Automatica 47 (2011) 16381645

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Explicit/multi-parametric model predictive control (MPC) of linear discrete-time

systems by dynamic and multi-parametric programming
K.I. Kouramas a, , N.P. Fasca b , C. Panos a , E.N. Pistikopoulos a
a
Centre for Process Systems Engineering, Department of Chemical Engineering, Imperial College London, London SW7 2AZ, UK
b
AspenTech, C1 Reading International Bus Park, Reading, RG2 6DT, UK

article info abstract

Article history: This work presents a new algorithm for solving the explicit/multi-parametric model predictive control
Received 10 May 2010 (or mp-MPC) problem for linear, time-invariant discrete-time systems, based on dynamic programming
Received in revised form and multi-parametric programming techniques. The algorithm features two key steps: (i) a dynamic
7 October 2010
programming step, in which the mp-MPC problem is decomposed into a set of smaller subproblems in
Accepted 16 February 2011
Available online 1 June 2011
which only the current control, state variables, and constraints are considered, and (ii) a multi-parametric
programming step, in which each subproblem is solved as a convex multi-parametric programming
Keywords:
problem, to derive the control variables as an explicit function of the states. The key feature of the
Explicit Model Predictive Control proposed method is that it overcomes potential limitations of previous methods for solving multi-
Model Predictive Control parametric programming problems with dynamic programming, such as the need for global optimization
Multi-parametric control for each subproblem of the dynamic programming step.
Multi-parametric programming 2011 Elsevier Ltd. All rights reserved.
Dynamic programming

1. Introduction systems, where the constraints are explicitly handled in the

optimization problem. Nevertheless, the key limitations of MPC
Model predictive control (MPC) is a control method which aims is its online computational effort due to the repetitive solution of
to derive the sequence of control actions over a finite time horizon, the underlying optimization and its implicit feedback nature: the
by seeking to optimize the controller performance based on the optimal values of the control action are determined numerically
predicted states of the system under control (Rawlings & Mayne, from the MPC optimization without any knowledge of the
2009). This is achieved by repetitively solving an online open- governing control laws.
loop control problem at each sampling instant, which minimizes These shortcomings can be overcome by using the so-called
the state and control deviations from the set point based on the explicit/multi-parametric MPC (also known as explicit MPC, multi-
prediction of the states given from a mathematical model of the parametric MPC or simply mp-MPC) methods (Bemporad, Morari,
system. Only the first control of the optimal control sequence is Dua, & Pisitkopoulos, 2002; Johansen, 2002; Pistikopoulos, Dua,
applied to the system, and the optimization is repeated in the Bozinis, Bemporad, & Morari, 2002; Pistikopoulos, Georgiadis, &
next time instant for the new values of the states; in that way, Dua, 2007a). In mp-MPC, the online optimization problem involved
an implicit feedback controller is obtained. Being a model-based in traditional MPC is solved with multi-parametric programming
control method, and since it relies on online optimization, MPC methods (Bank, Guddat, Klatte, Kummer, & Tammer, 1982;
is a popular method for the control of constrained multivariable Pistikopoulos, Bozinis, & Dua, 2000; Pistikopoulos, Georgiadis, &
Dua, 2007b) to derive the optimal controls as explicit functions
of the system states and the critical regions in the state-space
where these functions are valid; hence, a feedback control policy
The financial support of EPSRC (Projects GR/T02560, EP/047017/1), European
is derived which describes the governing control laws of the
Commission (PRISM ToK project, Contract No: MTKI-CT-2004-512233 and DIA-
MANTE ToK project, Contract No: MTKI-CT-2005-IAP-029544), European Research underlying MPC problem. The online optimization involved in
Council (MOBILE, ERC Advanced Grant No: 226462) and KAUST is gratefully ac- MPC can then be replaced by a sequence of function evaluations,
knowledged. This paper was not presented at any IFAC meeting. This paper was thus reducing the online computational effort. Explicit/multi-
recommended for publication in revised form by Editor Ber Rstem.
Corresponding author. Tel.: +44 20 75946632; fax: +44 20 75941129. parametric MPC has received extensive attention in the control
systems literature. An indicative list of key publications is given in
E-mail addresses: [email protected] (K.I. Kouramas),
[email protected] (N.P. Fasca), [email protected] Table 1, and the interested reader can look in Pistikopoulos (2009)
(C. Panos), [email protected] (E.N. Pistikopoulos). and Pistikopoulos et al. (2007a,b), and the references within.
0005-1098/$ see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2011.05.001
 K.I. Kouramas et al. / Automatica 47 (2011) 16381645 1639

Table 1
Key publications in explicit/multi-parametric Model Predictive Control.

Multi-parametric model predictive control (Bemporad et al., 2002; Pistikopoulos, 2000;

Pistikopoulos et al., 2002; Tondel et al., 2003)
Multi-parametric continuous time model (Kojima & Morari, 2004; Sakizlis et al., 2005)
predictive control
Hybrid multi-parametric model predictive control (Bemporad et al., 2000; Borrelli et al., 2005; Sakizlis
et al., 2001)
Robust multi-parametric model predictive control (Bemporad et al., 2001; Mayne et al., 2006;
Pistikopoulos et al., 2009; Sakizlis et al., 2004)
Multi-parametric dynamic programming (Borrelli et al., 2005; Fasca et al., 2008; Muoz de la
Pea et al., 2004; Pistikopoulos et al., 2009)
Multi-parametric nonlinear model predictive (Bemporad, 2003; Dominguez & Pistikopoulos, 2010;
control Johansen, 2002)

