Annals of Operations Research 136, 21–34, 2005

c 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

An Improved IP Formulation for the Uncapacitated

Facility Location Problem: Capitalizing on Objective
Function Structure∗
ODED BERMAN
DMITRY KRASS
Joseph L. Rotman School of Management, University of Toronto, 105 St. George Street, Toronto, Ontario,
Canada M5S 3E6

Abstract. In this paper we examine the Uncapacitated Facility Location Problem (UFLP) with a special
structure of the objective function coefficients. For each customer the set of potential locations can be parti-
tioned into subsets such that the objective function coefficients in each are identical. This structure exists in
many applications, including the Maximum Cover Location Problem. For the problems possessing this struc-
ture, we develop a new integer programming formulation that has all the desirable properties of the standard
formulation, but with substantially smaller dimensionality, leading to significant improvement in computa-
tional times. While our formulation applies to any instance of the UFLP, the reduction in dimensionality
depends on the degree to which this special structure is present.

Keywords: Facility Location Models, integer programming, cover problems

1. Introduction

One of the most common and practically useful location models is the Uncapacitated
Facility Location Problem (UFLP). Since many of the solution techniques for this model
are based on integer programming, obtaining an integer programming formulation that
is as solvable as possible is of theoretical and practical significance.
The UFLP was analyzed extensively in the last three decades, see Krarup and Pruzan
(1983) and Cornuejols, Nemhauser, and Wolsey (1990). There are several alternative
formulations of the UFLP (e.g., Cornuejols, Fisher, and Nemhauser, 1977; Erlenkotter,
1978; Galvão, 1980; Galvão and Raggi, 1989). The version we follow in this paper
is similar to the one given by Galvão and Raggi (1989) as well as the one used by
Cornuejols, Fisher, and Nemhauser (1977). The objective of the problem is to locate (at
most) m facilities on a network so as to maximize the total net benefit (or minimize the
total net cost) of providing service to n customers located on the nodes of the network.
The benefit (profit) of providing service to customer i from location x is given by ci x . In
addition, a fixed cost f x may be incurred for locating a facility at node x.
∗
This work was supported by grants from NSERC.
22 BERMAN AND KRASS

The integer programming formulation for UFLP we discuss in this paper (which we
will refer to as the “standard” formulation) has many desirable properties, including a tight
upper bound on the “integrality gap”, i.e., the difference between the value of the integer
programming formulation and its LP relaxation (Cornuejols, Fisher, and Nemhauser,
1977). In addition, it has been observed that the LP-relaxation of this formulation very
often yields an all-integer optimal solution, or – when the optimal solution is fractional –
integrality can be achieved in just a few branch-and-bound steps. Revelle (1993) called
formulations with this property “integer friendly”. Thus, even though UFLP is an NP-hard
problem, it is not considered to be a particularly difficult integer program. We note that,
in practice, one often encounters UFLPs of very large dimensionality. Thus, even though
LP relaxation can often provide an excellent starting (or even exact) solution, solving
LP relaxation can be quite computationally taxing, in spite of significant advances in the
efficiency of the commercial software in recent years.
The UFLP we consider in this paper has the following special structure: for each
customer node i, the set of potential facility locations can be partitioned into several
subsets such that the benefit of serving customer i from any facility within each subset
is identical, that is ci x is constant for all x belonging to a particular subset – we call such
subsets “level sets”. We note that such “level set” structure is quite common in many
practical applications. For example, the classical Maximum Cover Location Problem (see
Church and ReVelle, 1974) belongs to this class – in fact, as discussed below, all potential
facility locations can be grouped into just two subsets in this case (for each customer node
i). Other problems possessing this structure are described in Section 2 below. In fact,
any instance of UFLP can be thought of as having this structure, however the number of
level sets may be large (as large as the number of potential facility locations, in the worst
case).
To take advantage of the level set structure we develop an alternative integer pro-
gramming formulation. This alternative formulation often has much smaller dimension-
ality than the standard formulation (the “compression” in dimensionality is defined more
precisely in Section 2 below).
In principle, a reduction in dimensionality of the formulation is often not desirable
in integer programming (unlike the situation in linear programming, where the reduc-
tion in dimensionality is always advantageous). The formulation with more variables
can be “tighter” than the more compact one – where “tightness” is defined in terms of
how closely the convex hull of the feasible integer solutions is approximated by the
feasible region of the LP-relaxation. In many problems, the formulation with higher
dimensionality turns out to be the one with the best computational performance. The
Traveling Salesman Problem is a famous example of this effect, where the “formulation
of choice” has exponential number of constraints, while several alternative formulations
with polynomial number of constraints perform very poorly.
Fortunately, this is not the case for the new formulation developed in the current pa-
per. In spite of being of lower dimensionality, it is shown to be just as tight as the standard
one (in fact, the optimal values of the LP-relaxations for the two problems are shown to be
identical). Thus, there is no “trade-off” involved in using the new formulation – it retains
IMPROVED IP FORMULATION FOR THE UNCAPACITATED FACILITY LOCATION PROBLEM 23

