Proceedings of the European Control Conference 2009 • Budapest, Hungary, August 23–26, 2009 TuA15.

Model Predictive Controller Matching: Can MPC Enjoy Small Signal

Properties of My Favorite Linear Controller ?
S. Di Cairano, A. Bemporad

Abstract— Model predictive control (MPC) strategies can contrary to the linear controller, the resulting MPC is able to
efficiently deal with constraints on system states, inputs, and properly handle the constraints during transient operations,
outputs. However, in contrast with linear control techniques, and that stability of the constrained system can be enforced.
closed-loop frequency-domain properties of MPC such as sen-
sitivity and robustness to small perturbations are difficult We also provide some answers to the question posed in [5]
to enforce a priori. This paper considers the problem of as regards the inverse optimality problem in the MPC frame-
transforming a given linear feedback control design, referred to work.
as “favorite controller”, into a model predictive control one. In The paper is structured as follows. In Section II we
this way, the MPC controller inherits all the stability, robustness formulate the MPC control matching problem and briefly
and frequency properties of the given favorite controller in the
region around the equilibrium where the constraints are not discuss direct methods for solving it in Section III. In
active. The added value is that the constructed MPC controller Sections IV, V we introduce two inverse methods, where the
is able to properly handle constraints that may be activated MPC weights are designed so that the resulting MPC law
during the transient, and that global stability in the set of equals the favorite control law in the absence of constraints.
feasible initial conditions can be guaranteed. Global stability of the inverse methods is guaranteed. The
I. I NTRODUCTION proposed techniques are exemplified in Section VI, and the
results summarized in Section VII.
Modern control systems are required to satisfy a large
number of specifications: asymptotic closed-loop stability, A. Notation
robustness with respect to external disturbances and mod-
Q > 0 (Q ≥ 0) indicates that a symmetric matrix Q is
eling errors, closed-loop performance, and the capability of
positive definite (positive semidefinite). Relational operators
handling constraints on input and state variables. Industrial
between vectors are intended componentwise. R and R0+ are
control systems are mainly designed to provide stability,
the set of real and nonnegative real numbers, respectively,
and a certain degree of robustness and performance, but
Z, Z+ and Z0+ the set of integers, positive integers, and
in general they do not easily account for constraints. Some
nonnegative integers, respectively. Z[a,b] denotes the set {r ∈
modifications have been introduced in standard control sys-
Z : a ≤ r ≤ b}. For a given vector v, we indicate by [v]i
tems to properly handle constraints, such as anti-windup
its ith component; similarly for a matrix A, [A]j is its j th
schemes for input saturation [1], but these often work for
column, [A]i is its ith row, and [A]ij is the element at the
a restricted class of constraints, are complicated to design,
ith row, j th column. In denotes the identity matrix of order
and may yield to reduced closed-loop performance.
n, On,m ∈ Rn×m denotes a matrix entirely composed of
Model predictive control (MPC) [2] is a control strategy
zeros, where subscripts will be dropped when clear from the
that naturally deals with constraints on system states, inputs,
context. We denote the interior of a set X by int(X ), and the
and outputs by solving at every control cycle an optimization
origin of a vector space by 0. Given dynamics x(k + 1) =
problem. Stability of MPC was surveyed in [3] where it is
φ(x(k)), a set X is positively invariant (PI) for φ(·) if for
shown that MPC stability, robustness, and frequency-domain
all x ∈ X , φ(x) ∈ X .
behavior are sensibly more difficult to characterize with
respect to linear feedback controllers. This has generated II. T HE C ONTROLLER M ATCHING P ROBLEM
reluctance in engineers as regards the use of MPC [4] in
industry. Model predictive control is based on solving at every
In this paper we propose a set of techniques for solving control cycle k ∈ Z0+ the finite horizon optimal control
the following inverse problem: how to select the performance problem
index so that the resulting MPC controller behaves as a given N −1

favorite linear controller in a region around the equilibrium V(x(k)) = min x (i|k)Qx(i|k) + u (i|k)Ru(i|k)
where the constraints are not active. The advantage is that, U (k)
i=0
+x (N |k)P x(N |k) (1a)
A. Bemporad is with the Information Engineering Department, University
of Siena, Italy, [email protected]. S. Di Cairano was with the s.t. x(i + 1|k) = Ax(i|k) + Bu(i|k), i∈Z[0,N −1] (1b)
Information Engineering Department, University of Siena, Italy, at the time
of this work. He has recently joined Powertain Control R&A, Ford Motor xmin ≤ x(i|k) ≤ xmax , i∈Z[0,N ] (1c)
Company. This work was partially supported by the European Commission umin ≤ u(i|k) ≤ umax , i∈Z[0,N −1] (1d)
under the HYCON Network of Excellence, contract number FP6-IST-
511368, and under project WIDE, contract number FP7-IST-224168. x(0|k) = x(k), (1e)

