You are on page 1of 13

The Multi-Commodity Facilities Location Problem Author(s): J. Karkazis and T. B.

Boffey Source: The Journal of the Operational Research Society, Vol. 32, No. 9 (Sep., 1981), pp. 803814 Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society Stable URL: http://www.jstor.org/stable/2581396 . Accessed: 20/12/2013 15:38
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize, preserve and extend access to The Journal of the Operational Research Society.

http://www.jstor.org

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Soc.Vol.32.pp.803to 814,1981 Res. J. Opl reserved Allrights Britain. inGreat Printed

0160-5682/81/090803-12S02.00/0 Ltd Research ? 1981 Society Operational Copyright

The

Multi-commodity Problem

Facilities

Location

and T. B. BOFFEY J. KARKAZIS and Statistical of Science, University of Liverpool Department Computational The Multi-commodity Location Problem was one of the 'multi-lever location problems introduced by Warszawski and Peer in their paper on building sites. Methods of solution are developed for this problem via a dual-based approach and via a Lagrangean dual-based approach with hill-climbing. Numerical results are presented.

INTRODUCTION and Peer1 introduced a multi-commodity location problem with In 1973, Warszawski reference to building sites. They gave, as an example, the case of a site on which 152 in 7 four-storey residential buildings. Locations must be dwellings are to be constructed centre and a for a concrete mixing plant, a building blocks manufacturing determined and storing of reinforcing steel. Nine sites have been centre for cutting, bending identified as possible sites for the supply centres, and each site can accommodate only the dwellings are grouped together into 38 destinations one centre. For convenience for the various commodities. a 'multi-lever Warszawski and Peer argue that for such problems the is appropriate. model location thereon) (and variations Among was the Multi-commodity Location Problem (MLP) considered MLP: minimize *o = vv n m vv n Z Z Z <?uAi + Z Z fifi r= 1 i= 1 j= 1 r= 1 i = 1 Z Ai i=\ > 1 (i) fixed charge models they

subject

to

(2)

I f r= 1 ft ~

< 1

(3) (4) (5)

Aj > 0 = 0, 1

Aj, f

X\

r from of commodity customer /s requirements crtjis the cost of transporting = if i. 0 customer situated at the a facility (of j has type) appropriate crtj of commodity r. zero requirement a facility of type r at site i. gl is the fixed cost of establishing r from i fl if j is supplied with commodity (0 otherwise 1 if there is a facility 0 otherwise. at / for supplying commodity r

Vi =

and Peer, but is actually slightly different from that of Warszawski This formulation that g\ > 0 for all i and r, and for under the reasonable is equivalent assumption each r at least one customer has a non-zero requirement. 803

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Journal

of the Operational

Research

Society

Vol. 32, No. 9

is one, then MLP reduces to the Clearly, if the number, w, of different commodities Location Problem (SLP), which has been much studied familiar Simple (uncapacitated) (e.g. Spielberg2, Efroymson and Ray3, Bilde and Krarup4 and Erlenkotter5). for obtaining solutions for MLP. The first Warszawski6 put forward two methods method with (w + l)-way at was a branch-and-bound p on branching subproblem site should be unoccupied a whether or not a selected or occupied facility of by = r r w. the . . The lower is for bound at obtained ., 1,2, p type by relaxing Lp been not constraint for those sites whose has (3) yet non-interference occupancy these bounds were later for improving specified and solving several SLPs. Suggestions results have been given and and Neebe.7 No experimental put forward by Khumawala it is doubtful whether it would be a very effective method as it stands. second approach was heuristic. First, w SLPs are solved (one for each Warszawski's commodity); then, w if Z yrt ^ i r= 1 an optimal with solution to MLP has been found. f?r aU sites **' On the other hand, if there is a conflict at k

w I yi > 1, r= 1 in in C(k) = {r\yrk = 1 j are relocated then all but one of the facilities for commodities This bound the lower in increase process such a way as to lead to a minimal Lp. with this approach was repeated until no conflicts remained. experience Experimental later in this paper. It might be noted that this heuristic was given6 and is reproduced and of Khumawala bounds the improved method could be enhanced by employing Neebe.7 PRESENT APPROACH

