain mentioned previously can also be used for establish- for all x ! X and u ! U. In the case of a quadratic cost
ing the turnpike property for trajectories and controls that
, (x, u) = (x - x e) T Q (x - x e) + (u - u e) T R (u - u e),
are not strictly optimal but only satisfy the inequality
J T (x 0, u) # VT (x 0) + d, that is, they are near optimal. In other this amounts to requiring that Q is positive definite. For
words, a near-optimal turnpike property is obtained. For sys- LQ problems with dynamics f (x, u) = Ax + Bu, generalized
tems that are locally controllable in a neighborhood of xe quadratic cost
and controllable to this neighborhood from all x ! X, it was
, (x, u) = x T Qx + u T Ru + q T x + r T u
shown in [40] that in discrete time, strict dissipativity is
equivalent to the near-optimal turnpike property. To the with Q = C T C, and no state constraints (that is, X = R n ),
best of the author’s knowledge, results in this generality are strict dissipativity is equivalent to detectability of the pair
not yet known in continuous time. However, results under (A, C) (that is, all unobservable eigenvalues μ of A satisfy
the stronger assumptions that the turnpike is exact or that Re n 1 0 in continuous time or n 1 1 in discrete time). If X
the running cost , is locally positive definite with respect to is bounded with xe in its interior, then strict dissipativity is
(x e, u e) can be found in [36, Sec. 4]. Interestingly, the proof of equivalent to all unobservable eigenvalues μ of A satisfying
the equivalence in discrete time relies on a similar equiva- Re n ! 0 in continuous time or n ! 1 in discrete time (see
lence result for plain (that is, nonstrict) dissipativity. In [41] [20] and [21] for proofs in discrete and continuous time,
(for discrete-time systems) and in [36] (for continuous-time respectively). Note that the last criterion applies to (15).
systems), it was shown that, again, under a local controlla- Although previously cited results apply to problems in
bility assumption, dissipativity is equivalent to the fact that which the R-matrix is positive definite, some of the implica-
the average cost for all admissible u satisfies the inequality tions were also proven in the indefinite case [42]. This article
also extends the analysis in [20] to quadratic costs including
liminf 1 J T (x 0, u) $ , (x e, u e), (17) cross terms of the form xTSu by using a transformation of a
T"3 T
cost function with cross terms into one without such terms.
a property known as optimal operation at steady state. Strict For nonlinear costs, it can be shown that strict dissipativity
dissipativity also explains why optimal trajectories exhib- holds if f is affine and , is strictly convex, a criterion that
iting the turnpike property typically look as they do in applies to (14). In this case, the storage function can be chosen
Figure 2 and rarely as they do in Figure 1. The reason is that as a linear function m (x) = p T x, where p ! R n is the Lagrange

80 IEEE CONTROL SYSTEMS » APRIL 2022


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

multiplier of the optimization problem for determining the course) only an indication and not a rigorous check. A rigor-
optimal equilibrium, that is, in discrete time ous check can be obtained by computing a storage function
min , (x, u) s.t. f (x, u) = x. m numerically. Given the conceptual similarity between stor-
(x, u) ! X # U
age functions and Lyapunov functions, it seems plausible
A proof can be found in [37], and it is straightforward to that most computational techniques for Lyapunov functions
apply the statement and proof in continuous time. Note (for a survey see, for example, [45]) can be modified for
that this fact was known and used before [7]. However, an computing storage functions. A method for which this has
earlier reference to a formal proof could not be found. Note already been done is the sum-of-squares approach, which
that a linear function of the form m (x) = p T x is, in general, was applied for the computation of storage functions in [46]
not bounded from below, as required from a storage func- in continuous time and in [47] in discrete time. The main
tion. Hence, appropriate state constraints must be imposed bottleneck of this approach—and of most other numerical
to guarantee strict dissipativity. For fully nonlinear dis- approaches for computing Lyapunov-like functions—is the
crete-time systems, as shown in [22], strict dissipativity also fact that the computational effort of the methods grows rap-
follows from nonlinear detectability notions, such as the idly with the space dimension of the problem. This is the
one introduced in [43] or input–output-to-state stability so-called curse of dimensionality. This problem vanishes for
[44]. The former connection is briefly explained. The non- linear-quadratic problems because one can always resort to
linear detectability condition from [43] demands the exis- linear-quadratic storage functions, which can be efficiently
tence of a nonnegative function W : X " R $ 0 and K 3 computed using linear matrix inequality techniques [48].
functions ai, i = 1, f, 3 such that the inequalities Recently, a purely data-driven approach for verifying dissi-
pativity for this setting was developed [49].
W (x) # a 1 ^ x - x e h For general nonlinear problems, a promising way appears
W ( f(x, u)) - W (x) # - a 2 ^ x - x e h + a 3 (, (x, u)) to be the use of approximation architectures that can over-
come the curse of dimensionality, such as deep neural net-
hold for all x ! X and u ! U. In this setting, , (x, u) $ 0 is works. These architectures are provably able to compute
assumed for all x ! X and u ! U as well as , (x e, u e) = 0, approximate solutions of certain partial differential equa-
which is consistent with the fact that detectability applies tions in high space dimensions [50], [51], and it was recently
to optimal control problems that are designed for stabiliz- shown that the same is true for Lyapunov functions [52]. Of
ing a given equilibrium (here, xe). Under a growth assump- course, this cannot work in general but only under suitable
tion on a3 (which holds without loss of generality if , is structural properties of the problem under consideration.
bounded on X # U ), it was shown in [22] that for an appro- However, there is hope that these are related to well-known
priately chosen t ! K 3, m = t & W is a storage function cer- system theoretic properties (such as small-gain conditions in
tifying strict dissipativity for the supply rate s (x, u) = , (x, u) . the case of [52]), which are well understood. It is quite likely
Conversely, if the problem is strictly dissipative and that progress in this direction will soon be seen, particularly
e
m (x ) # m (x) for all x ! X, then the problem is also detect- in connection with reinforcement learning approaches.
able with W (x) = m (x) - m (x e) . As mentioned previously,
these results are in discrete time and we are not aware of AN APPLICATION: MODEL PREDICTIVE CONTROL
the existence of continuous-time counterparts, although we MPC, as described in “Model Predictive Control,” is a
surmise that such a result should be feasible. highly popular control method in which the computation-
ally challenging solution of an infinite-horizon optimal
Verifying (Strict) Dissipativity of control problem is replaced by the successive solutions of
Optimal Control Problems optimal control problems on finite time horizons. As des­­
As (strict) dissipativity of optimal control problems has cribed in “Model Predictive Control,” this induces an error
many useful applications, it is of course desirable to be able that can be analyzed by employing strict dissipativity of
to verify whether an optimal control problem has this prop- the optimal control problem.
erty. If none of the sufficient structural conditions discussed
in the previous section can be used for this purpose, then Stability Analysis
other methods must be applied. To this end, a couple of algo- Strict dissipativity is useful for two important aspects of
rithmic approaches were developed in the last few years. this analysis. The first aspect is stability analysis. In gen-
The conceptually simplest approach relies on the relation- eral, it is not clear that the MPC closed-loop solutions will
ship between strict dissipativity and the turnpike property. exhibit stability-like behavior. However, under the assump-
One can simply solve the optimal control problem numeri- tion that strict dissipativity holds, an auxiliary optimal
cally for different initial conditions and time horizons and control problem can be defined by using the modified or
check whether the turnpike property holds, which is very rotated running or stage cost
easy to observe. This already yields a good indication of
whether strict dissipativity may hold. However, it is (of ,u (x, u) := , (x, u) - , (x e, u e) - Dm (x) f (x, u)

