Академический Документы
Профессиональный Документы
Культура Документы
Leon Cooper
Southern Methodist University, Dallas, Texas
(1)
In what follows, let us designate the set TF as all tii = \wj} satisfying the constraints
of (1).
The relation of the problem given in (1) to the transportation problem and the
location-allocation problem is readily seen. If the 4'(XD,, yo,; Xi, y.) =o,,, a set of
constant costs, and if a,,;3, is designated as y,,, then (1) can be written as:
min 21= : : ? J2',-i y./^.,,
weW,
(2)
which is a usual form of the transportation problem. Hence, for a fixed set of costs,
problem (1) becomes a transportation problem.
Returning to equations (1), we may also note that (a) if the requirements are
stated in terms of the P, and (b) the capacities associated with each source are unlimited, then the set of m-\-n constraints of (1) are no longer explicitly required and
(1) now becomes:
2^ T
W E SHALL REFER to the problem stated in equations (1) as the general transportation-location problem,. There are a number of important observations that can
be made concerning this problem.
T H E O R E M 1. A necessary and sufficient cmdition for (1) to have a feasible set of w^
is
that-',-'r^^Zl^Ci.
Proof. This result is the same as for a transportation problem (see reference 8).
In what follows and throughout this paper, it can be shown that the functions
f(xD,, ym; *., yi) are such that (1) has a finite minimum. The next theorem is an
important result.
THEOREM 2.
weW,
96
Leon Cooper
THE TRANSPORTATION-LOCATION PBOBLEM
LET US NOW consider the problem of equations (1) for a particular form of the
function 4/:
^(xo,, yv,; X., y,) = [(xD,-x,)'+(y:>i-y,)']"\
(4)
Hence, what we shall now call the transportation-location problem is the following:
as z(x, y, iv). Consider the special case where all w,,=0 except Wu- Further,
let all Xi, y, be fixed and known for i>l and let 2/1 = 2/01 and xm>xi.
Hence z(x,
y, w) =zi(xi, Wn)+K, where K is the sum of all the constant terms. If 2, is neither
convex nor concave over all of the space defined by (x, y, w), then the same will be
true of 2. Hence, we examine 2i(xi, w^u) =/SIM'U(XDI-XI).
It is well known'"
that a function such as Zi will be neither convex nor concave if the Hessian matrix
associated with 21 is neither positive nor negative semidefinite. The Hessian matrix
for z is
(8)
which is indefinite. Hence 21 is neither convex nor concave in the portion of the
space defined by xi and ;. Therefore, z(x, y, w) is neither convex nor concave
over all of the space defined
hy(x,y,iv).
This proof shows that the objective function 2 is not a convex or a concave function over all possible values of x, y, iv. A more pertinent question is whether or
not the objective function is convex or concave over values of x, y, and w limited
by the constraint set of the transportation-location problem. The answer to this
is given in the following result.
THEOREM 4. The function 2= E S ^ E i - i " ;8,w.,[(xBy-x.)'+(2//>,-j/.)T rteither a convex nor a concave function of the vector (x,y,w) over the constraint set of (7).
Proof. Consider the following problem:
subject to
(9)
where D.y=l(xBy-x.)'+(2/B-2/i)'rIn order to simplify the argument but allow sufficient variation over the constrMnt set, we assume that (xj, j/,) = (x, j/) and further that v,=vm = Wn. and
XDI> XI, xta> Xl. This simplifies the objective function to:
2(Xi, Wu, toil, UI-i)
where
='Pl'Wll(xDlXi)+^Wu(XmXi
KD^XDIXM-
Since the Hessian matrix is indefinite for the problem given by (9), we see that, in
general, the objective function is neither convex nor concave over the constraint
set of the transportation-location problem.
A consequence of Theorem 4 is that the transportation-location problem, which
is given by equations (7), is a nonconvex nonlinear programming problem in which
we are minimizing a nonconvex objective function over a convex set. However,
because of Theorem 2, we can state the equivalent result for the transportationlocation problem.
T H E O R E M 5.
r,
some optimal solution wiU (x
to (7).
I extreme
point
w,W, (7)
X set of feasible
solution!
Proof. The problem above is a special case of the general formulation proven
in Theorem 2.
The importance of this result is that it makes it unnecessary to consider any
but basic feasible solutions to the constraints of (7). This will be important when
we consider computational approaches in the next section of this paper.
AN EXACT ALGORITHM FOR THE TRANSPORTATION-LOCATION PROBLEM
THE FIRST algorithm we shall consider is exact and relatively simple in concept.
However, its use will be limited, as will be evident, to relatively small problems.
