Ind. Eng. Chem. Res.

2007, 46, 5141-5151 5141

A New Methodology for the General Multiparametric Mixed-Integer Linear

Programming (MILP) Problems
Zukui Li and Marianthi G. Ierapetritou*
Dept. of Chemical and Biochemical Engineering, Rutgers UniVersity, Piscataway, New Jersey 08854

In this paper, a general algorithm is developed to address the multiparametric mixed-integer linear programming
(mpMILP) problem with uncertain parameters in the left-hand side (LHS), right-hand side (RHS), and objective
function coefficients simultaneously. The algorithm is based on a general multiparametric linear programming
(mpLP) algorithm, which is derived using the optimality conditions of standard linear programming (LP)
problem. The intent of the proposed framework is to propose a general framework to address different
uncertainties in the process engineering problems. In addition, the paper also discusses the solution of the
special problem where LHS uncertainty is included in the optimization model as a coefficient of continuous
variable, and it notes the high computational complexity needed to retrieve the rigorous solution of large-
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

scale problems, because of the nonlinearities of the objective function and nonconvex critical regions. Several
numerical examples are presented to illustrate the effectiveness and applicability of the proposed method.

1. Introduction Early studies on multiparametric linear programming (mpLP)

Uncertainty appears in all the different levels of industry from include the work of Gal and Nedoma,10 which addressed the
Downloaded via ITESM on November 7, 2018 at 17:52:38 (UTC).

the detailed process description to multisite manufacturing. The right-hand side (RHS) mpLP problem based on the simplex
existence of parameter variability in most real process design algorithm for deterministic linear programming (LP) problems.
and operation problems necessitates the consideration of un- They proposed to cover the parameter space by critical regions,
certainties in the mathematical programming models used to where a critical region is defined to be the set of parameters
simulate and optimize the design performance and process such that a given basis is optimal for the parametric program.
operations. The process system engineering community has The proposed procedure is a dual pivoting method, which
proposed several ideas in the literature to address these process searches a connected component of the graph whose vertices
design and operations problems under uncertainty, which involve are the dual feasible bases of the problem and whose edges are
stochastic programming,1 fuzzy programming,2 robust optimiza- the dual pivots.
tion,3 and reactive scheduling;4 a thorough review can be found Recently, multiparametric programming has received more
in the work by Sahinidis.5 attention from the field of system theory and optimal control.
An alternative approach to process design/operations opti- A geometric algorithm for mpLP is proposed by Borrelli et al.,11
mization under uncertainty is either to utilize the information which is based on a recursive subdivision of the parameter space
extracted from sensitivity analysis6 or to use parametric around a critical region and requires that the critical region be
programming. Although both approaches provide input regard- convex and be described by a set of linear constraints. Filippi
ing the effect of uncertainty variability in optimal solution and and Romanin-Jacur12 proposed a lexicographic network pivoting
objective function value, parametric programming has a unique method under which degeneracy is virtually eliminated. How-
advantage, because it results in the full characterization of the ever, the characteristics of the method are not extendable to a
entire range of solution space and provides a complete map of general multiparametric linear program. Recently, Filippi13
the various alternatives in the face of uncertainty. developed an algorithm to obtain the approximate solution of
Based on the number of the uncertain parameters, two an mpLP problem. For a given full-dimensional simplex in the
categories can be distinguished: the single parametric program- parameter space and an optimizer for each simplex vertex, the
ming methods, which address the presence of a one uncertain algorithm formulates the linear interpolation of the given
parameter; and the multiparametric programming, which consid- solutions as an explicit function of the parameters and gives a
ers more than two uncertain parameters. Original studies on primal feasible approximation of an optimizer in the simplex.
parametric programming were initiated to address the single However, this method still only considers the RHS uncertainty.
parametric case. Motivated by the need for technology to implement con-
The first method for single parametric linear programming strained optimal feedback control with a minimal amount of
was published by Gass and Saaty,7 where parametric objective real-time computations, multiparametric quadratic programming
function is analyzed using simplex algorithm. A vast number (mpQP) problem is studied by Seron et al.,14 Bemporad et al.,15
of publications followed this initial work, a detailed review of Baotic,16 and Tøndel et al.17 All these algorithms are based on
which can be found in the book by Gal.8 Parametric integer an iterative procedure of generating new polyhedral regions in
programming problems that consider a single parameter variation the parameter space at each step, and they differ in the way
are also widely studied. Those methods include implicit they explore the parameter space. As an extension, the more-
enumeration methods, branch and bound (B&B) methods, and general multiparametric nonlinear programming (mpNLP) prob-
cutting plane methods. A detailed review can be found in the lem is also studied recently. Johansen18 used local mpQP to
work by Jenkins.9 approximate the original nonlinear programming (NLP) prob-
lem. Bemporad and Filippi19 proposed a recursive algorithm to
* To whom correspondence should be addressed. Tel.: 732-445- approximate the optimal value function in a given suboptimality
2971. Fax: 732-445-2421. E-mail address: [email protected]. threshold. Acevedo and Salgueiro20 studied the RHS multipara-
10.1021/ie070148s CCC: $37.00 © 2007 American Chemical Society
Published on Web 06/23/2007
5142 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007