APRIL 2022 « IEEE CONTROL SYSTEMS 81


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

in continuous time or to uJ T (compare, for example, [53], [54], and [55, Ch. 2], [18,
Ch. 5]) or by imposing conditions that ensure that Vu T is
,u (x, u) := , (x, u) - , (x e, u e) + m (x) - m ^ f (x, u) h close to Vu 3 for sufficiently large T (compare, for example,
[43] or [18, Ch. 6]), it can be proven that Vu T is a Lyapunov
in discrete time. It is immediate from the strict dissipa- function for the MPC closed loop. However, this is true
tivity inequalities (11) or (12) that this modified stage only if ,u is used as the running or stage cost in the MPC
cost is positive definite at xe, that is, it satisfies u, (x e, u e) = 0 scheme. This, however, is often not practical because the
and ,u (x, u) $ a ^ x - x e h . Thus, if we consider the optimal storage function may not be known and is difficult to
control problems (1) or (3) with ,u in place of , , it forces compute. In this case, it is desirable to keep the original
the optimal solutions to approach xe. Denote the corre- cost , in the optimal control problem. In case the run-
sponding functional and optimal value function by uJ T ning or stage cost , , together with the terminal cost
and Vu T, respectively, and the optimal solution by xu ) ( ∙ ) F, is used, the trick is to define ,u as mentioned and
By either adding a terminal cost Fu with suitable properties Fu : = F + m. Then, the optimal control problems using ,

Model Predictive Control

M odel predictive control (MPC) is one of the most success-


x MPC(t )
ful optimization-based control techniques, with ample
applications in industry [73], [74]. MPC is described here in
discrete time, noting that the adaptation to continuous time
is relatively straightforward. For further reading, refer to the
monographs in [18] and [55]. In many regulation problems, the
goal is to solve the infinite-horizon optimal control problem

3
minimize J 3 (x 0, u) = / , (x (t), u (t)) (S1)
k= 0
t
with respect to u ! U, u (t) ! U and x (t) ! X for all t = 0, 1, 2, f, 0 1 2 3 4 5 6
where U is an appropriate space of sequences, X and U are FIGURE S1 A sketch of the solutions generated in the model
as mentioned, and x(t) satisfies x (0) = 0 and (4). Unless the predictive control (MPC) loop. The solid red line depicts the
problem has a very particular structure (as, for example, lin- MPC closed-loop solution, while the black dashed lines indi-
ear dynamics, quadratic costs, and no constraints), a closed- cate the predictions. The figure sketches the ideal case in
which the closed-loop dynamics exactly coincide with the
form solution of (S1) is usually not available, and (due to the
dynamics used to compute the predictions.
infinite time horizon) a numerical solution is very costly to
obtain, particularly if one wants to obtain the optimal control
in feedback form (that is, in the form u ) (t) = F (x (t)) for a suit- line depicts the MPC closed-loop solution, while the black
able map F). The key idea of MPC now consists of truncating dashed lines indicate the predictions. The figure shows the
the optimization horizon and, instead of (S1), solve (3) or a ideal case in which the closed-loop dynamics exactly co-
variant thereof. This variant may include additional terminal incide with the dynamics used to compute the predictions.
constraints of the form x (T) ! X 0 and/or a terminal cost of the Note that by means of the dynamic programming principle
form F (x (T)) as an additional summand in J T . The MPC loop (see [18, Th. 4.6]), the solution strategy “solve the optimal
then proceeds as control problem and use the first element of the resulting
1) Select an initial value x MPC (0) and a time horizon T. Set optimal control sequence as feedback control value” would
k := 0. yield an optimal feedback law if the infinite-horizon problem
2) Set x 0 := x MPC (k) and solve (3) with this initial value. De- (S1) in step 2 was solved. However, by resorting to the (nu-
note the resulting optimal control sequence by u)( ∙ ) and set merically much more easy to obtain) solution of the finite-
FMPC (x MPC (k)) := u ) (0) . horizon problem (3) in step 2, an error is made that must be
3) Apply FMPC (x MPC (k)), measure x MPC (k + 1), set k := k + 1, analyzed. The MPC for stage costs considered in this ar-
and go to 2). ticle, which do not merely penalize the distance of the state
The resulting solution x MPC (t) is called the MPC closed- from a desired steady state, is often denoted as economic
loop solution, while the individual open-loop finite-horizon MPC (although the term general MPC would probably be
optimal solutions computed by solving (3) in step 2 are more fitting). In any case, strict dissipativity of the optimal
called the predictions. Figure S1 presents a schematic control problem plays an important role in the analysis of
sketch of the resulting solutions loop. Here the red solid such MPC schemes.