We may note, according to Theorems 2 and 5, that an optimal solution t the
transportation-location probleni_will occur at an extreme point of the convex set of
solutions, W=\iv,,\Aw^b,w^O\.
We know that the extreme points of W correspond to basic feasible solutions of Aw^b, w^O. If the number of basic feasible
solutions is designate<l as Ners, for an m-source, n-destination problem, NarsS
( T^ ) "^^ follows from the basic theory of the transportation problem.'"
98
leon Cooper
\
1 is the number of basic solutions, most of which are infeasible.
TO+n-1/
Hence, it is usually the case that NBrs.("^^_^\
For example, in the simple
example we shall present, with m = 2, n = 4, f _ ^ _ j j = U j = 8!/5!3! = 56.
However, there are only nine different nondegenerate basic feasible solutions.
According to results of DEMUTH'" AND DOIG,'" the minimum number of basic
feasible solutions, for mgn, to a transportation problem of order mXn is w!/(n
m+1)! Form = 2, n=4, wehave4!/3! = 4. Hence, the number of bases we have
had to examine, 9, is much closer to 4, the minimum number, than to the unrealistic
upper bound, 56. This is somewhat encouraging.
In any case, Nara is a finite number. Suppose we generate all basic feasible
solutions. Let us designate any such solution t6={iol- (We consider, subsequently, how to do this.) For each such solution, we can then solve the problem:
by considering this problem as a set of problems of the form:
We can do this, since the {iJ,,| are simply known nonnegative constants or weights
An iterative technique for solving problems of the form given by equations (13)
was given by the author in references 2 and 3, and is repeated here for convenience.
The initial estimates of (z,, j/,) are given by
(16)
If we now designate the minimum value of z for the 2th basic feasible solution as z*,
then it is obvious that the optimal value of the objective function 2* for the transportation-location problem will be:
/ = minV,2i*.
(17)
If the minimum is taken on at l=s, then the optimal values of the variables are
(X,,, Vi.), t = l , ,m, and it>,y., t = l, ,m,j=l,
,n, where the designation
(xi,, 2/..) indicates (x,, y.) for the sth basic feasible solution and T.y. are the set of
?,j for this solution.
Let us now return to a point we g l o s ^ over, earlier, viz., generating all the
basic feasible solutions to the constraints of the transportation-location problem,
i.e., the constraints:
i:;:r...c.,
w.,gO,
.=i,...,.; Ei-o=r.
1=1,
,m, j = l, , n .
,=i,...,; ^^^^
99
As has already been mentioned, these are simply the constraints of the standard
transportation problem. We can make use of the transportation problem tableau
(see reference 8) to advantage in order to generate only basic feasible solutions.
The alternative would be to find all basic solutions and discard the infeasible solu-
70
70
10
40
90
40
10
80
90
30
120
90
40
70
70
80
80
80
120
40
40
30
90
70
90
80
70
10
40
70
90
40
40
40
70
50
70
90
120
80
120
40
80
120
40
T h e method described
As an example, consider a problem with m =2, n =3. There wiU be at most 2-1-3-1 = 4
noniero values of w.y in a basic feaaibk solution. Suppose c, =80, c = 120, r, =70, r, =90,
r. =40. An initial tableau is shown in Tableau 1 of Table I. The blank squares bave zero
values of u>,y, i.e., wwO, Wsi=O. It is a simple matter to find all basic feasible solutions
100
Leon Cooper
from this initial Tableau 1. We can use the standard 'loop' method for allowing a zero
variable to become positive and still remain feasible (see reference 8). This is merely the
application of the simplex method to the special case of the transportation problem. From
Tableau 1, we can generate two new tableaux. These are given in Tableaux 2 and 3. From
Tableau 2 we can generate one new tableau given in Tableau 4. From Tableau 3 we can
generate Tableau 5. From Tableau 4 we can generate Tableau 5, and from Tableau 5 back
to Tableau 4.
We have now generated all the basic feasible solutions. The relations between them for
this example can be represented as the directed graph shown in Fig. 1.
It can be seen that there are five basic feasible solutions. If all basic solutions had been
found, we would have had to solve four simultaneous equations in four variables, ( 1 = 15
times. The above procedure is simpler and very much less work.
We now state the algorithm we have been discussing for solving the transportation-location problem and then present a numerical example.
A sample problem. Let (xm, J/ci)=(O, 0); (x, ym)=(O, I ) ; (xm, VDZ) =(1, 1); (xw,
yot) = (1, 0); m =2, n =4; c, =50, Cj = 100; n =20, r, =40, r, =60, r, =30. A basic feasible
solution is given in Tableau 1 of Table I I . From Tableau 1 we can generate Tableaux 2, 3
and 4. From Tableau 2 we can generate Tableaux 5 and 6. From 3 we generate 7 and 4.
