You are on page 1of 7

Available online at www.sciencedirect.

com

Omega 31 (2003) 205 – 211


www.elsevier.com/locate/dsw

A simple heuristic for solving small xed-charge


transportation problems
Veena Adlakhaa;∗ , Krzysztof Kowalskib
a University
of Baltimore, 1420 N. Charles Street, Baltimore, MD 21201, USA
b Department of Transportation, State of Connecticut, 121 Cimarron Road, Middletown, CT 06457, USA
Received 14 December 2001; accepted 29 January 2003

Abstract
The xed-charge transportation problem (FCTP) is an extension of the classical transportation problem in which a xed
cost is incurred, independent of the amount transported, along with a variable cost that is proportional to the amount shipped.
The introduction of xed costs in addition to variable costs results in the objective function being a step function. Therefore,
xed-charge problems are usually solved using sophisticated analytical or computer software. This paper deviates from that
approach. It presents a simple heuristic algorithm for the solution of small xed-charge problems. We present numerical
examples to illustrate applications of the proposed method.
? 2003 Elsevier Science Ltd. All rights reserved.

Keywords: Fixed-charge; Transportation problem; Heuristic algorithm

1. Introduction on the optimal solution to the variable portion of the prob-


lem, which he admitted himself is a weak start for problems
The xed-charge transportation problem (FCTP) is a with big xed charge values. Sadagopan and Ravindran [2]
mathematical programming problem in which a xed cost, improved the situation with a dual approach by solving the
sometimes called a setup cost, is incurred if another related xed charge portion of the problem and switching to it if
variable assumes a nonzero value. The problem has a wide it gives better result than Murty’s method. As they demon-
variety of classic applications that have been documented strated, their technique provides a considerable improvement
in the scheduling and facility location literature. Two of the over Murty but still requires analyzing a sizable domain of
most common of these arise (1) in making warehouse or distributions.
plant location decisions, where there is a charge for opening One approach to solving FCTP involves a mixed integer
the facility, and (2) in transportation problems, where there programming formulation. Gray [3] has attempted to pro-
are xed charges for transporting goods between demand vide an exact solution to this problem by decomposing it
and supply points. into a master integer program and a series of transporta-
While even a sizable transportation problem (TP) can be tion sub-programs. In contrast Palekar et al. [4] and Stein-
solved relatively easily by hand, such is not the case with berg [5] attempted to provide exact algorithms based on
a FCTP. The best widely known exact method of ranking the branch-and-bound method. The exact branch and bound
extreme points requires analyzing a big domain of load dis- method is not practical for solving a FCTP by hand. Find-
tributions. The size of the domain depends on the initial dis- ing a solution to a 4 by 4 problem requires analyzing sev-
tribution. Murty [1] suggested an initial distribution based eral hundred distributions. There are many instances (e.g., in
the area of solid-waste management, Walker [6]) in which
∗ Corresponding author. Tel.: +1-410-837-4969; fax: +1-410- the dimensions of the problem are relatively small, and for
837-5675. which a solution via a simple heuristic procedure does in-
E-mail address: vadlakha@ubmail.ubalt.edu (V. Adlakha). deed have merit. Solving such problems manually has the

0305-0483/03/$ - see front matter ? 2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0305-0483(03)00025-2
206 V. Adlakha, K. Kowalski / Omega 31 (2003) 205 – 211

added advantage of providing the analyst with a thorough 


m
xij = bj for j = 1; 2; : : : ; n; (3)
understanding of how optimality is achieved.
i=1
This paper develops a relatively simple procedure to solve
xij ¿ 0 for all (i; j);
a FCTP. Managers without a strong mathematical back-
ground can master it. Although the proposed method is more
yij = 0 if xij = 0;
time consuming than the algorithms for solving a regular TP,
it provides a good foundation for solving small problems by yij = 1 if xij ¿ 0:
hand. It also introduces an optimality test. The method uses
elements of heuristic approaches developed by Cooper and Without loss of generality, we assume that
Drebes [7], Drenzler [8], Steinberg [5], Cooper [9], Walker m  n