82 IEEE CONTROL SYSTEMS » APRIL 2022


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

and F produce the same optimal solutions as the one constraint X 0 and costs F and Fu . In this case, however, the
using ,u and Fu , and the analysis described in the previ- trick of enforcing identical solutions of the optimal control
ous paragraph can be used, without the need to actually problems with , and with ,u by setting Fu : = F + m is no longer
compute ,u and Fu . This is because the conditions on the applicable. In fact, without terminal costs, the optimal trajec-
terminal cost needed to make Vu T a Lyapunov function tories and controls of the problem with , and with ,u no
can be formulated directly for F without having to use longer coincide. However, they may still coincide approxi-
(or even to know) m. The resulting required inequality mately in a suitable sense.
for F is of the form To establish this property, assume that the storage func-
tion m and the optimal value functions V T and Vu T are
DF (x) f (x, u) # - , (x, u) + , (x e, u e) (18)
continuous in xe, uniformly in T (the former is not very
in continuous time and restrictive as m is often a polynomial and the latter can again
be ensured by local controllability, see [10, Sec. 6]). Both the
F ^ f (x, u) h # F (x) - , (x, u) + , (x e, u e) (19) problem with cost , and the problem with cost ,u are strictly
dissipative and thus exhibit the turnpike property. From
in discrete time, which must hold for all x in the terminal this, it can be concluded that for any two optimal trajecto-
constraint set X 0 and a control value u (depending on x), ries x ) ( ∙ ) and xu ) ( ∙ ) starting in the same initial value x0,
which is such that the solution does not leave X 0 . As the there will be a time x such that they satisfy xu ) (x) . x e and
optimal control problems with ,u and Fu on one side and with x ) (x) . x e, where the error hidden in “ . ” tends to zero
, and F on the other side produce the same optimal solu- as the horizon T increases. Together with the continuity
tions, the MPC closed loops resulting from the two problems assumption on the optimal value function, this implies that
also coincide, and Vu T can again be used as a Lyapunov func- the cost of the two trajectories on the time interval [0, x] is
tion to conclude asymptotic stability. This trick was first pro- almost identical. This can finally be used to conclude that
posed in [7] and then refined in [8]. Figure 3 shows the MPC Vu T is an approximate Lyapunov function, from which prac-
closed-loop trajectories and the corresponding predictions tical asymptotic stability (that is, asymptotic stability of a
for (15) using the terminal constraint set X 0 = {0} and the ter- neighborhood of xe, which shrinks to {x e} when T increases)
minal cost F / 0. In many cases, it may be difficult to find a can finally be concluded. The details of this reasoning were
terminal cost F that meets the required condition (18) or (19). originally derived in [10] and [14] in discrete time and [15] in
Although the trivial choice X 0 = {x e} and F / 0 always satis- continuous time. A concise presentation can be found in [18,
fies (18) or (19), this choice of X 0 may cause problems in the Sec. 8.6] or [19, Sec. 4].
numerical optimization and result in a small set of feasible Figure 4 shows the MPC closed-loop trajectories and the
states for (1) or (3) when the terminal condition x (T) ! X 0 corresponding predictions for example (15) without any ter-
is added. It may thus be attractive to drop the terminal minal conditions. In comparison to Figure 3, the merely

2 2
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
x MPC(t )

x MPC(t )

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18 20
t t
(a) (b)

FIGURE 3 The model predictive control closed-loop trajectories (solid red) and predictions (black dashed) for (15) using the terminal
constraint set X 0 = {0}, terminal cost F / 0, and (a) horizon T = 3 and (b) T = 5. Note that the terminal constraint x(T) = 0 forces all pre-
dictions to end in x e = 0.