In most available methods for mp-MPC, the main approach each stage as a function of the systems states. By exploiting ba-
for explicit/multi-parametric MPC is to solve the underlying MPC sic properties of dynamic programming and the convexity of the
optimization problem as a single-stage off-line multi-parametric optimization problem of the mp-MPC formulation, the proposed
programming problem (Bemporad et al., 2002; Pistikopoulos algorithm only solves a convex, multi-parametric programming
et al., 2007a). It has been shown that, even for linear cases, the problem at each stage of the DP procedure, and hence avoids the
complexity of this multi-parametric program increases with the need for global optimization.
number of optimization variables, constraints and the prediction
horizon (Bemporad et al., 2002; Fasca, Kouramas, Saraiva, Rustem, 2. Problem definition
& Pistikopoulos, 2008; Pistikopoulos et al., 2002). In order to
2.1. Explicit/multi-parametric model predictive control
overcome this, dynamic programming (DP) techniques (Bellman,
2003; Bertsekas, 2005) have been introduced for solving multi-
Consider the case of linear, discrete-time time-invariant sys-
parametric programming problems such as the explicit/multi- tems
parametric MPC problem. The main idea is to reduce the original
multi-stage optimization problem to a set of smaller stage xt +1 = Axt + But , (1)
n m
optimization subproblems in which only the control variables, where xt X R is the state vector, ut U R is the system
states, and constraints at the current stage are considered input vector, A Rnn , B Rnm , and X and U are the state and
(Bertsekas, 2005). Dynamic Programming-based methods have input constraint sets, given as follows:
been introduced for a number of mp-MPC problems including X = {x Rn | Gx w} (2)
the constrained linear quadratic regulator problem (CLQR) (Fasca
et al., 2008; Muoz de la Pea, Alamo, & Bemporad, 2004), explicit U = {u R | Mu },
m
(3)
MPC of hybrid systems (Borrelli, Baoti, Bemporad, & Morari, where G R ,d R ,M R
ng n ng
, R . It is as-
mM n mM

2005), and robust explicit MPC (Bemporad, Borrelli, & Morari, sumed that both sets contain the origin in their interior. The linear,
2003; Pistikopoulos, Fasca, Kouramas, & Panos, 2009). discrete-time explicit/multi-parametric MPC problem is defined as
Although DP can be applied for a wide class of mp-MPC prob- the following multi-stage optimization problem:
lems, there are certain issues that might affect the complexity of VN (x) = min J (U, x)
UU
DP. Solving each stage problem in the DP procedure as a multi-
parametric programming problem generally results in nonlinear N 1

stage optimization subproblems (Pistikopoulos et al., 2007a). Even = min {xTt Qxt + uTt Rut } + xTN PxN (4)
UU
for the linear case, the objective functions of each stage opti- t =0

mization become piecewise quadratic (Borrelli et al., 2005; Pis- s.t. xt +1 = Axt + But , t = 0, . . . , N (5)
tikopoulos et al., 2007a). Then, either a multi-parametric nonlinear xt X, t = 0, . . . , N (6)
programming problem has to be solved (Pistikopoulos et al.,
ut U , t = 0, . . . , N 1 (7)
2007a), or a series of multi-parametric quadratic programming
(mp-QP) problems at each stage, which, however, results in the x = x0 , (8)
overlapping of critical regions, thus requiring the use of compar- where U UNm RNm is the control sequence of the cur-
ison techniques, in which the objective functions of the problem rent and past control inputs, N is the prediction horizon, Q , P
for each of the (overlapping) critical regions are compared (Borrelli are positive semi-definite matrices and R is a positive definite
et al., 2005). If the objective function of the MPC optimization prob- matrix of appropriate dimensions. It is easy to observe from
lem is linear (i.e. if it is described by an infinite or one-norm), the (4)(8) that the explicit/multi-parametric MPC is a multi-stage op-
comparison is obtained by solving a convex linear program. How- timization problem, in which the time instant t defines each stage
ever, if the objective function of the MPC optimization problem is of (4)(8), and the stage and terminal costs are quadratic (where P
quadratic, the comparison is in general a non-convex, optimization is usually obtained from solving the Ricatti equation for the uncon-
problem, and hence the use of global optimization methods is re- strained problem, (4)). The solution of (4) gives an optimal control
quired (Borrelli et al., 2005; Fasca et al., 2008); these are typically sequence U = {u0 , u1 , . . . , uN 1 } and the corresponding optimal
hard to solve. path x1 , x2 , . . . , xN which minimizes the stage additive cost in (4)
and satisfies the state and input constraints.
In this work, based on our earlier work of Fasca et al. (2008),
we develop a new method for explicit/multi-parametric MPC of Remark 1. In Bemporad et al. (2002) and Pistikopoulos et al.
linear discrete-time systems with state and input constraints. (2002), a method was presented for solving (4) as a multi-
The proposed method employs dynamic programming and multi- parametric quadratic programming (mp-QP) problem and deriving
parametric programming techniques to disassemble the optimiza- the control actions ut as explicit functions of the state x, u =
tion problem of the mp-MPC formulation into a set of smaller stage f (x) = Ki x + ci , if x CRi , where CRi is the critical region in
optimization subproblems and the optimal control variables for the state space in which the control u = f (x) is valid.
1640 K.I. Kouramas et al. / Automatica 47 (2011) 16381645

2.2. Dynamic programming which is a multi-parametric quadratic programming (mp-QP)

problem with uN 1 as the optimization variable and xN 1 as the
Dynamic programming (DP) (Bertsekas, 2005) is a well-docu- problem parameter. The solution to problem (12) is a piecewise
mented, powerful tool for solving multi-stage problems such as affine function of the states uN 1 = fN 1 (xN 1 ) = KNi 1 xN 1 +
(4). Based on the optimality principle, the multi-stage problem (4) cNi 1 , and the value function VN 1 () is a piecewise quadratic
can be disassembled into a set of lower-dimensional problems, in function of the states, i.e. VN 1 (xN 1 ) = xTN 1 Q i xN 1 . Then, the
which the value function for each stage t of (4) is given by Bertsekas objective function JN 2 (uN 2 , xN 2 ) = xTN 2 QxN 2 + uTN 2 RuN 2 +
(2005) VN 1 (xN 1 ) of (10) at stage N 2 is also a piecewise quadratic
N 1 function. Therefore, the objective function at each stage (10) of