[6], and Sun et al. [10]. These methods try to reach the opti- ai = bj
i=1 j=1
mum through simplex like iterations. Since there is no guar-
antee that the obtained minimum solution is global and not ai ; bj ; cij ; fij ¿ 0:
local, they experiment with substituting diHerent combina- Despite its similarity to a standard TP problem, FCTP
tions of variables in an eHort to obtain a better solution. None is signi cantly harder to solve because of the discontinuity
of these methods puts much emphasis on the initial solution. in the objective function Z introduced by the xed costs.
Other well-known heuristic approaches are those by Diaby Hirsch and Danzig [14] established that the feasible region
[11], and Kuhn and Baumol [12]. Sandrock [13] presents of FCTP is a bounded convex set with a concave objective
a simple algorithm for the solution of small, xed-charge function. An optimal solution occurs at an extreme point of
problems where the xed charge is associated with the sup- the constraint set, and for a non-degenerate problem with
ply points instead of the routes considered in this paper. all positive xed costs, every extreme point of the feasible
The proposed method tries to obtain the best solution in region is a local minimum. We will use this property when
the rst iteration and uses the simplex like iterations either checking the optimum conditions for a derived solution.
to verify the minimum conditions, or, eventually, to perform The proposed algorithm consists of two parts. In the rst
the nal convergence. When solving a problem by hand, the part, the algorithm nds a strong initial feasible solution, and
latter is done by using a simpli ed “one step pivot method” in the second part it develops steps for the search and veri-
introduced by Murty [1]. Those willing to accept a chance cation of the optimum. The procedure developed in the
that the achieved best solution in the rst iteration may not second part can be used independently with other crude ap-
be optimal, can skip the second part. It must be noted that proximation methods, but this can result in a signi cant
the rst part of the proposed method can also be used to increase in eHort. In addition, the same procedure can be
solve large xed charge transportation problems which can- used to nd a minimum for any concave function with
not be solved by other heuristic methods due to a shortage transportation-type constraints.
of computer memory or a limit on the computing time.

2. Fixed-charge transportation problem 3. Part I—initial feasible solution


The xed-charge transportation problem can be stated as
The process of nding a strong initial solution for our
a distribution problem in which there are m suppliers (ware-
method depends on Balinski’s method and the Vogel Ap-
houses or factories) and n customers (destinations or de-
proximation Method (VAM). We will also “Hungarize” the
mand points). Each of the m suppliers can ship to any of
pre-VAM matrix that is known to improve this method.
the n customers at a shipping cost per unit cij (unit cost for
The reader may refer to Hillier and Lieberman [15] and
shipping from supplier i to customer j) plus a xed cost fij ,
Wagner [16] for coverage of the VAM technique and cover-
assumed for opening this route. Each supplier i = 1; 2; : : : ; m
age of classical TP. Balinski [17] observed that there exists
has ai units of supply, and each customer j = 1; 2; : : : ; n has
an optimal solution to the relaxed version of FCTP (formed
a demand of bj units. The objective is to determine which
by relaxing the integer restriction on yij ), with the property
routes are to be opened and the size of the shipment on those
that
routes, so that the total cost of meeting demand, given the
supply constraints, is minimized. yij = xij =mij ; (4)

Standard FCTP formulation: where


m  n
mij = min(ai ; bj ): (5)
P: Minimize Z = (cij xij + fij yij ); (1)
i=1 j=1 So, the relaxed transportation problem (RTP) of an FCTP
Subject to would be simply a standard TP with unit transportation costs
n as Cij = cij + fij =mij . We refer to this problem as P  and use
xij = ai for i = 1; 2; : : : ; m; (2) this RTP formulation of FCTP to develop an initial feasible
j=1 solution. The optimal solution {Xij } to the RTP problem P 
V. Adlakha, K. Kowalski / Omega 31 (2003) 205 – 211 207

can be easily modi ed into a feasible solution of {Xij ; yij } Step 6: Reduce/adjust the FCTP matrix by deleting the
of P as follows: recent satis ed constraint/s.
yij = 0 if Xij = 0; Step 7: If the loading assignment is not complete, go to
Step 1.
and Step 8: Stop. Record the solution as the initial feasible
yij = 1 if Xij ¿ 0: solution (extreme point), x(1) .
In case of a tie in VAM coeNcients in Step 4, pick the lo-
Balinski shows that the optimal value of RTP provides a cation based on the other (row or column) VAM coeNcient.
lower bound on the optimal value Z ∗ (P) of FCTP and
  Remark 2. For the initial solution x(1) = {Xij(1) },
Cij Xij 6 Z ∗ (P) 6 (cij Xij + fij yij ): (6)  
Cij Xij 6 Cij Xij(1) (8)

