View metadata, citation and similar papers at core.ac.

uk brought to you by CORE

provided by Calhoun, Institutional Archive of the Naval Postgraduate School

NPS55-87-007

NAVAL POSTGRADUATE SCHOOL

Monterey, California

BOOLEAN AND GRAPH THEORETIC FORMULATION OF

THE SIMPLE PLANT LOCATION PROBLEM

P. DEARING
M.
P. HAMMER
L.
B. SIMEONE

AUGUST 1987

Approved for public release; distribution is unlimited.

Prepared for:
Naval Postgraduate School
Monterey, CA 93943-5000

FedDocs
D 208.1M/2
NPS-55-87-007
 NAVAL POSTGRADUATE SCHOOL
MONTEREY, CALIFORNIA

Rear Admiral R. C. Austin D. A. Schrady

Superintendent Provost

Reproduction of all or part of this report is authorized.

This report was prepared by:

SECURITY CLASSIFICATION OF ThiS PAGE
""""
DUDLg M^OA LIBKARV
\'

REPORT DOCUMENTATION PAGE ^V>CposTGf^^E SCHOOL

la REPORT SECURITY CLASSIFICATION ID RESTRICTIVE MARK NG^ieN T EKhV UA 9394S-510<
UNCLASSIFIED
2a SECURITY CLASSIFICATION AUTHORITY 3 DISTRIBUTION AVAILABILITY OF REPORT
.

Approved for public release; distribution

2b DECLASSIFICATION /DOWNGRADING SCHEDULE is unlimited

4 PERFORMING ORGANIZATION REPORT NUMBER(S) 5 MONITORING ORGANIZATION REPORT IMUMBER(S)

NPS55-87-007
6a NAME OF PERFORMING ORGANIZATION 6b OFFICE SYMBOL 7a NAME OF MONITORING ORGANIZATION
Naval Postgraduate School (If applicable)

Code 55
6c. ADDRESS {City, State, and ZIP Code) 7b ADDRESS (C/ty. State, and ZIP Code)

Monterey, California 93943-5000

8a. NAME OF FUNDING SPONSORING-
8b OFFICE SYMBOL 9 PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER
ORGANIZATION {If applicable)

8c. ADDRESS (Dry, State, and ZIP Code) 10 SOURCE OF FUNDING NUMBERS
PROGRAM PROJECT TASK WORK UNIT
ELEMENT NO NO NO ACCESSION NO

11. TITLE (Include Security Classification)

BOOLEAN AND GRAPH THEORETIC FORMULATIONS OF THE SIMPLE PLANT LOCATION PROBLEM
12 PERSONAL AUTHOR(S) „ .. \
(Rutgers University)
.
, ,,

DEARING, P.M. , Hammer, P.L. (Rutgers University )

^SIMEONE , B.

13a. TYPE OF REPORT 3b T,ME COVERED DATE OfREPORZ (Year, Month, Day) 15 PAGE COUN
Technical FROM TO August 18
16. SUPPLEMENTARY NOTATION

17 COSATI CODES 18 SUBJECT TERMS (Continue on reverse if necessary and identify by block number)
FIELD GROUP SUB-GROUP location theory, pseudo-Boolean functions, set covering
vertex packing, graph theory, integer programming

19 ABSTRACT (Continue on reverse if necessary and identify by block number)

The simple plant location problem is formulated as the minimization of a pseudo-Boolean

function. This form of the problem is then transformed into a set covering problem and also
into a weighted vertex packing problem on a graph. These formulations are compared to simil
formulations in the literature and to the "standard" integer programming formulation.

20 DISTRIBUTION /AVAILABILITY OF ABSTRACT 21 ABSTRACT SECURITY CLASSIFICATION

Q UNCLASSIFIED/UNLIMITED D
SAME AS RPT Q one USERS Unclassified
22a NAME OF RESPONSIBLE INDIVIDUAL EPHONE (Include Area Code) 22c OFF (

Prof. P.M. Dearing 408) 646-2594 Code 55

DD FORM 1473, 84 mar 83 APR edition may oe used until exhausted SECURITY CLA
ither editions are obsolete ft U.S. Government Printing IBS 606
 BOOLEAN AND GRAPH THEORETIC FORMULATIONS OF THE
SIMPLE PLANT LOCATION PROBLEM

by

P. M. Dearing*
Clemson University
and Naval Postgraduate School

P. L. Hammer**
Rutgers University

and

B. Simeone
Universita' di Roma "La Sapienza"
and Rutgers University

Abstract

The simple plant location problem is formulated as the minimization of a pseudo-Boolean

function. This form of the problem is then transformed into a set covering problem and also into a

weighted vertex packing problem on a graph. These formulations are compared to similar

formulations in the literature and to the "standard" integer programming formulation.