SLP is a well-solved are being problem in the sense that guaranteed optimal solutions time of a second or so, for problems with a 100 or found, within a computation more possible are by which this might be achieved facility sites (n ^ 100). Methods with the of the of Bilde and modifications dual-based Erlenkotter5, Krarup4 (a) approach in a similar and (b) the use of subgradient (which has been applied optimization the authors context, namely that of finding p-medians, by Narula et al.8). Recently, the 'commonsense' these two approaches: dual-based is used have combined approach to obtain optimal values of the dual variables and these used as or near optimal input to the subgradient optimization phase. The dual-based approach for SLP

The primal formulation of SLP, which will be denoted simply by P, is obtained r and the of MLP given earlier by dropping the superscript from the formulation P will the of P non-interference constraints denote relaxation obtained (3). by ignoring dual the integrality the dual of P. Associating and D will denote constraints, variables Xj9 p(j with the analogues of constraints (2) and (4) leads to D: maximize m I h j= 1 subject to m I fJttj < Gi (7) (6)

Ptj ^ 0, kj unrestricted. 804

(9)

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

J. Karkazis The choose

and T. B. Boffey?The

Multi-commodity

Facilities

Location

Problem to

does not contain objective small as possible to as be p^ pu

any ptj and so, for given kj, without violating (8). That is = max(0, kj Cij).

it is sensible

(10) {kj}. This can be achieved small values of kj and, at the fewest ptj, thus aiming the constraint

The problem thus reduces to that of finding a suitable set with by starting (Bilde and Krarup4 and Erlenkotter5) those kj which lead the increase of each step, increasing to make best use of gx when 'sharing between customers' via m X l*tj < 9t-

BKE and is that of Erlenkotter5 The formal version of this will be designated Algorithm with a somewhat and different of his 'dual-adjustment' without inclusion procedure A is first found set values for of scheme. by consistently using the scanning kj then set another of is order values found ..., ku k2, km; by consistently scanning using the scanning order km,. . ., ki. to the higher value The set of values of kj [and hence of ptj by (10)] leading of the objective of D is selected and will be denoted by [kj\. The output from algorithm BKE is the set of values jlyj, and the value dlb(A) of this solution values bound to the optimal P and P): of v(P), v(P) is a lower dlb(I) value i;(D) of D (and hence to the optimal

Setting

^ v(D) = v(P) ^ v(P) m A if X l*tj = 9i j=1 3t = \ otherwise (0 to a nearest open facility xtj by assigning each customer of P and hence to an upper bound pub(y) to v(P). By (11) dlb(I) ^ t,(P) ^ pub(y). leads

(11)

(12) to a

and determining feasible solution

(13)

e = pub(y) - dlb(A) is typically small and often zero. In the latter case to P. If 6 > 0, then the solution solution an optimal may be yh xtj provides Erlenkotter5 describes be that noted branch-and-bound. (It might by using completed ways in which the bounds might be improved before resorting to branch-and-bound.) The difference Dual-based As noted approach to MLP constraints effects while are still

earlier, MLP reduces to w SLPs if the non-interference can be taken of non-interference some account ignored. However, retaining the simplicity of solving only a set of w SLPs: vv n vv n xo = x0 + Z Z Wtft) Z Z (M) r= 1 i = 1 r= 1 i = 1 n vv n vv n m + Z Z (9i + 0<w - Z 0, > Z Z Z tjx'j i= 1 r= 1 i = 1 r= 1 i = 1 j = 1 provided vv I tf < 1. r= 1 a lower bound to the optimal n Z o, i= 1 from any lower bound to the optimal value of MLP 805 value of MLP is provided

Consequently

by subtracting

with g\ + 0,- replacing

g\.