ISBN 978-963-311-369-1
© Copyright EUCA 2009 2217
S. Di Cairano and A. Bemporad : Model Predictive Controller Matching: Can MPC Enjoy Small Signal Properties ... TuA15.1

where x ∈ Rn , u ∈ Rm , N is the prediction horizon, factor. Thus, since R > 0, it is not restrictive to require
U (k) = [u (0|k) . . . u (N − 1|k)] ∈ RN m is the vector to be R ≥ σI, where σ > 0 is a (small) positive constant. Such a
optimized1 , and V : Rn → R0+ is the value function. The constraint avoids numerically ill-conditioned solutions.
performance criterion to be optimized is defined by (1a), In industrial practice, MPC strategies are often used more
where Q, P ∈ Rn×n , R ∈ Rm×m , Q = Q ≥ 0, P = P  ≥ for their capability to handle constraints than for performance
0, R = R > 0. From now on, we will always assume that optimization, meaning that there are several unexploited
Q, R, P are symmetric matrices, even if not stated explicitly. degrees of freedom in choosing the weight matrices P , Q, R.
Equation (1b) defines the prediction model, a linear model On the other hand, these affect the robustness properties and
with state x and input u which is assumed to be completely the frequency-domain response for small signals of MPC [3],
reachable, and (1c), (1d) define the constraints. [4], which are very difficult to shape by design. This often
Given the current state x(k), the finite horizon optimal results in the reluctance of engineers to use MPC as a
control problem (1) can be reformulated as a quadratic substitute for simpler control schemes, such as PID or linear
program2 (QP) with respect to U (k), state feedback, that do not allow for constraint handling but
are well suited for enforcing robust stability and frequency
min U  (k)HU (k) + 2x (k)F U (k) (2a)
U (k) response specifications.
s.t. G U (k) ≤ W + M x(k). (2b) In this paper we want to obtain by design an MPC
controller that behaves as a pre-assigned “favorite” controller
In (2), G ∈ Rq×N m , M ∈ Rq×n and W ∈ Rq define the when the constraints are not active:
problem constraints, while the cost function is defined by Problem 1 (MPC matching): Given a favorite controller
H = (R + S  QS), F = T  QS, (3) ufv = Kx (6)
where S is the N -steps state reachability matrix, T is the
where K ∈ Rm×n , define the cost function (1a) such that
N -steps free state evolution matrix
⎡ B ⎤ ⎡ A ⎤ when the constraints are not active the MPC control (5)
0 ... 0
AB B ... 0 A2 based on (1) is equal to the favorite control (6). The problem
S=⎣ .. .. . . .. ⎦ , T = ⎣ .. ⎦ , is approximately solved if, when the constraints are inactive,
. . . . .
AN −1 B AN −2 B ... B AN
(5) is as close as possible to (6) in a given criterion.
Of course the MPC behavior in general will be different from
and Q ∈ RN n×N n and R ∈ RN m×N m are block-diagonal
the linear one during transients to properly deal with active
matrices
⎡ Q 0 0 ... 0 ⎤ constraints and possibly achieve global stability.
0 Q 0 ... 0
⎡ ⎤ Before discussing techniques that synthesize MPC feed-
R 0 ... 0
⎢. . . . .⎥ 0 R ... 0
Q=⎢
⎣
.. .. . . .. .. ⎥ , R = ⎣ .
⎦ .. .. . . .. ⎦ .
back laws that solve Problem 1, we briefly discuss techniques
0 0 ... Q 0 . . . based on tracking adequately generated references, called
0 0 ... 0 P 0 ... 0 R
direct techniques.
We denote by U ∗ (k) the optimizer of problem (2). III. D IRECT M ATCHING T ECHNIQUES
Definition 1: Given U ∗ (k), the ith constraint, i ∈ Z[1,q] ,
is active at optimality if [G U (k)∗ ]i = [W + M x(k)]i . A direct approach to solve Problem 1 is to use the behavior
If no constraint is active, the optimizer of problem (2) is of (1b) in closed loop with (6) to generate a reference
the unconstrained solution signal for the state and input vectors, and to modify (1)
⎡ ∗ ⎤ to be in the form of state and input tracking [2]. This
u (0|k)
amounts to consider the state reference dynamics rx (i +
U (k) = ⎣
∗ .. ⎦ = −H −1 F  x(k). (4)
. 1|k) = (A + BK)rx (i|k), rx (0|k) = x(0|k), the reference
u∗ (N −1|k)
input ru (i|k) = Krx (i|k), and to modify (1a) to weight
In this case, the MPC command at step k ∈ Z0+ is x(i|k)−rx (i|k) and u(i|k)−ru (i|k). However, this approach
has some drawbacks. It results in a time-varying reference
uMPC (x(k)) = u∗ (0|k) = −ΨH −1 F  x(k), (5) tracking formulation, while most of the stabilization results
where matrix Ψ = [ Im O ··· O ] extracts the first move apply to regulation. As the reference generation model is
actually applied to the process from the optimal sequence linear, such formulation can be transformed into a state-
U ∗ (k). For given system dynamics and prediction horizon regulation problem by extending the state vector, but in this
(i.e., for fixed S and T ), the matrix −H −1 F  obtained by case the overall extended system would not be controllable,
P̄ , Q̄, and R̄ is the same as the one obtained by σ P̄ , σ Q̄, although stabilizable. This, combined with the initialization,
and σ R̄, where σ ∈ R, σ > 0, is an arbitrary positive scaling rx (0|k) = x(0|k) makes the use of standard stabilization
approaches [3] more difficult.
1 The results of this paper can be extended to the case of a control horizon
Remark 1: Most of the approaches for stabilization by
Nu < N , u(i|k) = u(Nu − 1|k), i = Nu , . . . , N − 1. MPC [3] are based on the existence of a feasible solution
2 The reformulation of (1a) results also in a constant term Y = x(k) (Q+
T  QT )x(k) which is not included in (2a), since it does not affect the for (1) at time k obtained by extending the solution computed
optimizer. at time k − 1. However, whenever uMPC (x(k)) = ufv (x(k)),