all of the desirable properties of the standard formulation, while having the advantage of
lower dimensionality. As a result, the new formulation may lead to significant computa-
tional advantages over the standard formulations (the degree of improvement depends on
how successfully the dimensionality of the original formulation has been “compressed”).
Indeed, in our own computational experiments, the new formulation typically improved
solution times by over 100 times.
The plan for the remainder of the paper is as follows. The UFLP model and the level
set structure considered in this paper are formally defined in Section 2, where a number
of potential applications of models with this structure are discussed. The standard and
alternative formulations are presented and analyzed in Section 3. A set of computational
experiments is presented in Section 4. Concluding remarks are presented in Section 5.

2. The UFLP with level set structure of objective function

coefficients – Definitions and applications

Let N = {1, . . . , n} be a set of customers (or customer locations) that require service
from facilities and let X = {1, . . . , X̄ } be a finite set of potential facility locations. We
assume that opening a facility at location x ∈ X incurs a fixed cost of f x (the sign of
f x is unrestricted – negative values can be viewed as “fixed revenue”). Let m ≥ 1 be the
maximum number of facilities that can be opened. Let ci x , i ∈ N , x ∈ X be the benefit
(profit) for serving customer i from the facility location x.
The Uncapacitated Facility Location Problem (UFLP) can now be stated as follows:

n


max h(S) = ci x − fx . (1)
S⊂X,|S|≤m
i=1 x∈S x∈S

Our problem of interest has the following special structure of the objective function
coefficients. For any customer location i ∈ N we assume that there are L i distinct values
j
of objective function coefficients ci x , for x ∈ X . For j = 1, . . . , L i , let ci be the jth
j
distinct value of ci x . It is convenient to assume that values ci are sorted in non-decreasing
j
order. Let X i ⊂ X j = 1, . . . , L i be defined by:

j  j
X i = x ∈ X ; ci x = ci .

j  i j
We call X i the jth level set for customer location i. Note that Lj=1 Xi = X .
In principle, no generality is lost by assuming the level set structure – indeed, if all
objective function coefficients are distinct and non-zero, then each level set consists of a
single element and L i = |X | for all i ∈ N . However, the new IP formulation developed in
Section 3 only provides computational advantages over the standard UFLP formulation
24 BERMAN AND KRASS

if for at least one customer location i, L i < |X |. Let


Li
d =1− i∈N
.
n|X |

We call d the “level set compression index”. Note that 0 ≤ d < 1. As will be discussed
in Section 3 below, d represents the approximate percentage of reduction in the number
of decision variables obtained by using the new formulation that capitalizes on the level
set structure. In the remainder of the current section we point out that high values of d
can be expected in several important special cases of the UFLP.

