ISSN: 2469-9062 (print), 2467-9240(e)

Journal of Institute of Science and Technology, 2015, 20(1): 119-122, © IOST, Tribhuvan University

Mathematical Modelling of Robust Optimization for Integer

Programming Problem
Gyan Bahadur Thapa
Central Campus, Institute of Engineering, Tribhuvan University
P. O. Box 19758, Kathmandu, Nepal
E-mail: thapagbt@ioe.edu.np

ABSTRACT
Dealing with data uncertainty in mathematical programming has been recognized as a central problem in optimization
for a long time. There are two methods that have been proposed to address data uncertainty over the years, namely
stochastic programming and robust optimization. In this short paper, we present the brief review of mathematical
models of integer programming problem and robust optimization approach to solve them. Robust Binary optimization
is also briefly presented.
Keywords: Linear programming, optimization, integer programming, robust optimization

INTRODUCTION
Mathematical programming is a broader concept
representing actions that can be taken in the systems
being modelled and then optimized. It maps each
possible set of decisions into a single score that assures
the quality of the solution as well. The optimization
problems initially originated as linear programs to solve
many practical problems that emerged during and after
the World War II such as military logistics, diet
problems and other problems in economical analysis.
The linear programming problem (LPP) contains an
objective function in a set of decision variables assigned
with a set of constraints. The LPP model can be
observed in different literatures (Thapa 2005, 2006,
Bertsimas & Tsitsiklis 1997). The LPP deals with
minimization or maximization of the value of the
objective function satisfying the given constraints. Each
constraint requires that a linear function of the decision
variables is either equal to or less than or greater than a
scalar value. And each decision variable must be
nonnegative.
x j ≥ 0, j = 1, 2, L , n
The coefficients cj (j=1, 2, … n) in the objective
function (1) are referred as cost coefficients in
minimization problems and profit coefficients in the
maximization problems. The right-hand-side values
b1, b2 , L , bm represent the amounts of available resources
(especially for ≤ constraints) and requirements
(especially for ≥ constraints). The coefficient values aij
means how much of resource or requirement i is
consumed or satisfied by decision j. This formulation is
a linear program because the objective function and all
the constraints are linear. The LPP formulation (1) can
be expressed in its standard form as follows:
(LPP) min c' x (2)
subject to Ax ≤ b
x≥0
m
where c ∈ R , b∈ R are given vectors and A ∈ R m×n is a
n

If the set of decision variables is x1 , x2 ,L, xn , general m × n matrix. If some of the decision variables are
form of LPP is formulated as the following: restricted to be integers, then the linear program is
Minimize or maximize c1 x1 + c2 x2 + L + cn xn (1) known to be a mixed-integer linear program. If the
variables take only 0 or 1, then the LPP is known to be
subject to 0-1 integer programs. And if all the decision variables
a11 x1 + a12 x2 + L + a1n xn ≤ (= or ≥ ) b1 are restricted to be integers, then the problem is known
to be integer programming problem (IPP). This problem
a21 x1 + a22 x2 + L + a2 n xn ≤ (= or ≥ ) b2 is one of the difficult problems to solve for exact and
L L L L L algorithmic solutions. It arises when a large number of
L L L L L discrete organizational decisions have to be made under
am1 x1 + am 2 x2 + L + amn xn ≤ (= or ≥ ) bm some constraints and optimization criteria. When all or

119
1
Mathematical Modelling of Robust Optimization for Integer Programming Problem

some of the data are uncertain, then the problem pij , zi , y j ≥ 0 ∀i, j ∈ J i
becomes more difficult. One usual approach to handle
this problem is stochastic programming which yields − yj ≤ xj ≤ yj , ∀j
some sort of probabilistic solutions which is beyond this xi ∈ Z + , ∀i
paper. Note that robust solution for LPP has been richly
carried out in Ben-Tal and Nemirovski (1999) and Ben- Proof: Given a vector x* , we define
Tal and Nemirovski (2000). Convex programming βi (x* ) = max ~ * which equivalently is
∑ aij x j
approach for inexact linear programming has been {si : si ⊆ J i , si ≤ Γi }
j∈s i
dealt in Soyster (1973). Here we explain about (6)
βi (x ) = max ∑ a~ij x*j zij
*
deterministically feasible solutions for integer programs j∈J i
considering and redefining the data uncertainty. such that
ROBUST INTEGER PROGRAMMING
∑z
j∈ J i
ij ≤ Γi

The new robust approach is extended to discrete zij ∈ {0, 1}, ∀i, j ∈ J i
optimization problems (Bertsimass & Sim 2004). The
Since the polyhedron  z : 0 ≤ z ≤ 1, j ∈ J , 
nominal mixed integer programming (MIP) is  ij i ∑z ij ≤Γi 
formulated as follows:  j∈ J i 
is integral for all Γi ∈ Z , so zij ∈ {0,1} can be replaced by
+
minimize c' x (3)
such that Ax ≤ b 0 ≤ z ij ≤ 1 . Now (6) becomes
xi ∈ Z + , i = 1, 2,L , k β i (x* ) = max ∑ aˆij x*j zij (7)
where n
c∈ Z , b∈ Z , A∈ Z m m× n
. j∈J i

such that
Without loss of generality, assume that A and c (not b ) ∑z ij ≤ Γi
are affected by data uncertainty.
j∈J i