metric convex nonlinear problems based on an outer-ap-

proximation algorithm. Hale et al.21 proposed a framework for
numerically approximating mpNLP solutions based on numer-
ical methods for single-parameter nonlinear programming.
Existing multiparametric mixed-integer linear programming
(mpMILP) methods are generally based on B&B methods and
the solution of mpLP subproblems at the B&B nodes. Such an
approach is developed by Acevedo and Pistikopoulos22 for the
analysis of linear process engineering problems under uncer-
tainty. Dua and Pistikopoulos23 proposed an alternative algo-
rithm by extending the idea of Pertsinidis et al.24 The problem
is decomposed into an mpLP and an MILP subproblem. When
Figure 1. Flow chart for multiparametric mixed-integer linear programming
the values of the binary variable are fixed, an mpLP is solved, (mpMILP). (From Jia and Ierapetritou.25)
and its solution provides a parametric upper bound to the
objective function value. When the parameters in the original respond to the LHS, RHS, and objective parameters, respec-
problem are treated as free variables, an MILP is solved, which tively. In this paper, we focus on developing a general algorithm
provides a new integer vector. The algorithm is composed of to address the mpMILP problem with different types of
an initialization step, and a recursion between the solution of uncertainty. Future publications will present the application of
an mpLP subproblem and an MILP subproblem. Jia and the algorithm in scheduling problems.
Ierapetriou25 recently proposed a systematic framework to solve In particular, we propose a methodology for the mpMILP
the mpMILP problem, which is used in this work and will be problems with uncertain parameters in the LHS, RHS, and
described in detail in the next section; they also developed a objective function coefficients simultaneously. The algorithm
new mpLP method to address RHS uncertain parameters, which is based on a B&B framework and a general mpLP algorithm,
solves many nonlinear programming (NLP) problems iteratively. which is derived using the optimality conditions of a standard
All these existing mpMILP approaches only involve the linear programming (LP) problem. The objective of the proposed
uncertain parameters in RHS of the MILP problems. Crema26 framework is to propose a general framework to address
studied the multiparametric 0-1 integer programming problem, different uncertainties in the process engineering problems. In
which considers the perturbation of the constraint matrix, the addition, the paper also discusses the solution of the special
objective, and the RHS vector, but it is designed for the 0-1 problem where LHS uncertainty is included in the optimization
integer programming and cannot address the general MILP model as a coefficient of continuous variable, and it notes the
problem. high computational complexity needed to retrieve the rigorous
In the direction of addressing more-general mixed-integer solution of large-scale problems, because of the nonlinearities
nonlinear problems, Dua et al.27 studied the multiparametric of the objective function and nonconvex critical regions.
mixed-integer quadratic programming (mpMIQP), using a The article is organized as following: the general mpMILP
sequential procedure iterating between a mpQP problem and a framework is presented in next section, followed by the mpLP
mixed-integer nonlinear programming (MINLP) problem until algorithm in section 3. A few numerical examples then are
convergence is achieved. Bemporad28 studied multiparametric presented to illustrate the proposed method. The paper is
nonlinear integer programming problems where the optimization summarized in the last section, including a discussion of future
variables belong to a finite set and where the cost function and challenges.
the constraints are linear functions of the uncertain parameters
and analyze the main theoretical properties of the optimizer and 2. mpMILP Framework
of the optimum as a function of the parameters.
The algorithm proposed in this paper is based on the mpMILP
Many process design and operation problems (e.g., plant
framework proposed by Jia and Ierapetritou,25 as shown in
synthesis, process scheduling) are commonly formulated as
Figure 1, which addresses the mpMILP problem with RHS
mixed-integer linear programming (MILP) problems. Thus, an
uncertain parameters. Here, we extend the methodology to cover
efficient mpMILP algorithm can address different uncertainties the general case of mpMILP problems and present the basic
emerging in process engineering problems. Many examples can principles and steps of the framework for completeness. A small
be found in the literature, including the work of Acevedo and example is also given in this section to illustrate the proposed
Pistikopoulos,22 who formulated the plant synthesis problem as framework.
an mpMILP problem with demand uncertainty. Pistikopoulos Let us consider the general form of mpMILP problem:
and Dua31 studied the uncertain process planning problem, which
is also formulated as a mpMILP problem. Ryu and Pistikopou-
min z ) c(θ)x (P1)
los32 studied the zero-wait batch process scheduling problem,
considering processing time uncertainty, and formulated the s.t. A(θ)x ) b(θ)
problem as mpMILP with RHS uncertainty after linearization.
In Jia and Ierapetritou,6 the mpMILP problem with RHS where x is the vector of decision variables that involve
uncertainty is generated from a scheduling problem and a continuous variables and integer variables; θ is the parameter
framework is proposed that is extended in this paper. As the vector that generally varies within a fixed region, which is
previous literature review illustrate, most existing mpLP/ described using a set of linear constraints or, most commonly,
mpMILP approaches only involve uncertain parameters in RHS a set of bounds. The objective of mpMILP is to determine the
and objective coefficients of the LP/MILP problems. However, effect of the parameter variation on the optimal solution and
realistic problems require incorporating the left-hand side (LHS) the optimal objective value.
uncertainty. For example, in the problem of process scheduling To address this problem, a general mpMILP algorithm is
under uncertainty, processing time, demand, and price cor- developed that follows these steps:
 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007 5143

Step 1. The original problem is first solved at an initial set

of uncertain parameters (θ0), following a B&B solution proce-
dure. The node that gets the optimal solution is flagged as the
current optimal node.
Step 2. The mpLP is solved at each of the leaf nodes, which
include infeasible nodes, because they are just infeasible for
current parameter value. The details of the solution of the mpLP
problem are given in the next section, because this is the main
step of the proposed algorithm. By performing mpLP at leaf
node p, several critical regions (CR(1) (2) (K)
p , CRp , ..., CRp ) are
identified, and in each critical region CRp , k ) 1, 2, ..., K
(k) Figure 2. Branch and bound (B&B) tree of example 1. (Throughout the
figures, the infeasible nodes are identified by the symbol “x”.)
(where K is the number of critical regions for this node), the
corresponding optimal function (z(k) p ) value is determined. The The global optimization solver should be used to solve
critical region is defined as the range of parameter values where problem P2 for the case that it is a nonconvex NLP problem.
the same basis remains optimal. If the leaf node is infeasible The final parametric solution at the uncertain range of interest
for all the parameter values, it is fathomed and will not be can be represented by several critical regions that contain their
branched later. The optimal function value of the current optimal corresponding optimal functions and integer solutions. The
node is denoted as z(k) opt and its critical regions is denoted as following example illustrates the mpMILP framework.
CR(k)
opt . Example 1. mpMILP illustration.
Step 3. The B&B tree is then updated. The main idea of this
step is to compare the optimal functions between the current max 2x1 + x3
optimal nodes and other leaf nodes in the intersection of their s.t. θx1 + 0.75x2 e 0.75
critical regions to identify a set of new critical regions, their
corresponding objective function values, and optimal integer θx3 + 0.75x4 e 0.75
values. To perform this comparison between the leaf node i
and the current optimal node (opt), the following redundancy -θx1 - 0.75x2 e -0.25
test problem (P2) is solved in the intersection between CR(k1) i -θx3 - 0.75x4 e -0.25
and CR(k2)opt .
x1 - x2 + x3 - x4 e 1
max err ) (( z(k1)
i - z(k2)
opt ) (P2) x1 + x 2 e 1
s.t. θ ∈ CR(k1)
i ∩ CR(k2)
opt x3 + x4 e 1