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Journal

of the Operational

Research

Society

Vol. 32, No. 9

Table 1

MLP for which the data is given in Table 1. Example. Solve the two-commodity The number above a diagonal line refers to commodity 1 and the number underneath to commodity 2. Solution. By direct calculation of costs it can be verified that the values for all feasible solutions to the two SLPs which result from ignoring non-interference constraints are shown in Table 2. For r = 1 the optimal solution is y1 = (yj, yi, }'L vi) = (1, 0, 1, 0) (that is only sites 1 and 3 are occupied) with value 47, and for r = 2 the optimal solution is = 0,0,0, y2 = (yi,yl*yhyl) of 1) with value 44. This gives an initial lower bound 47 + 44 = 91 for MLP. However, the solution to MLP obtained by combining y1 and y2 is not feasible, there being a conflict at site 1. to lead to two alternative 0{ from 0 to 2 is just sufficient Increasing optimal solutions for r = 1, namely (1,0, 1,0) and (0, 1,0, 1) of value 49. (This is readily checked from Table 2 as the only changes are that the values of solutions in which site 1 is used are increased by 2.) For r = 2, (1,0,0, 1) is still the only optimal solution but its value is now 46 leading to a new lower bound of 49 + 46 - 2 = 93 for MLP. Thus the conflict at site 1 can be resolved by taking 0 = {0U02,03,0^) = (2,0,0,0) and the solution obtained there is by combining y1 = (0, 1,0, 1) and y2 = (1,0,0, 1). However now new 4 which was not present before. If fl4 is increased by 1 the are y1 = (1,0, 1,0) and y2 = (1,0,0, 1) with values 49 and 47 respectively, giving a lower bound of 49 + 47 - 3 = 93 for MLP. The values ofthe feasible solutions are now as shown in Table 3. Site 1 again gives rise to a conflict. Increasing solutions 8{ to 3 leads to optimal (1,0, 1,0) and (0, 1,0,1) with value 50 for r = 1, and to optimal solutions (1,0,0, 1) and (0,0,0, 1) with value 48 for r = 2. Selecting the combination optimal solutions y1 = (1,0,1,0) and y2 = (0,0,0,1) (14) a conflict at site

806

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

J. Karkazis

and T. B. Boffey?The

Multi-commodity

Facilities

Location

Problem

gives a feasible solution to MLP of value 47 + 47 = 94 (cf. Table 2) and also to a lower bound of 50 + 48 ? 4 = 94. It follows that the solution (14) must be optimal. In practice, for SLP (and other combinatorial a very good (and perhaps problems) can often be found much more quickly than a guaranteed optimal) solution optimal lower bounds for MLP will be obtained by subtracting solution. Consequently n 1 et i= 1 from the sum of lower bounds to w SLPs. As seen by the above example, the can give rise to another resolution of one conflict and this prompts the question as to how the 0t should be chosen. A procedure for finding suitable values for 0t was lines the the solution of the above example, and is by developed, along suggested the dual variables of below. For values particular presented k), prtj the Ij, ]x\j, conflict set CONF is conf where y{ is determined Algorithm MultiBKE = j; t{yri> i} of superscript r).

(15)

as in (12) (but with the addition

Step 1 Set Oi = 0 for variables k), p\j. Step 2 Decide Step 3 whether

all

i.

Use

algorithm

BKE

to

find

initial

values

of

the

dual

to terminate.

If yes then go to step 4.

Let lrj, prij be the values of the dual variables obtained BKE by using algorithm with gr( + 0t temporarily replaced by g\ + 0t + M for all k e CONF where M is a very large number. Set k) <? 1) for all r and j. \ / m m ~ 0k <- max X tij Z Aj r \J=1 j=\ J Return to step 2. Step 4 dlb(A) The termination = X Z *rj r= 1 j = 1 ~ I Ot is a lower bound j=1 for MLP.

test of step 2 will include a feasibility check of the solution in If this solution is feasible the current dual variables from obtained (12)). (as [yrh xri)} and the algorithm terminates. with value pub(yr) = dlb(xr), then it must be optimal Otherwise the execution a of the algorithm be terminated after specified number may of applications of step 3 (one application was found to be sufficient for the problems tested). values of 1) are clearly the maximal BKE no matter how much 0k is increased. BKE by using algorithm may be achieved the or more of constraints one blocked by m k) that can The smallest occurs be obtained using algorithm value of 0k for which this set when further increase in all k) is