APRIL 2022 « IEEE CONTROL SYSTEMS 83


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

practical asymptotic stability of the equilibrium x e = 0 is 3

clearly visible as the red closed-loop solution does not con-


J MPC
3 (x 0) := / , (x MPC (t), u MPC (t))
t= 0
verge to the equilibrium x e = 0 but only to a neighborhood
of this equilibrium (which becomes smaller as T increases). would be the “natural” measurement if MPC is considered
Note that under stronger assumptions than strict dissipa- an approximation to an infinite-horizon problem. How-
tivity, stronger stability statements can be made for MPC ever, as the infinite sum may not converge, other measure-
without terminal conditions. For instance, under the non- ments are also considered. We also consider the finite-horizon
linear detectability condition discussed in the previous closed-loop performance
section and suitable bounds on the optimal value function, S- 1
it was shown in [43] that true (as opposed to merely practi- J MPC
S (x 0) := / , (x MPC (t), u MPC (t)) (20)
t= 0
cal) asymptotic stability of xe for the MPC closed loop can
be expected for sufficiently large T. For positive-definite and the averaged infinite-horizon performance
costs, different ways of estimating the length of the horizon
T needed for obtaining asymptotic stability were proposed Jr MPC
3 (x 0) := limsup 1 J SMPC (x 0) .
S"3 S
in [56]–[59] (see also [18, Ch. 6]). For LQ problems, it was
shown in [60] that for indefinite problems that yield asymp- These last two performance measurements complement
totically stable optimal solutions, it is always possible to each other as the first measures the performance on finite
find an equivalent LQ problem with positive-definite costs, intervals [0, S], while the second measures the perfor-
to which these results are applicable. Finally, under suitable mance in the limit for S " 3. The derivation of estimates
conditions, true asymptotic stability can also be enforced for these quantities heavily relies on strict dissipativity and
using linear terminal costs, which are much simpler than the turnpike property (which are exploited for MPC, both
those in (18) or (19) and derived from the adjoints of the with and without terminal conditions) as well as on the
optimal control problem [61]. stability properties described in the previous section. The
key idea is to use the similarity of the initial pieces of
Performance Analysis optimal trajectories until they reach the optimal equi-
As MPC relies on the solution of optimal control prob­lems, librium to derive approximate versions of the dynamic
it seems reasonable to expect that the MPC closed loop also programming principle. From these, estimates on the
enjoys certain optimality properties. This is the second aforementioned quantities can be derived by induction
aspect of the analysis of MPC schemes where dissipativity over t. The following discrete-time estimates were origi-
is helpful. To this end, denote the MPC closed-loop trajec- nally developed in [10], [14], and [16]. Concise presenta-
tory by x MPC (t) and the corresponding control by u MPC (t) . tions can be found in [17] or [18, Ch. 8]. In continuous time
Then, different measurements for the closed-loop cost can (to the best of the author’s knowledge), only the results for
be defined: The infinite-horizon performance the averaged performance are available, which can be

2 2
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
x MPC(t )

x MPC(t )

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 18 20 0 5 10 15 20 25
t t
(a) (b)

FIGURE 4 The model predictive control closed-loop trajectories (solid red) and predictions (black dashed) for example (15) without terminal
conditions with (a) horizon T = 5 and (b) T = 10. Without any terminal constraints, all predictions end in x = 2, compare also Figure 2(a).

84 IEEE CONTROL SYSTEMS » APRIL 2022


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

found in [62]. For MPC with terminal conditions satisfy- with d1 and d2 of the same type as mentioned previously.
ing (19), the identity Note that the increase of the error term d 1 (T) in (23) to
Sd 1 (T) in (25) is not an effect of an insufficiently precise
Jr MPC
3 (x 0) = , (x e, u e) (21)
analysis, but instead a natural consequence of the mere
holds for all N ! N, and because of (17), this is the best possible practical asymptotic stability of xe without terminal con-
value the average performance functional can attain. In case ditions. If the closed-loop solution does not converge to xe,
V3 assumes finite values, there exists a function d 1 : N " [0, 3) the stage cost will typically not converge to zero but to a
with d 1 (T) " 0 as T " 3 such that the inequality nonzero residual value (which keeps accumulating over
the time S). To illustrate this effect, Figure 5 shows the
J 3MPC (x 0) # V3 (x 0) + d 1 (T) (22) values of J MPC
S (x 0) for (15) with terminal conditions (%) and
without terminal conditions (×) for different S. In Figure 5(a)
holds. In case V3 does not assume finite values, the estimate with horizon T = 5, the effect of the additional factor S in
front of d 1 (T) in (25) is clearly visible. In Figure 5(b) with
J SMPC (x 0) # inf S J S (x 0, u) + d 1 (T) + d 2 (S) (23) horizon T = 10, the error term d 1 (T) in (25) is already so
u!N
U
small that the effect of the factor S is no longer visible. For
can be obtained for each S ! N, where d2 is the same type this horizon, MPC with and without terminal conditions
of function as d1. Here U u S denotes the set of admissible yield almost the same performance. Just as for stability,
c o n t r o l s f o r w h i c h x u (S, x 0) - x e # x MPC (S, x 0) - x e stronger statements can be made under stronger assump-
holds. As x MPC (S, x 0) " x e holds for S " 3, for large S, the tions. For instance, for positive-definite cost ,, the finite-
quantity infu ! Uu S J S (x 0, u) measures the optimal transient ness of J MPC
3 (x 0) and an estimate of the form (22) can also
cost for trajectory going from x0 to a small neighborhood of be obtained for MPC without terminal conditions (see [58]
xe. Thus, the estimates show that MPC produces trajectories and [59]).
that yield optimal averaged cost and approximately opti-
mal transient cost. Without terminal conditions, the results SUMMARY, EXTENSIONS, AND OUTLOOK
become somewhat weaker. Particularly, J MPC 3 (x 0) is no Dissipativity and strict dissipativity are important sys-
longer guaranteed to be finite, even if V3 (x 0) is finite. How- tems-theoretic properties with multiple applications. This
ever, the counterparts of (21) and (23) can still be estab- article demonstrated that they naturally link to optimal
lished, namely, control, a fact that is already prominently present in Wil-
lems’ earliest publications on the subject. This article sur-
Jr 3
MPC
(x 0) # , (x e, u e) + d 1 (T) (24) veyed recent results in this direction, which establish a
and close link between strict dissipativity and the turnpike
property, including how these concepts can be used for
J MPC (x 0) # inf S J S (x 0, u) + Sd 1 (T) + d 2 (S), (25)
u!N
S
U analyzing the stability and performance of MPC schemes.

