Constrained optimal control: an application to

semiactive suspension systems

Tina Paschedag
Institute of Automatic Control Engineering (LSR), Technische Universitaet Muenchen
D-80290 Munich, Germany
Tel: +49-89-289-23404, Fax: +49-89-289-28340, [email protected].
Alessandro Giua

, Carla Seatzu
Dept. of Electrical and Electronic Eng., University of Cagliari
Piazza dArmi 09134 Cagliari, Italy
Tel: +39-070-675-5759, Fax: +39-070-675-5782, {giua,seatzu}@diee.unica.it.
Abstract
This paper applies three dierent control techniques to the design of a quarter car semi-
active suspension system. The three techniques, originally developed to solve a constrained
optimal control problem, are optimal gain switching, discontinuous variable structure control
and explicit model predictive control. All of them divide the state space into convex regions
and assign a linear or ane state feedback controller to each region. The partition of the
state space is computed o-line. During the on-line phase, the controller switches between
the subcontrollers according to the current state. All the above techniques gave satisfactory
results when applied to the design of semiactive suspension systems. A detailed comparison
in terms of computational complexity, performance and simplicity of the design is proposed
in the paper.
Published as:
T. Paschedag, A. Giua, C. Seatzu, Constrained optimal control: an application to semiactive
suspension systems, Int. Journal of Systems Science, Vol. 41, No. 7, pp. 797-811, July 2010.

Contact author to whom all correspondence should be addressed.

1
1 Introduction
The design of active suspensions for road vehicles aims to optimize the performance of the
vehicle with regard to comfort, road holding and rideability [13, 25]. In an active suspension
the interaction between vehicle body, the so-called sprung mass, and wheel (nonsprung mass) is
regulated by an actuator of variable length. The actuator is usually hydraulically controlled and
applies between body and wheel a force that represents the control action generally determined
with an optimization procedure.
In contrast to active suspensions, passive suspensions consist of dampers and springs and the
interaction between body and wheel is determined by their elastic constants and damping coef-
cients, that are constant.
A good tradeo between active and passive suspensions in terms of performance and costs is given
by semiactive suspensions [9, 19, 23]. A semiactive suspension system consists of a spring and a
damper: the spring has a constant stiness, while the damper has a characteristic coecient f
that can be modied within an interval [f
min
, f
max
]. In particular, the active suspension system
may be assumed as a target, and at any time instant, the value of f is chosen so as to make
the evolution of the semiactive suspension system as close as possible to that of the active
suspension.
The main objective of this paper is that of discussing the applicability of some optimal control
methods to the design of a semiactive suspension system. In particular, we consider three
dierent techniques to control linear systems with state or input constraints and, assuming the
semiactive suspension system as a case study, we compare them under several aspects, namely
range of applicability, computational complexity and performance. In particular, the considered
approaches are explicit model predictive control (eMPC) [3, 7], optimal gain switching (OGS)
[27], and discontinuous variable structure control (dVSC) [1, 2, 17].
As well known, the two most commonly used approaches to optimally control linear systems
with state or input constraints are based on anti-windup and model predictive control.
Anti-windup schemes are auxiliary to control schemes that are able to provide ecient
performance around nominal operating conditions. Their goal is that of preventing the
violation of certain constraints (windup phenomenon) using a control scheme that does
not modify the behavior of the original controller if saturation does not occur, while it
becomes active when certain bounds are exceeded. The most popular synthesis tools for
the design of anti-windup compensators are based on linear matrix inequalities (LMI).
However, even if several works have been presented in this framework [4, 18, 24, 28], a
systematic approach for the synthesis of anti-windup schemes that guarantee closed-loop
stability and optimal performance is still missing. In particular, this is the case when
dealing with complex constrained multivariable control problems.
2
The implicit model predictive control approach is based on the solution of an optimal
control problem at each sampling time, starting from the current state. The optimal
control problem is solved over a given time horizon that may either be nite or innite.
At the next time step the optimization is repeated considering an horizon of the same
length of the previous one, but shifted. This is the reason why such an approach is also
known as moving horizon control. The main drawback is that it requires the solution of
an optimization problem at each sampling time, whose complexity may be prohibitive for
large horizons and in the case of high sampling frequency.
An alternative to the classical implicit model predictive control has recently been proposed by
Bemporad et al. [3, 7]: this new approach is called explicit model predictive control (eMPC).
Its main advantage is that the most burdensome part of the procedure is moved o-line. By
appropriately converting the optimal control problem with constraints into a multi-parametric
programming problem, appropriate algorithms in this framework can be used, and the state
space is partitioned into polytopes described by linear inequalities
1
. A piecewise ane control
law is associated to each region, thus the on-line phase of the procedure simply consists in
determining the current region and assigning accordingly the appropriate control law.
There exist two other interesting approaches that share the same philosophy of eMPC, namely
that of solving the optimal control problem for constrained linear systems with a procedure
that requires both an o-line and an on-line phase. During the o-line phase the state space is
partitioned into a nite number of convex regions and to each region a linear subcontroller is
assigned. During the on-line phase the controller switches between these subcontrollers according
to the current system state. These approaches are called optimal gain switching (OGS) [27],
and discontinuous variable structure control (dVSC) [1, 2, 17].
All three approaches, i.e., OGS, dVSC and eMPC, guarantee closed-loop stability. However,
eMPC is the most general in the sense that it can take into account general constraints of the
form
Ex(t) +Lu(t) M (1)
where x(t) R
n
is the state, u(t) R
m
is the control input, and E, L and M are matrices of
suitable dimension. On the contrary, the other two approaches can only deal with symmetric
constraints of the form
|u(t) k
T
x(t)| u
max
(2)
where u(t) R is a scalar input and k R
n
is a constant gain vector. Constraints (2) are
obviously a particular case of (1).
The OGS procedure has been successfully applied in [12] to the design of quarter-car suspension
systems, where the constraint on the input takes the form |u(t)| u
max
, i.e., is a special case
of (2). The procedure is used for the design of active control laws, that can be approximated
1
A bounded polyhedron P R
n
, P = {x R
n
| Ax B} is called a polytope.
3
using a semiactive suspension system. In this paper we investigate if dVSC and eMPC can also
be applied to semiactive suspension design.
1.1 Relevant literature on semiactive suspensions
Several newly designed top cars already use semiactive suspension systems. As a result great
eorts are now devoted to improve their eciency, as demonstrated by the large number of
publications in this area.
A comprehensive survey on this topic is given in [14]. Here the author surveys applications of
optimal control techniques (basically focusing on linear quadratic optimal control) to the design
of active and semiactive suspensions, starting from the quarter-car model, but also considering
the half-car and the full-car models. Another more recent survey is given in [15] by Jalili.
In [6] the authors present a learning algorithm using neural networks which allows a vehicle
with semi-active suspension to improve continuously not only the ride comfort but also the
tyre/ground contact.
In [8] Canale, Milanese and Novara propose a solution that is based on MPC. In particular, they
try to overcome the problems typically related to the online implementation of MPC strategies
using a fast MPC implementation based on nonlinear function approximation techniques. The
proposed solution is also successfully compared with well-established control algorithms such as
two state Sky-Hook [14] and clipped strategy [21].
Giorgetti et al. in [10] modeled the constrained quarter-car semiactive suspension as a switching
ane system, where the switching is determined by the activation of passivity constraints,
force saturation, and maximum power dissipation limits. Explicit optimal control laws have
been derived with dierent nite horizons, and it has been shown that the optimal control is
piecewise ane in state.
Finally, the use of magnethoreological dampers has been considered in [11, 16].
1.2 Structure of the paper
The paper is structured as follows. In Sections 2 and 3 we provide some background on the
OGS and the VSC procedure, respectively. In Section 4 we recall the main features of the eMPC
approach and the main design parameters necessary to implement it using the MATLAB Multi-
Parametric Toolbox [20]. In Section 5 we present the linear dynamical models of the suspension
system considered in the paper. A comparison among the three considered approaches when
applied to the design of suspension systems is given in Section 6. Conclusions are nally drawn
in Section 7.
4
2 Optimal gain switching
Let us consider a linear and time-invariant system
x(t + 1) = Gx(t) +Hu(t), (3)
where t N = {0, 1, 2, . . .}, x R
n
, u R.
We want to determine the control law u