Thus.0fc need only be increased

just sufficiently m Z lij< 807

to satisfy the constraint 9k + 8k-

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Journal Now

ofthe

Operational

Research

Society

Vol. 32, No. 9 with equality by the

previous

since there was a conflict at k, this constraint was satisfied values of pkj and 0k and so 0k must be increased by m I j=i Kj m I Aj j= j algorithm.

and this is the value used in step 3 of the above Finding a feasible primal solution

there are still locations with conflicts, of multiBKE, If, after the application feasible primal solution is found as follows by removing conflicts (one at a time). Step 1 Set f according to (12), and hence x\j. vv / Set multiub For all k e CONF Step 2 = Z r= 1 Pub(A') = n m Z <?uAi + Z fift i=l Z Z r=l\i=lj=l

then

perform

step 2.

(Removal of conflicts) the the removal cost RCk of removing (2a) For each re C(k) = [5(3/^ = 1 j calculate is r This further conflicts. done of from site k without by facility introducing type of the constraints applying BKE with the temporary imposition fk = 0 FoTt*k * = {l U fiZl and Xv^l}PreVi0USly'

(2b) Select r0 e C(k) for which RQ is maximal. Set yrk?= 1, yfcs= 0 for all s ^ r0. Set multiub *- multiub + X RC* reC(k) _ RC*? ?

Step 3 The solution obtained contains no conflicts bound to the value of an optimal solution. Branch-and-bound If a proven solution

and is hence

feasible.

multiub is an upper

has not yet been found, then a branch-and-bound solution optimal tree is entered. at each node of the branch-and-bound be Branching may = = 1 ... or or or some selected site all for + k, l)-fold according as, (vv y\ = 1. yrk 0, y\ Sites are chosen for branching on in a pre-specified order in which those giving rise to conflicts are considered first. Lower bounds at nodes are determined by algorithm set r For let Of be the set of locations multiBKE modified. a suitably type commodity = e 1 if k and Add constraints be set set to 'closed'. Cr the of locations Of yk 'open' (dual variable <?k) and yk = 0 if keQ (dual variable wk) to the primal SLP, dualize routine and increase the dual objective value by

z^ keOr such that

max (0, krj - ckj) ^ ?J j

where

k*=u iikeQf c grk-(j>k gk- <Pk ? = < 00 if ke C ?J ^ gk otherwise. 808

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

J. Karkazis

and T. B. Boffey?The

Multi-commodity

Facilities

Location

Problem

As an immediate consequence be used as a starting feasible bound is calculated as follows:

the final feasible solution (kj9 ju^) of the root node can lower solution for the current node. The corresponding add the slackness m ~ Qk Z tij j=i

to the

and node lower bound, set g\ = 0 for all keO% gk = oo for all keC BKE. apply algorithm The development rule was branch from lowest bound. Solutions to MLP are generated at each node [according to (12)] and checked for feasibility. root

Lagrangean

relaxation in a manner is to MLP

the constraints By incorporating (2) and (3) into the objective function similar to that for SLP it is readily verified that a lower bound /(A, 0) the optimal value of w n m min ? I I ditfj y] \r= 1 i= 1 j= 1 *U? + X X rfyf + X X Xi 1 r= 1 i = 1 r= 1 j = 1 \ I Ai i= 1

?<?-!/)) subject to fi Aj > 0 (18) (19)

<">

= o, i. Ai, yri Then a Lagrangean dual of (l}-(5) is to find max/(A, Now, g\ > 0 for all i and r, (17H19) w f(l,0) = m V X X r=lj=l 0). is equivalent to

(20)

assuming

n / w n m I 0? + min I I X Wi W; i=l X-;= 0,1 \r=l i=l j=l +

te? + 0.O max x^ j

Setting

dlj

? = min (0, c^/J), and d? = Z drij

it may be verified

that /(A, 0) is obtained ^ 0

by setting

/. = 0, xrij = 0 if g\ + |^| = i tf = 1 and xru IJ (0 It follows 1

> ? lf^j otherwise

if yi tf + ' |^|'' < 0. by using the formula

that f(k, 0) can be computed w m Z ijj=l + n Z w 0, n min \ /

readily

/(M)=?