13 12.5

12.5 12
12
11.5
11.5
11
11
JSMPC (x0)

JSMPC (x0)

10.5 10.5

10 10
9.5
9.5
9
9
8.5
8 8.5
0 5 10 15 20 25 0 5 10 15 20 25
S S
(a) (b)

FIGURE 5 The model predictive control closed-loop cost J MPC


S (x 0) for (15) with x 0 = 2 for varying S. The solid line with circles shows the
values with terminal conditions X 0 = {0} and F / 0, and the dashed line with “xs” is the values without terminal conditions, with (a) hori-
zon T = 5 and (b) T = 10.

APRIL 2022 « IEEE CONTROL SYSTEMS 85


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

The present results have already been extended in various [6] R. Sepulchre, M. Jankovic, and P. Kokotović, Constructive Nonlinear Con-
trol. Berlin, Germany: Springer-Verlag, 1997.
directions. In particular, dissipativity has been extended to
[7] M. Diehl, R. Amrit, and J. B. Rawlings, “A Lyapunov function for eco-
optimal control problems that do not exhibit an optimal nomic optimizing model predictive control,” IEEE Trans. Autom. Control,
equilibrium but an optimal periodic orbit [63]–[66] or gen- vol. 56, no. 3, pp. 703–707, 2011, doi: 10.1109/TAC.2010.2101291.
[8] D. Angeli, R. Amrit, and J. B. Rawlings, “On average performance and
eral time-varying optimal trajectories [67], [68]. In this con-
stability of economic model predictive control,” IEEE Trans. Autom. Control,
text, the concept of overtaking optimality can also be used vol. 57, no. 7, pp. 1615–1626, 2012, doi: 10.1109/TAC.2011.2179349.
to define a meaning for optimality in the case where the [9] D. A. Carlson, A. B. Haurie, and A. Leizarowitz, Infinite Horizon Opti-
mal Control: Deterministic and Stochastic Systems, 2nd ed. Berlin, Germany:
infinite-horizon optimal value function V3 is not finite. In
Springer-Verlag, 1991.
the field of dissipativity and optimal control, there are still [10] L. Grüne, “Economic receding horizon control without terminal
a variety of open questions, some of which are mentioned constraints,” Automatica, vol. 49, no. 3, pp. 725–734, 2013, doi: 10.1016/j.
automatica.2012.12.003.
in the following. For instance, as noted throughout the arti-
[11] F. P. Ramsey, “A mathematical theory of saving,” Econ. J., vol. 38, no. 152,
cle, several important results are (thus far) available only in pp. 543–559, 1928, doi: 10.2307/2224098.
discrete time. It would be very interesting to see if and how [12] J. von Neumann, “Über ein ökonomisches Gleichungssystem und eine
Verallgemeinerung des Brouwerschen Fixpunktsatzes,” in Ergebnisse eines
they can be transferred to continuous time. Another major
Mathematischen Seminars, K. Menger, Ed., Köln: Kiepenheuer & Witsch,
open question is the relationship between strict dissipativ- 1938, pp. 172–181.
ity and detectability-like notions for infinite-dimensional [13] R. Dorfman, P. A. Samuelson, and R. M. Solow, Linear Programming and
Economic Analysis (A Rand Corporation Research Study Series). New York,
systems. Even for linear-quadratic optimal control prob-
NY, USA: McGraw-Hill, 1958.
lems, this relationship is not yet fully understood. Finally, [14] L. Grüne and M. Stieler, “Asymptotic stability and transient optimality
it is known that the turnpike property also appears in vari- of economic MPC without terminal conditions,” J. Proc. Control, vol. 24, no.
8, pp. 1187–1196, 2014, doi: 10.1016/j.jprocont.2014.05.003.
ous forms in stochastic optimal control problems (see, for
[15] T. Faulwasser and D. Bonvin, “On the design of economic NMPC based
example, the recent numerical study in [69]). Although on approximate turnpike properties,” in Proc. 54th IEEE Conf. Decision Con-
some results on turnpikes in the stochastic case are avail- trol — CDC, 2015, pp. 4964–4970, doi: 10.1109/CDC.2015.7402995.
[16] L. Grüne and A. Panin, “On non-averaged performance of economic
able (compare [70]–[72]), the connection to dissipativity
MPC with terminal conditions,” in Proc. 54th IEEE Conf. Decision Control
concepts is thus far largely unexplored. — CDC, Osaka, Japan, 2015, pp. 4332–4337, doi: 10.1109/CDC.2015.7402895.
[17] L. Grüne, “Approximation properties of receding horizon optimal con-
trol,” Jahresbericht der Deutschen Mathematiker-Vereinigung, vol. 118, no. 1, pp.
AUTHOR INFORMATION
3–37, 2016, doi: 10.1365/s13291-016-0134-5.
Lars Grüne (lars.gruene@uni-bayreuth.de) has been a pro- [18] L. Grüne and J. Pannek, Nonlinear Model Predictive Control. Theory and