Vt (xt ) = min {xTk Qxk + uTk Ruk } + xTN PxN . (9) the dynamic programming procedure is a nonlinear function of
ut ,...,uN 1 the inputs and states of the same stage, and the optimization (10)
k=t
at each stage becomes a nonlinear multi-parametric programming
By further applying the optimality condition (Bertsekas, 2005), the problem.
multi-stage problem (4) can be disassembled into the following
recursive optimization problem (Bertsekas, 2005): Remark 3. Since (10) is a nonlinear multi-parametric optimiza-
Vt (xt ) = min Jt (ut , xt ) tion problem, it is difficult for the available methods for multi-
ut U parametric programming to be applied. As it was shown in Borrelli
et al. (2005) and Muoz de la Pea et al. (2004), this issue can be ad-
= min {xTt Qxt + uTt Rut + Vt +1 (xt +1 )} (10)
ut U dressed by solving the problem at each stage t as a set of q simple
s.t. xt +1 = Axt + But mp-QP problems, which in general results in overlapping critical
regions. Hence, comparison methods have to be employed which
xt X, ut U, xt +1 X,
result in a, possibly, non-convex optimization problem, i.e. a global
where VN (xN ) = xTN PxN . The stage-wise problem (10) is solved for optimization problem (Fasca et al., 2008).
each stage t = N 1, . . . , 0 starting from stage N 1 and ending
The objective here would be to avoid the use of global op-
at stage 0 and only the control inputs ut and constraints on the
timization methods and solve the DP problem, and in addition
current stage t are considered. The cost-to-go function Vt +1 (xt +1 )
the explicit/multi-parametric MPC problem, by only using convex
is the value function for (10) for stage t + 1, and it is obtained
multi-parametric programming methods. Therefore, we present a
from the state and optimal solutions of the previous stages. Hence,
method for solving the DP problem (10) by only solving a con-
the current optimal solution ut of (10) depends on the optimal
vex multi-parametric programming problem for each stage of the
decisions obtained in the previous iterations.
DP procedure, and thus avoiding the need for global optimization
methods.
Remark 2. Note that V (x) = V0 (x0 ), i.e. dynamic programming
obtains the optimal solution for (4) when the last stage t = 0 is 3. Algorithm for explicit/multi-parametric MPC
solved (Bellman, 2003; Bertsekas, 2005).

In Borrelli et al. (2005), Fasca et al. (2008), and Muoz de la The proposed algorithm for solving the explicit/multi-parame-
Pea et al. (2004), a method was proposed for solving the DP tric MPC problem (4)(8) is realized in two main steps: (i) a
problem (10) with multi-parametric programming, to obtain the dynamic programming (DP) step, and (ii) a multi-parametric
optimal control decision at each stage t as an explicit function of programming step; these are described in the following sections.
its incumbent state xt :
3.1. Dynamic programming representation of the MPC problem
1
K1 xt + ct
if xt CR1
.
ut = t (xt ) = ..
.. (11) The multi-stage problem (4)(8) is decomposed into the follow-
. ing stage optimization problems:
q
if xt CR ,
q
Kq xt + ct

Vt (xt ) = min Jt (ut , xt )
ut U
where CRi are the critical regions in which the control law Ki xt +
cti , i = 1, . . . , q is valid, for all t [0, 1], and hence the N 1

control sequence U = {u0 (x0 ), . . . , uN 1 (xN 1 )}. By using multi- xTi Qxi + uTi Rui + xTN PxN

= min (13)
ut U
parametric programming to solve each stage problem (10), the i=t
control variables at the current stage are obtained as a piecewise s.t. xi+1 = Axi + Bui
affine function of the states (11), and in addition the value xt X , ut U , x t +1 X
function Vt +1 () as a piecewise quadratic function of the states
(Borrelli et al., 2005; Fasca et al., 2008; Muoz de la Pea et al., for all t = N 1, . . . , 0. Problem (13) is equivalent to prob-
2004). Hence, the objective function of (10) becomes a piecewise lem (9), where only the current control variable is considered
quadratic function, and thus problem (10) becomes a nonlinear in the optimization, since, based on the dynamic programming
multi-parametric optimization problem (Fasca et al., 2008). For principle, we assume that the future control variables ut +1 =
example, problem (10) at stage t = N 1 is given as the following t +1 (xt +1 ), . . . , uN 1 = N 1 (xN 1 ) have been obtained in the
optimization problem: previous stages of dynamic programming. In addition, problem
(9) is equivalent to (10), where the previous solutions ut +1 =
VN 1 (xN 1 ) = min JN 1 (uN 1 , xN 1 ) t +1 (xt +1 ), . . . , uN 1 = N 1 (xN 1 ) have not yet been substi-
uN 1 U
tuted in the cost-to-go function Vt +1 (xt +1 ).
= min {xTN 1 QxN 1 + uTN 1 RuN 1 + VN (xN )} (12)
uN 1 U
Remark 4. The stage-wise optimization problem (13) is convex
s.t. xN = AxN 1 + BuN 1 with respect to the control variables ut , . . . , uN 1 and the state
xN 1 X, uN 1 U, xN X, variable xt , since Q 0 and R > 0.
 K.I. Kouramas et al. / Automatica 47 (2011) 16381645 1641

