You are on page 1of 7

This article was downloaded by: [UQ Library]

On: 06 November 2014, At: 19:46


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

A I I E Transactions
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/uiie19

A Heuristic Programming Algorithm for Warehouse


Location
a b
Robert E. Shannon & James P. Ignizio
a
University of Alabama in Huntsville ,
b
Virginia Polytechnic Institute and State University ,
Published online: 09 Jul 2007.

To cite this article: Robert E. Shannon & James P. Ignizio (1970) A Heuristic Programming Algorithm for Warehouse Location,
A I I E Transactions, 2:4, 334-339, DOI: 10.1080/05695557008974773

To link to this article: http://dx.doi.org/10.1080/05695557008974773

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of
the Content. Any opinions and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied
upon and should be independently verified with primary sources of information. Taylor and Francis shall not be
liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities
whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of
the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
A Heuristic Programming Algorithm
for Warehouse Location
ROBERT E. SHANNON, Senior Member, AIIE
University of Alabama in Huntsville
JAMES P. IGNIZIO
Virginia Polytechnic Institute and State University

Abstract: A new heuristic programming method of solving a portation costs from factory to warehouse and from ware-
Downloaded by [UQ Library] at 19:46 06 November 2014

particular type of warehouse location problem is presented. house to demand center, plus the cost of building and
The problem is to allocate K or less facilities to N possible
locations so as to service M demand centers at minimum cost. operating the warehouses.
The algorithm presented is suitable for hand calculation of Several approaches have been proposed for attacking
medium-size problems (50x50) or when computerized will the problem as discussed here, as well as for solving the
readily solve large-scale problems of the order of (600 X600); problem when the maximum number of warehouses is not
i.e., 600 demand centers and 600 possible locations. specified. Theoretically, a linear programming approach
* * * using an all integer algorithm could be used to solve the
This article describes a new heuristic programming ap- problem (I), (2). In practice, however, the size and non-
proach to the problem of selecting locations for ware- linearities involved are such that even small sample prob-
houses in large-scale distribution networks in which the lems often involve so many variables and constraints that
maximum number of warehouses has been specified. One they cannot be handled even by large-scale third-
industry-wide survey (26) indicated that physical dis- generation computing equipment.
tribution costs absorb between 10 and 25 percent of net Analogue computer solutions have been proposed by
sales income. Thus, the reduction of distribution costs can several investigators (4), (5), (9), (1I), (15). Although the
constitute a substantial cost saving area. Perhaps, of analogue approach allows the nonlinearities to be readily
equal importance in highly competitive marketing situa- handled, computing equipment characteristics again
tions is the opportunity for improving customer service by present a severe limitation as to the size of the problem
achieving a better position from which to offer the which can be handled. Eilon and Deziel (9) note that to
customer what he wants, when he wants it (16). Other solve a problem with m customers and n warehouses re-
components of the distribution system such as marketing quires approximately 4+2n+5m amplifiers. Thus, a
strategy, sales, customer service, and inventory manage- sample problem such as that solved by Kuehn and Ham-
ment are directly influenced by these decisions related to burger (12) with 50 demand centers and 24 potential
the locations and sizes of warehouses. warehouse sites would require approximately 302 ampli-
Regional warehouses can serve a variety of functions in fiers, while the one discussed by Feldman (10) with 49
the distribution system. Among these are reduction of potential sites and 200 demand centers would require
transportation costs by permitting bulk or quantity 1102 amplifiers.
shipments of goods from factory to warehouse and im- Branch-and-bound techniques have also been proposed
provement of customer relations by reducing delivery (7), (S), (17), (IS), (22), (23) as well as dynamic pro-
time, thereby permitting customers to reduce their inven- gramming (6). Although these techniques work well for
tories. However, there are substantial costs associated medium-size problems, say of size (m X n = 50 X 200), they
with the construction and operation of a regional ware- become impractical for large-size studies. The only
house system. The problem of minimizing distribution realistic approaches suggested for analysis of practical
cost thus becomes one of balancing shipping against large-scale problems have either used approximation
warehousing costs. The objective is to locate and size the (I), (2), (14) or heuristic algorithms (lo), (12), (13), (20),
warehouses and to determine which demand centers are (25)-
supplied from which warehouses so as to minimize overall The definition of an algorithm generally used by math-
costs. The distribution costs to be considered are trans- ematicians is that it is a systematic method for computa-