2218
Proceedings of the European Control Conference 2009 • Budapest, Hungary, August 23–26, 2009 TuA15.1

rx (0|k) = x(0|k) = rx (1|k − 1), hence the input se- to add to the MPC optimal control problem (1) the linear
quence obtained at time k − 1 cannot be extended, due to constraint3
a different initial state. A solution in which the reference
x(k) Acl P B(u(0|k) − Kx(k)) ≥ 0. (10)
closed-loop dynamics evolves autonomously, i.e., rx (k) =
rx (1|k − 1), cannot guarantee that whenever (6) is feasible It is immediate to prove that a lower bound to the optimum
uMPC (x(k)) = ufv (x(k)), due to cumulated errors. is J(U ∗ (k), x(k)) = 0, achieved when d(k) = 0, i.e., when
An approach for which stability can be analyzed is based u∗ (0|k) = Kx(k) is a feasible input. Thus, when all the
on input tracking only. In particular, one can use original constraints of (1) are not active (while constraint (10)
is in general active), the MPC behaves as (6).
J(x(k), U (k)) = (u(0|k) − Kx(k)) QK (u(0|k) − Kx(k)), The MPC-based controller obtained by (9) does not
(7) provide global stability guarantee. However, Proposition 1
as the cost function in (1a), where QK > 0 has to be ensures that even when constraints are active, the MPC
determined. Cost function (7) provides u(0|k) = Kx(k) controller tries to optimize the decrease of the Lyapunov
for any QK > 0 whenever the constraints (1c)–(1d) are not function, and thus it tries at its best to stabilize the system.
active. The minimum value of the cost function is achieved The direct approaches for controller matching are based on
when the MPC algorithm produces the same input as the the reformulation of MPC problem (1) as a tracking problem.
favorite controller. Thus, whenever the favorite controller is Even though they are simple, global stability is hard to
feasible, the MPC controller behaves as such. However, by guarantee and they reduce optimization capabilities of MPC.
this approach we almost completely cancel the performance Also, direct approaches only force MPC to track a reference
optimization capabilities of MPC, since only the first input signal, but they do not give provide any relation between
is accounted for in the cost, and the objective is only related MPC feedback law and (6). To address these limitations we
to the favorite controller local approximation, even though investigate the capability of MPC to be locally functionally
constraints are still enforced along the whole prediction equivalent to the favorite controller. We aim at solving
horizon. Furthermore, cost function (7) is not strongly related Problem 1 by selecting appropriate weight matrices in (1a).
with closed-loop stability, but it shall be adjusted for this
purpose. Let P, Q ∈ Rn×n be matrices such that IV. M ATCHING BASED ON QP M ATRICES
Problem 1 is immediately solved if one can find weight
(A + BK) P (A + BK) − P ≤ −(1 + α)Q, P > 0, Q > 0 matrices Q, R, P in (1) such that
(8)
where α > 0 is a fixed (small) scalar, so that x(k) P x(k) is −ΨH −1 F  x(k) = Kx(k). (11)
a Lyapunov function for the system in closed-loop with the Unfortunately (11) is not trivial to solve, due to the non-
favorite controller. Modify cost function (7) into invertibility of matrix Ψ and the way the inverse of H
J(x(k), U (k)) = 2x(k) (A + BK)P B(u(0|k) − Kx(k))+ depends on Q, R, P .
To solve Problem 1, we start removing Ψ in (11) by setting
(u(0|k) − Kx(k)) B  P B(u(0|k) − Kx(k)) ⎡ κ0 ⎤
(9) κ1
H −1 F  x(k) = − ⎣ .. ⎦ x(k), (12)
Proposition 1: Let U ∗ (k) be the solution at time k ∈ Z0+ .
κN −1
of (1), where (1a) is replaced by (9). If there exists k̄ ∈ Z0+ m×n
such that for all k ≥ k̄ J(U ∗ (k), x(k)) ≤ x(k) Qx(k), the where κ0 = K, while κi ∈ R , i ∈ Z[1,N −1] are free
system is asymptotically stable. matrices. In (12) we account for the whole optimal input
Proof: Let d(k) = u∗ (0|k) − Kx(k) and Acl = sequence of problem (1), but we enforce the equality with the
A + BK. The evolution of the system in closed-loop with favorite control action only for the first control action. As the
the MPC controller is x(k + 1) = Acl x(k) + Bd(k). Since remaining optimal control functions κi , i = 1, . . . , N − 1 are
x (k + 1) P x(k + 1) − x(k) P x(k) = −x(k) P x(k) + left as free variables, they do not introduce further constraints
x(k) Acl P Acl x(k) + d(k) B  P Bd(k) + 2x(k) Acl P Bd(k), on the original problem.
and by (8) x (k + 1) P x(k + 1) − x(k) P x(k) ≤ −(1 + Proposition 2: Let (K̃, Q̃, R̃, P̃ ) be any feasible solution
α)x(k) Qx(k)+d(k) B  P Bd(k)+2x(k) Acl P Bd(k), hence of the following problem
for all k ≥ k̄, x (k + 1) P x(k + 1) ≤ −x(k) P x(k) − (1 + min J(K, Q, R, P ) (13a)
K,Q,R,P
α)x(k) Qx(k) + J(U ∗ (k), x(k)) ≤ −αx(k) Qx(k). Thus,
P is a Lyapunov function for all k ≥ k̄, which proves s.t. Q ≥ 0, P ≥ 0, R ≥ σI (13b)
asymptotic Lyapunov stability of the closed-loop system. (R + S  QS)K + S  QT = 0 (13c)
The condition in Proposition 1 is sufficient, but is not κ0 = K (13d)
necessary. The global minimum of the cost function (9) can
be smaller than zero, and it can be achieved for u∗ (0|k) = where K = [κ0 · · · κN −1 ] , and J(·) is any convex objective
Kx(k), which means that when the constraints are not active function. Then the MPC strategy based on the optimal
the MPC controller does not behave as the favorite controller. 3 At any control cycle in the MPC algorithm x(k) is fixed. Hence
In order to force the behavior required by Problem 1 we need constraint (10) is linear in the optimization variables.

2219
S. Di Cairano and A. Bemporad : Model Predictive Controller Matching: Can MPC Enjoy Small Signal Properties ... TuA15.1