The convexity of problem (13) can then be used to derive ut +1 , . . . , uN 1 have been already obtained from the previous
an algorithm that solves each of the problems (13) (and hence stages of the DP procedure, we can use them to eliminate
(10)) as a convex multi-parametric programming problem. More ut +1 , . . . , uN 1 from expression (15) and derive ut as an explicit
specifically, by function of the state xt , i.e. derive a state-feedback solution. We
i1 present next a procedure for obtaining ut = t (xt ), which for
(1) substituting the prediction of the state xi = Ai x + j =0 simplicity will be first demonstrated for stages N 1 and N 2
Ai1j Buj for all i = t + 1, . . . , N 1 in the objective function and then extended for every stage t.
and inequalities of (13), and Stage N 1: In stage N 1, the control variable is uN 1 , and
(2) considering ut as the optimization variable, and = [xt ut . . . the parameter is only the state xN 1 . Therefore, the solution of (14)
uN 1 ]T the vector of parameters for (13), for stage N 1 is obtained from (15) by using t = N 1 and
a multi-parametric quadratic programming (mp-QP) reformula- N 1 = xN 1 :
tion of problem (13) is obtained as follows (Fasca et al., 2008; Pis- uN 1 = fN 1 (xN 1 ) = KNi 1 xN 1 + cNi 1
tikopoulos et al., 2002):
if xN 1 CRiN 1 , i = 1, . . . , sN 1 (16)
1
Vt (xt ) = min uTt Hut + Fut + Y
T T
(14) where CRiN 1 = {xN 1 | AiN 1 xN 1 biN 1 },
ut U 2
s.t. GAxt + GBut w i.e. uN 1 is already an explicit PWA function of its incumbent state
x N 1 .
Gxt w Mut t ,
Stage N 2: In stage N 2, the optimization variable for (14)
where the matrices H, F , and Y are obtained after substituting is uN 2 , and the parameter vector is N 2 = [xN 2 uN 1 ]T . The
the systems linear model into the objective function, are of solution of the mp-QP problem (14) for stage N 2 is then given
appropriate dimensions, and H > 0. by
uN 2 = fN 2 (xN 2 )
Remark 5. Note that if traditional DP methods had been applied,
the values of the control variables ut +1 , . . . , uN 1 , obtained at = KNj 2 xN 2 + LjN 2 uN 1 + cNj 2
the previous stages of the DP procedure, would have been j
if xN 2 , uN 1 CRN 2 , j = 1, . . . , sN 2 , (17)
substituted in (14), hence leading to a nonlinear multi-parametric
programming problem with a piecewise quadratic function, as was where
described in Section 2 and Remark 5, that would have needed the
CRiN 2 = {xN 2 | AiN 2 xN 2 + BNi 2 uN 1 biN 2 },
use of global optimization methods.
i.e. uN 2 is an explicit expression of both xN 2 and uN 1 . Sub-
Problem (14) is a convex mp-QP problem, since H > 0, stituting xN 1 = AxN 2 + BuN 2 in (16), we derive uN 1 =
and therefore it can be solved by applying conventional multi- fN 1 (xN 2 , uN 2 ). Then, the expressions (16) and (17) are com-
parametric-programming methods (Pistikopoulos, 2009). The bined to obtain the following set of PWA expressions on xN 2 , uN 2
solution of problem (14) as an mp-QP problem is presented in the and uN 1 :
following section.
uN 1 = KNi 1 AxN 2 + KNi 1 BuN 2 + cNi 1 (18)
3.2. Multi-parametric programming solution strategy for constrained j j j
dynamic programming uN 2 = KN 2 xN 2 + LN 2 uN 1 + cN 2 (19)
j
x N 2 , uN 2 CRiN 1 , xN 2 , uN 1 CRN 2 . (20)
Since (14) is a convex mp-QP problem, its solution is given
by the following explicit linear piecewise affine (PWA) expression The control uN 2 can then be derived as an explicit function of xN 2
(Pistikopoulos et al., 2007a,b): by eliminating uN 1 either (i) by substituting uN 1 from (18) into
(19)(20) and then solving for uN 2 , or (ii) by using elimination
Kt t + ct if t CR1t
1 1
methods such as orthogonal projection or FourierMotzkin elimi-
ut = ft (t ) =
.. (15) nation. Both operations are linear, and hence, when performed on
. the set of linear equalities and inequalities (18)(20), the resulting
s
+ s
if t
s

Kt t t ct t CRt t

expression for uN 2 is a also linear PWA expression of xN 2 (Fasca
where CRit = {t | Ait t bit }, et al., 2008):

where Kti and cti are of appropriate dimensions and CRit , i = uN 2 = N 2 (xN 2 ) = KNi 2 xN 2 + cNi 2
1, . . . , st are the critical regions where the linear affine control
if xN 2 CRiN 2 , i = 1, . . . , sN 2 (21)
laws Kti t + cti are valid. Note from (15) that the control variable
ut at the current stage t is an explicit function of the state xt at where CRiN 2 = {xN 2 | AiN 2 xN 2 biN 2 }.
the current stage of the DP problem and also of the future control
variables.
Remark 7. Note that (18)(20) are obtained for all possible com-
Remark 6. Recall that, by using DP methods to solve the explicit binations i, j of control expressions and critical regions from (16)
MPC problem, the value of the control ut derived at each and (17). Each of these combinations corresponds to a set of values
stage t of the DP procedure depends on the values of the of xN 2 , uN 2 and uN 1 , described by (18)(20).
control inputs ut +1 , . . . , uN 1 derived in the previous stages by
means of the cost-to-go function (Bertsekas, 2005). This is also Remark 8. If a combination (18)(20) is feasible, i.e. the resulting
expressed in the relation (15), where, instead of replacing the set of xN 2 , uN 2 and uN 1 in (18)(20) is not empty, then the
controls ut +1 , . . . , uN 1 with their values, we considered them as corresponding control sequence {uN 2 , uN 1 } is a feasible solution
parameters for the stage problem (14). to (14), and hence to (13).