() that minimizes a performance index of the form:

J =

t=0
x
T
(t)Qx(t), (4)
(with Q positive semidenite) under the constraint
|u(t)| u
max
(t 0). (5)
It is well known that the optimal solution u

() does not correspond to a feedback control law

[26].
The OGS approach, rstly proposed by Yoshida in [27], approximates the optimal control law
u

() by switching among a certain number of feedback control laws whose gains can be computed
as the solution of a family of LQR problems. More precisely, to determine the OGS control law
u
OGS
we consider a family of performance indices
J

t=0
[ x
T
(t)Qx(t) +u
T
(t)Ru(t)], > 0, R > 0. (6)
For a given value of , the unconstrained control law that minimizes J

can be written as
u

(t) = k
T

x(t) (7)
where the gain vector k
T

is obtained by solving an algebraic Riccati equation. The resulting

controller then switches among dierent control laws in the form (7) depending on the current
value of the system state.
For a given value of it is possible to compute a linear region

in the state space such that

for any point x
0
within this region the following equation holds:
|u

(t)| |k
T

(GHk
T

)
t
x
0
| u
max
, (t 0). (8)
Thus, considering the system (3) controlled with u

and an initial state x

0

, we can be sure
that in its future evolution the value of the control input will always satisfy the constraint (5).
A nite set of m values of , namely {
1
, . . . ,
m
} should be rst selected. A good choice of the

i
s may inuence the performance of the OGS law. As m increases, the performance index J

decreases, but the procedure becomes computationally more intensive. The weighting coecient
5

1
should be determined such that the linear region

1
contains all the initial conditions of
interest. The weighting coecient
m
should be selected such that the region

m
covers small
disturbances or very small system noises. The coecients
2
, . . . ,
m1
should be chosen taking
into account the size of the linear region
i
. Once
1
,
m
and m are determined, the intermediate
values of can be chosen such that the ratios of the norm between two adjacent gains are
constant, i.e.,
||k

i
||
||k

i1
||
=

||k

m
||
||k

1
||
1
m
. (9)
Then, following a simple procedure given in [27], the regions

s are computed o-line. Such

a procedure, that is not reported here for brevitys requirements, is based on the solution of m
linear programming problems that provide appropriate vectors z

s. At each sampling time t

the on-line phase of the approach simply requires to determine the largest value v such that
v = max{ i | x(t)

i
, i = 0, . . . , m} (10)
and set (t) =
v
. The condition x(t)