334 AIIE Tr Volume I1 No. 4


tion, i.e., one which can be taught to an intelligent clerk formulation. This is the identical problem type as dis-
or programmed on a computer. Simon (21) has referred cussed previously by Curry and Skeith (6).
to heuristics as rules of thumb selected on the basis of In order to demonstrate the key steps of the basic
aiding in problem solving or contributing to a reduc- algorithm and to expose the rationale behind them, a
tion in the average time spent in searching for a solution. formulation will be used which has already appeared in
Making use of these definitions, a heuristic algorithm is the literature (6). Let M demand locations be considered,
defined here as a problem-solving program, explicitly each with a specified demand Qi per time period. It is
stated, which utilizes certain common-sense principles or desired to place K warehouses in N possible locations
devices to derive acceptable, although not necessarily (K 6 N) and assign the demands to the warehouses such
optimal solutions. that the total cost (quantity times distance traveled)
Heuristic techniques are most often used when the is minimized.
problem is of such size and complexity that exact opti- The following additional notations are used to formu-
mizing algorithms are not available or not practical. The late the mathematical model of the problem and in the
heuristic approach seeks solutions based upon accepta- steps of the algorithm:
bility characteristics rather than optimizing rules. It gives T = Total cost = (distance X demand)
explicit consideration to a number of factors (for ex- Dd3= Distance from demand center i to location j
ample, computer storage capacity and solution time) in
addition to the quality of the solution produced. In yj= {01ifif aa warehouse
warehouse is allocated to locationj
is not allocated to location j
addition, the evaluation of heuristic techniques is usually 1 if demand i is allocated to warehouse j
done by inductive rather than deductive procedures. 2i3=
{0 if demand i is not allocated to warehouse
Downloaded by [UQ Library] at 19:46 06 November 2014

j
Specific heuristics are justified, not because they attain A= ((ciaj)) =matrix (M XN) of cost coefficients
a n analytically verifiable optimum solution, but, rather, aij= Cost of supplying demand location i from ware-
because experimentation has proven that they are use- house location j, thus aij= (DijXQi).
ful in practice (24). 0 = ( j l when Yi=l, l ~ j 6 N ) .
The heuristic algorithm proposed in this article for
locating and sizing warehouses consists of two parts: The objective function to be minimized is then
(1) The main program which picks warehouse locations
one a t a time until either the maximum number of loca-
tions have been chosen or until additional warehouses
would not decrease the total distribution network costs.
(2) The improvement check and elimination subrou- Subject to the following three constraints :
tine which is designed to remove from solution those
locations chosen early which become uneconomical due
t o later choices.

PROBLEM FORMULATION
Several factors affect the way in which this problem is
treated. In most formulations of the problem, the num-
ber of potential warehouse sites is finite. Cooper ( 5 ) , Equation 2 insures that all demands are assigned to
however, has considered an infinite-site case, where the some warehouse location. Each demand location is as-
set of feasible sites is a two-dimensional continuum and signed to only one warehouse location, thus giving rise to
where some of the various analogue models are appro- Equation 3, while Equation 4 sets the maximum number
priate. Another variation in problem formulation deals of warehouses to be allocated.
with the treatment of the various costs; i.e., whether they
are fixed, linear or curvilinear. Most of the investigators
assume a fixed amount of potential sales for each demand
STEPS OF THE ALGORITHM
center, but Shenoy and Sifferd (20) have considered the As stated in the preceding section, a problem which
case where there is an inverse relationship between dis- has already appeared in the literature will be used for
tance from the warehouse and market share at each de- demonstration purposes. One of the problems that Curry
mand point. and Skeith (6) consider is that of finding a dynamic pro-
The specific problem dealt with in this article is that of gramming solution to a problem where three warehouses
locating and sizing a fixed number (K) of warehouses are to be allocated to four possible locations. The associ-
from out of N possible locations. Consequently, the maxi- ated distances and demands are presented in Table 1.
mum number of warehouses is fixed and thus not a The total demand from a demand center is to be met from
decision variable as would be the case in the more general only one warehouse. However, more than one demand