• Maximum Cover Location Problem (MCLP). In the MCLP customer location

i ∈ N is considered covered if there is at least one facility location within a given
radius R (see Church and Meadows, 1979; Church and ReVelle, 1974). It is assumed
that the total customer demand available at i is wi > 0. The objective is to locate m
(or at most m) facilities to maximize the total demand covered. The fixed costs are
normally ignored, i.e., f x = 0 for x ∈ X . Defining ci x = wi if x ∈ X and d(i, x) ≤ R
(where d(i, x) is some measure of distance between i and x), and ci x = 0 otherwise,
we observe that MCLP is a special case of UFLP see laso (Kolen and Tamir, 1990).
Since for any i ∈ N , there is only one non-zero value of ci x , we see that L i = 2.
Thus for MCLP the compression index d = |X|X|−2 |
, approaching 100% as the number
of potential facility locations grows.
• Generalized Maximum Coverage Location Problem (GMCLP). This generaliza-
tion of MCLP allows for several coverage levels – e.g., if there is a facility x “very
close” (i.e., within distance R1 ) to demand point i, the benefit of coverage is ci x = wi1 ;
if there is no “very close” facility, but there is one (located at x ∈ X ) that is “pretty
close” (within distance R2 > R1 ) the benefit of coverage is ci x = wi2 > wi1 , etc. This
type of coverage structure (typically with 2 or 3 levels of coverage) is commonly used
in estimating the market potential of retail facilities. This generalization of the MCLP
is examined in detail in Berman and Krass (2002). Let k be the number of coverage
radii, and assume that ci x = 0 for all locations x ∈ X outside the greatest coverage
radius Rk . Then the number of level sets L i = k + 1 for all i, and the compression
index is d = |X |−k−1
|X |
.
• Gradual Decay Coverage Model. This is another generalization of MCLP where
two coverage radii Rmin and Rmax > Rmin are specified, along with a non-increasing
“coverage decay function” f (s) for s ∈ [0, Rmax ] with f (s) = 1 for 0 ≤ s ≤ Rmin
and f (Rmax ) = 0. This function represents the proportion of total available demand
wi that is captured by a facility located at a distance s from i. Thus, a facility located
within the “inner radius” Rmin captures all of the available demand, and a facility
located outside the “outer radius” Rmax captures no demand. For further discussion
of this problem the reader can refer to Berman, Drezner, and Krass (2002). To see the
relationship to UFLP, set ci x = wi f i (d(i, x)), for i ∈ N , and x ∈ X .
IMPROVED IP FORMULATION FOR THE UNCAPACITATED FACILITY LOCATION PROBLEM 25

For a customer location i ∈ N , L i is bounded above by one plus the number of

facility locations in X at a distance between the outer and inner radii from i, depending
on the values of Rmin and Rmax , this number can be substantially smaller than |X |,
leading to high values of the compression index d.
• UFLP on a Grid Graph. In a “typical” application of the UFLP model, the objective
function coefficients ci x are inversely proportional to the distance between i ∈ N
and x ∈ X . An important special case occurs when the problem is solved on a grid
graph with Manhattan metric (i.e., l1 norm). Assume all the customer locations N
and potential facility locations X are contained in a rectangular region on a plane,
with sides a, b > 0, and that each grid element is a unit square. Then the number of
possible distinct values of the distances from any point i on the grid to any other point
is bounded by a + b. Thus, the compression index d ≤ |X |−a−b |X |
. If we assume that
any point on the grid can serve as a potential facility location, then |X | = ab, and d
approaches 100% as a, b → ∞.

3. Main results

We begin by stating the “standard” IP formulation for the UFLP defined by (1) above.
The formulation developed in Cornuejols, Fisher, and Nemhauser (1977) and Cornuejols,
Nemhauser, and Wolsey (1990) uses the following decision variables:

1 if facility is located at x ∈ X
Sx =
0 otherwise

and

1 if customer i ∈ N is served by the facility located at x ∈ X
Yi x =
0 otherwise

The formulation, which we will refer to as F1 can now be defined as follows:

n


max ci x Yi x − f x Sx (2)
i=1 x∈X x∈X
S.t.


Sx ≤ m (3)
x∈X
Sx ≥ Yi x , i ∈ N , x ∈ X (4)


Yi x ≤ 1, i ∈ N (5)
x∈X
Sx , Yi x ∈ {0, 1}, i ∈ N, x ∈ X (6)
26 BERMAN AND KRASS

Note that constraint (3) ensures that no more than m facilities are located, constraints
(4) ensure that a customer node can only be served from an open facility, while constraints
(5) guarantee that each customer is served by at most one facility.
To take advantage of the “level set” structure of the objective function coefficients
(as defined in the previous section) we now develop an alternative formulation. Let us
j
define a new set of binary decision variables Yi , i ∈ N , j = 1, . . . , L i as follows:
 j
j 1 if customer i is served by a facility in level set X i
Yi =
0 otherwise
Using this new set of decision variables, along with the location variables Sx , x ∈ X
defined earlier, we obtian the following alternative formulation, which we will call F2:


n

Li
j j


max ci Yi − f x Sx (7)
i=1 j=1 x∈X
S.t.