control problem (1) where we set Q = Q̃, P = P̃ , R = R̃, where Qi = Qi ≥ 0, for all i ∈ Z[1,N −1] , Ri = Ri ≥ σI,
solves Problem 1. for all i ∈ Z[0,N −1] . When K is fixed by (16), problem (13)
Proof: Equality (13c) represents (12) after multiply- straightforwardly becomes a problem with LMI constraints
ing both sides by H, which is invertible. Constraint (13d) in the variables Qi , i ∈ Z[1,N −1] , Ri , i ∈ Z[0,N −1] , and
enforces the equality between the MPC command (5) and P . Also, a problem similar to (15) can be formulated that
the favorite controller (6). Finally, (13b) ensures that the returns an MPC controller that approximates the favorite
obtained matrices define a valid cost function for the MPC controller (6) when its precise behavior cannot be obtained.
problem (1). Any solution satisfying (13b), (13c), (13d) Consider the problem
results in (11) when the constraints in (1) are not active.
min (R + S  QS)K̄ + S  QT  (17a)
Q,R
Problem (13) is nonconvex due to bilinear constraint (13c). s.t. P ≥ 0, Ri ≥ σI, i ∈ Z[0,N −1] (17b)
Different cost functions (13a) can be used to define an Qi ≥ 0, i ∈ Z[1,N −1] (17c)
optimal triplet of matrices (Q, R, P ) among those which
solves Problem 1. As an example, for a given reference triplet which is a convex problem subject to LMI constraints, where
(Q̄, R̄, P̄ ) set we have set ⎡ ⎤ K
K(A+BK)
J(K, Q, R, P ) = Q− Q̄+wR R− R̄+wP P − P̄  (14) K̄ = ⎣ .. ⎦. (18)
.
where wR , wP ∈ R0+ , and  ·  is any matrix norm. K(A+BK) N −1

In case it is impossible to find matrices Q, R and P that If the optimum is 0 for the optimizer Q∗i , i ∈ Z[1,N −1] , Ri∗ ,
satisfy (13b), (13c), (13d), Problem 1 does not have an exact i ∈ Z[0,N −1] , P ∗ , the MPC controller with time-varying cost
solution. Hence, we must accept a different behavior u = function (1a) defined by Qi = Q∗i , i ∈ Z[1,N −1] , Ri = Ri∗ ,
K̃x, where K̃ = K. Problem (13) can be modified to return i ∈ Z[0,N −1] , P = P ∗ and for any4 Q0 ≥ 0 behaves as the
matrices Q∗ , R∗ , P ∗ so that K̃ approximates K. Consider favorite controller (6) when the constraints are not active. If
the optimum is not 0, when constraints are not active the
min K + V S  QT  (15a)
K,Q,R,P,V MPC behaves as the best approximation of the controller in
s.t. Q ≥ 0, P ≥ 0, R ≥ σI, κ0 = K (15b) the norm  · , weighted by the positive definite matrix H.
V R + V S  QS = I (15c) Note that by (16) we require that the MPC behaves as the
favorite controller along the whole prediction horizon, which
and let K∗ , Q∗ , R∗ , P ∗ be the optimizer of problem (15), and is a tighter requirement than what is needed by Problem 1.