From 4 we generate 8 and 4. From 5 we generate 6 and 7. From 6 we generate 5 and 8.
TABLE II
20
30
10
20
40
50
60
30
60
30
100
10
2
40
10
20
40
Tableau 1
2
40
20
40
20
c.
30
30
100
60
30
30
40
20
40
50
30
20
2
20
100
30
40
60
40
60
c.
50
100
30
Tablea i4
4
10
c.
50
40
50
30
20
60
30
r.
20
Tableau 5
60
Tableau 3
20
30
rablea
4
50
60
40
10
50
60
20
100
60
30
Tablea
leon Cooper
TABLE n-Cont.
1
50
20
40
20
40
10
30
60
30
100
20
20
40
50
Tableau 7
20
30
60
60
50
100
30
Tableau 8
1
3
20
40
40
4c,
30
40
60
50
100
30
Tableau 9
From 7 we generate 9 and 5. From 8 we generate 6 and 9. From 9 we generate 4 and 8.
Hence we are done. All of these tableaux are shown in Table II. Figure 2 gives the graph
of the nine basic feasible solutions.
For each of these nine basic feasible solutions to the constraints of the transportationlocation problem, two location problems were solved with the results shown in Table III.
It can be seen that the minimum occurs for solution 2. Hence the solution to our problem is:
U)u = 10, 'is=40, wn=O, u)n=O; Wn = 10, teM=O, u>23=60, Wj = 30; (xi, i/i) =(0, 1), (xj, 1/2) =
( l , l ) ; z * =54.143.
103
Basic feasible
solution
(x<, y.)
Z,i
(0,1)
(1,1)
20
40
60
(0,1)
(1,1)
10
44.143
54.143
(1, 1)
(0.641, 0.792)
28.286
65.490
93.776
(1,0)
(1,1)
20
40
60
(0,1)
(1,1)
10
58.284
68.284
(0, 1)
(1,1)
14.142
48.286
62.428
(1, 1)
(0.287, 0.596)
0
69.126
69.126
(1,0)
(1,1)
28.286
48.286
76.572
(1,0)
(0.173,0.924)
20
59.579
79.579
m e t h o d for finding all t h e basic feasible solutions will result t h a t is more efficient
t h a n t h e usual 'loop' m e t h o d . T h i s would greatly improve t h e algorithm presented
here.
HEURISTIC ALGORITHM NO. 1 FOR THE TRANSPORTATIONLOCATION PROBLEM
THE HEURISTIC algorithm described in this section was a first attempt to devise a
rapid suboptimal method for solving transportation-location problems. It was
suggested by a heuristic method previously developed for the pure location-allocation problem and described in reference 3. This method, called the 'alternate
location-allocation method,' is as follows:
1. Select some subset consisting of m of the n destinations that are given and consider
these as source locations.
2. Allocate each of the remaining n -m destinations to the closest of the m sources given
in Step 1.
3. Within each of theTOsets of destinations determined in Step 2, use the iteration method
given in references 2 and 3 [also used in equations (14) and (15) of this paper] to find the
exact location of the optimal source location.
4. Determine for each destination whether or not it is closer to another of the sources
located in Step 3 than the one to which it is allocated. This defines a new grouping of m
subsets of destinations.
5. Repeat Steps 3 and 4 until no further changes are possible.
It is shown in reference 3 that this algorithm is a moderately successful one, but
by no means the best of the several heuristics tested for the pure location-allocation
problem. A modification of this method for the transportation-location problem
can be made as follows.
Alternating Transportation-Location Heuristic
1. Select arbitrarily m of the (XD,, yoi) and let these be the m initial source locations.
This then yields a set of distances between each of the destinations and the assumed sources.
2. Using these distances as cost coefficients 7,y [as in equations (2)], we can solve an
ordinary transportation problem tofinda set of [w,,\.
3. Using the |te,j| from Step 2, we can solve a location problem using equations (14)
and (15) andfinda new set of source locations.
4. We now iterate Steps 2 and 3 until no further changes, to within some tolerance, are
obtained in two successive cycles.
It can be seen that the general notion behind this approximate method is to
locate sources alternately given a pattern of allocations and to determine an allocation given a set of source locations. The location-allocation problem method and
the usual transportation-problem method are alternately applied to perform the
calculations. It can readily be seen that this iteration method yields a convergent
monotone nonincreasing sequence of values for z. However, there is no guarantee
that it will converge to the global maximum we seek. Still, what experience we
have with this and similar algorithms indicates that the result, when not optimal,
lies within -^lO percent and usually within 2-3 percent of the optinval solution.