The outcome of this problem results in one of the following xi ∈ {0,1} (for 0 e θ e 1)
three cases.
Case I: err* ) max{zi(k1) - zopt(k2)} e 0 or err* ) max- The first step is to solve the problem using a B&B
{zopt(k2) - zi(k1)} e 0, where err* is the optimal Value of problem methodology at the nominal value of θ ) 0, as shown in Figure
P2. For this case, one optimal function value is not always 2, which gives five leaf nodes, including four infeasible ones.
greater than the function at the leaf node i at the intersection Based on the B&B tree in Figure 2, the next step is to solve
region. If z(k1) is also an integer node (all the integer variables the parametric programming problem at the five leaf nodes. The
i
get integer values), z(k1) is updated as the current optimal node parametric LP at the leaf nodes (using the approach described
i
in that region. If node i is not an integer node (only a subset of in the next section) leads to the following results.
the integer variables get integer values), the node should be At node 1, the result is
branched. If z(k2) (k1) (k2)
opt is not greater than zi , zopt remains the z ) 2.5 (for 0.25 eθ e0.75)
current optimal at this intersection region.
Case II: err* ) max {zi(k1) - zopt(k2)} > 0 and err* ) max- At nodes 2 and 4, the result is
{zopt(k2) - zi(k1)} > 0. For this case, we divide the intersection
region into two parts, using the constraint z(k1) ) z(k2) infeasible (for all 0 eθ e1)
i opt . In the
region{CR(k1) i ∩ CR(k2)
opt ∩ (zi
(k1)
e z(k2)
opt )} , zi
(k1)
is updated as At node 3, the result is
current optimal; whereas, in the other part {CR(k1) i ∩ CR(k2)opt ∩
(k1) (k2) (k2)
(zi g zopt )}, zopt remains the current optimal. z)1 (for 0.25 eθ e0.75)
Case III: Infeasible Problem. In this case, there is no
At node 5, the result is
intersection between the two critical regions CR(k1) i and CR(k2) opt ,
so z(k2)
opt remains the current optimal solution at CRopt . CRi
(k2) (k1)
is z)0 (for 0 eθ e1)
updated as the current optimal in any part of critical regions of
node i that have no intersection with the critical regions of Thus, after this step, the current optimal node is node 1, because
current optimal nodes. it has the maximum objective function value (2.5).
Step 4. The mpLP is performed at the new leaf nodes and The third step is then to compare the optimal function values
the comparison procedure continues. In every iteration step of between nodes in the intersection of critical regions. We can
the procedure, the new leaf nodes in the updated B&B tree will fathom node 2 and node 4, because they are infeasible. We
be compared to the current optimal nodes, to determine the new fathom node 3 because it is always less than node 1 in the
optimal functions in the intersection regions. This procedure intersection region of 0.25 e θ e 0.75. The comparison of the
stops when no further branching is required. function values between the current optimal node 1 and node 5
5144 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007

we start by deriving the optimality condition of the standard

LP problem:

min z ) cTx (P3)

s.t. Ax ) b
xg0
The constraints Ax ) b can be partitioned in two parts by the
basic and nonbasic variables:

Figure 3. Updated B&B tree of example 2.

x
( )
Ax ) (B N ) xB ) BxB + NxN ) b
N
(1)

gives the following result: In the range of 0.25 e θ e 0.75, where B forms a square matrix, xB is the vector of basic
node 1 must be branched and then compared with node 5 {x ) variables, and xN is the vector of nonbasic variables. If we fix
(0, 1, 0, 1), z ) 0, for 0.25 eθ e0.75}. xN to some values, we can calculate xB, provided that B is
To update the B&B tree in the range of 0.25 eθ e0.75, we invertible:
select θ ) 0.5 and solve the B&B as shown in Figure 3. (Note
that node 2 has been studied and is infeasible for 0 e θ e 1, so xB ) B-1(b - NxN) (2)
it is not necessary to resolve the entire B&B tree.)
Parametric programming at new leaf nodes is depicted as The nonbasic variables are normally fixed at one of their
follows (node 2 and node 4 are already fathomed). (noninfinite) bounds (for example, zero when the other bound
At nodes 1, 3, 7, and 8, the result is is infinite). We can partition the objective similarly and derive
the following equation:
infeasible (for all 0 e θ e 1)
z ) cTBxB + cTNxN ) cTBB-1b + (cTN - cTBB-1N)xN )
At node 6, the result is
constant + ∑
j∈N
σjxj (3)
z)1 (for 0.25 e θ e 0.75)
where σj ) cj - cTBB-1Aj represent the reduced costs. The
At node 9, the result is optimality condition for the LP problem (P3) can be derived as
follows:
z)2 (for 0.25 e θ e 0.75)
xB ) B-1b g 0, σj ) cj - cTBB-1Aj g 0 (4)
Comparing these results with {x ) (0, 1, 0, 1), z ) 0, for
0.25 e θ e 0.75}, we have node 9 as being current optimal. The optimal solution then is
Until now, no further branch has been needed, so the final result
is x ) (xB xN ) ) (B-1b 0 )
and the optimal objective is
x ) (0, 1, 0, 1), z ) 0
(for 0 e θ < 0.25 or 0.75 < θ e 1) z* ) cTBB-1b
x ) (1, 0, 0, 1), z ) 2 (for 0.25 e θ e 0.75)
Defining AB as the matrix composed by the columns corre-
sponding to the basic variables, and AN as the matrix corre-
In the framework proposed by Jia and Ierapetritou,25 a mpLP
sponding to the nonbasic variables, the following result is
approach is developed for the RHS uncertainty, which solves
obtained describing the optimality conditions and the optimal
many NLP problems iteratively. The solution of this optimiza-
objective function:
tion problem results in identification of a point in the uncertainty
space where the objective value cannot be represented by the
Optimality conditions: A-1
B bg0 (5)
current optimal functions. The new optimal function is then
determined and included in the next iteration. However, the cTN - cTBA-1
B AN g 0 (6)
method is limited to the cases where the optimal objectives are
linear functions of the uncertain parameters. As shown in the Optimal objective: z* ) cTBA-1
B b (7)
next section however, when RHS and objective function
uncertainties appear simultaneously, or when LHS uncertainty For the general mpLP problem
exists, the optimal value functions are not linear, which
necessitates the need for a new methodology (which is presented min z ) c(θ)x (P4)
in the next section). The new approach can address the general
s.t. A(θ)x ) b(θ)
mpLP problem where uncertainties in the LHS, RHS, and
objective function can appear simultaneously. xg0
the derived optimality conditions (eqs 5 and 6) still hold. A
3. General mpLP Method
critical region of the mpLP problem is the set of parameters
The main idea of the proposed mpLP method is based on such that a given basis is optimal for the parametric program.
the optimality conditions of the standard LP problem. Thus, Therefore, when the uncertain parameter θ is included in eqs 5
 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007 5145