Remark 1. Since FCTP is a discrete function, the lower since {Xij(1) } is a feasible solution to P  and Xij is the optimal
bound in Eq. (6) from the optimal solution of the RTP can solution to problem P  .
be rounded up to the nearest interval.
Note that after deleting the recently satis ed constraint
3.1. A pre-screening algorithm in Step 6, we reformulate Balinski’s RTP matrix for the
remaining problem. By construction, successive mij values
Before we consider the RTP formulation of an FCTP by decrease as either row or column constraints are satis ed,
Balinski [17], we introduce a “Pre-screening algorithm” to resulting in increased cij + fij =mij values for the reduced
identify all xed costs fst∗ that will appear in every possible RTP matrix.
distribution (xij ; yij ) for a given FCTP. This phenomenon is
especially present in many small problems.
Remark 3. For the initial solution x(1) = {Xij(1) }
Reformulating Eq. (1), we get 

Z = fst∗ + min {cij xij + (fij yij | i = s Z ∗ (P) 6 (cij Xij(1) + fij yij(1) ) (9)

and j = t)}; (7) since {Xij(1) ; yij(1) } is a feasible solution to problem P where
yij(1) = 1 if Xij(1) ¿ 0 and Z ∗ (P) is the optimal value of the
where
  problem P.
fij = fst∗ if ai ¡ bt or b j ¡ as :
i=s j=t

After identifying a xed cost fst∗ , the value of fst is set to 4. Part II—improvement and veri"cation of the optimum
0 to solve the FCTP and this xed cost is added to the cost
of nal solution. Please see Adlakha and Kowalski [18] for The second part of the proposed method is based on the
more details on the pre-screening algorithm. concavity of the objective function. Let {x(1) ; x(2) ; : : : ; x(k) }
Hungarization is a well-known process used in optimiza- be a sequence of extreme points (alternatively referred to
tion. The process involves what is called matrix reduction as peaks) of the xed-charge problem ranked in decreasing
through row and column reduction. Hungarization consists order of value of objective function Z(x). Starting with the
of subtracting the smallest element in each row from every initial solution x(1) , the algorithm recursively investigates all
element in that row. Then, using the row-reduced matrix, adjacent peaks of the best solution obtained until that point.
the smallest element in each column is subtracted from ev- If Z(x(k) ) ¡ Z(x) for all adjacent peaks of x(k) , then solution
ery element in that column. The problem presented by this x(k) is considered the nal solution of the FCTP.
reduced matrix is equivalent to the original problem in the The adjacent peaks can be found through perturbing the
sense that the same solution will be optimal. load distribution obtained in Part I by using elements of
Murty’s “one step pivot operation”. This method is based
3.2. Procedure for the initial solution on the fact that the FCTP has the same constraints as regu-
lar TP with identical set of ai and bj values. In the case of a
Step 1: Pre-screen the FCTP to identify and separate any transportation tableau, the search simply translates into per-
xed costs fst∗ . turbing each load using only one stepping-stone move. Each
Step 2: Formulate Balinski’s RTP matrix with Cij = cij + perturbation should result in the complete relocation of at
fij =mij . least one load (emptying one location). It must be noted that
Step 3: Hungarize the RTP matrix. some loads can be perturbed several times while involving
Step 4: Identify critical location/s using VAM. a diHerent pair each time. Each obtained distribution will
Step 5: Load the VAM critical location, say cell (s; t), correspond to an adjacent peak. If any peak yields a lower
with the maximum feasible amount. value of Z, we move to this location and again analyze the
208 V. Adlakha, K. Kowalski / Omega 31 (2003) 205 – 211

adjacent locations. This process is repeated until the opti- Steps 3 & 4: The VAM coeNcients of the Hungarized
mality conditions are ful lled. RTP matrix are as follows, with ∗ identifying the critical
Speci cally, the steps for Part II are as follows: VAM location.
Step 1: Set i = 1.
Step 2: Determine all peaks adjacent to x(i) . Let x∗ denote VAM coeNcients 0 1.67 0
the peak with the smallest Z(x).
Step 3: If Z(x(i) ) ¡ Z(x(∗) ), go to Step 5. 1.67 0 1.67 3
Step 4: Set i = i + 1. Let x(i) = x∗ . Go to Step 2. 0 1.83 0 0
Step 5: Stop. Record x(i) as the optimal solution of FCTP. 1.92 0∗ 1.92 2.17
1.5 1.5 2.17 0