Sx ≤ m (8)
x∈X

j
Sx ≥ Yi i∈N j ∈ {1, 2, . . . , L i } (9)
j
x∈X i

Li
j
Yi = 1 i ∈ N (10)
j=1
j j
Sx , Yi = 0, 1 i ∈ N , x ∈ X i , j ∈ {1, . . . , L i } (11)
Here, constraints (9) specify that a customer can only be covered from level set j if
there is at least one open facility in this set, and constraints (10) ensure that a customer
is covered from some level set, the objective function will ensure that the customer is
covered at the best available level. Note that constraints (10) is an equality since the level
sets partition the potential facility set X , and thus any open facility must belong to some
level set.
Formulation F1 has n|X |+|X | binary decision variables and n|X |+n+1 constraints,
 the alternative formulation F2 has i∈N L i + |X | binary decision variables and
while
i∈N L i + n + 1 constraints. We summarize the key facts on the dimensionality of the
two formulations in the following Observation.

Observation 1. Let N V1 (N V2 ) be the number of binary decision variables in formulation

F1 (F2), and let N C1 (N C2 ) be the number of constraints in formulation F1 (F2). Then,
1. N V2 ≤ N V1 and N C2 ≤ N C1 . Moreover, the inequalities above are strict if L i < |X |
for some i ∈ N .
N V1 −N V2
2. The relative reduction in the number of decision variables N V1
is O(d), where d
is the compression index defined in Section 2 above.
IMPROVED IP FORMULATION FOR THE UNCAPACITATED FACILITY LOCATION PROBLEM 27

N C1 −N C2
3. The relative reduction in the number of constraints N C1
is O(d).

Part (1) above follows since L i ≤ |X | for all i ∈ N . Parts (2) and (3) follow
directly from the definition of the compression index d. We also note that formulation
F2 specialized to the MCLP problem described earlier yields the “standard” formulation
for the latter (see Church and ReVelle, 1974).
We now turn our attention to comparing the tightness of the two formulations for
UFLP presented above. Let LFP1 and LPF2 be the linear relaxations of F1 and F2,
respectively, obtained by dropping constraints (6) and (11). Let h LPF1 and h LPF2 be the
optimal values of the respective linear programs. We have the following key result.

Theorem 1. Formulations F1 and F2 are equally tight, in the sense that the optimal
values of their linear relaxations are identical, i.e.,

h LPF1 = h LPF2 .

The proof relies on two preliminary results. The first one establishes Theorem 1 in
one direction.

Lemma 1. h LPF1 ≤ h LPF2

Proof. Let (Sx , Yi x ), i ∈ N , x ∈ X be a feasible solution to LPF1. Define

j


Ỹi = Yi x for i ∈ N , j ∈ {1, . . . , L i }. (12)
j
x∈X i

j
Then, it is easy to check that (Sx , Ỹi ), i ∈ N , j = 1, . . . , L i , x ∈ X is a feasible
solution to LPF2. Indeed, constraints (3) and (8) are identical, constraint (9) is implied
by constraint (4) and constraint (10) is implied by constraint (5). Moreover, the objective
j
function values of (Sx , Yi x ) in LPF1 and (Sx , Ỹi ) in LPF2 are clearly identical. This
establishes the result.

Proving Theorem 1 in the opposite direction is more difficult. We first establish the
following lemma.
j
Lemma 2. Let (Sx , Yi ), i ∈ N , j = 1, . . . , L i , x ∈ X be a feasible solution to LPF2.
Then for each i ∈ N there exists at most one j(i) ∈ {1, . . . , L i } for which
j(i)


0 < Yi < Sx .
j(i)
x∈X i

j j 
For all j = j(i), j ∈ {1, . . . , L i } either Yi = 0, or 0 < Yi = j
x∈X i Sx and
constraint (9) holds as equality.
28 BERMAN AND KRASS

Proof. For a fixed location vector Sx , x ∈ X , the LPF2 model separates into a set of n
independent linear programs, one for each i ∈ N . For a particular i ∈ N , the resulting
linear program takes the following form:

Li
j j
max ci Yi
j=1
s.t.
j


Yi ≤ Sx , j = 1, . . . , L i (13)
j
x∈X i

Li
j
Yi ≤ 1 (14)
j=1
j
Yi ∈ [0, 1] j = 1, . . . , L i (15)
j
Recall that ci for j ∈ L i are assumed to be in non-decreasing order. It is clear that
j
the solution to the above LP is obtained by lexicographic maximization of Yi , starting
with j = L i . Thus, the optimal solution is given by the following procedure:

Step 0. Set T = 0, j = L i .