*Partially supported by NSF Grant number ISP-801 1451 (EPSCoR) and the National Research
Council

**Partially supported by NSF Grant number DMS-83-05569

INTRODUCTION

In this paper the simple plant location problem, SPLP, is formulated as the minimization of

a pseudo-Boolean function. This formulation is then transformed into two other discrete

optimization problems: a set covering problem and a weighted vertex packing problem on a graph.

These three formulations of the SPLP are compared to similar formulations that have appeared in

the literature and the differences are discussed.

The SPLP is described as follows. Let P= {i: i = 1, .
. , p} be the index set of potential

locations for plants (or plant sites) in some space such as a network or the plane. If a plant is

opened at location i, a fixed cost fj is incurred. Let D= {j: j = 1, . . , d} be the index set of

customers. The unit cost of transportation between customer j and plant location i is given by cj;

for each i s P and j e D. Each customer has a demand which must be met by the opened plants
and is assumed to be one unit. If a customer's demand is different from one unit, it is scaled to one

and the transport costs are scaled accordingly. Assume further that the capacity for each opened

plant is sufficient to meet the demand of all customers (hence the alternative name "uncapacitated

plant location problem"). The SPLP is then to choose a subset of locations from P at which plants
are opened and to specify the transportation between opened plants and all customers so as to meet
customer demand and to minimize the total fixed cost plus the total transportation cost.

As an example, consider the network in Figure 1 where there is a customer at each node and

a plant site at each of the nodes 1, 3 and 4. Each edge number is the transportation cost accross that

edge. For each plant site i and customer j, the transportation cost Cjj is the minimum cost over all

paths between i and j. Table 1 gives the fixed cost fj and the transportation costs Cy for each i e P

and j€ D.

© i
l

Figure 1: A tree network with customer and plant sites at nodes

 customers

1 2 3 4 5 fi

1 2 1 4 3 5

plant sites 3 1 1 3 2 6

4 4 4 3 1 4

Table 1: Costs data for the SPLP of Figure 1

The approaches discussed here are not limited to a SPLP defined on a tree network, nor even to a

network. The approaches apply to any SPLP defined by a matrix of nonnegative Cj; values and

nonnegative fixed costs fj.

In the literature there are several recent surveys of location problems that include a
discussion of the SPLP. The extensive surveys by Cornuejols, Nemhauser and Wosley [3] and by
Krarup and Pruzan [8] are devoted almost entirely to the SPLP. Each includes a thorough
discussion of the problem's origins and several of its formulations. References [2,3] consider the

"uncapacitated facility location problem" which is equivalent to the SPLP, but maximizes the total

revenue from satisfying customer demands minus the cost of opening plants at the chosen sites.

The formulations discussed below may be applied directly to the uncapacitated facility location

problem with only minor modifications.

A BOOLEAN FORMULATION

We formulate the SPLP directly as the minimzation of a pseudo-Boolean function. This

formulation follows naturally from the well known property that in some optimal solution, each

customer j will receive its entire unit of demand from one open plant, namely a plant with minimum
transportation cost to customer j.

For each i e P, define the variable yj to be 1 if a plant is opened at location i, and

otherwise. Let yj = 1 - yj be the complement of yj. For each customer j, let j(.) be a

permutation of the location indices i e P so that the transportation costs from each j(i) to j are in

nondecreasing order:

Cj(l)j ^ c £ .<Cj (1)

j(2)j ( p)j.
• •

Observe that in meeting the demand of customer j, the cost c:(]\j is incurred iff a plant is
open at j(l), i.e., yw\\ = 1; or the cost Cj(2)j is incurred iff a plant is open at j(2) and no plant is

open at j(l), i.e., yjmyjO) = 1; or in general, the cost Cjaaj is incurred iff a plant is open

at j(k) and no plant is open at locations with smaller cost: j(l),j(2), .. ,j(k-l), i.e.,
yj(i)yj(2) • • •

= i- Finall y' tf n ° P lam is °P en i- e -» yj(i)yj(2) = *• tne demand of

yj(k-i)yj(k) > • • •
yj(p)
customer j is not met. This corresponds to an infeasible solution and is avoided by including in the

problem a penalty term M;yjMWj/2) • • •

yj(n) where M; is a large cost. Suitable values for the

M: are discussed subsequently.