Table 1. Projection Process To Identify the Feasible Region

polytope vertices, [x1 x2 θ1 θ2] projected vertice, [θ1 θ2] convex hull
-x1 - x2 + θ1 + 3θ2 e 0 [0 2500 1000 500] [1000 500] -θ1 - 1.27θ2 e 0
-2x1 + 3x2 - 8θ1 + θ2 e 0 [0 2614 1000 157.89] [1000 157.89] -θ1 - 11.25θ2 e 0
[0 1000 1000 0] [1000 0]
2x1 + 3x2 - 5θ1 - 18θ2 e 0 [0 1133 1000 -88.89] [1000 -88.89] -θ1 e 1000
x1 - 2x2 + 2θ1 - 3θ2 e 0 [0 0 0 0] [0 0] θ1 e 1000
-x1 e 0 [3000 0 0 1000] [0 1000] θ2 e 1000
[2500 0 -500 1000] [-500 1000]
-x2 e 0 [7000 3000 1000 1000] [1000 1000]
θ1 e 1000 [4000 5000 1000 1000] [1000 1000]
-θ1 e 1000 [1000 3000 1000 1000] [1000 1000]
[3000 1000 1000 1000] [1000 1000]
θ2 e 1000 [4400 0 -1000 800] [-1000 800]
-θ2 e 1000 [4489 64 -1000 787.2] [-1000 787.2]
[5857 428.6 -1000 1000] [-1000 1000]
[4500 0 -1000 1000] [-1000 1000]
[5500 666.7 -1000 1000] [-1000 1000]
[5000 0 -1000 1000] [-1000 1000]

and 6, the optimality conditions define the critical region of The following is an example of identifying the feasible region
the mpLP problem and the optimal objective described by eq 7 for an mpLP problem:
provides the optimal objective function for the mpLP. Based
on this idea, we develop the general mpLP algorithm, the max z ) x1 + 2x2
objective of which is to explore the entire feasible region to
identify all the critical regions. s.t. x1 + x2 g θ1 + 3θ2
The mpLP algorithm is composed of the following steps: -2x1 + 3x2 e8θ1 - θ2
Step 1. Fix the parameter values and solve the relaxed LP 2x1 + 3x2 e5θ1 + 18θ2
problem. Determine the current basis from the solution with
the idea that nonbasic variables must become zero (their lower x1 - 2x2 e -2θ1 + 3θ2
bound). For the initial step, the uncertain parameters can be
x1 g 0, x2 g 0
fixed at the nominal value.
Step 2. Identify the critical region from the optimality
To form a polytope that describes the feasible region based
conditions described by eqs 5 and 6, and get the optimal value
on the original constraints of the problem, we must add some
function using eq 7.
additional bounds on the uncertain parameters, that are large
Step 3. Generate a new point outside the identified critical enough so that they will not affect the final result. For this
regions and go to step 1. This is achieved by generating evenly example, the following constraints are added: -1000 e θ1 e
distributed points in the feasible region and testing whether they 1000, -1000 e θ2 e 1000. In addition, the feasible region is
have already been covered by the existing critical regions. obtained as shown in Table 1.
For the case when LHS uncertainty is included, because the The following example is used to illustrate, step by step, the
feasible region cannot be retrieved, the evenly distributed points proposed mpLP algorithm.
are generated in the original boundary. Example 2. mpLP illustration.
For the case without LHS uncertainty, the evenly distributed
min z ) θ1x1 + x2
points are generated in the feasible region, which can be obtained
using the following method. Because the critical regions are s.t. -x1 + x2 + x3 ) θ2
formed by linear constraints characterized by eqs 5 and 6, the
entire feasible region of the original problem can be determined x1 - θ3x2 + x4 ) 1
using a projection method. The feasible region for problem P4 xi g 0
is given by
-5 e θi e 5 (for i ) 1, 2, 3)
{θ|A(θ)x ) b(θ), x g 0}
This mpLP example involves three uncertain parameters, in
To get the feasible region of uncertain parameters, we can the RHS, LHS, and objective functions. The optimal solution
project the polytope formed by {θ|A(θ)x ) b(θ), x g 0} to the of the problem with θ ) (0, 0, 0) is x ) (1, 0, 1, 0), where x1
θ space, using the projection algorithm. Many effective projec- and x3 are basic variables; then,

[ ]
tion algorithms can be used; for example, a vertex enumeration/
convex hull-based projection algorithm that was proposed by -1 1
AB )
Avis and Fukuda,29 which first enumerates the vertices of the 1 0
polytope and then projects these vertices onto the sub-hyper-
plane. Finally, the projection is evaluated by calculating the
convex hull from these projected vertices. These computations
1
AN ) -θ
3
[ ] 0
1

are completed via the double description method proposed by cTN ) [1 0 ]

Fukuda and Prodon.30 In this way, the complete description of
the feasible region is determined. cTB ) [θ1 0 ]
5146 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007

Figure 4. Critical regions of example 2.

Using the optimality conditions described by eqs 5 and 6,

we get the following inequalities to describe the first critical
region {θ2 g -1, θ1 e 0, θ1θ3 g -1}. The corresponding
optimal function is derived from eq 7: z1 ) θ1. By generating
evenly distributed points in the boundary of parameters and
checking whether they are covered by the identified regions,
we get the final solution: Figure 5. B&B tree of example 3.