December 1970 AIIE Transactions 335


center can be assigned to the same warehouse. The
mathematical formulation of this problem was presented
in the preceding section.
For each column Ai where jee, calculate Sf
= CE1 min(ai,-aij*, 0). I
f S i < O for some j. then find
The steps of the algorithm are then as follows: min Xi. For this j, set Y, = 1 and place j in the next posi-
tion of 8 (see Table 4). (If all S i 2 0 , then this indicates
that additional warehouses will not reduce transportation
Establish the problem in matrix form, where for each costs. If this occurs the procedure is terminated as follows.
column vector Aj, the coefficients aij are the costs associ- Let R be the matrix with columns Aol, AO2, . . . , Aok,
ated with supplying demand i from source j (see Table 2). where k is the number of warehouses thus far selected.
Go to step 9.)

Table 1 : Distances and demands for the example problem

Facility Locations
L I I1 I11 IV
0
D C
E A 1 1 9 17 24
M T 2 10 2 8 15 171
Downloaded by [UQ Library] at 19:46 06 November 2014

A I 3 16 8
N O 4 20 12
D N 5 24 16 10
S

Table 2 : Step 1
Remove the column vector with j E 9 (as selected in
Location I I1 I11 IV step 3). Replace A* with a new column vector A*, where
1 75 675 1,275 1,800 a;* = minjee ( a~ ] with i= 1, 2, . . . , M (see Table 5).
2 1,710 342 1,368 2,565 This step serves to indicate the results obtained with the
3 2,448 1,224 306 1,683 combination of warehouses thus far obtained.
4 2,740 1,644 548 685 If xE1Yj=2, repeat steps 3 and 4. Otherwise, pro-
5 19,320 12,880 8,050 805 ceed to step 5 .

Let k represent the number of warehouses thus far


selected. Thus, 8= (81, 02, . , B k ) . Let R be the matrix
STEP2 (SELECTION
OF FIRST LOCATION) with columns AB,, A B ~ ,. , Aok, A*. For each column
For each column vector Aj, calculate Ti= xE1
ad* of R, define E = t h e elimination effect. Then calculate
Select the column vector A, with the minimum Tias the
first location and designate i t A*. Set Y, = 1,place j in the
ordered set 8 (see Table 3).
Table 5 : Step 4

Location I
-
II I11
Best A*
Combination
Table 3 : Step 2
1 75 1,257 675
Location I I1 I11 IV(A*) 2 1,710 1,368 342
3 2,448 306 1,224
1 75 675 1,275 1,800 4 2,740 548 685
2 1,710 342 1,368 2,565 5 19,320 8,050 805
3 2,448 1,224 306 1,683 ---
4 2,740 1,644 548 685 Repeat of -600 - 1,058 Step 4
5 19,320 12,880 8,050 805 Step 3 -+

Ti 26,293 16,765 11,547 7,538 Note: Location 111 selected when Step 3 is repeated.

336 AIIE Transactions Volume I1 No. 4


EA, =
..