To simplify the notation, let Fly;^) denote the product yi(i)yi(2) • • •

yj(k)
f°r eacn

j e D and k = 1, . .
, p. Furthermore, let riyj/Q\ = 1 in subsequent expressions. Then the

transportation cost to customer j from the set of open plants is given by

T =-
Xcj(0j
v vny v yv
'j(i-iyjO)
. .
1N ., + M.ny v ,. (2)
j pj *j(p) j

In the example problem, for customer 2, we have

T2 = ly
3
+ 2y 3yi + 4y iy 3y 4 + M2y 1 y3 y4 .

The total fixed cost incurred by the open plants is given by the sum of the f^yj. Then, a
pseudo-Boolean function F (see reference [7]) of the variables yj, y^, . ,y is defined as the total
p
.

fixed cost plus the sum of the T-'s for each j in D:

P dp
F(y
r • •
.
p
) -
gtyi + X Ic [
j(l)j
nyj(i . 1)yj(i) + Mj ny
j(p)
] .

The SPLP may be written as

PI: min { F( yi , . .
,y ) :
ft
= 0,1, ie P }.
p

A formulation equivalent to PI is reported by Hammer [6] and also included in the survey [8].

In a specific instance of the SPLP, the function F can generally be simplified by adding

together those terms with identical variables. Two customers j and j' will have terms with identical
variables, say ny:^. j Wj^) in Tj, and riyjv^. j WiVM m If, whenever the sequence of plant
sites, ordered by j(i) and j'(i), are identical for i = k, k+1, . . ,p for some k < p. This ability to
aggregate like terms is a desirable feature of the pseudo-Boolean function F that is not found in

most other formulations of the SPLP. In particular, since each penalty term ny^ identical for is

each customer j, all these may be added together to yield one penalty term. Let M denote the sum

of the M; over j e D. Then the function F for the example problem, simplified by adding together
like terms, is given as follows:

FCy1.y3.y4) = 5 yi + 7 y3 + 5 y4 + y iy3 + 3 y3yi + 5 W3 + ^3^4 + ^wi + Myi^y*

1

PENALTY VALUES

It is convenient to rewrite the function F, using the common penalty term M, as:

P dp
F(y = I + Ic ny + MRy
J
r . . .
, y ) f.
y .

j(i)j j(M) yj(i)

.

p j(p)

The following discussion shows how the value of F depends on the penalty term M. Observe that

a set of variable values (y j, . . .

, yp ) is feasible to PI if and only if FIyj^p\ = 0. Furthermore, if

(y 1' • • • . yp) *s an infeasible solution, then F(yj, . . . , yp) = M, the penalty value. Consider a

minimum feasible solution (y*\, . . .

,
y*
D ).
and let z* = F(y*^, . . .
,
y* ) denote the optimal

objective function value.

The first M < z*, each minimum solution to F infeasible.

claim is that with penalty value is

This follows since each infeasible solution has the objective function value M < z* < F(y^, . . .
,

y p for each feasible solution (yj,

) yp Alternatively, with a penalty cost of M > z* each . . . , ).

minimum solution to F has optimal value z*. This follows since for each feasible solution

(yi»
• • • . yp)' z * — F(yj, • • • . y p ). and for each infeasible solution (y'j, . . .
,y' ), z* < M=
F(y'j, . . . ,
y'
D ).
This same argument shows that with penalty cost M > z*, each minimum
solution to F is feasible.