z1 ) θ1, CR1 ) {-5 e θ1 e 0, -1 e θ2 e 5, θ1θ3 g -1}

when LHS uncertainties appear, which, thus, leads to a much
z2 ) -θ1θ2, CR2 ) {0 e θ1 e 5, -1 e θ2 e 0} more complex problem in regard to exploring the parameter
space.
z3 ) 0, CR3 ) {0 e θ1 e5, 0 e θ2 e 5}
-[θ1(θ2θ3 + 1) - (θ2 + 1)] 4. Case Studies
z4 ) ,
θ3 - 1
In this section, we present several numerical examples to
CR4 ) {0 e θ1 e 5, -5 e θ2 e -1, 1 e θ3 e 5} illustrate the effectiveness of the proposed general mpMILP
algorithm. These examples involve parameter uncertainties in
It should be noticed from the result of example 2 that the the LHS, RHS, and objective functions simultaneously.
optimal objective function z4 takes the fractional nonlinear form, Example 3. ObjectiVe Uncertainty. This example involves
which is caused by the inverse of basis matrix, where uncertain uncertainty in the objective function coefficients that is modeled
parameters is included. We can predict that, for large-scale with the introduction of two uncertain parameters, θ1 and θ2:
problems with LHS uncertainty, great computational effort is
required to retrieve the rigorous solution of the problem.
min z ) (-2 + θ1)x1 + (-1 + θ2)x2 + y1 + y2
However, there is a special type of mpMILP with LHS
uncertainty that can be addressed efficiently, which is the case s.t. x1 + 3x2 - y1 e 9
when LHS uncertain parameters only appear as the coefficient
of integer variables, because, in the relaxed mpLP problem, the 2x1 + x2 - y2 e 8
LHS uncertainty is actually transformed to RHS. Also, from x1 - y1 + y2 e 4
Figure 4, we can see that one critical region is nonlinear and
actually nonconvex, which is due to the presence of LHS xi g 0, yi ∈ {0,1}, 0 e θi e 10, ∀ i ) 1, 2
uncertainty and is a problem that is not addressed in most mpLP
approaches in the literature. An intuitive method that addresses
the problem of exploring the parameter space is the geometric Step 1: The B&B solution. At θ ) (0, 0), the B&B method
method proposed by Borrelli et al.,10 which, however, is only results in the tree shown in Figure 5.
effective when the critical region is formed by linear constraints. Step 2: Solving the mpLP. At the leaf nodes, following the
Almost all the examples studied in the past involve RHS and general mpLP method proposed in the previous section, the
objective uncertainties, where the critical region is formed by following result is obtained. Let f1 ) θ1 - 2, f2 ) θ2 - 1, f3 )
linear constraints. However, the critical region can be nonconvex 3θ1 - θ2 - 5, and f4 ) θ1 - 2θ2:
 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007 5147

Figure 7. B&B tree of example 4.

test problem P2 is not solved, because the problem of comparing

the critical regions is trivial.
Example 4: LHS Uncertainty. This example involves uncer-
tainty in the constraint matrix coefficients, which is modeled
with the introduction of two uncertain parameters, θ1 and θ2:

min z ) -2x1 - x2 + y1 + y2
s.t. x1 + (3 + θ1)x2 - y1 e 9
Figure 6. Critical regions for node 4.
(2 + θ2)x1 + x2 - y2 e 8
At node 1, the mpLP result is
x1 - y 1 + y 2 e 4
1 ) 0, {f1 g 0, f2 g 0}; z1 ) 3θ2 - 3, {f2 e 0, f3 g 0}
z(1) (2)
xi g 0, yi ∈ {0,1}, 0 e θi e 10, ∀ i ) 1, 2
z(3)
1 ) -8 + 3θ1 + 2θ2, {f3 e 0, f4 g 0};
Step 1: The B&B solution. Performing the B&B method at
1 ) -8 + 4θ1, {f1 e 0, f4 e 0}
z(4) θ ) (0, 0) results in the B&B tree shown in Figure 7. The leaf
nodes for this problem are nodes 1, 2, 3, and 4. Node 1 has the
At node 2, the mpLP result is minimum objective (-8) and corresponds to the current optimal
node.
2 ) 1, {f1 g0, f2 g0}; z2 ) 3θ2 - 2, {f2 e0, f3 g0}
z(1) (2)
Step 2: Solving the mpLP. The mpLP is solved at the leaf
nodes.
2 ) -5 + 3θ1, {f1 e 0, f2 g 0};
z(3) At node 1, the mpLP result is
2 ) -7 + 3θ1 + 2θ2,{f2 e 0, f3 e 0}
z(4)
-40 - 16θ1 - 9θ2
At node 3, the mpLP result is z1 )
5 + 2θ1 + 3θ2 + θ1θ2
{0 e θ1 e 10, 0 e θ2 e 10}
3 ) 1, {f1 g 0, f2 g 0};
z(1)
3 ) -2.333 + 3.333θ2, {f2 e 0, f3 g 0}
z(2) At node 2, the mpLP result is

3 ) -7 + 4θ1, {f1 e 0, f4 e 0};

{ }
z(3) -21 - 5θ1 1 + θ1
2 )
z(1) 0 e θ1 e 10, 0 e θ2 e
3 ) -7 + 2.8θ1 + 2.4θ2, {f3 e 0, f4 g 0}
z(4) 3 + θ1 3 + θ1

At node 4, the mpLP result is -45 - 18θ1 - 9θ2

2 )
z(2) +1
5 + 2θ1 + 3θ2 + θ1θ2
4 ) 2, {f1 g 0, f2 g 0}; z4 ) 4θ1 - 6, {f1 e 0, f2 g 0}

{ }
z(1) (2)
1 + θ1
0 e θ1 e 10, e θ2 e 10
z(3) ) -1.333 + 3.333θ2, {f2 e 0, f3 g 0} 3 + θ1
4

4 ) 3.4θ1 + 2.2θ2 - 7; {f3 e 0, f4 g 0};

z(4) At node 3, the mpLP result is
z(5)
4 ) 4θ1 + θ2 - 7 {f2 e 0, f4 e 0} -40 - 16θ1 - 10θ2
z3 ) +1
(the critical regions are shown in Figure 6). 5 + 2θ1 + 3θ2 + θ1θ2
Step 3: Comparing the optimal functions. From the mpLP {0 e θ1 e 10, 0 e θ2 e 10}
results, there are five intersection regions between the critical
regions of the four nodes, which are the same as the critical At node 4, the mpLP result is
region of node 4, as shown in Figure 6. We must compare the
relevant optimal value functions in these regions. For example,
in CR(1)
4 of node 4, we must compare the optimal value
1 ) 0, z2 ) 1, z3 ) 1, and z4 ) 2; clearly, z1 )
functions z(1) (1) (1) (1) (1)
4 )
z(1)
-24 - 6θ1
3 + θ1 { 0 e θ1 e 10, 0 e θ2 e
θ1 - 3
12 + 4θ1 }
0 is optimal. Similarly, we see that node 1 is optimal in the -45 - 18θ1 - 10θ2
4 )
z(2) +2
other four regions. Thus, finally, the integer solution of the 5 + 2θ1 + 3θ2 + θ1θ2

{ }
problem is (y1,y2) ) (0, 0) and the critical regions and θ1 - 3
corresponding optimal value functions are the same as the mpLP 0 e θ1 e 10, e θ2 e 10
result of node 4. Note that, for this problem, the redundancy 12 + 4θ1
5148 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007

Figure 8. Intersection between critical regions of example 4. Figure 10. Critical regions of leaf node 2.