is true i
u
max
z
T

x
0
u
max
(11)
where z
T

= k
T

(GHk
T

). Thus the control law at time t is chosen equal to

u
OGS
(t) = k
T

v
x(t). (12)
It has been shown by Yoshida that if no disturbance is acting on the system, (t) is a nonde-
creasing function of t.
3 Discontinuous Variable Structure Control
The basic ideas of discontinuous VSC (dVSC) have been rstly proposed by Kiendl and Schneider
[17]. Most of the literature on this topic is in German, but a good survey in English is available
[2].
The variable structure controller, depending on the systems state, either switches between a
nite number of linear subcontrollers (discontinuous VSC) or changes the controller parameters
continuously (soft VSC) with the objective of obtaining a better performance in terms of shorter
settling times while avoiding violation of control signal constraints.
The dVSC method makes use of a set of nested, positively invariant sets each with a dedicated
linear controller. During the regulation cycle, the trajectory runs from a positively invariant
region in the state space into the next smaller one, simultaneously activating the assigned con-
troller.
6
Here we briey outline the general structure of the dVSC.
Consider the linear time-invariant plant in continuous time
x(t) = Ax(t) +Bu(t) (13)
where t R
0
, x R
n
, u R, under the control signal constraint
|u(t)| u
max
. (14)
The control input is chosen according to
u
dV SC
(t) = F(x(t), p) (15)
where F is an operator
2
that depends on the systems state x and a selection parameter p,
that is computed by a selection strategy or supervisor, i.e., p = S(x), dened by a discontinuous
function S. The selection strategy switches between a nite number m of dierent subcontrollers
so as to optimize the systems performance in terms of settling times.
Note that in the following we consider only bounded sets X
0
R
n
of possible initial vectors
x(t = 0), since X
0
= R
n
is usually not of practical interest. The three major steps of the dVSC
design procedure are:
(D1) Choose a family of m linear state controllers u(t) = k
T
p
x(t) leading to stable control
loops
x(t) = (ABk
T
p
) x(t) =

A
p
x(t), p = 1, . . . , m (16)
whose response times decrease with increasing index p.
(D2) According to each control loop (16) construct a Lyapunov region
G
p
= {x | v
p
(x) < c
p
} (17)
where c
p
determines the size of G
p
. Moreover, G
p
should be such that all x G
p
satisfy
the constraint |u
dV SC
| = |k
T
p
x| u
max
.
(D3) The Lyapunov regions should be nested one inside the other in accordance with
G
p+1
G
p
, p = 1, . . . , m1 (18)
with an increasing index p.
The dVSC method follows an approach similar to the OGS design, and consists of an o-line
and an on-line phase. The three steps mentioned above represent the o-line phase. During the
on-line phase the controller determines the smallest Lyapunov region that contains the current
2
A common practice, as we do in this section, is that of choosing u
dV SC
(t) = k
T
p
x(t).
7
systems state and activates the subcontroller corresponding to this region. As soon as the
trajectory enters a smaller region, the controller switches to the corresponding subcontroller.
In the rst step the subcontrollers vectors k
p
are determined utilizing pole placement such that
the n eigenvalues
p,j
of

A
p
conform to

p+1,j
= h
p,j
, h > 1 (19)
and lead to a stable closed loop, i.e. Re{
p
} < 0. These controllers thus accelerate the con-
trol systems behavior, while simultaneously causing a similar behavior, since the eigenvalue
conguration remains the same.
In a second step the Lyapunov regions are constructed employing quadratic Lyapunov functions
v
p
(x) = x
T
R
p
x, where the matrix R
p
is the solution of the Lyapunov equation

A
T
p
R
p
+R
p

A
p
=
Q
p
.
The matrices Q
p
have to be positive-denite: Q
p+1
= Q
p
is frequently a reasonable choice.
Thus, the Lyapunov regions will be ellipses determined by the matrices R
p
. Since the condition
|k
T
p
x| u
max
has to be satised for all x G
p
, to ensure the regions G
p
are as large as possible
the constant c
p
in (17) are chosen such that the hyperplanes k
T
p
x = u
max
are tangent to the
elliptical Lyapunov regions. A suitable c
p
is
c
p
=
u
2
max
k
T
p
R
1
p
k
p
. (20)
and it can be determined solving a quadratic optimization problem [2].
The largest Lyapunov region G
1
has to be determined such that X
0
G
1
, i.e. the rst region
includes all possible initial states.
Finally, in a third step we verify that all regions G
p
s are nested: if all points of interest satisfy
x
T
R
p
x
c
p
<
x
T
R
p+1
x
c
p+1
< 1 (21)
then G
p+1
G
p
is ensured. To check whether (21) is true or not it is sucient to make sure
that the matrices
R
p+1
c
p+1

R
p
c
p
(22)
are positive denite for p = 1, . . . , m1.
We conclude pointing out that the most burdensome step in this design procedure is the de-
termination of a suitable matrix R
1
(from which all other matrices R
p
for p > 1 can also be
derived). Several techniques to do this, based on nonlinear optimization, have been presented
in the literature [1, 2].
8
4 Explicit model predictive control
Model Predictive Control (MPC) [5], also referred as moving horizon control or receding horizon
control, is an advanced control method that has become an attractive feedback strategy, espe-
cially for linear and time-invariant systems of the form (3) under the constraint (5), that are
those of interest here.
The basic idea of MPC is the following: at every time step, the control action is chosen solving
an optimal control problem, minimizing a performance criterion over a future horizon. Only
the rst control command will be applied and after one time step other measurements will
be got and the optimization problem is repeated. This is an on-line procedure and in many
cases it is dicult (or even impossible) to implement because the on-line solution of a linear or
quadratic program, depending on the performance index, is required. Various MPC algorithms
use dierent cost functions to obtain the control action. In this paper we consider the following
standard form:
J
N
(x(t)) =
N