z=1
1 min at,, - at*]
ZEB,l#~
for j = el, e,, . . , 0,
min {ao). See circled costs in Table 7.
Thus, for our example problem, the solution is to assign
demands 1and 2 to warehouse location 11,demands 3 and
(see Table 6). 4 to warehouse location 111, and demand 5 to warehouse
location IV. The total cost (distancexdemand) for this
. solution is the sum of the elements in the final column
vector A* or 2676. The solution is identical to that of
Table 6 : Combination improvement check-matrix R
Curry and Skeith (6).
Location IV I1 I11 A* The rationale or heuristics behind the algorithm are
-- fairly self-evident. The solution at the end of step 2 is
1 1,800 675 1,257 675 obviously the best answer if only one warehouse location
2 2,568 342 1,368 342
3 1,683 1,224 306 306 is chosen. The sum of the column vector A* at each step
4 685 1,644 548 548 is the value of T or the total cost at that stage. Steps 3 and
5 805 12,880 8,050 805 4 are merely calculating in a straightforward manner
Step 6 ---------
which of the remaining warehouse locations would im-
Elimi- prove the solution most if picked next. Steps 5 and 6 may
nation 7,245 1,608 1,055 -
Effect not be so readily apparent.
(EAJ)
Note: In Table 6 , column A* is formed from 0 =
Downloaded by [UQ Library] at 19:46 06 November 2014

Table 7 : Assignment-Step 9
(IV, 11, I11 ] in the first repeat of Step 4.
Location IV I1 111 A*
-
1 1,800 1,257 675
2 2,568 1,368 342
If min EA,= E A ~proceed
~, to step 8. If not, remove the 3 1,683 1,224 306
Aj with min E Afrom
~ matrix R (i.e., from the solution), 4 685 1,644 548
remove j from 0. Set Y j= 0. Proceed to step 7. 5 12,880 8,050 805
STEP7 (FORMATION
OF IMPROVED BEST COMBINATION)
Solution 2,676
Construct a new column vector A*, where ai* Cost
-
=minigo (aij) with i = 1, 2, . - , M. Use this new A*
to replace the previous A* of step 4. Return to step 3.
As more and more warehouse locations are added to
STEPS (CHECK) the solution, it is possible for a location chosen earlier to
Determine if x:,
Y j = K . If not, return to step 3.
Otherwise, proceed to step 9.
no longer be significant to the combined solution. Step 5
serves to calculate the effect of eliminating a particular
chosen location from the solution. If the location with the
smallest elimination effect was one chosen earlier, there
From matrix R, find min(aii) for i=l, 2, . - , M is a chance that there is a better location at this stage.
and j= 01, 02, - . . , Oh. Assign demand i to be supplied We test this by eliminating it from the solution and
from warehouse j for those i and j corresponding to each choosing another location. If it was not insignificant or

Table 8 : Results
Number of Average Maximum Average Problems Method of
Problem Problems Time Error Error Solved Problem
Size Solved (Seconds) (Percent) (Percent) Without Error Generation
40 X 100 3 2.3 4 1.3 2 C
50 X50 6 1.9 3 0.5 3 C
28 X70 1 1.0 3.6 3.6 0 C
16 X34 1 0.3 0 0 1 C
20 X20 29 0.3 2.1 0.2 25 A=25, B = 4
15x15 54 0.2 1.9 0.2 44 A=48, B = 6
10 X 10 224 0.1 2.8 0.1 195 A=174, B = 5 0
-
Totals 318 4 percent 0 . 2 percent 270

December 1970 AIIE Transactions 337


redundant, it will come back into the solution a t a later lems that are simply too large for any exact methods.
iteration, and only one iteration computing time will have Finally, not only is a final solution calculated, but also,
been lost. This procedure is analogous to Kuehn and in the process, the best solution for 1, 2, 3, .-
. , K-1,
Hamburger's "Bump and Shift" routine (12). and K warehouses. This is of interest if funds are avail-
able to build only a portion of the K warehouses initially,
and gives the priority ranking under which warehouses
should be added.
The method described here evolved from a study, the
Because this is a heuristic algorithm, it is known that initial purpose of which was to obtain an efficient and fast
an exact optimum is not guaranteed. Several hundred method of generating an approximate, feasible solution
problems of different sizes have been run and checked by that could be used as a "starter" (i.e., a good starting
other exact techniques where feasible. In 85 percent of point) for exact methods. The heuristic method of this
the cases, an optimum solution was obtained, and in no article was the end result of this study. However, in test-
case has the error exceeded four percent. It should be ing this heuristic, it was found that, not only was the
pointed out that, to date, no one has reported a truly method fast and efficient, but it was also much more ac-
practical method for obtaining a guaranteed optimum for curate than might be expected. The method is in essence
large problems (larger than 50x200) with existing com- a straightforward search technique with the additional
puter equipment. aspect of the elimination routine. The fact that the
Table 8 represents the results thus far obtained with method obtains such very good results for this type of
this algorithm. The technique was programmed in problem has not been explained, and is presently under
Downloaded by [UQ Library] at 19:46 06 November 2014