Step 2. Solving the mpLP. The mpLP are solved at the four
leaf nodes. The two infeasible nodes prove to be infeasible for
all 0 eθ e5 values, so they are fathomed. For node 1 and node
2, we introduce several functions to describe their critical
regions. Let f1 ) 4 - 1.5θ1 + θ2, f2 ) θ2 - 2, f3 ) 4θ1 - 3θ2
- 14, f4 ) θ1 + θ2 - 5, f5 ) 2θ1 - θ2 - 4, and f6 ) θ1 - 3;
the following mpLP results are obtained.
Figure 9. B&B tree of example 5.
For node 1, the mpLP result is
Step 3: Comparison. We can see that there are three
intersection regions between the critical regions of four leaf 1 ) (0.5θ1 + 6)(θ2 - 2) + 5
z(1) {f1 g 0, f2 e 0}
nodes, so we must compare the optimal value functions in these
regions. After solving several redundancy test with problem P2, 1 ) (-θ1 + θ2 + 10)(θ2 - 2) + 5
z(2) {f1 e 0, f2 e 0}
we found that node 1 is always the optimal node. Thus, the
integer solution is (0, 0) and the critical region and optimal value 1 ) 4θ2 - 3
z(3) {f2 g 0}
functions are the same as the mpLP result at node 1.
As discussed previously, because of the existence of LHS For node 2, the mpLP result is
uncertainty, the optimal value function takes a fractional form
that involves a quadratic term in the denominator. Some of the 2 ) (θ1 - 3)(1.333 + 0.333θ1) + (θ2 - 2)(5.333 +
z(1)
critical regions in the mpLP results also become nonconvex, as 0.333θ1) + 15 {f3 e 0, f4 g 0, f5 e 0}
shown in Figure 8.
Example 5: RHS and ObjectiVe Uncertainties. This example 2 ) (θ2 - 2)(0.5θ1 + 6) + 15
z(2) {f1 g 0, f2 e 0, f5 g 0}
involves uncertainty in both the objective function coefficients
2 ) (θ1 - 3)(3θ1 - 2θ2 - 8) + (θ2 - 2)(-θ1 + θ2 +
z(3)
and the RHS that is modeled with the introduction of two
uncertain parameters, θ1 and θ2: 10) + 15 {f1 e 0, f3 e 0, f5 g 0, f6 e 0}
min z ) (-3 + θ1)x1 + (-2 + θ2)x2 + 10y1 + 5y2
2 ) (θ2 - 2)(-θ1 + θ2 + 10) + 15
z(4)
s.t. x1 e 10 + θ1 + 2θ2 {f1 e 0, f2 e 0, f6 g 0}
x2 e 10 - θ1 + θ2
2 ) 4θ2 + 7
z(5) {f4 g 0, f2 g 0}
x1 + x2 e 20 - θ2
Step 3: Comparison. The critical regions of node 2 are plotted
x1 + 2x2 e 12 + θ1 in Figure 10. There are a total of six intersection regions between
the critical regions of node 1 and node 2. The redundancy test
x1 - 20y1 e 0
problem P2 is solved in these six regions. After comparison,
x2 - 20y2 e 0 the final solution of the mpMILP problem is

x1 - x2 e -4 (y1, y2) ) (0, 1)

-y1 - y2 e -1 z1 ) (0.5θ1 + 6)(θ2 - 2) + 5
x g0, y ∈ {0,1} ,0 eθ e5 {4 - 1.5θ1 + θ2 g 0, 0 e θ1 e 5, 0 e θ2 e 2}

Step 1: The B&B solution. Performing the B&B method at z2 ) (-θ1 + θ2 + 10)(θ2 - 2) + 5
θ ) (0, 0) results in the B&B tree shown in Figure 9. Node 1 {4 - 1.5θ1 + θ2 e 0, 0 e θ1 e 5, 0 e θ2 e 2}
is the current optimal node, and two leaf nodes get infeasible
solutions. z3 ) 4θ2 - 3 {0 e θ1 e 5, 2 e θ2 e 5}
 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007 5149

max  ) z(1)
1 - z2 >
(1)

s.t. f1 e 0, f2 e 0, f3 e 0, f5 e 0, f6 e 0 (for 0 e θ e 10)

The optimal objective is -58, so z(1) (1)

1 e z2 . Similarly, we
get
Figure 11. B&B tree of example 6.
1 e z2 in CR1 ∩ CR2
z(1) (2) (1) (2)

As we discussed in the previous section, because the RHS and

1 e z2 in CR1 ∩ CR2
z(2)
objective uncertainties appear simultaneously, the optimal value (1) (2) (1)
functions are nonlinear quadratic functions and the critical
regions are formed by linear constraints and, thus, form convex
1 e z2 in CR1 ∩ CR2
z(2) (2) (2) (2)
regions.
Example 6: LHS, RHS, and ObjectiVe Uncertainties. This
1 ∩ CR2 ) φ
CR(3) (1)
example involves the uncertainty in objective function coef-
ficients, RHS, and constraint matrix coefficients simultaneously
1 e z2 in CR1 ∩ CR2
that is modeled with the introduction of three uncertain z(3) (2) (3) (2)
parameters, θ1, θ2, and θ3.

1 e z2 in CR1 ∩ CR2
z(4) (1) (4) (1)
min z ) (-3 + θ1)x1 - 8x2 + 4y1 + 2y2
s.t. x1 + x2 e 13 + θ2
1 e z2 in CR1 ∩ CR2
z(4) (2) (4) (2)