j=0
x(t +j)
T
Qx(t +j) +u(t +j)
T
Ru(t +j) (23)
where Q and R are positive denite matrices.
The main limitation of the implicit MPC is that the computations are executed on-line, so that
it is only applicable to relatively slow and/or small problems.
The explicit MPC approach is based on multi-parametric programming. It moves all the burden-
some computations o-line and partitions the state space into polytopic regions, so that during
the on-line phase of the control procedure according to the current state the actual subcontroller
can be found out of a table. The on-line phase of the eMPC is similar to that one of the other
approaches presented above (OGS and dVSC).
Algorithms to compute the polytopic regions are given in [3, 7]. Moreover, the eMPC controller
can be computed using the Multi-Parametric Toolbox (MPT) [20], a free and user-friendly
MATLAB toolbox for design, analysis and deployment of optimal controllers for constrained
linear and hybrid systems.
As already pointed out in [3, 7], the main drawback of the eMPC is that it may easily lead
to controllers with prohibitive complexity, both in runtime and solution. In particular, there
are three aspects which are important in this respect: performance, closed-loop stability and
constraint satisfaction. The MPT toolbox provides several possibilities to compute the controller
and the partition of the state space, which are specied below and that we have investigated.
Finite Time Optimal Control (FTOC). This method yields the nite time optimal con-
troller, i.e. the performance will be N-step optimal but may not be innite horizon optimal.
The complexity of the controller depends strongly on the prediction horizon N, the larger
9
N the more complex the controller is. Furthermore, within this method, the MPT toolbox
provides two dierent modes.
probstruct:Tconstraint=0: The controller will be dened over a superset of the
maximum controllable set (i.e. all states, which are controllable to the origin), but
no guarantees on stability or closed-loop constraint satisfaction can be given. As
the prediction horizon N is increased the feasible set of states will converge to the
maximum controllable set from the outside-in, i.e., the controlled set will shrink as
N increases
3
.
probstruct:Tconstraint=1: The resulting controller will guarantee stability and
constraint satisfaction for all time, but will only cover a subset of the maximum
controllable set of states. By increasing the prediction horizon, the controllable set of
states will converge to the maximum controllable set from the inside-out, i.e. the
controlled set will grow larger as N increases.
Innite Time Optimal Control (ITOC). This method yields the innite time optimal con-
troller, i.e. the best possible performance for the control problem. Asymptotic stability
and constraint satisfaction are guaranteed and the maximum controllable set will be cov-
ered by the resulting controller. However, the controllers complexity may be prohibitive
and the computation may take a very long time.
Two other options are possible when designing the eMPC controller using the toolbox MPT,
namely Minimum Time Control and Low Complexity Control. However, we do not discuss these
cases here because we have not been able to apply them to our application: in both cases the
computation did not nish in adequate times.
5 Dynamical model of the suspension system
In this paper we consider two dierent dynamical models of a quarter-car suspension system.
The rst one is a two-degrees of freedom fourth-order model [12]. The second one is a reduced
one-degree of freedom second-order model that neglects the dynamics of the tire.
Since the reduced model does not describe the interaction of the tire with the suspended mass and
the ground, it cannot be used to evaluate features like road holding and rideability. However, as
it will be discussed later, it allows a signicant comparison among the dierent design techniques
we considered.
3
Even though closed loop stability and constraint satisfaction are not guaranteed, MPT provides a function
to extract the set of states which satisfy the constraints for all time and another function to analyze these states
for stability.
10
M
s
u
act

t
(a)
M
s
(c)
M
s

s
f
(d)
two-degrees of freedom scheme
one-degree of freedom scheme
M
w
(b)
M
s
M
w

s
f
x
2
x
1
x
2
x
1
x
1
x
1
u
act
Figure 1: Scheme of the fourth-order suspension model: (a) active; (b) semiactive. Scheme of the
second-order suspension model: (c) active; (d) semiactive.
5.1 Fourth-order model
The fourth-order model is depicted in Fig. 1.a (active suspension) and Fig. 1.b (semiactive
suspension), where we used the following notation:
M
w
is the nonsprung mass consisting of the wheel and its moving parts;
M
s
is the sprung mass, i.e. the part of the whole body mass and the load mass pertaining
to only one wheel;
x
1
(t) is the nonsprung mass displacement at time t with respect to a xed reference;
x
2
(t) is the sprung mass displacement at time t with respect to a xed reference;
x
3
(t) = x
1
(t) is the velocity of the nonsprung mass at time t;
x
4
(t) = x
2
(t) is the velocity of the sprung mass at time t;
11
u
act
(t) is the active control force at time t;

t
is the elastic constant of the tire, whose damping characteristics have been neglected.
This is in line with almost all researchers who have investigated synthesis of active sus-
pensions for motor vehicles as the tire damping is minimal [11, 12, 25];

s
is the elastic constant of the spring of the semiactive suspension;
f(t) is the adjustable damper coecient of the semiactive suspension at time t.
5.1.1 Active suspensions
The state equation of the active suspension system is
x(t) = Ax(t) +Bu(t) (24)
where
A =

0 0 1 0
0 0 0 1

t
M
w
0 0 0
0 0 0 0

, B =

0
0

1
M
w
1
M
s

.
In such an active suspension system it is required that the control force satisfy a constraint of
the form [12]:
|u
act
(t)| u
max
. (25)
This constraint bounds the acceleration of the sprung mass at least in nominal operating
conditions, i.e., when the linear model of the suspension is valid so as to ensure the comfort of
the passengers and to avoid loss of contact between wheel and road. Furthermore, this constraint
limits the maximal force required from the controller, i.e., it leads to the choice of a suitable
actuator.
The control laws all require the knowledge of the systems state x. Since not every component of
x(t) is directly measurable, we reconstruct the state through an appropriate state observer. To
do this, we choose a suitable output y(t) = Cx(t), with C = [1 1 0 0; 0 0 0 1], which corresponds
to measuring the suspension deformation and the sprung mass velocity. The resulting system is
thus observable and controllable
4
.
Moreover, as discussed in the following, both the OGS and the eMPC approaches make use of a
discrete-time state space model. Therefore, we choose a sampling interval T and discretize the
model (24), thus getting the new model
x(t + 1) = Gx(t) +Hu(t) (26)
4
Note that an observer causes performance loss, notably with respect to the actuator saturation. This problem
is not considered here when comparing the three dierent approaches.
12
where
G = e
AT
, H =

