Professional Documents
Culture Documents
JOURNAL
OF OPERATIONAL
RESEARCH
ELSEVIER European Journal of Operational Research 97 (1997) 571-579
Abstract
An important generalization of the traveling salesman problem called the traveling purchaser problem is considered. A
branch and bound algorithm which solves a related simple plant location problem for calculating the bounds is proposed for
this problem. Computational experiments with this algorithm provide the evidence that moderate size problems of upto 25
cities and 100 commodities can be solved in reasonable computation time. The problem is shown to have applications in the
areas of scheduling, warehousing and routing. © 1997 Elsevier Science B.V.
Keywords: Traveling purchaser problem; Traveling salesman problem; Simple plant location problem; Heuristic; Algorithm
k~l
r,,! i = 1 . . . . . m;
3.1 Lower bounds
such that (7) ensure that in the optimal tour a city i / all "~
must have a single immediate predecessor city and a tours ]
single immediate successor city. Now constraints
(3)-(5) imply
Xij ~ Yi (8) [ i, k
\
and
/
{o, 1} (9) /
' 'START )
, Initialize
l
If = Travel Cost Matrix
C = Commodity Cost Matrix
P = 1 (all tours), 7,0 = ~
2 1
Calculate fi ~ min { hlk I-
Solve SPLP k
Label P by w(P) = Objective
function value of SPLP
Level = 1
.~o
()
~3
[
Choose city-pair (t, p) for tree extension
such that node with this Pair excluded or Solution is I
( I ~ will yield as large a bound as Z0 = w (O)
optimal
possible. Save tour
5" 1
Make a branch from P t o ~ 7 the CP node.
ft = f't, the next higher cost. Solve SPLP.
10
Set up H for P :
No
1
Make a branch from P to Q, the t, p node.
Delete row t and column p in H. Make
necessary changes in fi values. Prevent
subtour of connected path if required.
Solve SPLP and label
Q by w (O) = Objective function value of
SPLP.
Level = Level + 1
()
Yes
Yes
No
F i g . 2. F l o w c h a r t o f t h e a l g o r i t h m .
K.N. Singh, D.L. van Oudheusden / European Journal of Operational Research 97 (1997) 571-579 575
ing i.e., when the bound exceeds the cost of the best
solution obtained so far or on getting a feasible
solution. For a detailed discussion of each aspect of where a is an integer whose value does not influence
the algorithm the reader may refer to Singh (1985). the optimal solution as all the commodities are to be
purchased. The proportion of commodity costs dif-
ferent from M is around 50%. A problem size is
4. Computational results specified by the number of cities, m, and the number
of commodities, n. A total of 65 problems (5 prob-
A Computer program in FORTRAN, called TRA- lems of each size) with asymmetric travel costs, and
PUR, of the above algorithm is written and tested on of 10 to 25 cities and 10 to 100 commodities are
an IBM 3031 computer. Since standard test problems generated and solved. As the data on computation
are not available, most of the cost data are randomly times for any other approach are not available, nu-
generated. As some of the applications of TPP can merical comparison was not possible.
be found in job sequencing on machines with se- Table 1 presents the mean computation time (t~)
quence-dependent changeover costs and since for each size of problem solved by TRAPUR. The
changeover costs need not be symmetric, we experi- table also includes the standard deviation (s) of
ment with asymmetric travel costs first. Later on, we computation times, and the mean CPU time Oz) for
experiment with symmetric travel costs as well. The obtaining the first feasible solution. The first feasible
results of these experiments are summarized in Table solution values were always within 5% and in many
1 and Table 2. cases within 1% of the respective optimal solution
Asymmetric travel costs for problems of various values. The results indicate that 10-city problems
size are generated in the same manner as has been with upto 100 commodities could be solved in a few
done for a similar scheduling problem in Hansen and seconds. Similarly, problems with 15 to 25 cities and
van Oudheusden (1972). Travel costs are integers 10 commodities could be solved in small computa-
randomly generated from a uniform distribution in tion times. Computation times start increasing when
the range (15, 30). Commodity costs different from problems of size (15 × 50) or bigger are solved. For
arbitrary large value, M, are also integers generated problems with same number of commodities, addi-
from a uniform distribution in the range (a, a + 10), tion of 5 cities increases the computation time by
576 K.N. Singh, D.L. van Oudheusden / European Journal of Operational Research 97 (1997) 571-579
Table 2
Computational results for problems with symmetric travel costs
Problem No Problem size t I (sec) t 2 (sec) Average relative Average relative
travel cost commodity cost
m n
TRAPUR/GS TRAPUR/GS
about 2 to 3 times, and for problems with same and 50 commodities. For problems of larger dimen-
number of cities, addition of 50 commodities in- sions, computation times may be comparatively
creases the computation time by about 3 to 4 times. higher.