5. Illustrative numerical examples Step 5: Load cell (3, 1) with 20.


Step 6: The adjusted reduced FCTP is as follows:
The proposed two-part algorithm constitutes the rst
“down to earth” heuristic method for nding a good solution
b2 b3 Supply
for a FCTP. As demonstrated by the following numerical
examples, the minimum found by the proposed method will a1 (30, 3) (20, 4) 10
most likely be the global solution for small problems. To a2 (30, 2) (20, 1) 30
illustrate steps of the algorithm, we consider the example a3 (30, 4) (20, 3) 20
from Balinski [17] (see Table 1). a4 (30, 5) (20, 2) 20
Demand 50 30
5.1. Part I—:nding the initial solution
Step 7: Since the loading assignment is not complete, we
Steps 1 & 2: No need to carry out Step 1 here. Formulating go back to step 1.
this FCTP as an RTP following Balinski [17], we obtain The VAM coeNcients of the Hungarized RTP matrix of
Table 2. the reduced FCTP are as follows:
An optimal solution to this RTP is X12 =10; X22 =20; X23 =
10; X31 =20; X32 =20, and X43 =20, to ship 100 units at a cost VAM coeNcients 1.33 0
 
Cij Xij = $321:70. Since (cij Xij + fij yij ) = 360,
0 0 0
Eq. (6) implies 321:70 6 Z 6 360. Note that the minimum
1.33 1.33 0
interval between the xed cost value fij is equal to 10 and
1.5 1.5 0
the minimum load is also a multiple of 10. Therefore, using
3.5 3.5 0∗
Remark 1, we can round up the lower bound to the optimal
value Z of the FCTP to 330.
Load cell (4, 3) by 20. The adjusted reduced FCTP is as
Table 1 follows:
Cost matrix (fij ; cij ) for numerical Example 1
b2 b3 Supply
b1 b2 b3 Supply
a1 (30, 3) (20, 4) 10
a1 (10, 2) (30, 3) (20, 4) 10 a2 (30, 2) (20, 1) 30
a2 (10, 3) (30, 2) (20, 1) 30
a3 (30, 4) (20, 3) 20
a3 (10, 1) (30, 4) (20, 3) 40
a4 (10, 4) (30, 5) (20, 2) 20
Demand 50 10
Demand 20 50 30
Using the Pre-screening algorithm, since a1 + a2 ¡ b1 and
a1 + a3 ¡ b1 , the modi ed cost matrix is

(30, 3) (20, 4)
Table 2
Balinski cost matrix Cij = (cij + fij =mij ) for numerical Example 1
(0, 2) (20, 1)
(0, 4) (20, 3)
b1 b2 b3 Supply

a1 3 6 6 10 Continuing the steps of Part I, we obtain an initial solution


a2 3.5 3 1.67 30 x(1) of the FCTP of Table 1 as X13 = 10; X22 = 30; X31 =
a3 1.5 4.75 3.67 40 20; X32 = 20, and X43 = 20, to ship 100 units at a cost of
a4 4.5 6.5 3 20 Z(x(1) ) = 350. Therefore, using Remark 3 and this solution,
Demand 20 50 30 330 6 Z ∗ 6 350.
V. Adlakha, K. Kowalski / Omega 31 (2003) 205 – 211 209

Table 3
Cost matrix (fij ; cij ) for Example 2

b1 b2 b3 b4 Supply

a1 (900, 760) (1000, 71) (700, 283) (800, 594) 50


a2 (900, 594) (300, 64) (700, 170) (600, 564) 15
a3 (600, 594) (200, 69) (400, 79) (0, 202) 5
Demand 25 20 15 10

Table 4
Modi ed cost matrix (fij ; cij ) for numerical Example 2

b1 b2 b3 b4 Supply

a1 (0, 760) (1000, 71) (700, 283) (800, 594) 50


a2 (900, 594) (300, 64) (700, 170) (600, 564) 15
a3 (600, 594) (200, 69) (400, 79) (0, 202) 5
Demand 25 20 15 10