fessor of applied mathematics at the University of Bayreuth, Algorithms, 2nd ed. London, U.K.: Springer-Verlag, 2017.
[19] T. Faulwasser, L. Grüne, and M. A. Müller, “Economic nonlinear model
Bayreuth, Germany, since 2002, where he also holds the
predictive control,” Found. Trends Syst. Control, vol. 5, no. 1, pp. 1–98, 2018,
chair of applied mathematics. He received the diploma doi: 10.1561/2600000014.
degree (M.Sc. degree) and Ph.D. degree in mathematics in [20] L. Grüne and R. Guglielmi, “Turnpike properties and strict dissipativ-
ity for discrete time linear quadratic optimal control problems,” SIAM J.
1994 and 1996, respectively, from the University of Augs-
Control Optim., vol. 56, no. 2, pp. 1282–1302, 2018, doi: 10.1137/17M112350X.
burg and the habilitation degree from Goethe University in [21] L. Grüne and R. Guglielmi, “On the relation between turnpike proper-
2001. He held visiting positions at the Universities of Rome, ties and dissipativity for continuous time linear quadratic optimal control
problems,” Math. Control Related Fields, vol. 11, no. 1, pp. 169–188, 2021, doi:
Sapienza (Italy), Padova (Italy), Melbourne (Australia), Paris
10.3934/mcrf.2020032.
IX Dauphine (France), and Newcastle (Australia). He is the [22] M. Höger and L. Grüne, “On the relation between detectability and
is editor-in-chief of Journal Mathematics of Control, Signals strict dissipativity for nonlinear discrete time systems,” IEEE Control Syst.
Lett, vol. 3, no. 2, pp. 458–462, 2019, doi: 10.1109/LCSYS.2019.2899241.
and Systems and associate editor of various other journals,
[23] C. I. Byrnes and W. Lin, “Losslessness, feedback equivalence, and the
including Journal of Optimization Theory and Applications, global stabilization of discrete-time nonlinear systems,” IEEE Trans. Autom.
Mathematical Control and Related Fields, and IEEE Control Control, vol. 39, no. 1, pp. 83–98, 1994, doi: 10.1109/9.273341.
[24] T. Faulwasser and C. M. Kellett, “On continuous-time infinite horizon
Systems Letters. His research interests lie in mathematical
optimal control—dissipativity, stability and transversality,” Automatica,
systems and control theory, with a focus on numerical- and vol. 134, p. 109,907, Dec. 2021, doi: 10.1016/j.automatica.2021.109907.
optimization-based methods for nonlinear systems. [25] L. Grüne, C. M. Kellett, and S. R. Weller, “On the relation between turn-
pike properties for finite and infinite horizon optimal control problems,”
J. Optim. Theory Appl., vol. 173, no. 3, pp. 727–745, 2017, doi: 10.1007/s10957
REFERENCES -017-1103-6.
[1] J. C. Willems, “Least squares stationary optimal control and the algebra- [26] W. A. Brock and L. Mirman, “Optimal economic growth and uncer-
ic Riccati equation,” IEEE Trans. Autom. Control, vol. 16, no. 6, pp. 621–634, tainty: The discounted case,” J. Econ. Theory, vol. 4, no. 3, pp. 479–513, 1972,
1971, doi: 10.1109/TAC.1971.1099831. doi: 10.1016/0022-0531(72)90135-4.
[2] J. C. Willems, “Dissipative dynamical systems part I: General theory,” [27] A. Porretta and E. Zuazua, “Long time versus steady state optimal
Arch. Rational Mech. Anal., vol. 45, pp. 321–351, Jan. 1972. control,” SIAM J. Control Optim., vol. 51, no. 6, pp. 4242–4273, 2013, doi:
[3] J. C. Willems, “Dissipative dynamical systems part II: Linear systems with 10.1137/130907239.
quadratic supply rates,” Arch. Rational Mech. Anal., vol. 45, pp. 352–393, Jan. 1972. [28] M. Gugat, E. Trélat, and E. Zuazua, “Optimal Neumann control for
[4] P. J. Moylan and B. D. O. Anderson, “Nonlinear regulator theory and an the 1D wave equation: Finite horizon, infinite horizon, boundary tracking
inverse optimal control problem,” IEEE Trans. Autom. Control, vol. 18, no. 5, terms and the turnpike property,” Syst. Control Lett., vol. 90, pp. 61–70, Apr.
pp. 460–465, 1973, doi: 10.1109/TAC.1973.1100365. 2016, doi: 10.1016/j.sysconle.2016.02.001.
[5] R. A. Freeman and P. V. Kokotović, Robust Nonlinear Control Design: State- [29] E. Zuazua, “Large time control and turnpike properties for wave equa-
Space and Lyapunov Techniques (Systems & Control: Foundations & Applica- tions,” Ann. Rev. Control, vol. 44, pp. 199–210, Apr. 2017, doi: 10.1016/j.arcon
tions Series). Boston, MA, USA: Birkhäuser, 1996. trol.2017.04.002.