(5 + θ3)x1 - 4x2 e 20
Therefore, the final integer solution of this problem is (y1, y2)
-8x1 + 22x2 e 121 ) (1, 1), and the critical regions and corresponding optimal
value functions are the same as the result of node 2, as shown
-4x1 - x2 e -8 in Figure 12.
x1 - 10y1 e 0 Example 7: LHS, RHS, and ObjectiVe Uncertainties. This
example also involves the uncertainty in objective function
x2 - 15y2 e 0 coefficients, RHS, and constraint matrix coefficients. Three
uncertain parameters (θ1, θ2, and θ3) appear in the LHS, RHS,
x g 0, y ∈ {0,1}, 0 e θ e 10 objective, respectively.
Step 1: The B&B Solution. The result from solving the max 4x1 + θ3x2 + x3
problem using the B&B method at θ ) (0,0,0) is shown in
Figure 11; node 1 is infeasible, and node 3 is the current optimal s.t. 3x1 + θ1x2 + x3 e θ2
node.
Step 2: Solving the mpLP. The mpLP at the leaf nodes can x1 - x2 - x3 e 0
be solved as follows. Let f1 ) θ1 - 5.909, f2 ) θ2 - 6.136, f3
) θ3 - {-78θ2 + 675}/{22θ2 + 165}, f4 ) θ3 - 0.655, f5 ) -x1 - x2 + x3 e 0
θ1 - 3, and f6 ) θ3 - 5. Node 1 is infeasible for all parameter
values. Therefore, it is fathomed. The results of nodes 2 and 3 -x1 - x2 - x3 e -1
are as follows.
At node 2 ((y1, y2) ) (1, 1)), the mpLP result is xi ∈ {0,1}, 4 e θ1 e 6

1 ) (-3 + θ1)(5.5 + 0.7333θ2) - 2.1333θ2 - 54

z(1) 3 e θ2 e 6, 3 e θ3 e 6
{f1 e 0, f2 e 0, f3 e 0}
Step 1: The B&B Solution. Solve the problem using the B&B
-4446 + 462θ1 - 484θ3 method at (θ1,θ2,θ3) ) (4,3,3). Because all the leaf nodes are
z(2) ) +6 {f1 e 0, f3 g 0}
1
11θ3 + 39 infeasible, as shown in Figure 13, we move the parameter to
(θ1,θ2,θ3) ) (4,6,3) and get the B&B tree shown in Figure 14.
1 ) 0.573θ1 - 41.385
z(3) {f1 g 0} Step 2: Solving the mpLP. At the leaf nodes, node 1, node
2, and node 5 are infeasible for all parameter values, so they
1 ) 10θ1 - 97.091
z(4) {f1 e 0, f2 g 0, f4 e 0} are fathomed. The mpLP result for node 3 is z ) θ3 (for θ1 -
θ2 e 0); the mpLP result for node 4 is z ) θ3 + 1 (for θ1 - θ2
At node 3 ((y1, y2) ) (1, 0)), the mpLP result is + 1 e 0); and for node 6, z ) 5 (for θ2 g 4). The current
optimal node is dependent on the value of parameter θ3.
20(θ1 - 3)
2 )
z(1) +4 {f5 e 0, f6 e 0} Step 3: Comparison. Because nodes 1, 2, and 5 are fathomed,
θ3 + 5 no further branching is needed. The next step is then to compare
the optimal value functions. The results between nodes 3 and 4
2 ) -2 + 2θ1
z(2) {f5 g 0, f6 e 0} are
Step 3: Comparison. To compare the optimal value functions θ1 - θ2 + 1 e 0, z ) θ3 + 1, (x1, x2, x3) ) (0,1,1)
in the intersection between CR(1) (1)
1 and CR2 , the following
redundancy test is solved, corresponding to problem P2: - 1 < θ1 - θ2 e 0, z ) θ3, (x1, x2, x3) ) (0,1,0)
5150 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007

general to address all different uncertainties and has been proved

to work well for small problems.
The mpLP method with RHS and objective uncertainty is
studied, mostly in the literature, because of the relative smaller
complexity. In these cases, the critical region is formed by linear
constraints, which can be directly predicted from the optimality
conditions of the LP problem; thus, it is also a convex region.
When LHS uncertainty is included in the mpLP problem, the
constraints of critical region become nonlinear and the region
becomes nonconvex. The optimal function also takes a fractional
nonlinear form, which brings more complexity for mpLP
problems in calculating the rigorous solution of the problem.
As illustrated in the paper, the proposed method can handle
LHS uncertainty for small mpMILP problems. For the special
case when the LHS uncertainty in the MILP formulation only
appears as coefficients of integer variables, the problem can be
Figure 12. Critical regions of example 6. transformed to one with RHS uncertainty and solved without
the complications of the LHS case, as shown with example 7.
However, for large-scale problems, the handling of LHS
uncertainty to get an exact solution of the mpMILP problem is
computationally complex, as shown with example 4.
Further improvement of the proposed method will focus on
more-effective ways to address degenerated LP problems, which
Figure 13. B&B tree of example 7 at (θ1,θ2,θ3) ) (4,3,3). can result in an increase in computational effort, because of
basis multiplicity. Among the work in the literature, most papers
have addressed uncertainty in process design and relatively less
attention has been devoted to the issue of uncertainty in process
planning and scheduling, mainly because of the increased
complexity of the deterministic problem. Therefore, another
direction on improving the proposed approach is to develop
computationally efficient numerical techniques for routinely
solving large-scale problems.

Figure 14. Updated B&B tree of example 7 at (θ1,θ2,θ3) ) (4,6,3). Acknowledgment

Table 2. Solution of Example 7 The authors gratefully acknowledge financial support from
z ) θ3 + 1 z2 ) 5 z ) θ3 z4 ) 5 the National Science Foundation (under Grant Nos. CTS