Thus for all M > z*, the optimal objective function value of PI is z*, and any upper bound

on z* will suffice for a value of M. Values for the penalty coefficients M: may be obtained from
any feasible solution (y* j, . y* ). Define
p
. . ,

P
P^fyV and c^ = mm { c
j(i)j
ny*
j(M)
y*
j(i)
: i = 1, . . .
, p
1

},

that is, f* is the total fixed cost of this solution and c*; is the minimum transport cost to customer j

for this solution. Then each penalty cost M: may be chosen as any value greater than c*j + P/d.
A SET COVERING FORMULATION

Problem PI is next transformed into a set covering problem. This formulation has fewer

variables and constraints than a set covering formulation of the SPLP by Kolen [9], while it retains

the combinatorial properties of his formulation. It is also related to a covering formulation for the

uncapacitated facility location problem reported in reference [2] but obtained by a different

approach.

For each variable y,vj\ which appears in uncomplemented form in expression (2) substitute

y,vj\ = 1 - y;Q\ and simplify. Expresion (2) then becomes, for customer j,

p-1
t. = Cviv +
j(i)j
y(c 7 1+1
l1v - - c./. v)ny 7 .v + (m.
v - cv v )ny v v 0)
j
-ti
j( )j J(0r j(0 j j(p)r j(p)

In the example problem, for customer 2, we have

T'
2
= l + y 3 + 2yjy3 + (M 2 - 4)y l y 3 y 4 .

Then the function F has an equivalent form, denoted by F', which is the sum of the fixed costs plus

the sum of the T: for each j in D:

F'(y
1
.--.yp) = I
1=1
f
1 yi
+ IT
j=l

The function F' has a natural interpretation. For each customer j e D, the cost Cjq\j is a

constant that is incurred by any solution to the SPLP. That is, the minimum transport cost to

customer j, c;q\j, must always be incurred. This cost may be zero if a plant site and a customer

location coincide. Each term (ci(j

+ i \j
-
Cj(i)i) is the incremental transportation cost incurred by
customer j if plants at j(l), j(2), . . ,j(i) are not opened. In this case, the transportation cost to j

is at least Cj/j+i);, and if j(i+l) is open the cost is exactly Cj(j

+ mj.
As above, the function F' may be simplified by adding together like terms. For the

example problem this yields:

 F '(yi.y3.y4) = 2 + 5 yi + 6 y3 + 4 y4 + yi + 2 y?> + 4 y4 + 7 yiy3 + 2 y3V4 + (
M - 18 )yiy3y4 •

A further reduction is possible for F between pairs of complementary variables yj and yj

using the identity yj + yj = 1. Any two terms of the form ayj+ byj may be rewritten as follows:

ayj+ byj = { (a-b)yj + b, if a > b; or Oyj + (b-a)yj + a, if a <b } (4)

In the second case, an uncomplemented variable yj with zero cost may be given the value one, and

all terms containing the complementay variable yj will have value zero and may be eliminated from

F. For the example problem the reduction (4) yields:

F'(yi, v 3,y4) = 9 + 4yj + 4y 3 + 0y 4 + 7^3 + 2y 3 y 4 + (M - 18)y 1 y 3 y4.

Here y 4 is retained in the function F' in order to illustrate the next transformation.

Next, the SPLP is transformed by replacing each term ny:/j\ in F by a variable znj(i)'

where nj(i) = (j(l)j(2). . j(i)}, and by adjoining the constraints

z nj(i) ^ n yj(i) i e p, j
e d. (5)

But each constraint of (4) is equivalent, for zero-one variables and their complements, to the

constraint

^/j( k ) nj(i)

These transformations, applied to the general expression for F, yield the set covering

problem:

p d p-1 d
mm gfc+gl I c + M + C
( cj(i + l)j
J
c
P2:
" j( i)j) ^(i) <
j
"
j(p)PW J(DJ

S-t-
gyj(k )
+ ^j(i) ^ 1 - P, JE D

yi' z nj(i) = O' 1 iE p, je d

7
Problem P2 will have at most d(p-l) + p + 1 variables and d(p-l) + 1 constraints.

When F' is simplified by adding together like terms and by applying the reduction (4), the

number of variables zjjjq) and the number of constraints in P2 will be decreased accordingly. For

the example problem, with F' simplified and M' = (M - 18), problem P2 becomes:

min 9 + 4yj + 4y3 + 0y4 + Iz^ + 2Z34 + M'z^

s. t.
yi + y 3 +z 13 >1
y 3 + y4 + z 34 > 1

yi+ Y3 + Y4 +z 134 ^ l

all variables zero - one.