For generating symmetric problems, hik values
are taken from the travel cost data for the 33-city
TSP given in Karg and Thompson (1964). cij values 5. Applications and extensions
different from M are integers randomly generated
from a uniform distribution in the range (0-500). A The structure of the traveling purchaser problem
total of 40 problems with 10 to 20 cities and 15 to 50 applies to many real-life problems. We discuss here
commodities are solved. These problems are also some of its potential applications to job scheduling,
solved by a generalized savings (GS) heuristic devel- warehousing and routing. In the later portion of this
oped by Golden et al. and the solutions given by the section we describe a possible extension of this
GS heuristic are compared with the optimal solutions problem.
obtained by TRAPUR. Table 2 provides the data on
mean computation time (t l) for TRAPUR, mean 5.1. Job scheduling
computation time (t z) for the GS heuristic and the
average relative travel and purchase costs. As mentioned earlier, the job scheduling problem
The results of Table 2 indicate that solution times considered by Burstall (1966) and later by Buzacott
for problems with symmetric travel costs are gener- and Dutta (1971) has the same structure as the TPP.
ally higher as compared to those for asymmetric cost hik, in this case will represent the changeover cost
problems. The reason may be that for every feasible from state i to state k, and cij, the cost of processing
solution to such a problem there is always another job j when the machine is in state i. The home city
solution in reversed sequence but having the same (city 1) of the TPP may represent a dummy state of
objective value. This results in more number of the machine in which a dummy job, (n + 1), is
alternatives to be examined. As regards the quality of processed at cost zero. The changeover costs from
solutions produced by the heuristic it is seen that this dummy state to other states of the machine may
quality deteriorates with the increase in problem size, be set equal to zero. Then the solution to this prob-
the only exception being the (10 × 15) problems. lem (as obtained by TRAPUR) will give a sequence
The heuristic produced solutions which are 10 to of selected states in which different jobs are to be
35% off the respective optimal solutions. processed.
The overall conclusion that can be drawn from However, in this procedure the initial bound given
these computational results is that the code, TRA- by the SPLP will be loose as all f, values will be
PUR, can efficiently solve TPP with upto 20 cities zero due to the presence of zeros in a row and a
K.N. Singh, D.L. van Oudheusden / E u r o p e a n Journal o f Operational Research 97 (1997) 5 7 1 - 5 7 9 577
duced in the problem. All the cost parameters change network design problems may be topics of further
from one time period to another. This can also be study.
formulated in the same manner as its static counter-
part and the same solution procedure can be applied.
The only change needed is that a dynamic SPLP is to Acknowledgements
be solved for determining the bound. A dynamic
SPLP can be transformed into an equivalent static The work reported in this paper was carried out
SPLP and it may be attempted by the same dual-based while the first author was a doctoral student at the
method of Erlenkotter (1978). However, a slightly Asian Institute of Technology, Bangkok. The finan-
improved procedure which incorporates primal-dual cial support provided for the study by the Govern-
adjustment in the dual ascent phase has been devel- ment of Japan is thankfully acknowledged. The au-
oped by Van Roy and Erlenkotter (1982). Their thors are indebted to the anonymous referees for
D Y N A L O C code based on this algorithm has been their comments which considerably improved the
able to solve (15 × 30) problems of 8 periods in presentation of the paper.
fractions of a second on an IBM 3033 computer. So
this procedure can be embedded to the code, TRA-
PUR, for the solution of a dynamic TPP. Appendix
Table 3
(a) Travel cost matrix, H and (b) Commoditycost matrix, C of a 5-city, 8-commodityproblem.
i\k 1 2 3 4 5 i\j 1 2 3 4 5 6 7 8 9
I - 32 54 47 48 1 M M M M M M M M 0
2 18 - 22 25 16 2 M M 17 26 85 M 32 M M
3 39 41 - 28 19 3 42 82 63 94 60 M 12 55 M
4 26 8 30 - 24 4 60 10 57 8 M 95 96 9 M
5 35 17 6 9 - 5 M M M 80 21 16 M 77 M
(a) (b)
K.N. Singh. D.L. van Oudheusden / European Journal of Operational Research 97 (1997) 571-579 579
{(1, 3), (3, 5), (5, 4), (4, 2), (2, 1)} with a cost 243.
Now the left-hand branches are searched till the tree
is complete (see Fig. 3).
An optimal solution to the above problem gives a
2 tour of 1 - 3 - 5 - 4 - 2 - 1 at a travel cost of 108 units.
The total commodity cost is 135 units. Commodity 9
is to be purchased from city 1; commodity 3 from
city 2; commodities 1 and 7 from city 3; commodi-
ties 2, 4, 8 from city 4; and commodities 5, 6 from
city 5.
References