Expression (15) gives the control variable ut as an explicit Remark 9. It is possible that not all combinations of (16) and (17)
function of xt , ut +1 , . . . , uN 1 . However, since the control inputs are realizable and that the set of xN 2 , uN 2 and uN 1 in (18)(20)
1642 K.I. Kouramas et al. / Automatica 47 (2011) 16381645

is empty, which further implies that no feasible control exists for Lemma 1 (Fasca et al., 2008). If xt , ut (xt ), . . . , uN 1 (xN 1 ) is the
(13). The combinations corresponding to an empty set in (18)(20) solution for each stage t of the DP problem (13), then the following
are then removed, and only the ones corresponding to non-empty inequalities are satisfied:
sets are kept.
Vt (xt , ut (xt ), ut +1 (xt +1 ), . . . , uN 1 (xN 1 ))
The same procedure can be applied for any stage t. At any stage t,
the optimal control for (13) is given by (15), and the solutions of Vt (xt , ut (xt ), ut +1 (xt +1 ), . . . , uN 1 (xN 1 ))
the previous stages are given by Vt +1 (xt , ut (xt ), ut +1 (xt +1 ), . . . , uN 1 (xN 1 ))
ut +1 = ft +1 (xt +1 ) uN 1 = fN 1 (xN 1 ). (22) Vt +1 (xt , ut (xt ), ut +1 (xt +1 ), . . . , uN 1 (xN 1 )).
i t
i1 i1j
The linear system model xi = A xt + j =t A Buj is then
substituted in (22) for all i = t + 1, . . . , N 1 to express all Proof. The proof of the lemma is given in Fasca et al. (2008) and
equalities in (22) in terms of xt , ut , . . . , uN 2 . Then, (15) and (22) results from the convexity of problem and the properties of the DP
are combined to obtain the following set of PWA expressions: procedure. 

uN 1 = fN 1 (xt , ut , . . . , uN 2 ) (23) The above lemma shows that the control sequence ut (xt ), . . . ,
.. uN 1 (xN 1 ), obtained at each stage t of the dynamic programming
. procedure, is optimal for all stages i t + 1. The following result
ut +2 = ft +2 (xt , ut , ut +1 ) (24) is a direct consequence of Lemma 1.

ut +1 = ft +1 (xt , ut ) (25) Theorem 1. The optimal control sequence U = {0 (x0 ), . . . , N 1

ut = ft (xt , ut +1 , . . . , uN 1 ). (26) (xN 1 )}, obtained by Algorithm 1, is an optimal control sequence for
the explicit/multi-parametric MPC problem (4).
An explicit control ut = t (xt ) as an explicit function of xt can
be derived from (23)(26), if the control variables of the previous Proof. Each control law t (xt ) satisfies the state and input
stages ut +1 , . . . , uN 1 are eliminated either constraints of (13) and hence the state and input constraints of (4);
(1) by substituting (23)(25) into (26) and then solving with therefore, U = {0 (x0 ), . . . , N 1 (xN 1 )} is a feasible solution for
respect to ut and xt , or (4). Moreover, from Lemma 1, we have that U is an optimal solution
(2) by using elimination methods such as orthogonal projection for (13) for all t 0. Since V0 (x0 ) = VN (x0 ), U is also an optimal
or FourierMotzkin elimination on the combined linear solution for (4). 
piecewise affine expressions (23)(26).
The above theorem shows that Algorithm 1 obtains the explicit
As was shown for the case of stage N 2, the resulting control ut solution to the explicit MPC problem (4).
will be give as a function of the state xt :

ut = t (xt ) = Kti xt + cti , if xt CRit (27) Algorithm 1 Algorithm for explicit/multi-parametric MPC
Data: A, B, Q , R, P , N , X, U, T
where CRit = xt | Ait xt bt . i

(28)
Output: A control sequence of explicit control
laws {ut , . . . , uN 1 } for the explicit/multi-
Remark 10. Note that, in the proposed procedure, the control vari- parametric MPC problem
ables of the previous stages (23)(25) are not substituted directly 1: Set t = N 1 (start at stage N 1)
in (13) and then the equation is solved. Instead, (13) is solved 2: Repeat for each stage t
first as a convex multi-parametric program and then (23)(25) 3: Solve (14) with ut being the optimization
are substituted in the solution of (14) to derive the final solution variable and t = {xt ut +1 . . . uN 1 } being
(27). This allows for (14) to be solved as a convex mp-QP problem the parameter vector. Obtain ut = ft (t ) from
and avoids global or combinatorial optimization. (15).
4: Eliminate ut +1 , . . . , uN 1 to obtain ut =
Remark 11. The set of feasible parameters (i.e. the states for which t (xt ) (Equation (27))
a feasible control for (14) exists) for (13) at each stage t can be 5: Set t = t 1
st i
obtained from (27) as Xt = i= 0 CR t . In addition, the set of 6: Until final stage t = 0 has been reached
feasible parameters for (13) for t = 0 is the set X0 , which also
implies that X0 is the set of feasible states for (4).
The main characteristics and advantages of the proposed
3.3. Novel multi-parametric model predictive control algorithm procedure are as follows.

Based on the procedure described in Sections 3.1 and 3.2, we

The proposed procedure (Algorithm 1) solves the explicit/multi-
parametric MPC problem (4), based on the solution from dy-
can now present an algorithm for the solution of the explicit MPC
namic programming via multi-parametric programming.
problem. The steps of the proposed algorithm are given in detail in
Algorithm 1. The algorithm starts at stage t = N 1, and iterates The solution of the multi-stage optimization (4) is replaced by
through Steps 26 of Algorithm 1 for each stage t, until t = 0. In solving a sequence of problems of smaller dimensions (14),
Step 3, the mp-QP problem (14) is solved for the current stage t i.e. with a smaller number of optimization variables and a
to derive ut = ft (t ) as a function of the state and future control smaller number of constraints.
variables. Then, in Step 4, the elimination procedure described in Algorithm 1 derives the sequence of optimal control laws U =
Section 3.2 is applied to obtain the explicit control ut = t (xt ). {0 (x0 ), . . . , N 1 (xN 1 )} for the explicit/multi-parametric
After the algorithm terminates at the last stage t = 0, a sequence MPC problem (4) (see Theorem 1).
of explicit feedback control laws U = {0 (x0 ), . . . , N 1 (xN 1 )} is At each stage of the proposed algorithm, only a convex multi-
computed; each control law t (xt ) is given by a PWA expression, parametric quadratic programming problem (14) (Step 3 of
(27). Algorithm 1) is solved, and only linear algebraic manipulations
We now recall the following result from Fasca et al. (2008). are used (Step 4 of Algorithm 1) to obtain the optimal solution.
 K.I. Kouramas et al. / Automatica 47 (2011) 16381645 1643