The above example suggests further simplifications that may apply to specific problems.

Any yj variable with a zero cost coefficient may be set to one in any optimal solution. The
corresponding column and all rows with a one in that column may be eliminated. In the example
problem, setting V4 = and eliminating columns and rows leaves one constraint: yj + y 3 + z^ 3 >1,
1

from which it is easily seen that the two alternative optimal solutions are to open plants at sites

{ 1,3} or { 1,4} with a cost of 13. Thus the set covering formulation may be interpreted as
preprocessing of the problem data that may specify some plant sites to be opened a priori .

The well known set covering 'reduction rules' [10] will generally apply to problem P2. For

example, since the column corresponding to a variable yj always dominates a column

corresponding to a variable zni(k) ^or * G nj(k)> the column of zni(k) can ^ e eliminated if the cost

coefficient of yj is not greater than that of zj-n^v In particular, since the coefficient of yj is less

than M, the variable zni(p) and its column can always be eliminated.
Row reductions may follow. If the variable zxii(k) nas ^>een eliminated from a row by a
column reduction, this row may now be dominated by some other row. In this case, the dominating
row may be eliminated.

Problem P2 is quite similar to Kolen's set covering formulation of the SPLP [9] with the

following differences. First, Kolen's formulation does not allow the simplifications we get from

adding together like terms. Each such simplification eliminates at least one variable and constraint.

His formulation has a penalty variable zni(p) an(i i ts corresponding constraint for every j in D,

8
whereas these terms are combined into one term and constraint in problem P2. These observations
imply the constraint matrix of problem P2 is a proper subset of the constraint matrix of Kolen's

formulation. The example problem is taken from Kolen [9]. His set covering formulation of the

example problem requires 17 constraints and 20 variables.

The set covering reduction rules also have more limited application to Kolen's formulation.

In general, these rules would eliminate the columns corresponding to the variables zni(p) ^or eacn J

in D, but since other terms are not agregated, fewer columns may be eliminated. For the example
problem, only the rows and columns corresponding to the zni(p) vai aD es could be eliminated. "i l

Furthermore, in Kolen's formulation, no y^ variables have zero coefficients so that no plants can be

set open as in the example problem.

Kolen showed that the constraint matrix of his set covering formulation was totally balanced

[9] when the transportation costs were weighted distances in a tree network and the nodes served
both as customer and plant locations. Since the constraint matrix of P2 is a submatrix of the matrix

of Kolen's formulation, and since a submatrix of a totally balanced matrix is totally balanced, this

same property holds for the formulation P2.

Problem P2 is closely related to a formulation reported in reference [2] for the uncapacitated

facility location problem and obtained by a canonical reduction of the matrix of q; values. This
formulation also aggregates terms where possible and so reduces the number of variables and

constraints.

A WEIGHTED DOMINATION PROBLEM

The set covering problem P2 is equivalent to a weighted domination problem on a bipartite

graph. To see this equivalence, problem P2 is expanded by adding redundant variables and
constraints so that the expanded matrix of constraint coefficients is the vertex-vertex incidence

matrix of a bipartite graph.

Consider the function F, which has been simplified by adding together like terms and

reduced by expression (4). Let S be the subset of indices i in P corresponding to those terms in F'

that contain only yj. For each i in S, yj represents the aggregation of one or more variables yu^
over those indices j such that j(l) = i. To obtain problem P2 each variable y-
{
was replaced by the

variable z x and the constraint yj + Zj > 1 is included.

A new problem, called P2', is obtained by expanding problem P2 as follows. For each i in

P-S, add the variable zj to problem P2 with a zero cost in the objective function and add the

constraint yj + Zj > 1. Problem P2' now has a variable z x and a constraint y { + z x > 1 for each i in
 A

P. This version of P2' is equivalent to P2 since any feasible solution to P2 may be amended by Zj

= 1 for i in P - S to obtain a feasible solution to P2' with no change in the objective function.

Conversely, deleting Zj = 1 from any feasible solution to P2' gives a feasible solution to P2 with

no change in the objective function.