5.2. Part II—improvement and veri:cation Table 5


Balinski matrix Cij = (cij + fij =mij ) of RTP of Example 2
The solution x(1) obtained in Part I is degenerate
b1 b2 b3 b4 Supply
with 5 loaded cells. Now we investigate adjacent peaks
to verify if an improvement to the initial solution is a1 760 121 329.66 674 50
possible. a2 654 84 216.66 624 15
The following solutions are obtained. a3 714 109 159 202 5
Peak 1: X12 = 10; X22 = 20; X23 = 10; X31 = 20; X32 = 20, Demand 25 20 15 10
and X43 = 20, with Z(peak 1) = 360.
Peak 2: X12 = 10; X22 = 30; X31 = 20; X32 = 10; X33 = 10,
and X43 = 20, with Z(peak 2) = 360.
Peak 3: X11 = 10; X22 = 30; X31 = 10; X32 = 20; X33 = 10, algorithm, since b2 +b3 +b4 ¡ a1 , the cost matrix in Table 3
and X43 = 20, with Z(peak 3) = 360. is modi ed as in Table 4. Formulating this FCTP as an RTP
Peak 4: X13 = 10; X21 = 20; X22 = 10; X32 = 40, and following Balinski [17], we obtain Table 5. Cell (3, 4) is
X43 = 20, with Z(peak 4) = 430. critical using VAM coeNcients of the Hungarized RTP ma-
Peak 5: X13 = 10; X22 = 30; X32 = 20; X33 = 20, and trix. Set X34 = 5. Delete the third row. The reduced/adjusted
X41 = 20, with Z(peak 5) = 430. RTP matrix after pre-screening is:
Peak 6: X13 = 10; X22 = 30; X31 = 20; X33 = 20, and
b1 b2 b3 b4 Supply
X42 = 20, with Z(peak 6) = 390.
Peak 7: X13 = 10; X22 = 10; X23 = 20; X31 = 20; X32 = 20, a1 (0, 760) (0, 71) (700, 283) (800, 594) 50
and X42 = 20, with Z(peak 7) = 420. a2 (900, 594) (300, 64) (700, 170) (600, 564) 15
Demand 25 20 15 5
Since all extreme points adjacent to initial solution x(1)
have higher value of the FCTP objective function, we stop with the corresponding Balinski RTP Matrix
and consider solution x(1) as the optimum solution. As
veri ed by the authors using the Ranking Extreme Points b1 b2 b3 b4
Method [1], the solution x(1) is indeed the global optimum.
It should be noted that Balinski’s method also yielded a a1 760 71 329.66 754
quite satisfactory result with a cost of 360 (only 10 more a2 654 84 216.66 684
than the best solution).
Now cell (2, 3) is VAM critical from the Hungarized RTP
matrix. Set X23 = 15 and delete the second row and third
column. We obtain an initial solution x(1) of FCTP of Table
5.3. Illustrative numerical Example 2 3 as X34 = 5; X23 = 15; X11 = 25; X12 = 20, and X14 = 5
at a cost of Z(x(1) ) = 30350. The value of RTP solution
We next consider the example from Sadagopan and (Table 5) at x(1) is 29950 (29050+900). Using this solution
Ravindran [2] (see Table 3). Using the Pre-screening 29950 6 Z ∗ 6 30350.
210 V. Adlakha, K. Kowalski / Omega 31 (2003) 205 – 211

Table 6 Note that those practitioners willing to accept a chance


Comparison of computational eNciency that the achieved best solution in the rst iteration may not
Ratio of xed to Number of stages of ranking
be the optimal one can skip the second part. This latter part
variable costs at optimum is done by using a simpli ed “one step pivot method” intro-
Murty’s S&R’s Proposed duced by Murty [1], who develops a pattern of search among
method method method the extreme points of FCTP for determining the global min-
imum. His pattern of search is essentially a ranking of ex-
0.126 21 12 1 treme points in increasing order of the variable costs. The
0.219 24 14 1
xed costs are then added to each ranked vertex in deter-
0.312 26 14 1
0.404 26 14 1
mining the optimal solution.
0.590 27 16 1 For many problems of relatively small dimensions, a so-
lution via a simple “hands on” procedure does have merit.
The typical solid-waste management problem, in which op-
timum strategies are sought regarding the location and de-
gree of activity with tip sites, transfer stations and collection
An investigation of the adjacent 6 peaks provides solu-
depots, falls into such a category. Solving such xed-charge
tions with the value of the objective function ranging from
problems manually has the added advantage of providing
31265 to 38055. Hence we stop with x(1) as the solution of
the analyst with a through understanding of how the opti-
FCTP. As shown by Sadagopan and Ravindran [2], after 12
mality was achieved. Future work should seek the devel-
stages of ranking, x(1) is the optimal solution.
opment of an eNcient code to implement this approach for
computational and comparative studies.