J ∗ be the corresponding optimal value. Problem 1 is exactly When the constraints are inactive, the MPC controller
solved only if J ∗ = 0. Otherwise, the solution matrices Q∗ , obtained as the solution of Problem 1 behaves as the fa-
R∗ and P ∗ define the MPC controller that, when constraints vorite controller (6), which is reasonably assumed to yield
are nor active, corresponds to the static linear state-feedback asymptotically stable closed-loop dynamics. Thus, if the
controller that is closest to (6) in the matrix norm  · . constraints are not active in a neighborhood of the origin, the
Problem (15) is also nonconvex due to the bilinear closed-loop dynamics are locally asymptotically stable. The
constraints (15c). Even though nowadays solvers exist for exact region of the state space where the matching occurs
bilinear problems [6], the convergence to a solution is not (i.e., where the constraints are inactive) is the polyhedron
guaranteed. In problems (13), (15), the main cause of P = {x ∈ Rn : −(GH −1 F + M )x ≤ W } where the
bilinearity are the free variables κi , i ∈ Z[1,N −1] . In fact, unconstrained optimizer −H −1 F x satisfies the constraints
these variables are free because only the first step of the of the QP problem (2) [7].
optimal input sequence is applied by the MPC and must be As for general MPC [3], global stability is more com-
equal to the command of the favorite controller (6). plicated to guarantee, especially when the local equivalence
In order to formulate problem (13) as a linear matrix with the favorite controller has to be maintained. The ap-
inequality (LMI), one can fix the whole vector proach described in [3] based on terminal cost and terminal
⎡ K ⎤ set can be specialized for this purpose.
κ̄1 Theorem 1: Let XT ⊆ Rn be a polyhedral PI set for (1b)
K = K̄  ⎣ .. ⎦. (16)
. in closed loop with (6) such that 0 ∈ int(XT ) and XT ⊆
κ̄N −1 {x ∈ Rn : x ∈ [xmin , xmax ], Kfv x ∈ [umin , umax ]}. Add
This obviously further constrains the design problem, be- constraint x(N |k) ∈ XT to (1) with Q, R, P are computed
cause constraints are imposed on the whole sequence of by (13) (or by (17), with Qi = Q, Ri = R, i ∈ Z[0,N −1] ,
command laws. In order to recover some degrees of freedom, and the optimum is J ∗ = 0), and denote by Xfeas ⊆ Rn the
the cost function in problem (13) can be made time-varying, set of states x ∈ RN such that (1) is feasible when x(k) = x.
⎡ Q1 0 0 ... 0 ⎤ ⎡ ⎤ Let P ≥ 0 be such that
R0 0 ... 0
0 Q2 0 ... 0
⎢ .. ⎥
0 R1 ... 0
(A + BK) P (A + BK) + K  RK + Q − P ≤ 0. (19)
Q = ⎣ .. .. . . .. ⎦ , R = ⎣ .. .. . . .. ⎦,
. . . . . . . . .
0 0 ... QN −1 0 0 ... 0 RN −1 4 Since x(0|k) is fixed in (1), the optimizer does not depend on Q0 .
0 0 ... 0 P