T
0
e
A
d

B. (27)
It is well known [22] that a system that is observable and controllable in the absence of sampling
maintains these properties after the introduction of sampling if and only if, for every eigenvalue
of A for the continuous time control system, the relationship Re{
i
} = Re{
j
} implies Im{
i

j
} =
2n
T
, n = 1, 2, . . .. The problem at hand results in the following set of eigenvalues:

0, 0,

t
M
w
,

t
M
w

. Under these conditions it is necessary to choose a sampling period

T, such that: T = n

M
w

t
.
5.1.2 Semiactive suspensions
The eect of the semiactive suspension which is composed of a spring and a damper with an
adjustable damper coecient (see Fig. 1.b) leads to the semiactive control law
u
sem
(t) = [
s

s
f(t) f(t)] x(t).
Note that, as f may vary, u
sem
(t) is both a function of f and of x(t).
In general, f may only take values in a real set [f
min
, f
max
]. We propose to choose at each step t
the value of f(t) to minimize the dierence F[f, x(t)] = (u
act
(t) u
sem
(t))
2
. Note that dierent
approaches can be used to derive the active force u
act
assumed as a target for the semiactive
suspension system. Nevertheless, the following rule to update the value of f is applicable in all
such cases.
Let us rst assume x
3
(t) = x
4
(t), then the value f

(t) such that F[f

(t), x(t)] = 0 is
f

(t) =
u
act
(t) +
s
x(t)
v(t)
(28)
where x(t) = x
2
(t) x
1
(t) is the suspension deformation and v(t) = x
4
(t) x
3
(t) is its rate
of change.
As the admissible values of f lie in the interval [f
min
, f
max
] the adjusted damper coecient
becomes
f(t) = minarg
f[f
min
,f
max
]
F[f, x(t)] =

f
max
if f

(t) > f
max
f

(t) if f

(t) [f
min
, f
max
]
f
min
if f

(t) < f
min
(29)
When x
3
(t) = x
4
(t), regardless to the values of f, the damper does not give any contribution to
u
sem
(t). Thus, in this case we assume f(t) = f
max
, which we choose also as the initial value for
the damper coecient f(0) = f
max
.
13
5.2 Second-order model
The second-order model of the suspension system is shown in Fig. 1.c (active suspension) and
Fig. 1.d (semiactive suspension), where we used the following new notation:
x
1
(t) is the sprung mass displacement at time t with respect to a xed reference;
x
2
(t) = x
1
(t) is the velocity of the sprung mass at time t.
5.2.1 Active suspensions
The continuous-time state space model of the active suspension is in the form (24) with constraint
(25), and
A =

0 1
0 0

, B =

0
1/M
s

.
while the discrete-time model can be obtained using eq. (27).
5.2.2 Semiactive suspensions
The eect of the semiactive suspension is equivalent to that of a control force
u
sem
(t) = [
s
f(t)] x(t).
Thus, minimizing (u
act
(t) u
sem
(t))
2
under the assumption x
2
(t) = 0, results in a damper
coecient
f

(t) =
u
act

s
x
1
(t)
x
2
(t)
. (30)
As the damper coecient has to be chosen out of the set [f
min
, f
max
], f(t) is determined consid-
ering (29).
6 A comparison among the dierent approaches
In this section we compare the three control design methods above applying them to the sus-
pension system illustrated in Section 5.
Following [12], we take: M
w
= 28.58 Kg, M
s
=288.90 Kg,
t
= 155900 N/m,
s
= 14345
N/m. We assume the sampling time equal to T = 0.01 s, to which it corresponds the sampling
frequency
s
= 2/T 6 10
2
rad/s.
The above choice of
s
is essentially due to the following reasons. Firstly, the bandwidth of
the passive suspension system described by (24) is
b
< 2 10
2
rad/s. A sampling frequency of
14

s
6 10
2
rad/s is in good agreement with Shannons theorem [22] that requires
s
> 2
b
.
Moreover, this choice of sampling interval ensures that the system will maintain the properties
of controllability and observability. Finally, to change f the controller must change the opening
of the damper valve. Present technologies impose a limit of about 10
2
Hz on the updating
frequency of the damper coecient.
Moreover, we take u
max
= 3000 N that is slightly less than the total weight resting on one
wheel. Note that this does not prevent loss of contact between wheel and road. Furthermore,
this constraint also limits the acceleration of the sprung mass and this is a necessary condition
for the comfort of passengers.
Finally we choose f(t) [800, 3000] Ns/m.
Other parameters are given in the following.
OGS. When dealing with the fourth-order model we assume Q = [11 1 0 0; 1 1
0 0; 0 0 0 0; 0 0 0 0] and R = 0.8 10
9
, that lead to a good performance in terms of
road holding and passengers comfort. Finally, as in [12] we choose the parameters
i
s
as follows:
1
= 0.01,
2
= 0.1,
3
= 0.5,
4
= 1,
5
= 4,
6
= 20,
7
= 50,
8
= 100,

9
= 1000,
10
= 10
5
. When dealing with the second-order model we assume Q = [1 0; 0 0],
R = 0.8 10
9
, and
1
= 0.5,
2
= 1,
3
= 4,
4
= 20,
5
= 50.
dVSC. In the case of the fourth-order model controller we assume that the set of eigenvalues
of the rst controller (p = 1) are placed at

1,2
= 4.5893 73.7067j
3,4
= 2.9243 2.2560j
while the remaining nine (here the total number of regions is m = 10) are chosen assuming
h = 1.1 in (19).
We assumed
5
a matrix R
1
equal to
R
1
=

352.0693 1.4000 0.0246 1.6864

1.4000 11.9693 0.0216 4.1430
0.0246 0.0216 0.0667 0.0071
1.6864 4.1430 0.0071 1.8548