Step 1. Let UB = x∈X j Sx .
i

j j
Step 2. If UB < 1 − T , then set Yi = U B, T = T + Yi , j = j − 1 and repeat Step 1.
Else Go To Step 3.
j
Step 3. Set Yi = 1 − T , Yir = 0 for all r = 1, . . . , j − 1 and STOP.

This procedure keeps constraint (13) tight at all steps except for the last one, thus estab-
lishing the result.

We are now ready to complete the proof of Theorem 1.

Proof. [Theorem 1]
j
In view of Lemma 1, we only need to show that h LPF2 ≤ h LPF1 . Let (Sx , Yi ), i ∈ N , j =
1, . . . , L i , x ∈ X be a feasible solution to LPF2. We will construct a feasible solution to
LPF1 with the same objective function value.
Consider i ∈ N . From Lemma 2 we know that there exists at most one j(i) ∈
j
{1, . . . , L i } for which constraint (9) is not tight and the associated Yi > 0. For such j(i)
we have

j(i)


Li
j


Yi =1− Yi < Sx . (16)
j= j(i)+1 x∈X i
j(i)
IMPROVED IP FORMULATION FOR THE UNCAPACITATED FACILITY LOCATION PROBLEM 29

j j
For all other j ∈ {1, . . . , L i }, either Yi = 1 and constraint (9) is tight, or Yi = 0. Let
  j 
JiZ = j ∈ {1, . . . , L i }  Yi = 0 ,

and

j


JiT = j ∈ {1, . . . , L i } | 0 < Yi = Sx .
j
x∈X i

Note that if j ∈ {1, . . . , L i } and j = j(i), then j ∈ JiT ∪ JiZ .

j
Similarly, we define X iZ = {x ∈ X | x ∈ X i and j ∈ JiZ } and X iT = {x ∈ X | x ∈
j j j
X i and j ∈ JiT }. Note that if x belongs to some level set X i and x ∈ X iT , then X i ⊂ X iT .
j
Similarly, x ∈ X iZ implies that X i ⊂ X iZ (i.e., no level set can be split between X iT and
Z
X i ).
We are now ready to define vector Ỹi x that will be used to construct a feasible
solution for LPF1. For i ∈ N and x ∈ X , let


 Sx if x ∈ X iT


0 if x ∈ X iZ
Ỹi x =

 Yi
j(i)

 S  if x ∈ X i
j(i)
 x j(i) Sx
x∈X i

We first verify that (Sx , Ỹi x ), i ∈ N , x ∈ X is a feasible solution to LPF1. Constraint

j(i)
(3) holds since Sx satisfies (8) in LPF2. For x ∈ X i , constraint (4) holds trivially. For
j(i)
x ∈ X i this constraint holds since (16) and the definition of Ỹi x imply that Ỹi x < Sx .
Finally, to verify (5) we observe that if j(i) does not exist then



j
Ỹi x = Sx = Yi ≤ 1.
x∈X x∈X iT j∈J T

On the other hand, if j(i) exists, then

Yi
j(i)
j j(i)
Ỹi x = Sx + Sx  = Yi + Yi = 1,
x∈X x∈X i
j(i) Sx
x∈X iT j(i)
x∈X i j∈J T

where the last equality follows by (16).