m 0,^ + 6, + 809 X min(0,tf; i= 1 -

\ AJ)). / (21)

X Z r= 1 i = 1

o.r.s.32/9-e

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Journal It remains now

of the Operational values

Research

Society

Vol. 32, No. 9

to determine

by using subgradient (asymptotically) tested that an alternative approach,

for k), 6t to satisfy (20). This may be achieved However, it was found for the problems optimization. was computationally faster. namely hill-climbing,

Hill-climbing for the moment, that dual (20), (21). Suppose, We again start with the Lagrangean vv = 1, in which case superscripts and parameters 0, may be omitted. Then from (21) m /(A) = X j=i ^ + n X ;=i min m + 0,0* X min (<),<:,?; j=i V / by \ kj) . / (22)

Right and left partial derivatives =

are defined

?Q^dAj ^-^ CAj

lim \(f(k t-^0 + lim \(f(k r?0"

+ re,) - f(k))/t)

(23)

+ tej) - f(k))/t\

(24)

where e, is the unit vector whose components other than the y th are zero. Theorem. If min* denotes minimization over strictly positive terms only m Pi(t) = 9t + + df (k) = 1 (a) ^? CAj where kj ^ 0 and p((0) ^ 0 _ j 1 if ctj {o otherwise for {b)^k-=-^k--^wij 0 ^ t ^ min*(|min(Pi(0), where ~ = ? Aj 1 lif{C? kj < 0 [ctj 0 otherwise. Proof. cf+(k) ^-r? CAj The proof of part (a) only is given. From (22), (23) = m " 1 f / ' lim - < t + X min 0, g{ + X min(0, cu - kj r-^0+ t l i=i V j=\ minl iin( 0, gff. + where d;IJ0 if = < {0 ? otherwise] 810 >is the Kronecker delta function. and and Pi{0) ^? Pi(0) = 0 0)|, |min(0, c0 kj)\) = TJo n X It) for 0 ^ r ^ min*(max(p,(0),0) ,= i ,+ max(0,cokj)) = Tjo X min(clV j= i (k} + t))

13

5ijot)

X min (?? cu ^j)

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

J. Karkazis Now

and T. B. Boffey?The

Multi-commodity

Facilities

Location

Problem

minfo,^.

min(0,co

kj

Sijot))

- minf unless CtJa or Pt(t) = gt + in which case it is zero. if (1) (2) (3) (4) ciJo kjo ^ 0 and Pt(0) ^ cijo kjo > 0 and P-(0) ^ ciJo kjo ^ 0 and Pt(0) > ciJo kj0 > 0 and Pt(0) > for 0 < t ^ cijo ? kjo and for t > cij{, - kj, + Pf(0). these four results shows ~ Ijo ~ t> m Z min(0, cij 0

0, gt +

Z min (?> cu

= *j))

~l

kj

tSijo) ^ 0

(26)

Now

0, then (26) holds for all t > 0 0, then (26) holds for t > cijo - kjo 0, then (26) holds for t > P,(0) 0, then Pi(t) = Pffl) hence (26) holds that the left hand side of (25) is ? niojo if ciJo kjo) + max(0, PIO(0)).

Combining

0 < f < min* (max(0, i + io It follows that

f(k -

+ tej) - f(k)

= t^rjij^j

= if 0 < f ^ min* (max(0, cijo Tjo and hence the result of part (a). kjo) + max (0, Pio(0)) The proof of part (b) is similar. The above result leads to the following algorithm for finding lower bounds for SLP. Algorithm (HC)