2220
Proceedings of the European Control Conference 2009 • Budapest, Hungary, August 23–26, 2009 TuA15.1

Then (i) the closed-loop dynamics are asymptotically stable Q̃, R̃, P̃ that are a feasible solution of (21), K is the LQR
and remain in Xfeas , for all x(0) ∈ Xfeas ; (ii) there exists gain that optimizes the LQR cost, where Q = Q̃, R = R̃.
a set Xfv ⊇ XT such that the MPC behaves as the favorite Theorem 2 guarantees that the optimal control problem (1)
controller; (iii) if Q > 0, Xfv is reached in a finite time where we set Q = Q̃, P = P̃ , R = R̃ results in an
k(x(0)), for all x(0) ∈ Xfeas and, if in addition Xfeas is MPC command (5) that is equal to the one of the favorite
bounded, there exists a finite k̄ = maxx(0)∈Xfeas k(x0 ). controller (6) whenever the constraints in the MPC quadratic
Proof: Due to space limitations, the proof is omitted program (2) are not active.
here and can be found in [8]. Cost function (21) can be used to define the reference
In order to satisfy the assumption of Theorem 1, the performance criterion of the MPC problem that comes into
LMI (19) is to be added to the cost design problems, e.g., play when the constraints are active, for instance as in (14).
(13) or (17), while XT can be computed for instance as the Also, Q can be removed from (21), by formulating (21c) as
maximum positively invariant set as described in [9]. an inequality and evaluating Q from the obtained P and R.
Problem (21) is not guaranteed to be feasible, because it
V. M ATCHING BASED ON I NVERSE LQR
is not true that any K ∈ Rn×m is the LQR gain for some
In this section we propose an alternative to solve Prob- choice of matrices P ∈ Rn×m , Q ∈ Rn×m , R ∈ Rn×m that
lem 1, which is computationally simpler. Instead of solv- satisfies (21b).
ing (11), we use the following theorem [10]. The problem of checking whether a given controller is
Theorem 2: Given Q̄ ∈ Rn×n , Q̄ = Q̄ ≥ 0, and R̄ ∈ optimal with respect to some performance criterion (inverse
m×m
R , R̄ = R̄ > 0, let P̄ ∈ Rn×n , P̄ = P̄  > 0 be the optimality problem) was first introduced by Kalman [11], in
solution of the Riccati equation the sixties. In the late eighties the inverse LQ design problem
P̄ = A P̄ A − A P̄ B(B  P̄ B + R̄)−1 B  P̄ A + Q̄. (20) has been studied [12], where the conditions for a linear
state-feedback controller to be an LQR gain were analyzed.
Set Q = Q̄, R = R̄, and P = P̄ in (1a). For any prediction The works [11]–[13] focused on continuous-time systems,
horizon N ∈ Z+ , when constraints are not active, the MPC and provided algebraic conditions for a given linear state
command (5) obtained by solving (1) is uMPC (x(k)) = feedback law to be an LQR. Equation (21) provides a way
KLQR x(k), where KLQR = −(B  P̄ B + R̄)−1 B  P̄ A is the to solve the inverse LQR design for discrete-time systems
LQR gain. by convex optimization.
Proof: See e.g. [10] and the references therein. Working with LQR gains introduces additional constraints
Theorem 2 establishes a relation between the linear quadratic but reduces problem complexity, since the size of (21), is not
regulator (LQR) and the model predictive controller. In related to the MPC horizon N . As a result, problem (21) is
fact, under the hypothesis of Theorem 2, the MPC con- significantly less complex than (15) when N is large. On the
troller that has no (active) constraints behaves as the other hand because of Theorem 2, whenever (21) is feasible,
linear state feedback gain that optimizes the LQR cost a solution of (17) exists with zero cost, while the opposite
∞
minu(·) k=0 x(k) Qx(k) + u(k) Ru(k) for the linear dy- is not guaranteed.
namics x(k + 1) = Ax(k) + Bu(k). MPC cost based on (21) ensures that when the constraints
We look for weights Q, R and P such that the favorite are not active the MPC is equivalent to an LQR, and hence
controller (6) is the LQR gain with respect to them. In this yields (locally) asymptotically stable closed-loop dynamics.
way by exploiting Theorem 2, the MPC controller, which Global stability in the set of feasible initial conditions in the
behaves as the LQR for any horizon N ∈ Z+ , will be also presence of constraints can be achieved by choosing N large
equal to the favorite controller (6). enough [7], [14], [15], since Corollary 1 is independent on
Corollary 1: Consider the optimization problem the value of N . Also, Theorem 1 can be applied using the
min J(Q, R, P ) (21a) Riccati matrix P as terminal weight and KLQR as auxiliary
Q,R,P controller, since P computed from (21) satisfies (19).
s.t. P ≥ 0, R ≥ σI, Q ≥ 0 (21b)
VI. E XAMPLES
P = A P A + A P BK + Q (21c)
B  P A = −(B  P B + R)K (21d) We present some examples of inverse matching tech-
niques.
where J(·) is a convex cost function of its arguments Example 1 (Matching based on QP matrices): Consider
(e.g., as in (14)) and (21b), (21c), (21d) are linear matrix the unstable linear system x(k + 1) = Ax(k) + Bu(k),
(in)equalities. Let Q̃, R̃, P̃ be any feasible solution (i.e., where
not necessarily the optimal one) of (21). Then the MPC 0.675 0.923 0.014 0.5 x ∈ R3
strategy based on the optimal control problem (1) where we A= −0.315 0.215 −0.750 , B= 0 ,
1.05 0.00 1.50 1 u ∈ R,
set Q = Q̃, P = P̃ , R = R̃, solves Problem 1.
Proof: Equalities (21c), (21d), and constraint (21b) and favorite controller (6), where K = [ −0.918 0.347 −0.806 ]
enforce P and K to be the solution of the Riccati equation is designed through pole-placement so that the closed-loop
and the corresponding LQR gain, respectively, and Q, R, to matrix A + BK has eigenvalues 0.35, 0.375, 0.40. Assume
be the corresponding cost function matrices. Thus, given any that the input constraints −1.5 ≤ u ≤ 1.5 must be enforced.