.
This enables us to determine matrix Q
1
= (A
T
1
R
1
+R
1
A
1
), where A
1
= ABk
T
1
and k
1
is the feedback controller that imposes the desired set of eigenvalues
1,2
and
3,4
. Finally,
assuming Q
p
= Q
1
and solving Q
p
= (A
T
p
R
p
+ R
p
A
p
) for all p = 2, . . . , 10 we are also
able to determine all other matrices R
p
.
5
In both cases (fourth order and second order model) suitable values for matrix R
1
leading to good performance
were provided by an anonymous reviewer.
15
In the case of the second-order model we assume
1,2
= 7.8224 7.8224j, h = 1.5 and
m = 5. Matrix R
1
has been chosen equal to
R
1
=

0.566 0.0617
0.0617 0.0079

.
Matrices Q
p
, p = 1, . . . , 5, and R
p
, p = 2, . . . , 5, have been determined following the same
procedure as in the fourth-order model.
For the eMPC we considered the same weighting matrix on the states (Q) as in the OGS
case. The weight on the input is taken equal to the weight on the input for the OGS case
divided by
max
, i.e., R
eMPC
= R/
max
. This guarantees for both approaches the same
level of optimality.
6.1 Some remarks on eMPC
In this section we highlight some problems we encountered when applying the eMPC to the
fourth-order model. Let us rst observe that in order to reduce the run times we determined
the partition of the state space for the fourth-order suspension model considering: X = {x
R
4
| |x
i
| 1, i = 1, . . . , 4}.
In order to clarify which kind of problems we get into, we reported in Fig. 2 some of the
resulting partitions, where a cut at x
3
= x
4
= 0 is done. The results relative to the FTOC case
with probStruct.Tconstraint=0 are shown in Fig. 2.a and Fig. 2.b: increasing the prediction
horizon N from 10 to 15 the controlled set converges towards the maximum controllable set from
the outside inwards. As expected the partition in Fig.2.a cover a larger set than the partition
in Fig. 2.b.
The partitions for the FTOC employing probStruct.Tconstraint=1 are illustrated in Fig. 2.c
and Fig. 2.d for N = 10 and N = 15, respectively. As mentioned above, by increasing the
prediction horizon N the controllable set should converge to the maximum controllable set from
the inside outwards. Clearly, this is not occurring in this case because parts of the state space
that have been covered by the partition with N = 10 are not covered by the partition obtained
with N = 15. Thus, we conclude that some numerical error should have occurred: it is obviously
not possible that a state is controllable under a given prediction horizon, but does not maintain
this property after increasing the latter.
In the case of the ITOC the unfeasability of the result is even more evident as illustrated in
Fig. 3. Only a subset of the state space that was identied to be controllable (see Fig. 2.c) is
covered by the ITOC partition: hence a suitable control law exists only in this restricted subset.
These results point out that some of numerical routines of the toolbox MPT [20] are not nu-
merically stable and require to be handled with care. This problem, however, is not addressed
in this paper.
16
Figure 2: The resulting partitions in the fourth-order case (cut through x
3
= x
4
= 0). (a)
FTOC, probstruct:Tconstraint=0, N = 10, 557 regions; (b) FTOC, probstruct:Tconstraint=0,
N = 15, 1038 regions; (c) FTOC , probstruct:Tconstraint=1, N = 10, 2195 regions; (c) FTOC ,
probstruct:Tconstraint=1, N = 15, 3852 regions.
Figure 3: The resulting partitions in the fourth-order case (cut through x
3
= x
4
= 0) for the ITOC.
17
6.2 A comparison among partitions
6.2.1 Second-order model
0.2 0.1 0 0.1 0.2 0.3
1
0
1
x
1
x
2
(a)
0.2 0.1 0 0.1 0.2 0.3
1
0
1
x
1
x
2
(c)
(a): OGS regions
(b): dVSC regions
(c): eMPC regions
0.2 0.1 0 0.1 0.2 0.3
2
0
2
x
1
x
2
(b)
Figure 4: Partition of the state space for the second-order model.
Fig. 4 shows the dierent state space partitions in the case of the second-order suspension model:
the Yoshida regions for the OGS are depicted in Fig. 4.a; Fig. 4.b and Fig. 4.c illustrate the
regions resulting from dVSC and eMPC, respectively.
Note that to limit the run times of the computation of the polytopic regions employing eMPC
we considered the following bounded polyhedron for the second-order suspension model: X =
{x R
2
| |x
i
| 1, i = 1, 2}. Furthermore, we considered the FTOC with a prediction horizon
N = 10 and we set probStruct.Tconstraint=1 to obtain a controller that guarantees closed
loop stability and constraint satisfaction for all times.
By looking at Fig. 4 we realize that all procedures provide regions that are large enough to cover
the state space region of interest for the considered application. Moreover, the OGS regions are
nested, thus we can use them to design a controller. Finally, we observe that the eMPC regions
are constrained in the x
2
-direction by the assumptions we made in order to reduce the run time
(i.e., x X), but also in the x
1
-direction they are smaller than the OGS regions.
6.2.2 Fourth-order model
Fig. 5 depicts a cut through x
3
= x
4
= 0 of the partitions obtained with the fourth-order
model resulting from the OGS and the eMPC. Here the dierence on the size of two state space
partitions is even more evident.
18
Figure 5: A cut at x
3
= x
4
= 0 of the regions obtained for the fourth-order model: (a) OGS (b) eMPC.
Fig. 6 shows a cut thought x
3
= x
4
= 0 of the dVSC regions. Comparing this result with those