Let m be the number of variables znj(i) * n ^> iricluding the Zj added above. Note that m is
also the number of constraints in P2'. The coefficient matrix of P2' may be written as (I
m ,A),
where I
m is an mxm identity matrix with a column corresponding to each variable zni(i)' an( A is *

an mxp matrix with a column corresponding to each variable y v Observe that A contains the pxp
identity matrix I as a submatrix since for each i in P, there is a constraint Zj + yj > 1.
p
Next, the set of constraints with coeficient matrix (
r
, I
p ),
where A 1
is the transpose of A,

is added to P2'. Then the coefficient matrix of P2' becomes:

c - (
V
Im
At
M
'pJ

For each i in P, the i

tn
constraint of (AMp) contains Zj and yj (since A contains I p ) and is therefore

satisfied by any solution that satisfies the constraint Zj -h yj > 1 in (I

m ,A). Thus the constraints of

(A^Ip) are redundant and problem P2' is equivalent to P2.

Since the matrix C is symmetric with ones on the diagonal, it is a vertex-vertex adjacency

matrix of some graph G. The graph G is seen to be bipartite with one set of vertices corresponding

to variables yj and the other to the variables znj(i)- ^ach vertex z ni(i) * s adjacent to those vertices

yj where i is in the set Ilj(i). Each vertex is weighted by the cost coefficient of the respective
variable. Therefore, the set covering problem P2' is a weighted domination problem on the

bipartite graph G. Figure 2 shows the bipartite graph of the example problem.

10
 Figure 2: Bipartite Graph of the Example Problem

A VERTEX PACKING FORMULATION

The third formulation obtained for the SPLP is that of a weighted vertex packing problem

on a graph. For each customer j and its permuation j(i) for i = 1, . .

. , p, we have the following

Boolean identity:

Solving for ny^p\, substituting into expression (2) and simplifying yields

P
T"- = V (c
v v
.

v -M.)ITy
J v y
j(i-irj(i)
+ M. .
(7)
j gj jO)j j j

for each j in D. For the example problem and customer 2,

T" 2 = (1-M 2 )y 3 + (2-M 2 )y 3 y 1 + (4-M 2 )y 1 y 3 y4 + M2 .

Next substitute y^= 1 -

yj in each term fjy^ of F. Making these substitutions in the function F
gives an equivalent form:

F(y,, • • • ,y
p
> = -
.f^i
-
l£ (Mj -
j(i)j
)ny
j(i . 1)yj(i)
m + gfj

Define the function G(yj, . . . , yp ) = - F(y 1? . . . , y p ), which is a pseudo-Boolean function in

posiform. Then problem PI may be written in the equivalent form as:

P3: max {G(yi, . .

. , y p ): y, = 0,1, ieP}.

The function G for the example problem, with M; = 5 and without simplifications, is:

G( yi , v 3 ,y 4 ) = 5 yi + 4 yi y 3 + y iy3 y 4 + 4y 3 + 3y 3yi +yiy 3 y4 +

+6y 3 +4y 4 + 5y 1
5y 3 + 4y 3yi + 2y iy3 y 4 + 5y 4 + 2y 4 y 3 + y 3 y 4 yj + 4y 4 + 3y 4 y 3 + 2y 3 y 4 y 1 - 40.

11
 Hammer and Rudeanu [7], and Ebenegger, Hammer, and de Werra [5] have shown that the

problem of maximizing a pseudo— Boolean function given in a posiform may be transformed into a

maximum weighted vertex packing problem on a graph. In reference [5], the graph corresponding

to the posiform pseudo-Boolean function is called a "conflict graph". The conflict graph has a

vertex for each term in the function, and an edge connecting two vertices if the corresponding terms

contain at least one complementary pair of variables. That is, if yj is in some term of G and its

complement yj is in some other term, there is an edge between the two vertices. The vertex weights

of the graph are the coefficients of the terms in G. We call the conflict graph of the function G
derived from a SPLP a "plant location graph" (PLG). The plant location graph for the expample
problem, without any simplifications is shown in figure 3.

It is shown in reference [5] that the maximum value of G is equal to the value of a maximum
weighted vertex packing of the conflict graph. We denote the problem of determining a maximum
weighted vertex packing in the plant location graph of the function G as problem P4.