It follows that (Sx , Ỹi x ), i ∈ N , x ∈ X is a feasible solution to LPF1. To complete
j
the proof we need to show that the objective function value of (Sx , Yi ), i ∈ N , x ∈
X, j ∈ {1, . . . , L i } in LPF2 is identical to that of (Sx , Ỹi x ), i ∈ N , x ∈ X in LPF1.
30 BERMAN AND KRASS


It is clear that the values of the fixed cost terms x∈X f x Sx are identical. We now
turn our attention to the first term in each objective function. Observe that for i ∈ N ,

Li


Li
j


ci x Ỹi x = ci x Ỹi x = ci Ỹi x (17)
x∈X j=1 x∈X j j=1 x∈X i
j
i

j j
As noted earlier, for every j ∈ {1, . . . , L i }, j = j(i), either X i ⊂ X iT or X i ⊂ X iZ . By
j   j
definition of X iT it follows that if X i ⊂ X iT then x∈X j Ỹi x = x∈X j Sx = Yi . Similarly
j  j i i
if X i ⊂ X iZ , then x∈X j Ỹi x = 0 = Yi . Finally, if j(i) exists, then
i

Yi
j(i)
j(i)
Ỹi x = Sx  = Yi .
j(i) j(i) x∈X i
j(i) Sx
x∈X i x∈X i

 j
We see that for any j ∈ {1, . . . , L i }, j
x∈X i Ỹi x = Yi . Substituting this into (17) we
obtain for all i ∈ N ,

Li
j j
ci x Ỹi x = ci Yi ,
x∈X j=1

which completes the proof.

The result above has several important consequences. First, it implies that the
“standard” formulation F1 does not have any advantages over the alternative formulation
F2 – it may be significantly larger, but no tighter than F2. In fact, Theorem 1 shows
that the projections of the polyhedra induced by the two formulations onto the space
spanned by the Sx variables are identical. Second, the equivalence between LPF1 and
LPF2 allows us to bound the “gap” between the optimal solutions of LPF2 and F2 by
applying a similar result for UFLP from Cornuejols, Fisher, and Nemhauser (1977).

Corollary 1. The worst case bound for the difference between the optimal values of F2
and its LP-relaxation LPF2 is given by

(h LPF2 − h ∗ )/ h LPF2 ≤ [(m − 1)/m]m < 1/e

.37,

where h ∗ is the optimal value of UFLP (i.e., F1 or F2). Moreover, the bound given above
is tight.

Proof. The bound follows from Theorem 2 in Cornuejols, Fisher, and Nemhauser (1977)
(which establishes the same bound for F1) and Theorem 1 above. The tightness of the
bound follows from Theorem 4 in Cornuejols, Fisher, and Nemhauser (1977), which
constructs a family of UFLP instances for which the bound above is tight.
IMPROVED IP FORMULATION FOR THE UNCAPACITATED FACILITY LOCATION PROBLEM 31

The “integer-friendly” nature of formulation F1 has been discussed above - its linear
relaxation often gives an all-integer solution, and even when the solution is fractional
the integrality can usually be restored with just a few branches of the Branch and Bound
method. The results above indicate that formulation F2 is just as “integer-friendly”, but
may be substantially easier to solve in view of its lower dimensionality. This facts are
borne out in computational experiments presented in the following section.

4. Computational results

The theoretical results derived in the previous section strongly suggest that formulation
F1 (and its linear relaxation LPF1) should hold significant computational advantages for
UFLP when the objective function coefficients can be grouped into a relatively small
number of level sets, as described in Section 2. To verify these results we have conducted
a number of computational experiments, which are described in the current section.
The controllable parameters in our experiments were: the number of nodes n, the
maximum number of facilities m, and the number of level sets for objective function
coefficients L = L i , i ∈ N (since we used the same value of L i for all nodes i ∈ N ,
hence we will drop the subscript i from this parameter for the remainder of the section).
For each value of n, m, and L, the random instances of UFLP were generated as follows.
We first generated a random n × n distance matrix with elements generated from a
uniform distribution on [0, 200]; we then applied the shortest path algorithm to convert it
to the shortest distance matrix for the network. Next, for each node i ∈ N , we generated
L random numbers from the uniform distribution on [0, 10]. These numbers were sorted
in non-increasing order and then assigned to the objective function coefficients cil , l =
1, . . . , L. Each node j was also used as a potential facility location (i.e., X = N ). Each
j
j ∈ N was randomly assigned to one of the L level sets for node i, with ci = cil if
j is assigned to level set l ∈ 1, . . . , L. The fixed costs were set to 0 for all j (i.e.,
f j = 0, j ∈ N ).
We used four settings for the number of nodes n = 100, 200, 300, 400, three
settings for the number of facilities m = 2, 5, .1 × n (the last setting was used to capture
the case of “many” facilities – i.e., one facility per 10 customer nodes), and four settings
for the number of levels L = 2, 3, 4, 5. For each combination of parameters, 5 different
problem instances were generated, yielding a total of 240 instances. For each instance
we formulated two LP relaxations, LPF1 and LPF2, and two IP formulations, F1 and F2.
All problems were solved using the CPLEX LP/IP solver on a Pentium V PC. We set a
time limit of 30 minutes for each problem.
For each problem instance, three performance measures were tracked: (1) solution
times (CPU seconds) for LPF1 and LPF2 (2) whether instance was solved successfully
within the 30 minute time limit by formulations F1 and F2, and (3) the IP solution times
for formulations F1 and F2 (the solution times were compared only for instances solved
to optimality by both formulations). The results for each of these performance measures
are discussed below.
32 BERMAN AND KRASS