Implementation: The optimal control sequence U is implemented Table 2

to the system in an MPC fashion, i.e. by only applying the first Explicit solution u1 = f1 (x1 ) for stage 1Example 1.

control u0 = 0 (x0 ) of the sequence. In that way, an explicit/multi- Control law Critical region
parametric controller is obtained 1 u1 = 2 0.9961x11 + x21 0.3292
u(x) = 0 (x) 0.9961x11 x21 0.3292
2 u1 = 6.051x11 6.074x21
which consists of a linear PWA state feedback control law. +0.9961x11 + x21 0.3292
Alternatively, the whole optimal control sequence U can be applied 3 u1 = 2 0.9961x11 x21 0.3292
in an open-loop control fashion, by applying sequentially each of
the controls ut = t (xt ) at each time instant t, although this will
0.7326 0.0861 0.0609
[ ] [ ]
result in the loss of the prediction and feedback properties of the A= , B= , N =2
0.1722 0.9909 0.0064
MPC controller.
1.8588 1.2899
[ ] [ ]
Computational complexity: An analysis of the computational 1 0
Q = , P = , R = 0.01
complexity of the proposed procedure and a comparison with the 0 1 1.2899 6.7864
methods presented in Bemporad et al. (2002) and Pistikopoulos
et al. (2002), which solve (4) as a single mp-QP problem, is provided
X = R2 , U = {u R | 2 u 2}. (29)
next. It was shown in Pistikopoulos et al. (2002) and Pistikopoulos Algorithm 1 is applied to solve the above problem and each of the
et al. (2007a) that, if (4) is solved as a single multi-stage mp-QP steps is described below in detail. The algorithm starts at stage
problem, then an upper bound to the number CR of critical regions N 1 = 1.
of the explicit solution to (4) is given by
Stage 1: In Step 3 of Algorithm 1, the stage optimization problem
1
(13), i.e.
mpQP
CRr k!N k (mM + ng )k ,
V1 ( x 1 ) = min {xT1 Qx1 + uT1 Ru1 + xT2 Px2 }
k=0 u1 U

where s.t. x2 = Ax1 + Bu1 , 2 u 1 2 ,

N (mM + ng )!
N m
is solved as an mp-QP problem, with u1 being the optimization
= . variable and x1 being the parameter. The mp-QP algorithm of
i=0
(N mM + N ng i)!i!
Pistikopoulos et al. (2002) is used to derive the explicit solution
Note that this upper bound on the number of critical regions u1 = 1 (x1 ), which is given in Table 2 and Fig. 1. Since u1 is
depends on the prediction horizon N. This implies that the number already a function of x1 , there is no need to implement Step 4 of
of critical regions may increase significantly if a large prediction the algorithm.
horizon is used. Stage 0: Step 3 of Algorithm 1 is applied and the problem
Following similar arguments in this proposed approach, we can
derive an upper bound for each stage problem (14). Since, in (14), V0 (x0 ) = min {xT0 Qx0 + uT0 Ru0 + xT1 Qx1 + uT1 Ru1 + xT2 Px2 } (30)
the number of optimization variables for each stage optimization u0 U

(13) is m, the number of parameters is n + (N t 1)m, and the s.t. x2 = Ax1 + Bu1 , x1 = Ax0 + Bu0
number of linear inequality constraints is mM + ng , an upper bound 2 u 0 2
to the number of critical regions for (14) is given by
is solved as an mp-QP problem, with u0 being the optimization
1
m
(mM + ng )! variable and x0 , u1 the parameters. The explicit solution of (30)
DP
k!(mM + ng ) k
where = . is a PWA function u0 = f0 (x0 , u1 ) of x0 , u1 that consists of three
ineq
k=0 i=0
(mM + ng i)!i!
critical regions; due to space limitations, we present here only
The above relations shows that the number of critical regions for the mathematical representation of the second critical region and
stage-wise optimizations (13) does not depend on the prediction corresponding control law:
horizon N, and hence the complexity of the stage-wise problems
remains the same even if N increases. This implies that, even if u0 = 7.104x10 7.131x20 0.3494u1 ,
the multi-stage problem (4) becomes harder or even intractable to if x0 , u1 CR20 = {0.9962x10 x20 0.049u1
solve for large N, solving each smaller problem (13) will still remain
a much more tractable procedure. 0.2805, +0.9962x10 + x20 + 0.049u1 0.2805}. (31)
On the other hand, the number of parameters at each stage Then, in Step 4 of Algorithm 1, the system model x1 = Ax0 + Bu0
increases since the number of future inputs considered in each is incorporated in the expression given in Table 2 to obtain u1 =
stage increases. Hence, the number of parameters that have to be f1 (x0 , u0 ). The new expression u1 = f1 (x0 , u0 ) is incorporated
eliminated in (15) and (23)(26) increases, in order to derive the in the piecewise affine expressions of (31) to form equations
explicit solution (27), thus affecting the computational complexity (18)(19), and u1 is eliminated from the expressions of u0 by direct
of the algebraic computations in (23)(26) that are needed to substitution, to obtain the final solution u0 = 0 (x0 ). This final
derive the explicit control (27). Nevertheless, the operations solution is given in Table 3 and Fig. 2. Since only the first control
performed at this step of the proposed algorithm rely only on will be applied to the system, the solution to the explicit/multi-
linear operations such as FourierMotzkin or Gauss elimination parametric MPC problem in this example is the explicit control law
(see Schrijver, 1986). u0 = 0 (x0 ). Note that the solution u0 = 0 (x0 ), obtained with
the proposed algorithm, is the same as the solution obtained with
4. Illustrative examples the method of Bemporad et al. (2002), where the explicit/multi-
parametric MPC problem is solved as a single mp-QP problem.
4.1. Example 1 Finally, Table 4 presents the execution times and number of
parameters for an indicative number of stages of Algorithm 1, when
The proposed methodology is first illustrated for the following the horizon for problem (29) is N = 50. The total execution time
example, (taken from Bemporad et al., 2002): for Algorithm 1 is 433.2 s. Note that the number of parameters for
1644 K.I. Kouramas et al. / Automatica 47 (2011) 16381645