2221
S. Di Cairano and A. Bemporad : Model Predictive Controller Matching: Can MPC Enjoy Small Signal Properties ... TuA15.1

3
Figure 2, where the solid line is the MPC command and the
2 dashed line with circle markers is the input that would be
applied by the favorite controller.
1

VII. C ONCLUSIONS
0
u

To exploit both the frequency-domain properties of a given

−1 linear controller for signals small enough not to activate con-
−2
straints, and the ability of model predictive control to handle
constraints, in this paper we provided criteria to convert
−3
0 5 10 15 the existing linear control law into an MPC controller. The
k
approach can be seen as a way of automatically generating
Fig. 1. Example 1, matching based on QP matrices. MPC command an anti-windup scheme, which is a piecewise-affine function
uMPC compared to favorite controller command, ufv = Kx, K = in case the resulting MPC controller is considered into its
[−0.918 0.347 − 0.806] explicit form [7]. Finally, we have shown how to obtain
50
global stability of the MPC scheme in the set of feasible
initial conditions, by adequately adapting the MPC stabiliza-
tion results to preserve the controller matching objective.
25

R EFERENCES
0 [1] K. Åstrom and L. Rundqwist, “Integrator windup and how to avoid
u

it,” in American Control Conference, 1989, pp. 1693–1698.