in Fig. 5 it is easy to conclude that the size of the state space partitions obtained using dVSC is
intermediate between that of the regions obtained using OGS and that obtained using eMPC.
In any case, even the eMPC cover the portion of the state space that is of interest in practical
cases.
Let us also observe that in the eMPC case the size of the covered state space is related to
dierent issues, namely, the constraints on the states (x X) we introduce to implement the
procedure, the options we choose (see the discussion above relative to the setting of parameters
in the MPT toolbox), and, in the case of FTOC, the prediction horizon N.
We nally remark that the number of regions obtained using eMPC is signicantly higher than
the number of regions obtained with the other controllers and their computation is much more
burdensome.
19
0.1 0.05 0 0.05 0.1 0.15
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
x
1
x
2
Figure 6: A cut at x
3
= x
4
= 0 of the regions obtained for the fourth-order model using dVSC.
6.3 The control performance
6.3.1 Second-order model
In the case of active suspensions we only present the results of numerical simulations carried out
on the second-order model, because in such a case all the considered techniques provide state
space partitions that are large enough to deal with realistic cases.
We computed the systems evolution for the initial state x
0
= [0.01 0.1]
T
. The simulation results
are summarized in Fig. 7. We can observe that the OGS and eMPC controllers determine
practically the same system evolution. The dVSC controller provides comparable results in
terms of sprung mass position, but with a lower sprung mass velocity.
Finally, in the right bottom graph of Fig. 7 we have pointed out the variation of the index
denoting the current region of the state space in the OGS case and in the dVSC case. Here 1
denotes the largest region and 5 the smallest one.
6.3.2 Fourth-order model
In this section we compare the simulation results for the fourth-order suspension model.
Assume that the initial state is x
0
= [0.015 0.1 0 0]
T
.
Fig. 8 shows the evolution of the semiactive suspension system compared to that of the active
suspension in the OGS and in the eMPC case. Note that in the two top gures the two contin-
20
0 0.05 0.1 0.15 0.2 0.25 0.3
0.002
0
0.002
0.004
0.006
0.008
0.01
0.012
Time [s]
x
1
[m]
0 0.05 0.1 0.15 0.2 0.25 0.3
0.2
0.1
0
0.1
Time [s]
x
2
[ms
1
]
0 0.05 0.1 0.15 0.2 0.25 0.3
3000
2000
1000
0
1000
2000
Time [s]
u [N]
0 0.05 0.1 0.15 0.2 0.25 0.3
3
3.5
4
4.5
5
Time [s]
region
OGS
eMPC
dVSC
Figure 7: Active suspension and initial state x
0
= [0.01 0.1]
T
. Top-left: sprung mass position x
1
. Top-
right: sprung mass velocity x
2
. Bottom-left: control force u. Bottom-right: index denoting the current
region.
uous lines (OGS active and eMPC active) practically coincide, and the two dotted lines (OGS
semiactive and eMPC semiactive) also practically coincide.
Again we can conclude that the OGS and the eMPC performances are very similar, and the
states at each time instant only dier by an order of magnitude of |x
i,OGS
x
i,eMPC
| 10
10
.
In the bottom left of Fig. 8 we have reported the evolution of the target control laws computed
with the OGS and the eMPC, and the control laws that are really applied to the system by the
semiactive suspension when appropriately adjusting the damping coecient f (whose variation
is shown in the bottom right of Fig. 8). Note that the values of f are undistinguishable in the
two cases.
Even better performances are obtained with the dVSC controller. The numerical simulations
are reported in Fig. 9. Note that to better compare the performances of the dVSC controller
with the previous ones, in the same gure we have reported again the results obtained using
eMPC.
This gure enables us to conclude that in the fourth order case the dVSC approach provides
the best performances in terms of sprung mass velocity, while comparable results are obtained
in terms of sprung mass position.
21
0 0.2 0.4 0.6 0.8 1
0.02
0
0.02
0.04
0.06
0.08
0.1
Time [s]
x
2
[m]
0 0.2 0.4 0.6 0.8 1
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1
Time [s]
x
4
[ms
1
]
0 0.2 0.4 0.6 0.8 1
4000
3000
2000
1000
0
1000
2000
3000
Time [s]
u [N]
0 0.2 0.4 0.6 0.8 1
500
1000
1500
2000
2500
3000
Time [s]
f [Nsm
1
]
OGS applied u
eMPC target u
OGS target u
eMPC applied u
OGS semiactive
OGS active
eMPC semiactive
eMPC active
Figure 8: Semiactive suspension with x
0
= [0.015 0.1 0 0]
T
using eMPC and OGS. Top-left: sprung
mass position x
2
. Top-right: sprung mass velocity x
4
. Bottom-left: control force u. Bottom-right: value
of f in the semiactive suspensions.
7 Conclusions
In this paper we considered three dierent design techniques, namely optimal gain switching,
discontinuous VSC and explicit MPC. All these approaches are based on the computation of an
o-line partition of the state space, and guarantee the closed-loop stability and the satisfaction
of bounds on the input magnitude. To each convex region a linear or an ane control law is as-
sociated, and the on-line phase of the approaches simply consists in selecting the current region.
A detailed comparison among these techniques, when applied to the design of semiactive suspen-
sion systems, is provided, both in terms of magnitude of the resulting state space partitions, and
in terms of the system behaviour. As a result, in this application the three considered approaches