Step 1 Set kj = min ctj for all j. i Step 2 Decide whether to terminate. For j0 = 1, 2, .. ., m perform step 3. Step 3 If * > 0, then set k+-k + aTjoejo

otherwise

if < 0 then set k +- k -

^^-

bTjoejo

f(k)^f(k)-dJ-^-bTh. 811

CAjo

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Journal

ofthe

Operational

Research

Society

Vol. 32, No. 9

? o o ? o ^_ r_ o rt Ont~~] ? <*~\ 0s; O ? <n^n on

?- \o r- <n B c^ ri ?jf~ O ??? oo ? ^n ? oo H vo *-> <nc^ un O Tt On ?; OO ? m *-^ ^t un <n ? Tt ? \o un ? r-j r-jTt O Tt

un O r- un <no r- O o r~ON Tt ? on Tt OQ ? ^ OO vO 2 O O Tt OO ? Tt ? H' unm ? r-jr-jO Tt CO ON ON \Q oo oo r~^ on ? ? oc ? co ? uo un ? c^ c^ O C/3 P-d-o 3 o o a-9 -? O C UC a al : x> X) 81 o o *ea a c

<nco co co \D\0 r-~on OfN M OC

- M fOTj-

812

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

J. Karkazis The

and T. B. Bojfey?The

Multi-commodity

Facilities

Location

Problem were

values taken for a and b were 1 if the corresponding partial positive and 0.3 and 0 if the derivatives were zero. For the multi-commodity problem the following algorithm was used. Algorithm multi-HC

derivatives

Step 1 Set multilb = f(k, 0), where 0 = 0 and k) are found using algorithm Perform step 2 for i = 1, .. ., n. Step 2 Let S = {r\Pri < 0) where m Pri = gri+ lf\S\> X mm&Cijkj).

HC.

1, then set 0,- = max|Pf| reS

multilb <? multilb +

X Prt

Oi.

NUMERICAL

RESULTS

above were tested. Results are presented in Table 4 Several variations of the approaches for the following two: The algorithm is terminated at step 2 if step 3 has been executed once. (1) MultiBKE. HILLCLIMB. This is as described in the last section with initial values of (2) BKE for each r. kj, prij provided by has not been found If, in each of the above cases, a guaranteed optimal solution at this stage, then the solution is completed by branch-and-bound. The programs were coded in Algol 68R and run on an ICL 1906S computer. In order to Warszawski's results it should be noted that MultiBKE to be directly comparable and HILLCLIMB compiled in not more than 5 seconds each.

CONCLUSION The SLP is a well solved problem. This suggests the possibility of methods for solving based on the use of an SLP-solving subroutine. Dualother, less restricted, problems this idea were developed above for MLP based and hill-climb approaches incorporating It was found, for the problems results presented. and numerical tested, that both was the better of methods were computationally quite effective and that HILLCLIMB 4 it be noted that data set referred to the the two. Finally might building site problem this problem was considered in the introduction; and Peer1 mentioned Warszawski by to be a realistic size for a typical building site application. ACKNOWLEDGEMENT The authors wish to express their gratitude to A. Warszawski for supplying the data for the problems of Table 4. REFERENCES 1A. Warszawski and S. Peer (1973) Optimizing the location of facilities on a building site. Opl Res. Q. 24, 35-44. 2K. Spielberg (1969) Algorithms for the simple plant location problem with side constraints. Opns Res. 17,85-111. 3M. A. Efroymson and T. L. Ray (1966) A branch-bound algorithm for plant location. Opns Res. 14, 361-368. 40. Bilde and J. Krarup (1976) Sharp lower bounds for the simple location problem. Ann. Discr. Math. 1, 79-97. 813

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

Journal

ofthe

Operational

Research

Society

Vol. 32, No. 9

5D. Erlenkotter (1979) A dual-based procedure for the uncapacitated facility location problem. Opns Res. 26, 992-1009. 6 A. Warszawski (1973) Multi-dimensional location problems. Opl Res. Q. 24, 165-179. 7B. Khumawala and A. Neebe (1978) A note on Warszawski's multi-commodity location problem. J. Opl Res. Soc. 29, 171-172. 8S. C. Narula, V. I. Ogbu and H. Samuelson (1977) An algorithm for the p-median problem. Opns Res. 23, 709-713. 9A. M. Geoffrion (1974) Lagrangean relaxation for integer programming. Mathemat. Progr. Stud. 2, 82-114.

814

This content downloaded from 163.178.101.228 on Fri, 20 Dec 2013 15:38:32 PM All use subject to JSTOR Terms and Conditions

You might also like