FORTRAN and run on the Univac 1108 computer. At investigation.


present, the program is in a preliminary form, and should In addition, the method is even more accurate for
be capable of being improved in both speed and scope. realistic problems (type B) than the figure of 85 percent
Three problem types have been considered. Problem type given above for all types of problems. Where more
A has been generated by simply using random numbers as realistic data and problems were used, the method solved
matrix elements. A check on the heuristic results on type 57 out of 60 cases exactly and was in error less than one
A problems was made with a branch-and-bound program percent on the remaining three cases. Efforts are now being
developed in conjunction with the heuristic. Problem directed to programming the technique for large-scale
type B was generated to simulate the more realistic prob- problems (600 to 800 possible locations and 600 to 800
lem by choosing cities and distances from maps. Again, the demand centers).
branch-and-bound method was used to verify results.
Problem type C represents somewhat larger problems
that were generated with known solutions.
The heuristic algorithm presented in the preceding (I) BALINSKI, M. L., "Fixed-Cost Transportation Problems,'
Naval Research Logistics Quarterly, Volume 8, 1961.
section has several distinct advantages over previously (2) BAUMOL, W. J. AND WOLFE,P., "A Warehouse Location
suggested procedures. The first of these advantages is its Problem," Operations Research, Volume 6, March-April,
simplicity. The only operations involved are addition 1958.
(3) BRINK,E. L. AND DecAN1, J. G., "An Analogue Solution to
and subtraction. Knowledge of Lagrange multipliers, the General Transportation Problem with Specific Applica-
dynamic programming, difference equations, and branch- tions to Marketing Location," Proceedings of the First Inter-
and-bound techniques is not required. Secondly, hand national Conference on Operations Research, Oxford, England,
1957.
solutions of reasonably large problems are feasible. The (4) BURSTALL, R. M., LEAVER,R. A., AND SUSSAMS, J. E.,
solution of the example problem, including setting up the "Evaluation of Transportation Costs for Alternative Fac-
problem, took seven minutes by hand. Medium-size prob- tory Sites-a Case Study," Operations Research Quarterly,
Volume 13, No. 4, 1962.
lems (50 X50) have been solved in 30 to 40 minutes with (5) COOPER, L., "Location-Allocation Problems," Operations
only a desk calculator. Thirdly, because it is an additive Research, Volume 11, No. 3, May-June, 1963.
algorithm, the solution time goes up linearly as problem (6) CURRY,G. L. AND SKEITH,R. W. "A Dynamic Program-
ming Algorithm for Facility Location and Allocation,"
size increases, rather than exponentially as do most of the AIIE Transactions, Volume 1, No. 2 (June, 1969).
other algorithms proposed. Fourthly, the nonlinear objec- (7) DAVIS,P. S. AND RAY,T. L., "A Branch-Bound Algorithm
tive function, nonlinear constraints, and discrete (zero- for the Capacitated Facilities Location Problem, Naval Re-
search Logistics Quarterly, Volume 16, No. 3, September,
one) domains of the decision variables can be handled 1969.
with ease. Fifthly, the algorithm may be used to generate (8) EFROYMSON, M. A. AND RAY,T. L. "A Branch-Bound Al-
a very good (and usually exact) initial solution which can gorithm for Plant Location," Operations Research, Volume
14, No. 3, May-June, 1966.
then be used as a starting point for other more elaborate (9) EILON,S. AND DEZIEL,D. P., "Siting a Distribution Centre,
methods. Sixthly, the algorithm is general in nature and An Analogue Computer Application," Management Science,
can be used for solving a rather broad class of zero-one, Volume 12, No. 6, February, 1966.
(10) FELDMAN, E., LEHRER,F. A., AND RAY,T. L., "Warehouse
nonlinear covering (location-allocation) problems (19). Location Under Continuous Economics of Seale," Manage-
Seventhly, it is capable of providing a solution to prob- ment Science, Volume 12, No. 9, May, 1966.