• LP Solution Times. The detailed results are presented in Table 1. It can be seen that
while the solution times for LPF1 vary between 0 and 1800 seconds, those for LPF2
do not exceed 17.5 seconds. The reduction in solution times due to the improved
formulation LPF2 is very dramatic – over 100 times, on average. Chart 1 provides
more detail on the % improvement in the solution time due to LPF2 (computed
as 1 minus the ratio of LPF2 and LPF1 solutions times). It can be seen that the
comparative advantage of LPF2 increases with the number of nodes n and decreases
with the number of level sets L. Both of these results are in line with the theoretical
expectations outlined in Observation 1 above. It is interesting to note that the longest
solution times (for both LPF1 and LPF2) are observed when the number of facilities is
small (m = 2 or m = 5), while the problems with a large number of facilities appear
to be much simpler to solve. This is in line with the results observed by us elsewhere
for similar problems (e.g., see Berman and Krass, 2002).
• Percentage of IP’s Solved Successfully. The results are summarized on Chart 2. It is
clear that the improved formulation F2 provides a significant advantage in obtaining
the solution to the original UFLP model within the 30 minute time limit – the per-
centage of problems solved to optimality is significantly higher for F2 for all values
of n. The difference in success rates of F1 and F2 is even more dramatic when one
considers an important additional observation made during the computational exper-
iments: for instances with 200 or more nodes, the only cases where F1 was able to
solve the IP successfully were the ones where the LP relaxation yielded an all-integer
solution (it should be noted that whenever an all-integer solution was obtained by
LPF1, it was also obtained by LPF2). In contrast, F2 was often able to solve the IP
successfully even when the LP relaxation did not provide an all-integer solution – in
fact, the success rate was over 80% for problems with less than 400 nodes. It is inter-
esting to note that the LP relaxation yielded an all-integer solution for roughly 35%
of all instances. Also, as noted above, the problem becomes significantly easier when
the number of facilities m increases, for the “large” number of facilities, all instances
generated were successfully solved to optimality by both formulations; the percentage
was much lower for the 2-facility case.
• IP Solution Times. The difference in IP solution times for formulations F1 and
F2 was similar to the difference in LP solution times, with F2 providing 99% or
higher reductions, on average. However, the number of cases for which meaningful
comparisons could be made was quite small, since, as noted above, F1 failed to
solve any of the IP instances with over 100 nodes for which LP relaxation yielded
a fractional solution. For the 100 node instances, there were 23 instances for which
LPF1 yielded a fractional solution. For 8 of these instances, LPF2 yielded an all-
integer solution. For the 15 remaining instances (where IP solver actually had to
be invoked), the reduction in solution time for F2 ranged from 95% to 98%. It is
interesting to note that not only did the F2 formulation yield a large advantage for
instances where IP solver had to be invoked, but its LP relaxation LPF2 appears to be
IMPROVED IP FORMULATION FOR THE UNCAPACITATED FACILITY LOCATION PROBLEM 33

more “integer-friendly”, as it yielded all-integer solutions in a larger proportion of the

cases.