Table IV indicates the results with this heuristic method for the first seven
problems given in reference 2. These are location-allocation problems with m=2,
n = 7. They were solved as transportation-location problems by using a set of
capacities and requirements such that each destination was supplied by only one
source. The starting points in each of these problems were completely random.
For problems 3 and 4 the optimal solutions were obtained. In the others, results
of varying degrees of closeness to optimality are obtained.
Problem
Allocations
obtained
Z
obuined
Optimal
allocations
(1, 2, 3, 4, 5)
(6,7)
(1,2,4,5)
(3, 6, 7)
52.118
50.450
(1, 3, 4, 6, 7)
(2,5)
(1,3,4,6)
(2, 5, 7)
81.764
72.000
(1,2,3,4)
(5, 6, 7)
(1, 2 , 3 , 4 )
(5, 6, 7)
38.323
38.323
(1. 2, 3, 7)
(4, 5, 6)
(1,2,3,7)
(4, 5, 6)
48.850
48.850
(1, 2, 3, 4, 5)
(6,7)
(1, 2 , 3 , 5 )
(4, 6, 7)
38.560
38.033
(1, 2,3)
(4, 5, 6, 7)
(1, 2,3, 4)
(5,6,7)
44.564
36.175
(1,5,6,7)
(2, 3, 4)
(1,3,4,5,6,7)
(2)
61.935
59.716
In order to study how the results obtained by the heuristic are related to
initial choice of the sources, 17 different (randomly selected) starting values w
chosen and the heuristic then applied for Problem 1 of Table IV. Table V ii
cates the results obtained. It will be noted, from Table IV, that, for problem ni
R PROBLKM 1
Allocation
obtained
obt4ied
(1, 2,3,4,5)
(6,7)
52.118
(1, 2, 4, 5)
(3, 6, 7)
50.450
3 (optimal solution)
(1, 2, 3, 4)
(5, 6, 7)
52.001
(1,3,4,5,6,7)
(2)
59.705
(1, 2, 3, 7)
(4, 5, 6)
60.128
57.672
(1,2,4,5,6)
(3,7)
106
Leon Cooper
ber 1, the optimal allocations are ( 1 , 2, 4, 5), (3, 6, 7) and the optimal value of z
is .50.450.
From the results of Table V we can see that in three of the 17 trials we obtained
the optimal solution. In roughly 60 percent of the trials we obtained a solution no
worse than about 3 percent of optimal and usually better than this. The maximum
error was about 15 percent. However, it can be seen that repeated use of the
heuristic method, with varying starting values, is apt to give a reasonably good
approximation to the optimal solution, if not the optimal solution itself.
""'"lo','
o!!=0.'
3. Using the original matrix A of allocations, we now derive two new matrices, analogous
to Wl of Step 2 above, as follows. Find the pair of points being served by the same source
such that the distance between them is a maximum. Let the sources be 8,,, u = l, ,m,
and let Q be the subsets of destinations. Then we wish to find a pair of points p,, p, as
follows:
Let the subset in which this occurs be Qi, with source . We then eliminate source 8^ and
replace it with points p . , p,. We now have a set of m + 1 sources (, u = l, . , m; u^h),
p., p,. Therefore, we have one more source than is desired. Let ft = ((, u = l,
, m;
U}^h), p., p,\. The source coordinates are (z., ;/) for u.,ih, (XD,, yo,), (XD,, yot). For
notational simplicity, we rename the sources as ,. Therefore, R = [sJ\v = \, , T O + 1 1 ,
where the first m 1 sources are ,,(u^h) and where s = p , , s+i =pt. We now wish to find
the pair of sources s. and % that are closest together. Let us designate the set of indices
V, corresponding to ,eK, as V, i.e., V = l|,eSl. The pair of sources and i is now de-
107
terminedby
Ux.-Xt)*+(y.-yt)']'i'=imiii.,.T-t^,Uxt-x,)'+(yt-y,)']^".
Having found the pair of sources ,that are closest toother, first we eliminate s, and apply
the 'alternate' method of successive location and allocation until convergence, as referred to
under step 1 above. This will determine a second solution. Next we restore source .
and eliminate sj, apply the 'alternate' procedure and obtain a third solution. We calculate
two new matrices Wi and W) corresponding to these solutions, as we did in step 2.
4. Sum the requirements for each subset of destinations served by one source, i.e., calculate E . . . . I-/for each subset Si, i = l,
, m. We now calculate the differences, c.E^.-i r,,
1 = 1,
,m. A negative difference implies a capacity deficit and a positive difference
indicates a capacity surplus.