0 ≤ z ij ≤ 1 , ∀ i , j ∈ J i
Each a → a~ ∈ [ a − aˆ , a + aˆ ],
ij ij ij j ∈ N = {1, 2,L, n}
ij ij ij

and each c j → c~j ∈ [ c j , c j + d j ]. The dual conversion of (7) is

β i (x * ) = min ∑ pij + Γi zi
Introduce Γi ∈ [0, J i ], J i = {j : aˆ ij > 0} which adjusts the j∈J i
(8)
robustness of the data against the level of conservatism. such that
If less than Γi coefficients aij are subject to change, then pij + zi ≥ aˆij x*j , ∀j ∈ J i

the robust solution will deterministically be feasible. pij ≥ 0, ∀j ∈ J i

The robust MIP is formulated as follows: zi ≥ 0, ∀i
minimize c ' x + (4)
max
{so : so ⊆ J o , so ≤ Γo }
∑d
j∈so
j xj Since (7) is feasible and bounded for all Γi ∈ [0, J i ], so
(8) is also feasible and bounded by strong duality with
such that the same objective values. Using the same argument in
ai x + max
{si : si ⊆ J i , si ≤ Γi }
∑ aˆ ij x j ≤ bi , ∀ i penalty term of objective function of robust MIP (4), we
j∈si get
xi ∈ Z + , i = 1, 2, L, k . β o (x * ) = max ∑ d j x *j
{s o : s o ⊆ J o , s o ≤ Γo }
j∈ s o
Theorem 1: (Bertsimas & Weismantel, 2005) The
⇒ β o (x *
) = max ∑ d j x*j zoj
Robust MIP (4) can be reformulated as an equivalent
j∈J o
MIP
such that
minimize c' x + zo Γo + ∑ poj (5) ∑z
j∈J o
oj ≤ Γo
j∈J o

such that 0 ≤ z oj ≤ 1, ∀j ∈ J o
ai x + zi Γi + ∑p ij ≤ bi , ∀i
j∈ J i ⇒ β (x* )= min p + Γ z
∑
zo + poj ≥ d j y j , ∀j ∈ J o
o oj o o
j∈J o

zi + pij ≥ aˆij y j , ∀i ≠ 0, j ∈ J i such that poj + z o ≥ d j x *j

120

2
 Gyan Bahadur Thapa

poj ≥ 0, ∀ j ∈ J o such that x ∈ X .

zo ≥ 0 Proof: Robust binary operation (10) can be written as
  (12)
Now, putting the values of β i (x * ) and βo (x* ) in (4), we Z R = min  c ' x + max ∑ d j x j u j 
x∈X
 j∈N 
get (5). This completes the theorem. such that
If more than Γi of aij are subject to change, then the
∑u
j∈N
j ≤Γ

robust solution will be feasible with very high u j ∈ {0, 1}, j ∈ N .

probability with probabilistic bound as follows: Since the polyhedron  
u : 0 ≤ u j ≤ 1, j ∈ N , ∑u j ≤ Γ  is
*
Theorem 2: Let x be optimal solution of the robust  j∈N 
MIP, then integral for all Γ ∈ Z + , so u j ∈ {0,1} can be replaced by
(i)  
P ∑ a~ij x*j > bi  ≤ B(n, Γi ) = 0 ≤ u j ≤ 1. Therefore (12) becomes
 j   
  Z R = min c' x + max ∑ d j x j u j 
1
(1 − µ ) ∑  n  + µ ∑  n  
n n
x∈X
n 
 j∈N 
2  l = v  l  =l v +1  l   such that
∑u j ≤Γ
where n = J i , v = Γi + 1 and µ = v − v  . This j∈N

2 0 ≤ u j ≤ 1, j ∈ N .
bound is tight.
Taking dual of inner maximization problem and
(ii) For Γi =θ n , lim B(n, Γ=
i) Φ(θ ),
n→∞ applying strong duality, we have
 
where Φ (θ ) =
1
∫
θ  y2 
exp  − dy is the Z R = min c ' x + min  Γ θ +
x∈ X
∑y j


2π −∞
 2   j∈ N 
cumulative distribution function of a standard normal. such that y j + θ ≥ d j x j
Note that the bound on (i) is best possible but difficult to y j , θ ≥ 0, j ∈ N
compute the sum of combination functions for large n.
 
The bound on (ii) is simple to compute and very tight. ⇒ Z R = min c' x + Γθ + ∑ y j  (13)
ROBUST BINARY OPTIMIZATION  j∈N 
The nominal case of binary optimization is formulated such that yj +θ ≥ d jxj
as x∈ X , yj, θ ≥ 0.
minimize c' x (9)
Clearly, an optimal solution (x* , y* , θ * ) of (13) satisfies
such that x ∈ {0, 1}
n

y *j = max (d j x *j − θ * , 0 ) .
where feasible set X is fixed, only the cost vector c is
Therefore (13) becomes
uncertain.
 