provided similar results in terms of performances. However, the computational complexity of
the eMPC is surely higher with respect to OGS and dVSC.
22
0 0.2 0.4 0.6 0.8 1
0
0.05
0.1
Time [s]
x
2
[m]
0 0.2 0.4 0.6 0.8 1
0.8
0.6
0.4
0.2
0
0.2
Time [s]
x
4
[ms
1
]
0 0.2 0.4 0.6 0.8 1
2000
1000
0
1000
Time [s]
u [N]
0 0.2 0.4 0.6 0.8 1
500
1000
1500
2000
2500
3000
f [N sm
1
]
Time [s]
dVSC applied u
dVSC target u
eMPC semiactive
dVSC semiactive
dVSC active
Figure 9: Semiactive suspension with x
0
= [0.015 0.1 0 0]
T
using eMPC and dVSC. Top-left: sprung
mass position x
2
. Top-right: sprung mass velocity x
4
. Bottom-left: control force u. Bottom-right: value
of f in the semiactive suspensions.
As a limitation of the dVSC we point out the diculty in determining appropriate values of the
design parameters.
Acknowledgments
We would like to thank Michal Kvasnica for his prompt and very helpful support concerning the
use of the MPT toolbox.
We also would like to thank an anonymous referee for his/her help in determining appropriate
values of the design parameters for the dVSC controller.
References
[1] J. Adamy. Strukturvariable Regelungen mittels impliziter Ljapunov-Funktionen. Ph.D dissertation,
University of Dortmund, Germany, 1991.
[2] J. Adamy and A. Flemming. Soft variable-structure controls: a survey. Automatica, 40:18211844,
2004.
23
[3] A. Bemporad, M. Morari, V. Dua, and E.N. Pistikopoulos. The explicit linear quadratic regulator
for constrained systems. Automatica, 38:320, 2002.
[4] A. Bemporad, A.R. Teel, and L. Zaccarian. Anti-windup synthesis via sampled-data piecewise ane
optimal control. Automatica, 40(4):549562, 2004.
[5] A. Bemporad, F.D. Torrisi, and M. Morari. Performance analysis of piecewise linear systems and
model predictive control systems. Sydney, Australia, December 2000.
[6] M.J.L. Boada, B.L. Boada, B. Munoz, and V. Diaz. Neural control for a semi-active suspension of
a half-vehicle model. Int. J. of Vehicle Autonomous Systems, 3(2/3/4), 2005.
[7] F. Borrelli. Discrete time constrained optimal control. Ph.D dissertation, ETH Zurich, 2002.
[8] M. Canale, M. Milanese, and C. Novara. Semi-active suspension control using fast model-predictive
techniques. IEEE Trans. on Control Systems Technology, 14(6):10341046, 2006.
[9] K.C. Cheok, N.K. Loh, H.D. McGree, and T.F. Petit. Optimal model following suspension with
microcomputerized damping. IEEE Trans. on Industrial Electronics, 32(4), 1985.
[10] N. Giorgetti, A. Bemporad, H.E. Tseng, and D. Hrovat. Hybrid model predictive control application
towards optimal semi-active suspension. Int. J. of Control, 79(5):521533, 2006.
[11] A. Giua, M. Melas, C. Seatzu, and G. Usai. Design of a predictive semiactive suspension system.
Vehicle System Dynamics, 41(4):277300, 2004.
[12] A. Giua, C. Seatzu, and G. Usai. Semiactive suspension design with an optimal gain switching
target. Vehicle Systems Dynamics, 31:213232, 1999.
[13] A. Hac. Suspension optimisation of a 2-dof vehicle model using a stochastic optimal control technique.
Journal of Sound and Vibration, 100(3):343357, 1985.
[14] D. Hrovat. Survey of advanced suspension developments and related optimal control applications.
Automatica, 33(10):17811817, 1997.
[15] N. Jalili. A comparative study and analysis of semi-active vibration-control systems. J. of Vibration
and Acoustics, 124(4):593650, 2002.
[16] M.A. Karkoub and M. Zribi. Active/semi-active suspension control using magnetorheological actu-
ators. Int. J. of Systems Science, 37(1):3544, 2006.
[17] H. Kiendl and G. Schneider. Synthese nichtlinearer Regler f ur die Regelstrecke const/s
2
aufgrund
ineinander geschalchtelter abgeschlossener Gebiete beschrankter Stellgrosse. Regelungstechnik und
Prozess-Datenverarbeitung, 20(7), 1972.
[18] M.V. Kothare and M. Morari. Multiplier theory for stability analysis of anti-windup control systems.
Automatica, 35:917918, 1999.
[19] E.J. Krasnicki. Comparison of analytical and experimental results for a semi-active vibration isolator.
The Shock and Vibration Bulletin, pages 6976, 1980.
[20] M. Kvasnica, P. Grieder, and M. Baotic. Multi-parametric toolbox (mpt).
http://control.ee.ethz.ch/ mpt/, 2004.
[21] J. Lu and M. DePoyster. Multiobjective optimal suspension control to achieve integrated ride and
handling performance. IEEE Trans. on Control Systems Technology, 10(6):807821, 2006.
24
[22] K. Ogata. Discrete-time control systems. Prentice Hall International Editions, 1972.
[23] C. Poussot-Vassal, O. Sename, L. Dugard, R. Ramirez-Mendoza, and L. Flores. Optimal skyhook
control for semiactive suspensions. In 4th IFAC Sym. on Mechatronics Systems, Heidelberg, Ger-
many, 2006.
[24] A.R. Teel, L. Zaccarian, and J.J. Marcinkowski. An anti-windup strategy for active vibration isola-
tion systems. Control Engineering Practice, 14(1):1727, 2006.
[25] A.G. Thompson. An active suspension with optimal linear state feedback. Vehicle System Dynamics,
5:187203, 1976.
[26] W.M. Wonham and C.D. Johnson. Optimal bang-bang control with quadratic performance index.
ASME Journal of Basic Engineering, 86:107115, 1964.
[27] K. Yoshida, Y. Nishimura, and Y. Yonezawa. Variable gain feedback control for linear sampled-data
systems with bounded control. Control Theory and Advanced Technology, 2 (2)(2), 1986.
[28] L. Zaccarian and A.R. Teel. Nonlinear scheduled anti-windup design for linear systems. IEEE Trans.
on Automatic Control, 49(11):20552061, 2004.
25