The SPLP has been formulated as a weighted vertex packing problem by Cho, Johnson
Padberg, and Rao [1] and by Cornuejols and Thizzy [4]. Our formulation is closely related theirs

but with some interesting differences. An obvious similarity is the highly structured form of the
plant location graph. For each j in D, the set of vertices corresponding to the terms in T'j form a
clique ('rows' in Figure 3). For each i, a star is determined with the vertex corresponding to yj at

the center and an adjacent vertex corresponding to the term containing yj in each expression T":

('columns' in Figure 3).

Several differences are observed. First, the plant location graph of P4 contains all the edges

of the graph constructed in references [1,4] but has the additional edges (between cliques)

corresponding to complementary variables. For example, there is an edge between y ^3 and yjy3

in the example problem. These additional edges strengthen the vertex packing formulation.
Second, if no simplifications are made in G by adding together like terms, the plant location

graph has the same number (dp+p) of vertices as does the graph constructed in [1,4]. However,
each simplification of G by adding like terms reduces the number of vertices in the plant location

graph. This reduction in vertices is not possible for the graph constructed in references [1,4].

Third, the reduction given by expression (4) may be applied to the function G. For each i

consider the terms (M; - Cj(j)j)y;(j\ + fjyj, such that j(i) = i. For the function G, the penalty cost
M; may be chosen sufficiently large so that Mj - Cj(j\; > fj. Then these two terms reduce to (Mj -

c i(i)i "
^i^i(i)
+ ^i
w ^ich eliminates the vertiex corresponding to to yj from the plant location graph.

With these reductions, the plant location graph has at most (dp) vertices. The conflict graph

resulting from the function G, after all simplifications and reductions, is called the reduced plant

12
location graph.

For the example problem without simplifications in the function G, the plant location graph
contains 18 vertices and 73 edges. However applying all possible simplifications and reductions,

the function G becomes:

G(y!, y 3 y 4 ) = 3y 3 + 5y 4 + 4 yi y 3 + 7yiy 3 + 5y 3 y 4 + 4y iy3 y 4 + 3y!y 3 y 4

,
- 25.

Figure 4 shows the reduced plant location graph for the function G as simplified above.

Finally, for all SPLP with p plant sites and d customers, the vertex and edge sets of the

graphs constructed in references [1,41 are identical. Only the vertex weights change. However, a

plant location graph may differ from one SPLP to another with the same sets P and D since the
vertex set and the edge set depend directly on the costs Cj; and fj. In particular, the plant location

graph for a SPLP arising from a tree network may differ from a SPLP arising from a general
network of the same size. This raises the following question corresponding to Kolen's result for

the SPLP on a tree network: Does the plant location graph arising from a SPLP on a tree network

have some special struture?

Figure 4: Reduced Plant Location Graph for G

RELATIONSHIP TO STANDARD INTEGER FORMULATIONS OF THE SPLP

Problems PI and P3 are equivalent to the well known integer programming formulation of

the SPLP, often called the "strong formulation" [3] which is given below. The variables yj are the

13
same as defined above. Define xy to be a binary variable indicating whether customer j is served

by a plant at i (xj; = 1) or not (x^ = 0). Then the strong formulation may be written as follows:

P5: minXXcx + Y fpr (8)

1=1 j = l
J J
1=1

p
s. t.
I
i=l
x..
1J
= 1 j e D (9)

Xy < yi
je D, ie P (10)

Xij = 0,1 yi
= 0,1 (11)

Given next is a relationship between the variables xj; and the Boolean expressions in the yj

variables that were defined earlier. For any set of values for the yj variables, xjj = 1 if and only if a

plant is open at site i and the transportation cost Cjj between customer j and site i is minimum over
all other open plant sites. That is, for each i € P and j e D with the j(.) defined by expression (1)

andj(k) = i,

x ij = yj(i)yj(2) • • •
yj(k-nyj(k) = n yj(k-i)jyj(k). (12)

Expression (12) is substituted for each xj; in problem P5. Thus each constraint of (10)
becomes

nyj(k-i)jyj(k) ^ yi

which is always satisfied for y^ equal to either or 1. With these substitutions, the constraints (10)

can be discarded since they are always satisfied.