Each c j → c~ j ∈ [ c j , c j + d j ], j ∈ N , d j ≥ 0 . The robust Z R = min  Γ θ + c ' x + ∑ max (d j x j − θ , 0 )
x∈X , θ ≥ 0
 j∈N 
binary optimization is formulated as follows:

i.e., Z = min  Γθ + c ' x + max (d − θ , 0 )x  ,
Z R = min c' x + max ∑d x (10) R
x∈ X , θ ≥0 
∑ j j 
{s : s⊆ N , s ≤Γ}
j∈s
j j
 j∈N 
since X ⊆ {0,1}.
such that x ∈ X .
To find optimal value for θ , we decompose the set of
Without loss of generality, d1 ≥ d 2 ≥ L ≥ d n and positive real numbers R + in the intervals
dn+1 = 0. [0, d n ] , [d n , d n−1 ], L , [d 2 , d1 ], [d1 , ∞ ] and we get
Theorem 3: (Bertsimas & Sim 2003), the robust binary  ∑ (d j −θ )x j , if θ∈[d1 , dl −1 ]=
l −1
, l n +1, L, 2

max(d j − θ , 0)x j = 
 j =1
optimization problem can be solved by solving n + 1 ∑
nominal problems: 0 if θ ∈ [d1 , ∞ )
j∈N

Z R = min Gl (11) Therefore, Z R = min Z l (14)
l =1, 2 ,L, n +1 l =1, 2, L, n+1
 l 
where G= Γd l + min c' x + ∑ (d j − d l ) x j  where Z =  l −1 
l
 j =1  l min  Γθ + c' x +
x∈ X , θ ∈[d l , d l −1 ] 
∑ (d j − θ )x j 
 j =1 

121

3
Mathematical Modelling of Robust Optimization for Integer Programming Problem

Since we are optimizing a linear function of θ over the REFERENCES

intervals [d l , d l −1 ] , the optimal solution is obtained for Ben-Tal, A. and Nemirovski, A. 1999. Robust Solutions
to Uncertain Programs. Operations Research
θ = d l or θ = dl −1 , and thus for l = 1, 2,L, n + 1, we have Letters 25: 1-13.
  l −1   l −1 
Ben-Tal, A. and Nemirovski, A. 2000. Robust solutions
Z l = min Γd l + min c' x + ∑ (d j − d l )x j , Γd l −1 + min c' x + ∑ (d j − d l −1 )x j   of Linear Programming Problems
 x∈X x∈X 
  j =1   j =1 
Contaminated with Uncertain Data.
i.e.,

Mathematical Programming 88: 411-424.
 l   l −1 
Z l = min Γd l + min c' x + ∑ (d j − d l )x j , Γd l −1 + min c' x + ∑ (d j − d l −1 )x j   Bertsimas, D. and Sim, M. 2003. Robust discrete
 x∈X x∈X 
    
optimization and network flows. Mathematical
j = 1 j =1

Now (14) becomes Programming 98: 49–71.

  l 
Z R = min Γd1 + min c' x, LL, Γd l + min c' x + ∑ (d j − d l )x j  , L, min c' x + ∑ d j x j  
 

x∈X x∈X
 j =1
This completes the theorem.
For further detail of robust discrete optimization and
network flows, we refer to Bertsimas and Sim (2003)
and Bertsimas and Weismantel (2005).
CONCLUSIONS
Mathematical programming has a very wide and
important role in real life problems. There are two types
of optimization problems, namely continuous
optimization and discrete optimization. Data uncertainty
is obvious in any case of the optimization. Therefore the
only deterministic solution cannot address the data
uncertainty and sometimes the solution does not
guarantee the exactness. To address the uncertainty in
data, there are two approaches, the first is stochastic
programming and the other is robust programming. The
robust optimization defines the robustness of uncertain
data taking some intervals and handle the problems
deterministically. The general case of integer
programming is NP-hard problem and the existence of
exact solution is not possible. However, there is always
a possibility of finding near to optimal solution, for
example heuristic approaches.
 n 
Bertsimas, D. and Sim, M. 2004. The Price of

x∈X
 j =1  Robustness. Operations Research 52(1): 35-53.
Bertsimas, D and Weismantel, R. 2005. Optimization
Over Integers. Dynamic Ideas, Massachusetts,
USA. ISBN: 0-9759146-2-6.
Bertsimas, D and Tsitsiklis, J.N. 1997. Introduction to
Linear Optimization. Massachusetts Institute of
Technology, Athena Scientific, Belmont,
Massachusetts, USA.
Soyster, A.L. 1973. Convex Programming with set-
inclusive Constraints and Applications to
inexact linear programming. Operations
Research 21: 1154-1157.
Thapa, G.B. 2005. Basic Concepts on Integer
Optimization. Central Dept. of Mathematics,
Tribhuvan University, Nepal. The Proceedings
of The Two days Seminar on Application of
Mathematics 3: 33-43.
Thapa, G.B. 2006. Computational Complexity and
Integer Programming. Mathematical Sciences
and Applications 60-70, Kathmandu University,
Nepal.

122
4