Table 3
Explicit solution u0 = 0 (x0 ) for stage 0Example 1.
Control law Critical region

1 u0 = 2 0.9961x10 + x20 0.5119

0.9961x10 x20 0.5119

2 u0 = 2
+0.9961x10 + x20 0.3292
0.9961x10 x20 0.3292
3 u0 = 6.051x10 6.074x20
+0.9961x10 + x20 0.3292
0.9961x10 x20 0.3292
4 u0 = 2
+0.9961x10 + x20 0.5119
5 u0 = 2 0.9961x10 x20 0.5119

Fig. 3. Critical regions for Example 2.

x1
-5
-10
0 5 10 15 20 25 30 35 40 45 50
10

x2
5
0
0 5 10 15 20 25 30 35 40 45 50

0
x3 -1
-2
0 5 10 15 20 25 30 35 40 45 50
1
x4

0.5
Fig. 1. Critical regions for the solution u1 = 1 (x1 ) at stage 1Example 1.
0
0 5 10 15 20 25 30 35 40 45 50

1
u1

0.5
0
0 5 10 15 20 25 30 35 40 45 50
1
0.5
u2

0
0 5 10 15 20 25 30 35 40 45 50
Time Instants

Fig. 4. State and input time profiles for the system simulation starting at the initial
state x0 = [10 10 0 0]T Example 2.

0.40 0.0227

0.012 0.3055
B= ,
0 0.2159
Fig. 2. Critical regions for the solution u0 = 0 (x0 ) at stage 0Example 1. 0.1437 0

Table 4 with state and input constraints

Execution times and number of parameters for each stage of Algorithm 1
Example 1.
X = {x R4 | 10 xi 10, i = 1, . . . , 4}
Stages 49 48 26 25 1 0 U = {u R2 | 1 ui 1, i = 1, 2}.
Execution time (s) 0.01 0.109 1.61 1.70 7.47 6.26 The explicit/multi-parametric MPC (4) is then formulated with
# of parameters 1 2 24 25 49 50 Q = I4 , R = 0.01 I2 , P = 10 I4 , and prediction horizon N = 3,
where In denotes the n n identity matrix. By applying Algorithm
each stage increases, and hence the execution time for each stage 1 to this problem, we obtain an explicit MPC controller which
increases (due to the increasing number of parameters that have consists of 661 critical regions and corresponding affine control
to be eliminated in each stage). Nevertheless, the execution time laws: note that the critical regions of the explicit MPC are given at
for each stage is not significant, and the algorithm can be extended stage 0 of Algorithm 1. A two-dimensional projection of the critical
to larger horizons. It is also expected that the execution time will regions on the space of states x1 and x2 is given in Fig. 3. Finally,
Fig. 4 shows the time profiles of the four system states and two
increase for large values of the system inputs and constraints.
system inputs.

4.2. Example 2
5. Concluding remarks
We now consider the linear, discrete-time system (1) with
In this work, we have presented a method for solving the
0.9233 0.1812 explicit/multi-parametric MPC problem based on dynamic pro-

0 0
0 0.9461 0 0.1492 gramming with multi-parametric programming. The method
A= (32)
0 0 0.8111 0 features a two-step algorithm which, based on DP methods, de-
0 0 0 0.8464 composes the explicit/multi-parametric MPC problem into a set
 K.I. Kouramas et al. / Automatica 47 (2011) 16381645 1645

of smaller stage optimization problems and solves each of these Pistikopoulos, E. N., Georgiadis, M., & Dua, V. (2007a). Process systems engineering
stage problems as a simple convex multi-parametric program- series: Vol. 2. Multi-parametric model-based control: theory and applications.
Weinheim: Wiley-VCH.
ming problem. The key advantage of the proposed algorithm is Pistikopoulos, E. N., Georgiadis, M., & Dua, V. (2007b). Process systems engineering
that it overcomes the limitation of previous methods for dynamic series: Vol. 1. Multi-parametric programming: theory, algorithms and applications.
programming by using multi-parametric programming techniques Weinheim: Wiley-VCH.
Rawlings, J. B., & Mayne, D. Q. (2009). Model predictive control: theory and design.
and avoids the need for global optimization at each stage of the DP Madison, Wisconsin: Nob Hill Publishing.
procedure. Extensions of the proposed procedure for the case of hy- Sakizlis, V., Dua, V., Kakalis, N., Perkins, J. D., & Pistikopoulos, E. N. (2001). The
brid systems (corresponding to mixed-integer programming prob- explicit control law for hybrid systems via parametric programming. In IFAC
workshop on manufacturing, modelling, management and control (pp. 128133).
lems), continuous-time systems (where the mp-MPC optimization Sakizlis, V., Kakalis, N. M. P., Dua, V., Perkins, J. D., & Perkins, E. N. (2004). Design of
becomes a multi-parametric dynamic optimization problem) and robust model-based controllers via parametric programming. Automatica, 40,
nonlinear system (involving multi-parametric nonlinear program- 189201.
Sakizlis, V., Perkins, J. D., & Pistikopoulos, E. N. (2005). Explicit solutions to op-
ming methods) pose significant challenges, due to the more com-
timal control problems for constrained continuous-time linear systems. IEE
plex optimization problems involved in mp-MPC problems. These Proceedings: Control Theory and Applications, 152(4), 443452.
extensions are part of an on-going research effort. Schrijver, A. (1986). Theory of linear and integer programming. John Wiley & Sons Ltd..
Tondel, P., Johansen, T. A., & Bemporad, A. (2003). An algorithm for multi-parametric
quadratic programming and explicit MPC solutions. Automatica, 39, 489497.
References