86 IEEE CONTROL SYSTEMS » APRIL 2022


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

[30] A. Porretta, “On the turnpike property for mean field games,” Minimax [52] L. Grüne, “Computing Lyapunov functions using deep neural net-
Theory Appl., vol. 3, no. 2, pp. 285–312, 2018. works,” J. Comput. Dyn., vol. 8, no. 2, p. 737–752, 2021, doi: 10.3934/jcd.2027006.
[31] E. Trélat, C. Zhang, and E. Zuazua, “Steady-state and periodic ex- [53] H. Chen and F. Allgöwer, “A quasi-infinite horizon nonlinear model
ponential turnpike property for optimal control problems in Hilbert predictive control scheme with guaranteed stability,” Automatica, vol. 34,
spaces,” SIAM J. Control Optim., vol. 56, no. 2, pp. 1222–1252, 2018, doi: no. 10, pp. 1205–1217, 1998, doi: 10.1016/S0005-1098(98)00073-9.
10.1137/16M1097638. [54] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Con-
[32] L. Grüne, M. Schaller, and A. Schiela, “Sensitivity analysis of opti- strained model predictive control: Stability and optimality,” Automatica,
mal control for a class of parabolic PDEs motivated by model predictive vol. 36, no. 6, pp. 789–814, 2000, doi: 10.1016/S0005-1098(99)00214-9.
control,” SIAM J. Control Optim., vol. 57, no. 4, pp. 2753–2774, 2019, doi: [55] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive Control:
10.1137/18M1223083. Theory, Computation and Design. Madison, WI, USA: Nob Hill Publishing, 2017.
[33] T. Breiten and L. Pfeiffer, “On the turnpike property and the receding- [56] S. E. Tuna, M. J. Messina, and A. R. Teel, “Shorter horizons for model
horizon method for linear-quadratic optimal control problems,” SIAM J. predictive control,” in Proc. 2006 Amer. Control Conf., Minneapolis, MN,
Control Optim., vol. 58, no. 2, pp. 1077–1102, 2020, doi: 10.1137/18M1225811. USA, pp. 863–868, doi: 10.1109/ACC.2006.1655466.
[34] L. Grüne, M. Schaller, and A. Schiela, “Exponential sensitivity and [57] L. Grüne, “Analysis and design of unconstrained nonlinear MPC
turnpike analysis for linear quadratic optimal control of general evolution schemes for finite and infinite dimensional systems,” SIAM J. Control Op-
equations,” J. Differ. Equ., vol. 268, no. 12, pp. 7311–7341, 2020, doi: 10.1016/j. tim., vol. 48, no. 2, pp. 1206–1228, 2009, doi: 10.1137/070707853.
jde.2019.11.064. [58] L. Grüne, J. Pannek, M. Seehafer, and K. Worthmann, “Analysis of uncon-
[35] T. Faulwasser and L. Grüne, “Turnpike properties in optimal control: strained nonlinear MPC schemes with time varying control horizon,” SIAM
An overview of discrete-time and continuous-time results,” 2021. [Online]. J. Control Optim., vol. 48, no. 8, pp. 4938–4962, 2010, doi: 10.1137/090758696.
Available: https://arxiv.org/pdf/2011.13670.pdf [59] M. Reble and F. Allgöwer, “Unconstrained model predictive control and
[36] T. Faulwasser, M. Korda, C. N. Jones, and D. Bonvin, “On turnpike and suboptimality estimates for nonlinear continuous-time systems,” Automat-
dissipativity properties of continuous-time optimal control problems,” Au- ica, vol. 48, no. 8, pp. 1812–1817, 2011, doi: 10.1016/j.automatica.2012.05.067.
tomatica, vol. 81, pp. 297–304, Jul. 2017, doi: 10.1016/j.automatica.2017.03.012. [60] M. Zanon, S. Gros, and M. Diehl, “Indefinite linear MPC and approxi-
[37] T. Damm, L. Grüne, M. Stieler, and K. Worthmann, “An exponen- mated economic MPC for nonlinear systems,” J. Process Control, vol. 24, no.
tial turnpike theorem for dissipative discrete time optimal control prob- 8, pp. 1273–1281, 2014, doi: 10.1016/j.jprocont.2014.04.023.
lems,” SIAM J. Control Optim., vol. 52, no. 3, pp. 1935–1957, 2014, doi: [61] M. Zanon and T. Faulwasser, “Economic MPC without terminal con-
10.1137/120888934. straints: Gradient-correcting end penalties enforce asymptotic stability,” J.
[38] E. Trélat and E. Zuazua, “The turnpike property in finite-dimensional Process Control, vol. 63, pp. 1–14, Mar. 2018, doi: 10.1016/j.jprocont.2017.12.005.
nonlinear optimal control,” J. Differ. Equ., vol. 258, no. 1, pp. 81–114, 2015, [62] A. Alessandretti, A. P. Aguiar, and C. N. Jones, “On convergence and
doi: 10.1016/j.jde.2014.09.005. performance certification of a continuous-time economic model predictive
[39] T. Faulwasser, L. Grüne, J.-P. Humaloja, and M. Schaller, “The interval control scheme with time-varying performance index,” Automatica, vol. 68,
turnpike property for adjoints,” 2020, arXiv:2005.12120. pp. 305–313, Jun. 2016, doi: 10.1016/j.automatica.2016.01.020.
[40] L. Grüne and M. A. Müller, “On the relation between strict dissipativity [63] M. Zanon, S. Gros, and M. Diehl, “A Lyapunov function for periodic
and the turnpike property,” Syst. Contr. Lett., vol. 90, pp. 45–53, Apr. 2016, economic optimizing model predictive control,” in Proc. 52nd IEEE Conf.
doi: 10.1016/j.sysconle.2016.01.003. Decision Control — CDC, Florence, Italy, 2013, pp. 5107–5112, doi: 10.1109/
[41] M. A. Müller, D. Angeli, and F. Allgöwer, “On necessity and robustness CDC.2013.6760691.