338 AIIE Transactions Volume I1 No. 4


(11) HALEY,K . B., " T h e Siting of Depots," International Journal (24) T O N G E ,F. M., " T h e U s e o f Heuristic Programming in
Product Research, Volume 2, 1963. Management Science," Management Science, Volume 7,
(12) K U E H N A. , A., A N D HAMBURGER, M . J., " A Heuristic Pro- April, 1961.
gram for Locating Warehouses," Management Science, (25) W A N T Y ,J. R., " A Practical Application o f t h e Warehouse
Volume 10, July, 1963. Problem: Location of Depots i n a Petroleum Company,"
(13) M A N N EA, . S., "Plant Location Under Economics o f Scale-- operations Research Quarterly, Volume 9, No. 2, 1958.
Decentralization and Computation," Management Science, (26) "Physical Distribution Costs," Distribution Age, Decembea
Volume 11, November, 1964. 1963.
(14) MARANZANA, F . E., " O n t h e Location o f Supply Points t o
Minimize Transportation Costs," Operations Research
Quarterl?), Volume 15, No. 3, 1964.
(15) M I E H L E ,W . , "Link Length Minimization i n Networks,"
Operations Research, Volume 6 , March-April, 1958.
(16) OPENGART,A., " T h e Challenge i n Physical Distribution,"
The Journal of Industrial Engineering, Volume 19, No. 4,
April, 1968.
(17) RANDOLPH, P. H., S W I N S O NS. , E., and W A L K E R ,M . E .
" A Nonlinear Programming Warehouse Allocation Prob- Dr. Shannon i s associate professor and chairman of industrial
lem," paper presented a t T I M S / O R S A joint meeting, San
Francisco, California, M a y , 1968. and systems engineering at the University of Alabama in Hunts-
(18) SA, G., "Branch-and-Bound and Approximate Solutions t o ville. H e i s also head of the Systems and Computer Sciences Sec-
t h e Capacitatee-Plant-Location Problem," Operations Re- tion of the University of Alabama Research Institute, which i s
search, Volume 17, No. 6 , November-December, 1969. engaged in military and urban operations research. H e holds an
(19) S H A N N O NR., E . and ~ G N I Z I O J.
, P., " A Method t o Achieve M S E from the University of Alabama and a B S I E and P h D
O p t i m u m Air Defense Sensor Allocation," University o f from Oklahoma State University. Dr. Shannon is a member of
Alabama Research Institute, Report No. 51, April, 1968. O R S A , T I M S , and A S E E .
Downloaded by [UQ Library] at 19:46 06 November 2014

(20) S H E N O YB, . V . and SIFFERD,D. W., " A Heuristic Simula-


tion Approach for Locating Warehouses and Plant-Facili-
ties," paper presented at T I M S / O R S A joint meeting, San
Francisco, California, M a y 1968.
(21) S I M O N H., A., "Modeling H u m a n Mental Processes," Pro- M r . Ignizio i s currently on a leave of absence from the Univer-
ceedings of the 1961 Western Joint Computer Conference. sity of Alabama Research Institute, where he was a Senior En-
(22) SPIELBERG, K., " A n Algorithm for t h e Simple Plant Loca-
tion Problem w i t h Some Side Conditions," Operations Re- gineer engaged in military operations research, to pursue gradu-
search, Volume 17, No. 1, January-February, 1969. ate study at Virginia Polytechnic Institute. Previously, he was a
(23) SPIELBERG,K., "Plant Location w i t h Generalized Search project engineer with North American Aviation. M r . Ignizio
I
Origin," Management Sciences, Volume 16, No. 3, Novem- holds a B S E E from the University of A k r o n and a n M S E
ber, 1969. from the University of Alabama in Huntsville.

December 1970 AIIE Transactions 339

You might also like