[2] J. Maciejowski, Predictive control with constraints. Englewood Cliffs,
−25 NJ: Prentice Hall., 2002.
[3] D. Mayne, J. Rawlings, C. Rao, and P. Scokaert, “Constrained model
predictive control: Stability and optimality,” Automatica, vol. 36, no. 6,
−50
0 5 10 15 pp. 789–814, June 2000.
k
[4] S. Qin and T. Badgwell, “A survey of industrial model predictive
control technology,” Control Engineering Practice, vol. 93, no. 316,
Fig. 2. Example 2, matching based on inverse LQR. MPC command uMPC pp. 733–764, 2003.
compared ufv = [−5.38 − 2.84 − 0.25 − 2.37]x(k). [5] M. Bares, M. Diehl, and I. Necoara, “Every continuous nonlinear
control system can be obtained by parametric convex programming,”
IEEE Trans. Automatic Control, vol. 53, no. 8, pp. 1963–1967, 2008.
[6] M. Kocvara and M. Stingl, PENBMI User’s Guide, Tomlab Opti-
We solve problem (17), where we have set N = 3, σ = 10−3 , mization, 2003, latest version available at http://www.penopt.com/doc/
K̄ as in (18), and we have imposed Qi = Q ≥ 0, ∀i ∈ penbmi2 1.pdf.
[7] A. Bemporad, M. Morari, V. Dua, and E. Pistikopoulos, “The ex-
Z[1,N −1] , Ri = R > σI, ∀i ∈ Z[0,N −1] . Problem (17) is plicit linear quadratic regulator for constrained systems,” Automatica,
solved with an optimal cost of 0, resulting in matrices Q∗ , vol. 38, no. 1, pp. 3–20, 2002.
R∗ , P ∗ . Thus, when Q = Q∗ , R = R∗ , P = P ∗ in (1), the [8] S. Di Cairano and A. Bemporad, “Model predictive controller tun-
ing by controller matching,” IEEE Trans. Automatic Control, 2009,
MPC command uMPC equals ufv whenever the constraints submitted.
are not active. The simulation of the closed-loop system from [9] I. Kolmanovsky and E. G. Gilbert, “Theory and computation of dis-
initial state x(0) = [ 0 5 5 ] is shown in Figure 1, where the turbance invariant sets for discrete-time linear systems,” Mathematical
Problems in Engineering, vol. 4, pp. 317–367, 1998.
solid line is the MPC command applied to the system and the [10] P. Scokaert and J. Rawlings, “Constrained linear quadratic regulation,”
dashed line with circle markers is the hypothetical command IEEE Trans. Automatic Control, vol. 43, no. 8, pp. 1163–1169, 1998.
that would be issued by the favorite controller for the same [11] R. Kalman, “When is a Linear Control System Optimal?” Trans.
ASME, Jour. of Basic En., pp. 51–60, 1964.
state and that does not account for constraints. Note that in [12] T. Fujii, “A new approach to the LQ design from the viewpoint of the
order to have uMPC = ufv the constraints need not to be inverse regulator problem,” IEEE Trans. Automatic Control, vol. 32,
active along the whole MPC horizon N = 3. no. 11, pp. 995–1004, 1987.
[13] K. Sugimoto, “Partial pole placement by LQ regulators: an inverse
Example 2 (Matching based on inverse LQR): Consider problem approach,” IEEE Trans. Automatic Control, vol. 43, no. 5,
the linear system x(k + 1) = Ax(k) + Bu(k), where pp. 706–708, 1998.
[14] D. Chmielewski and V. Manousiouthakis, “On constrained infinite-
1.8
1
1.2
0
0
0
1
0
0
0 x ∈ R4 time linear quadratic optimal control,” Systems & Control Letters,
A= 3.6 2.4 1 2 , B= 0 ,
0 0 0 0 1 u ∈ R, vol. 29, no. 3, pp. 121–130, Nov. 1996.
[15] J. Rawlings and K. Muske, “The stability of constrained receding-
and constrained input −24 ≤ u ≤ 24, and the favorite horizon control,” IEEE Trans. Automatic Control, vol. 38, pp. 1512–
1516, 1993.
controller (6), K = [ −5.38 −2.84 −0.25 −2.37 ]. The inverse
LQR-based matching problem (21) where σ = 10−3 , and
with objective function (14), where Q̄ = 6.5I, R̄ = 1,
wR = 1, wP = 0 was solved obtaining Q∗ , R∗ , P ∗ . We
have implemented the MPC strategy (1) with N = 3 and
Q = Q∗ , R = R∗ , P = P ∗ so that uMPC = ufv whenever
the constraints are not active. The simulation of the closed-
loop system from initial state x(0) = [ 5 5 0 0 ] is shown in

2222