{ { { {
(x1,x2,x3) ) (0,1,1) (x1,x2,x3) ) (1,0,1) (x1,x2,x3) ) (0,1,0) (x1,x2,x3) ) (1,0,1) 0625515 and 0224745).
θ1 - θ2 + 1 e 0 θ1 - θ2 + 1 e 0 - 1 < θ1 - θ2 e 0 - 1 < θ1 - θ2 e 0
4 e θ1 e 6 4 e θ1 e 6 4 e θ1 e 6 4 e θ1 e 6
4 e θ2 e 6 4 e θ2 e 6 4 e θ2 e 6 4 e θ2 e 6 Literature Cited
4 e θ3 e 6 3 e θ3 e 4 5 e θ3 e 6 3 e θ3 e 5
(1) Ierapetritou, M. G.; Pistikopoulos, E. N. Batch plant design and
operations under un-certainty. Ind. Eng. Chem. Res. 1996, 35, 772.
Continuing the comparison of this result with node 6, we get (2) Balasubramanian, J.; Grossmann, I. E. Scheduling optimization under
the final result shown in Table 2. uncertainty - an alternative approach. Comput. Chem. Eng. 2003, 27, 469.
(3) Lin, X.; Janak, S. L.; Floudas, C. A. A new robust optimization
5. Summary and Future Work approach for scheduling under uncertaintysI. bounded uncertainty. Comput.
Chem. Eng. 2004, 28, 2109.
Motivated by the fact that, in real-world applications of (4) Honkomp, S. J.; Mockus, L.; Reklaitis, G. V. A framework for
schedule evaluation with processing uncertainty. Comput. Chem. Eng. 1999,
process design and operations, uncertainty is involved in most
23, 595.
parameters in model formulation, a multiparametric mixed- (5) Sahinidis, N. V. Optimization underuncertainty: State-of-the-art and
integer linear programming (mpMILP) method was proposed opportunities. Comput. Chem. Eng. 2004, 28, 971.
to examine the left-hand side (LHS), right-hand side (RHS), (6) Jia, Z.; Ierapetritou, M. G. Short-term Scheduling under Uncertainty
and objective function uncertainties based on a general multi- Using MILP Sensitivity Analysis. Ind. Eng. Chem. Res. 2004, 43, 3782.
parametric linear programming (mpLP) approach that was (7) Gass, S.; Saaty, T. The Computational Algorithm for the Parametric
developed based on standard linear programming (LP) opti- Objective Function. NaV. Res. Logist. Q. 1955, 2, 39.
(8) Gal, T. Postoptimal Analyses, Parametric Programming, and Related
mality conditions. The objective of this paper was to present Topics, 2nd Edition; de Gruyter: Berlin, Germany, 1995.
the algorithm such that it is only illustrated through a few (9) Jenkins, L. Parametric methods in integer linear programming. Ann.
numerical examples. Future publications will describe the Oper. Res. 1990, 27, 77.
application of the mpMILP approach for the solution of (10) Gal, T.; Nedoma, J. Multiparametric Linear Programming, Manage.
scheduling under uncertainty. Sci. 1972, 18, 406.
The framework proposed in this paper for general mpMILP (11) Borrelli, E.; Bemporad, A.; Morari, M. A geometric algorithm for
multi-parametric linear programming. J. Opt. Therm. Appl. 2003, 118, 515.
problems is based on a branch and bound (B&B) framework,
(12) Filippi, C.; Romanin-Jacur, G. Multiparametric Demand Linear
where, at the leaf nodes, the mpLP problem is solved using a Transportation Problem. Eur. J. Oper. Res. 2002, 139, 206.
new methodology that explores the primal and dual optimality (13) Filippi, C. An Algorithm for Approximate Multiparametric Linear
conditions for the deterministic LP problems. The approach is Programming. J. Opt. Therm. Appl. 2004, 120, 73.
 Ind. Eng. Chem. Res., Vol. 46, No. 15, 2007 5151

(14) Seron, M.; De Dona, J. A.; Goodwin, G. C. Global analytical model (25) Crema, A. The multiparametric 0-1 integer linear program-
predictive control with input constraints. In Proceedings of the IEEE ming problem: a unified approach. Eur. J. Oper. Res. 2002, 139,
Conference on Decision and Control, Sydney, Australia, 2000; p TuA05- 511.
2. (26) Jia, Z.; Ierapetritou, M. G. Uncertainty Analysis on the Right-Hand-
(15) Bemporad, A.; Morari, M.; Dua, V.; Pistikopoulos, E. N. The Side for MILP problems. AIChE J. 2006, 52, 2486.
explicit linear quadratic regulator for constrained systems. Automatica 2002, (27) Dua, V.; Bozinis, N. A.; Pistikopoulos, E. N. A multiparametric
38, 3. programming approach for mixed-integer quadratic engineering problems.
(16) Baotic, M. An Efficient Algorithm for Multiparametric Quadratic Comput. Chem. Eng. 2002, 26, 715.
Programming. Technical Report AUT02-04, Automatic Control Laboratory, (28) Bemporad, A. Multiparametric nonlinear integer programming and
ETH Zurich, Switzerland, February 2002. explicit quantized optimal control. In Proceedings of the IEEE Conference
(17) Tøndel, P.; Johansen, T. A.; Bemporad, A. An algorithm for multi- on Decision and Control, 2003; pp 3167-3172.
parametric quadratic programming and explicit MPC solutions, Automatica (29) Avis; Fukuda, K. A pivoting algorithm for convex hulls and vertex
2003, 39, 489. enumeration of arrangements and polyhedra. Discrete Comput. Geom. 1992,
(18) Johansen, T. A. On multi-parametric nonlinear programming and 8, 295.
explicit nonlinear model predictive control. In Proceedings of the 41st IEEE
(30) Fukuda, K.; Prodon, A. Double description method revisited. In
Conference on Decision and Control, Las Vegas, NV, 2002; pp 2768-
Combinatorics and Computer Science; Deza, M., Euler, R., Manoussakis,
2773.
I., Eds.; Lecture Notes in Computer Science 1120; Springer: Berlin, New
(19) Bemporad, A.; Filippi, C. An Algorithm for Approximate Multi-
York, 1996; p 91.
parametric Convex Programming. Comput. Optim. Appl. 2006, 35, 87.
(31) Pistikopoulos, E. N.; Dua, V. Planning under uncertainty; a
(20) Acevedo, J.; Salgueiro, M. An efficient algorithm for convex
parametric optimization approach. In Proceedings of the Third International
multiparametric nonlinear programming problems. Ind. Eng. Chem. Res.
Conference on Foundations of Computer Aided Operations (FOCAPO),
2003, 42, 5883.
Snowbird, UT, July 1998; Pekny, J. F., Blau, G. E., Eds.
(21) Hale, T. E.; Qin, S. J. Multi-parametric nonlinear programming
and the evaluation of implicit optimization model adequacy. In Proceedings (32) Ryu, J. H.; Pistikopoulos, E. N. A novel approach to scheduling of
of the 7th International Symposium on the Dynamics and Control of Process zero-wait batch processes under processing time variations. Comput. Chem.
Systems, Cambridge, MA, 2004. Eng. 2007, 31, 101.
(22) Acevedo, J.; Pistikopoulos, E. N. A multiparametric programming
approach for linear process engineering problems under uncertainty. Ind.
Eng. Chem. Res. 1997, 36, 717. ReceiVed for reView January 25, 2007
(23) Dua, V.; Pistikopoulos, E. N. Algorithms for the solution of ReVised manuscript receiVed April 24, 2007
multiparametric mixed integer linear optimization problems. Ann. Oper. Accepted May 22, 2007
Res. 2000, 99, 123.
(24) Pertsinidis, A.; Grossmann, I. E.; McRae, G. J. Parametric
optimization of MILP programs and a framework for the parametric
optimization of MINLPs. Comput. Chem. Eng. 1998, 22, S205. IE070148S