of dissipativity in economic model predictive control,” IEEE Trans. Autom. [64] M. A. Müller and L. Grüne, “Economic model predictive control with-
Control, vol. 60, no. 6, pp. 1671–1676, 2015, doi: 10.1109/TAC.2014.2361193. out terminal constraints for optimal periodic behavior,” Automatica, vol. 70,
[42] J. Berberich, J. Köhler, F. Allgöwer, and M. A. Müller, “Indefinite lin- pp. 128–139, Aug. 2016, doi: 10.1016/j.automatica.2016.03.024.
ear quadratic optimal control: Strict dissipativity and turnpike proper- [65] M. Zanon, L. Grüne, and M. Diehl, “Periodic optimal control, dissipa-
ties,” IEEE Control Syst. Lett., vol. 2, no. 3, pp. 399–404, 2018, doi: 10.1109/ tivity and MPC,” IEEE Trans. Autom. Control, vol. 62, no. 6, pp. 2943–2949,
LCSYS.2018.2842142. 2017, doi: 10.1109/TAC.2016.2601881.
[43] G. Grimm, M. J. Messina, S. E. Tuna, and A. R. Teel, “Model predic- [66] J. Köhler, M. A. Müller, and F. Allgöwer, “On periodic dissipativity no-
tive control: For want of a local control Lyapunov function, all is not lost,” tions in economic model predictive control,” IEEE Control Syst. Lett., vol. 2,
IEEE Trans. Autom. Control, vol. 50, no. 5, pp. 546–558, 2005, doi: 10.1109/ no. 3, pp. 501–506, 2018, doi: 10.1109/LCSYS.2018.2842426.
TAC.2005.847055. [67] L. Grüne, S. Pirkelmann, and M. Stieler, “Strict dissipativity implies
[44] C. Cai and A. R. Teel, “Input-output-to-state stability for discrete- turnpike behavior for time-varying discrete time optimal control problems,”
time systems,” Automatica, vol. 44, no. 2, pp. 326–336, 2008, doi: 10.1016/j. in Control Systems and Mathematical Methods in Economics (Lecture Notes in
automatica.2007.05.022. Economic and Mathematical Systems Series), vol. 687, G. Feichtinger, R. M.
[45] P. Giesl and S. Hafstein, “Review on computational methods for Lyapu- Kovacevic, and G. Tragler, Eds. Cham: Springer-Verlag, 2018, pp. 195–218.
nov functions,” Discrete Contin. Dyn. Syst. Ser. B, vol. 20, no. 8, pp. 2291–2331, [68] L. Grüne and S. Pirkelmann, “Economic model predictive control for
2015, doi: 10.3934/dcdsb.2015.20.2291. time-varying system: Performance and stability results,” Opt. Control Appl.
[46] J. Berberich, J. Köhler, F. Allgöwer, and M. A. Müller, “Dissipativity Methods, vol. 41, no. 1, pp. 42–64, 2019, doi: 10.1002/oca.2492.
properties in constrained optimal control: A computational approach,” [69] R. Ou, M. H. Baumann, L. Grüne, and T. Faulwasser, “A simulation
Automatica, vol. 114, pp. 108–840, Apr. 2020, doi: 10.1016/j.automatica. study on turnpikes in stochastic LQ optimal control,” IFAC-PapersOnLine,
2020.108840. vol. 54, no. 3, pp. 516–521, 2021, doi: 10.1016/j.ifacol.2021.08.294.
[47] S. Pirkelmann, D. Angeli, and L. Grüne, “Approximate computation [70] R. Marimon, “Stochastic turnpike property and stationary equilib-
of storage functions for discrete-time systems using sum-of-squares tech- rium,” J. Econ. Theory, vol. 47, no. 2, pp. 282–306, 1989, doi: 10.1016/0022
niques,” IFAC-PapersOnLine, vol. 52, no. 16, pp. 508–513, 2019, doi: 10.1016/j. -0531(89)90021-5.
ifacol.2019.12.012. [71] S. Sethi, H. M. Soner, Q. Zhang, and J. Jiang, “Turnpike sets and their
[48] C. Scherer and S. Weiland, “Linear matrix inequalities in control,” analysis in stochastic production planning problems,” Math. Oper. Res., vol.
Lecture Notes, Dutch Inst. for Systems and Control, Delft, The Nether- 17, no. 4, pp. 932–950, 1992, doi: 10.1287/moor.17.4.932.
lands, 2000. [72] V. Kolokoltsov and W. Yang, “Turnpike theorems for Markov games,” Dyn.
[49] A. Romer, J. Berberich, J. Köhler, and F. Allgöwer, “One-shot verifica- Games Appl., vol. 2, no. 3, pp. 294–312, 2012, doi: 10.1007/s13235-012-0047-6.
tion of dissipativity properties from input-output data,” IEEE Control Syst. [73] S. Qin and T. Badgwell, “A survey of industrial model predictive con-
Lett., vol. 3, no. 3, pp. 709–714, 2019, doi: 10.1109/LCSYS.2019.2917162. trol technology,” Control Eng. Pract., vol. 11, no. 7, pp. 733–764, 2003, doi:
[50] J. Darbon, G. P. Langlois, and T. Meng, “Overcoming the curse of di- 10.1016/S0967-0661(02)00186-7.
mensionality for some Hamilton-Jacobi partial differential equations via [74] M. G. Forbes, R. S. Patwardhan, H. Hamadah, and R. B. Gopaluni,
neural network architectures,” Res. Math. Sci., vol. 7, no. 3, 2020, Art. no. 20. “Model predictive control in industry: Challenges and opportuni-
[51] M. Hutzenthaler, A. Jentzen, T. Kruse, and T. A. Nguyen, “A proof that ties,” IFAC-PapersOnLine, vol. 48, no. 8, pp. 531–538, 2015, doi: 10.1016/j.
rectified deep neural networks overcome the curse of dimensionality in the ifacol.2015.09.022.
numerical approximation of semilinear heat equations,” SN Partial Differ.
Equ. Appl., vol. 1, no. 2, p. 34, 2020, doi: 10.1007/s42985-019-0006-9. 

APRIL 2022 « IEEE CONTROL SYSTEMS 87


Authorized licensed use limited to: Universite de Rennes 1. Downloaded on January 01,2024 at 12:55:09 UTC from IEEE Xplore. Restrictions apply.

You might also like