Bank, B., Guddat, J., Klatte, D., Kummer, B., & Tammer, K. (1982). Non-linear K.I. Kouramas is a Research Associate in the Centre for
parametric optimization. Birkhuser Verlag. Process Systems Engineering at the Department of Chem-
Bellman, R. (2003). Dynamic programming. Dover Publications. ical Engineering, Imperial College London. He obtained
Bemporad, A. (2003). Multiparametric nonlinear integer programming and explicit a diploma in Electrical Engineering and Computer Tech-
quantized optimal control. In Proceedings of the 42nd IEEE conference on decision nology from Patras University in 1997, and an M.Sc. in
and control, Hawaii, USA. Vol. 3 (pp. 31673172). Control Systems and a Ph.D. in Control Systems from the
Bemporad, A., Borrelli, F., & Morari, M. (2000). Piecewise linear optimal controllers Department of Electrical and Electronic Engineering, Im-
for hybrid systems. In Proceedings of American control conference, Chicago, USA. perial College London. His research interests include
Bemporad, A., Borrelli, F., & Morari, M. (2001). Robust Model predictive control: theory and computational algorithms for control and op-
piecewise linear explicit solution. In Proceedings of European control conference, timization, and their applications in process, energy, and
Porto, Portugal, September (pp. 939944). automotive/aeronautics systems.
Bemporad, A., Borrelli, F., & Morari, M. (2003). Minmax control of constrained
uncertain discrete-time linear systems. IEEE Transactions on Automatic Control,
48, 16001606. N.P. Fasca is a Senior Consultant at AspenTech. He
Bemporad, A., Morari, M., Dua, V., & Pisitkopoulos, E. N. (2002). The explicit linear obtained a Diploma in Chemical Engineering from the
quadratic regulator for constrained systems. Automatica, 38, 320. University of Coimbra, Portugal, in 2002, and his Ph.D.
Bertsekas, D. P. (2005). Dynamic programming and optimal control. Athena Scientific. in Chemical Engineering from Imperial College London in
Borrelli, F., Baoti, M., Bemporad, A., & Morari, M. (2005). Dynamic programming for 2008. His research interests are in the area of operational
constrained optimal control of discrete-time linear hybrid systems. Automatica, research and process systems engineering applications.
41, 17091721.
Dominguez, L. F., & Pistikopoulos, E. N. (2010). Multiparametric programming
based algorithms for pure integer and mixed-integer bilevel programming
problems. Computers & Chemical Engineering, 34, 20972106.
Fasca, N. P., Kouramas, K. I., Saraiva, P. M., Rstem, B., & Pistikopoulos, E. N.
(2008). A multi-parametric programming approach for constrained dynamic
programming problems. Optimization Letters, 2, 267280.
Johansen, A. (2002). On multi-parametric nonlinear programming and explicit
C. Panos graduated from the Aristotle University of Thes-
nonlinear model predictive control. In Proceedings of the 41st IEEE conference on
saloniki, Greece, in 2004, with a Diploma in Chemical En-
decision and control, Las Vegas, Nevada, USA.
gineering. He received his M.Sc. degree in Process Systems
Kojima, A., & Morari, M. (2004). Lq control for constrained continuous-time
Engineering from Imperial College London in 2005, and is
systems. Automatica, 40, 11431155.
currently working towards a Ph.D.degree in Chemical En-
Mayne, D. Q., Rakovc, S. V., Vinter, R. B., & Kerrigan, E. C. (2006). Characterization
gineering at Imperial Colege London. His research interests
of the solution to a constrained h-infty optimal control problem. Automatica,
include the development of algorithms and computational
42(3), 371382.
tools for optimal control, and mathematical modelling, op-
Muoz de la Pea, D., Alamo, T., Bemporad, A., & Camacho, E. F. (2004). A dynamic
timization, and control of energy systems.
programming approach for determining the explicit solution of linear MPC
controllers. In Proceedings of the 43rd IEEE conference on decision and control
2004, CDC04. Vol. 3 (pp. 24792484).
Pistikopoulos, E. N. (2000). On-line optimization via off-line optimization!A
guided tour to parametric programming. In AspenWorld 2000.
Pistikopoulos, E. N. (2009). Perspectives in multiparametric programming and E.N. Pistikopoulos, Director of the Centre for Process Sys-
explicit model predictive control. AIChE Journal, 55, 19181925. tems Engineering (CPSE) from 2002 to September 2009,
Pistikopoulos, E. N., Bozinis, N. A., & Dua, V. (2000). A MATLAB ( The Math Works, is a Professor of Chemical Engineering at Imperial College
Inc.) implementation of multi-parametric mixed-integer linear programming London. He obtained a Diploma from the Aristotle Univer-
algorithm. Centre for Process Systems Engineering, Imperial College. April. sity of Thessaloniki in 1984, a Ph.D. from Carnegie Mellon
Pistikopoulos, E. N., Dua, V., Bozinis, N. A., Bemporad, A., & Morari, M. (2002). University, in 1998, and was with Shell Chemicals (Ams-
On-line optimization via off-line parametric optimization tools. Computers & terdam) before joining Imperial in 1991. His research in-
Chemical Engineering, 26, 175185. terests focus on the development of theory, algorithms,
Pistikopoulos, E. N., Fasca, N. P., Kouramas, K., & Panos, C. (2009). Explicit robust and computational tools for multi-parametric program-
model predictive control. In Proceedings of the international symposium on ming and explicit model predictive control, and their ap-
advanced control of chemical processes, ADCHEM09. plications in biomedical, energy, and process systems.