Substituting expression (12) for each xj; into constraint (9) yields

JnWjW = * je D (13)

Comparing (13) to the Boolean identity (6) implies that each constraint (13) is equivalent to

requiring that ny:/

p
\ = for each j in D. Thus the constraints (9) may be eliminated from P5 if

each term ny;/ \ is forced to be zero in an optimal solution. This is accomplished by adding to the
p

14
objective function the terms M;nyj/ \ where M: is a large penalty cost, for each j in D.
p
Finally, the Cjj may be reindexed by the permutations j(.) to yield c;n\;, and expression
(12) is substituted for each x^ in the objective function. With these changes, problem P5 is written

as

dp
min
J [
J>)j
nyj(i-i) yj(i) + Mn
W +
p

S
f
i
yi

s. t. yj = 0,1

which is precisely problem PL

Since problems PI and P3 are equivalent, it follows that P5 is equivalent to P3. However,

it is interesting to note that P3 may be interpreted as a Lagrangian relaxation of P5. Expression

(12) is again substituted for xj: in the objective function (8) and in each constraint (9) and (10) of

problem P4. The constraints (10) are deleted as above since they are always satisfied. For the

resulting problem, consider a Lagrangian relaxation with respect to the constraints (9) and with
multipliers M;. This yields the following problem.

d d

1^^)
p p p
min 11 c
j(i)j
ny
j(i)j
yj(k) +
gty. + gM^l -

yi
= 0,1.

Problem P3 is obtained by substituting y^ = 1 - yj in each term f^y^ collecting terms, and

changing to a maximization problem by multiplying through by -1.

15
Figure 3: Conflict graph of G without simplifications

16
REFERENCES

[1] D. C. Cho, E. L. Johnson, M. Padberg, and M. R. Rao, "On the uncapacitated plant location

problem I: Valid inequalities and facets," Math, of Opns. Res. 8, 579-589 (1983).

[2] G. Cornuejols, G. L. Nemhauser, and L. A. Wolsey, "A canonical representation of simple

plant location problems and its applications," SIAM J. Alg. Disc. Math. 1, 261-272 (1980).

[3] G. Cornuejols, G. L. Nemhauser, and L. A. Wosley, "The uncapacitated facility location

problem," Tech Report 605, School of OR & IE, Cornell Univ., (1984). Also, Ch. 3,

Discrete Location Theory (Francis and Mirchandini, eds.) Interscience, (forthcoming).

[4] G. Cornuejols, and J. M. Thizy, "Some facets of the simple plant location polytope," Math.

Prog. 23,50-74(1982).

[5] C. Ebenegger, P. L. Hammer, and D. de Werra, "Pseudo-Boolean functions and stability of

graphs," Algebraic and Combinatorial Methods in Operations Research R. E. Burkard,

. et. al.

(eds.), North Holland, 83-97 (1984).

[6] P. L. Hammer, "Plant Location- a pseudo-Boolean approach," Israel J. Tech. 6, 330-332

(1968).

[7] P. L. Hammer and S. Rudeanu, Boolean Methods in Operations Research Springer- Verlag, .

New York, 1968.

[8] J. Krarup and P. M. Pruzan, "The simple plant location problem: survey and synthesis,"

European Journal of Operations Research 12, 36-81 (1983).

[9] A. Kolen, "Solving covering problems and the uncpacitated plant location problem on trees,"

European Journal of Operations Research 12. 266-278 (1983).

[10] R. S. Garfinkel and G. L. Nemhauser, Integer Programming John Wiley, . New York, 1972.

17
 DISTRIBUTION LIST

NO. OF COPIES

Library (Code 0142)

Naval Postgraduate School
Monterey, CA 93943-5000

Defense Technical Information Center

Cameron Station
Alexandria, VA 22314

Office of Research Administration (Code 012)

Naval Postgraduate School
Monterey, CA 93943-5000

Center for Naval Analyses

2000 Beauregard Street
Alexandria, Va 22311

Library (Code 55)

Naval Postgraduate School
Monterey, CA 93943-5000

Operations Research Center, Rm E40-164

Massachusetts Institute of Technology
Attn: R. C. Larson and J. F. Shapiro
Cambridge, MA 02139

Koh Peng Kong

Ministry of Defence
Blk A, Stockport Road
SINGAPORE 0511
 DUDLEY KNOX LIBRARY
-

nil || ii mill IIH II

3 2768 00302767 3