6. Concluding remarks
References
We have developed a simple, eNcient heuristic procedure
for solving small xed-charge transportation problems. The
[1] Murty KG. Solving the xed charge problem by ranking the
proposed method obtains the best initial solution in Part I and extreme points. Operations Research 1968;16:268–79.
uses simplex like iterations in Part II to improve that solution [2] Sadagopan S, Ravindran A. A vertex ranking algorithm for the
and to verify its optimality. It must be noted that the rst xed-charge transportation problem. Journal of Optimization
part of the proposed method can also be used to solve large Theory and application 1982;37:221–30.
xed charge transportation problems which cannot be solved [3] Gray P. Exact solution of the xed-charge transportation
by other heuristic methods due to a shortage of computer problem. Operations Research 1971;19:1529–38.
memory or a limit on the computing time. As demonstrated [4] Palekar US, Karwan MH, Zionts S. A branch-and-bound
by the numerical examples of small problems, the initial method for the xed charge transportation problem.
solution of Part I is most likely the global solution. Management Science 1990;36:1092–105.
To investigate the computational eNciency of the pro- [5] Steinberg DI. The xed charge problem. Naval Research
Logistics Quarterly 1970;17:217–35.
posed method, we follow the comparison model used by
[6] Walker WE. A heuristic adjacent extreme point algorithm for
Sadagopan and Ravindran [2]. We solved the numerical Ex-
the xed charge problem. Management Science 1976;22:587
ample 2 illustrated above by increasing all the xed costs –96.
by 500, 1000, 1500, and 2500. Murty’s ranking method [1] [7] Cooper L, Drebes C. An approximate algorithm for the
was used to determine the optimal solution. The results re- xed charge problem. Naval Research Logistics Quarterly
ported by Sadagopan and Ravindran [2] as compared to the 1967;14:101–13.
proposed method are presented in Table 6. [8] Drenzler DR. An approximate method for the xed charge
The proposed heuristic delivered the optimal solution us- problem. Naval Research Logistics Quarterly 1969;16:411–6.
ing only one stage of ranking regardless of the ratio of xed [9] Cooper L. The xed charge problem—I: a new heuristic
costs to variable costs. Murty [1] presents an example of method. Computers & Mathematics with Applications
FCTP of size 5 by 7 and ranks all extreme points to nd the 1975;1:89–95.
optimal solution. The proposed method delivered the ini- [10] Sun M, Aronson JE, McKeown PG, Drinka D. A
tial solution obtained in Part I in the following sequence: tabu search heuristic procedure for the xed charge
transportation problem. European Journal of Operational
X27 =26; X12 =9; X43 =35; X14 =14; X35 =8; X47 =9; X54 =
Research 1998;106:441–56.
40; X46 = 31, X36 = 24; X31 = 6, and X51 = 16. This matches
[11] Diaby M. Successive linear approximation procedure for
the optimal solution obtained by Murty. Though any de nite generalized xed-charge transportation problems. Journal of
conclusion needs an extensive computational study, the pro- Operational Research Society 1991;42:991–1001.
posed method demonstrates signi cant improvement over [12] Kuhn HW, Baumol WJ. An approximation algorithm for the
existing heuristic methods for small xed-charge transporta- xed charge transportation problem. Naval Research Logistics
tion problems. Quarterly 1962;9:1–15.
V. Adlakha, K. Kowalski / Omega 31 (2003) 205 – 211 211

[13] Sandrock K. A simple algorithm for solving small xed-charge [16] Wagner HM. Principles of operations research. Englewood
transportation problems. Journal of Operational Research CliHs, NJ: Prentice-Hall, 1969.
Society 1988;39:467–75. [17] Balinski ML. Fixed cost transportation problems. Naval
[14] Hirsch W, Danzig GB. The xed charge problem. Naval Research Logistics Quarterly 1961;8:41–54.
Research Logistics Quarterly 1968;15:413–24. [18] Adlakha V, Kowalski K. On the xed-charge transportation
[15] Hillier FS, Lieberman GJ. Introduction to operations research, problem. OMEGA: The International Journal of Management
4th ed. San Francisco, CA: Holden-Day, 1986. Science 1999;27:381–8.

You might also like