In summary, we can conclude that the computational results completely support the
theoretical results of Section 3: formulation F2 is far preferable to formulation F1 in UFLP
problems where the objective function coefficients possess the level-set structure. In fact,
the results strongly suggest that for such problems, using the improved formulation F2
may be the only way to obtain an optimal solution in a reasonable time.
We also refer the reader to Berman and Krass (2002) for the results of another
extensive series of computational experiments analyzing the quality of the bounds pro-
vided by LPF2 for GMCLP model with several coverage levels. In these experiments, as
in the ones presented above, LPF2 yielded all-integer solutions in all cases when such
solutions were obtained from LPF1. In addition, the solution times for LPF2 were sev-
eral orders of magnitude smaller than those for LPF1 – confirming the results presented
above.

5. Concluding remarks

In this paper we developed a new formulation for the UFLP which capitalizes on the
“level set” structure of the objective function coefficients. The new formulation retains
all of the attractive features of the “standard” formulation, but may have substantially
smaller dimensionality. In our computational experiments, the new formulation decreased
the solution times for UFLP by several orders of magnitude compared to the standard
formulation.
We note that while the computational advantage provided by the new formulation
depends on the degree to which the “level set” structure is present in a particular instance
of UFLP (as measured by the compression index d introduced in Section 2), its dimen-
sionality is bounded from above by that of the “standard” formulation. Thus, we would
recommend that the new formulation F2 should always be the formulation of choice for
UFLP – at worst, if each level set is a singleton, the formulation will be equivalent to the
“standard” formulation F1. On the other hand, if any “compression” is possible through
level sets, formulation F2 will automatically take advantage of the available structure.
We note that the fact that the two formulations are equally tight is somewhat sur-
prising. As shown in Lemma 1, the decision variables in F2 can be thought of as sums of
decision variables in F1. Under such circumstances, F2 would usually be regarded as the
“weaker” formulation, and the “common wisdom” in integer programming would lead
one to select the stronger formulation F1. However, Theorem 1 suggests that this logic
does not apply in our case.
It would be interesting to investigate whether the technique employed here in de-
veloping a new formulation for the UFLP model, by capitalizing on the structure of
the objective function coefficients, can be applied to other combinatorial optimization
models as well.
34 BERMAN AND KRASS

Finally, we note that there might be connections, which could be investigated,

between the procedure used in this paper and “aggregation” in linear programming, see
Zipkin (1980).

References

Berman, O. and D. Krass. (2002). “The Generalized Maximal Cover Location Problem.” Computers and
Operations Research 29, 563–591.
Berman, O., Z. Drezner, and D. Krass. (2003). “The Gradual Covering Decay Location Problem on Net-
works.” EJOR 151, 474–480.
Church, R.L. and M.E. Meadows. (1979). “Location Modeling Utilizing Maximum Service Distance Crite-
ria.” Geographical Analysis 11, 358–373.
Church, R.L. and C. ReVelle. (1974). “The Maximal Covering Location Problem.”Papers of Regional Science
Association 32, 101–118.
Cornuejols, G., M.L. Fisher, and G.L. Nemhauser. (1977). “Location of Bank Accounts to Optimize Float:
An Analytic Study of Exact and Approximate Algorithms.” Management Science 23(8), 789–810.
Cornuejols, M.L., G.L. Nemhauser, and L.A. Wolsey. (1990). “The Uncapacitated Facility Location Prob-
lem.” In R.L. Francis and P. Mirchandani (eds.), Discrete Location Theory New York: Wiley Interscience.
Erlenkotter, D. (1978). “A Dual-Based Procedure for Uncapacitated Facility Location.” Operations Research
26, 992–1009.
Galvão, R.D. (1980). “A Dual-Bounded Algorithm for the p-Median Problem.” Operations Research 28,
1112–1121.
Galvão, R.D. and L.A. Raggi. (1989). “A Method for Solving to Optimality Uncapacitated Location Prob-
lems.” Annals of Operations Research 18, 225–244.
Kolen, A. and A. Tamir. (1990) “Covering Problems.” In P. Mirchandani and R.L. Francis (eds.), Discrete
Location Theory. Wiley, pp. 263–304.
Krarup, J. and P.M. Pruzan. (1983). “The Simple Plant Location Problem: Survey and Synthesis.” Europena
Journal of Operational Research 12, 36–81.
Revelle, C. (1993). “Facility Siting and Integer-Friendly Programming.” European Journal of Operational
Research 65, 147–159.
Zipkin, P. (1980). “Bounds for Row-Aggregations in Linear Programming.” Operations Research 28, 903–
916.