5. The iterative calculation begins with this step. Let / _ be any set with a capacity
deficit and let /+ be any set with a capacity surplus. Choose a destination point pt in the
.let / _ such that the difference in the distances from pt to the source of / _ and from pt to
the source of 7+ is a minimum. Symbolically, if pt has coordinates {x^t, y,,t) and the source
location of / _ is (a:_, yJ) and that of 7+ is {x+, y+), then we choose p* as the index j from 7_
such that
e'=imn,.,JUx,,-x..y-t(yD,-y-)'V"-l(xD,-x+)'-i-(yo,-y^)'V"\.
6. We now reallocate part or all (if possible) of the requin
7_ to the source in I+. The amount reallocated depends on t
the surplus at 7+. The four cases are
(a) If the (requirement at pt) S (deficit for 7_) and the (surplus for 7^.) (deficit for 7_),
then we reallocate the deficit from 7_ by supplying this amount to p* from 7+ instead of 7_.
(b) If the (requirement at p*)fe(deficit for 7_) and the (surplus for 7+) <(deficit for 7_),
then we reallocate the amount of the surplus at 7+ by supplying this amount to p* from 7+
instead of 7...
(c) If the (requirement at pt) < (deficit for 7_) and the (surplus at 7+)fe(requirement at
Pk), then we reallocate the entire requirement at pt from 7_ to 7+.
(d) If the (requirement at pt) < (deficit for 7_) and the (surplus at 7+) < (requirement at
Pt), then we reallocate the amount of the surplus at 7+ by supplying this amount to pt from
7+ instead of 7...
7. With the new allocations, new source locations are computed for each subset of destinations S, by the use of equations (14) and (15), using the allocations as weights.
8. Each subset of destination points S, is now examined for the points that might be
closer to a different source with an excess capacity. If possible, we reallocate part or all of
its requirement, depending on the amount of the surplus at this source.
9. We now repeat steps 7 and 8 (exact source location and subsequent reallocation to
satisfy requirements), until no further change in the allocation matrix occurs.
10. The entire reallocation process (steps 5-9) is now repeated until all capacity deficits
are removed. The value of z for this allocation matrix Wi is computed.
11. The procedure of steps 4-10 is repeated for allocation matrices W, and W,.
12. The minimum of the three values of z obtained is chosen as the solution, together with
its source locations and destination allocations.
This heuristic method was tested on the eight two-source, seven-destination
problems listed in reference 2. The requirements were generated randomly. The
capacities were chosen all equal, with their sum 5 percent higher than the sum of the
Leon Cooper
requirements. In order to have exact solutions to compare the heuristic against,
the exact extremal equations (14) and (15) were used to solve for all possible allocations that, for these test problems, included the optimal solutions. For these eight
problems, the heuristic method produced the optimal solution.
In reference 4 this basic method was also applied to 100 randomly generated
problems with apparently good results, although the correct solutions were not
known in advance.
1. M. BALINSKI, "An Algorithm for Finding All Vertices of Convex Polyhedral Sets," J.
Soc. Ind. & Appl. Math 9, 72-88 (1961).
2. L. COOPER, "Location-Allocation Problems," Opns. Res. U, 331-343 (1963).
3. , "Heuristic Methods for Location-Allocation Problems," SIAM Review 6, 37-52
(1964).
4.
, "Solutions of Generalized Locational Equilibrium Models," J. of Regional Science
7, 1-18 (1967).
5. 0 . DEMUTH, "A Remark on the Transportation Problem," Casopis pest. mat. 86,103-110
(1961).
6. A. DoiG, "The Minimum Number of Basic Feasible Solutions to a Transport Problem,"
Opnat. Res. Quart. U, 387-391 (1963).
7. W. H. FLEMING, Functions of Severat Variables, Addison-Wesley, Reading, Mass., 1965.
8. G. HADLEY, Linear Programming, Addison-Wesley, Reading, Mass., 1962.
9. D . A. KoHLEE, "Projections of Convex Polyhedral Sets," ORC 67-29, University of
California, Berkeley, 1967.
10. T. S. MoTZKiN, H. RAIFPA, AND R . M. THRALL, "The Double Description Method,"
in H. W. KuHN AND A. W. TUCKER (eds.), ContribiUions to ihe Theory of Games, Vol.
II, Annals of Mathematics Study No. 28, Princeton U. Press, Princeton, 1953.
11. C. WiTZGALL, AND R. J. WETS, Journal of Research of the Naiimal Bureau of Standards
71B, 1-7 (1967).