You are on page 1of 19

A Multiphase Approach to the

Period Routing Problem


Robert A. Russell
Quantitative Methods and MIS, The University of Tulsa,
600 South College, Tulsa, Oklahoma 74104

Dave Gribbin
Gribbin Development Co., Jenks, Oklahoma 74037

In this article, we present a multiphase approach to the period routing problem. The
period routing problem involves the design of effective vehicle routes that satisfy cus-
tomer service frequencies over a specified planning horizon. The first phase of analysis
consists of a generalized network approximation to achieve an efficient initial solution.
The second phase involves an interchange heuristic that reduces distribution costs by
solving a surrogate traveling salesman problem. The third phase consists of an inter-
change heuristic that further reduces the distribution costs by addressing the actual
vehicle routes of the period routing problem. A fourth phase utilizes a 0-1 integer model
to attempt further improvements. Computational results on test problems indicate that
the multiphase approach yields improvements over previous best solutions.

1. INTRODUCTION
The period routing problem is a generalization of the ordinary vehicle routing
problem (VRP). In the standard VRP, a given set of customers must be serviced
(goods picked up or delivered) by vehicles operating from a central depot, the
objective being to minimize the cost of distribution or vehicle travel. In the
period routing problem (PVRP), the problem is to design a set of routes for each
day of a given T-day period. The planning period T is typically 5-10 days.
Solution methods for the period routing problem must address not only the
spatial considerations arising from customer geographic locations, but also the
temporal aspects deriving from constraints on the timing of deliveries. Thus, a
customer may require a number of visits during the T-day period, but only in
certain allowable day combinations. For example, if a customer requires ser-
vice twice during a period of T = 5 days, the allowable 2-day combinations for
visits might be Monday-Wednesday, Tuesday-Thursday, or Wednesday-Fri-
day; no other combinations would be acceptable.

NETWORKS, Vol. 21 (1991)747-765


0 1991 by John Wiley & Sons, Inc. CCC 0028-3045/91/070747-19$04.00
748 RUSSELL AND GRlBBlN

There are many application areas that represent period routing problems.
Some of the prominent applications include refuse collection [4, 161, fuel oil
delivery and industrial gas distribution [ l , 3, 81, and soft drink distribution [12].
The objective in the period routing problem is to design routes in order to
minimize the sum of the solutions to the resulting vehicle routing problems
occurring on each of the T days of the planning period. The PVRP is more
combinatorially complex than the VRP and is therefore an NP-hard combinato-
rial problem that is very difficult t o solve exactly. All previous work reported
on the PVRP has involved heuristics. The different heuristics all attempt to
generate efficient routes by determining customer-day combinations that clus-
ter customers in the same approximate geographic area on the same day of the
week.
The first work pertaining to a period type of routing problem was reported by
Beltrami and Bodin [4] in their analysis of hoist compactor trucks in New York
City. Their problem was limited to points requiring service either three or six
times per week. Russell and Igo [ 161 developed heuristics based on a modified
Clarke and Wright [6] savings approach and a k-opt approach that was based on
the traveling salesman heuristic of Lin and Kernighan [13]. Christofides and
Beasley [5] developed heuristics based on the solution of surrogate problems
for the PVRP. The two surrogate problems include a median problem and
period traveling salesman problem (PTSP). In their two-phase approach, the
median problem is used to establish an initial assignment of customers to deliv-
ery-day combinations that meets customer service requirements.
For each day in the planning horizon, a traveling salesman tour is generated
that represents a surrogate solution to the VRP. An interchange procedure is
then used to make changes in customer-delivery-day assignments. The total
length of the traveling salesman tour is the surrogate objective to be minimized.
The final solution to the PVRP (and actual objective) is obtained by applying
standard vehicle routing procedures to the points assigned to each day of the
planning horizon.
More recently, Tan and Beasley [I71 have proposed a heuristic based on the
VRP method of Fisher and Jaikumer [ 101. The heuristic solves the initial alloca-
tion of customers to delivery-day combinations. Seed points are used to spa-
tially represent each route on each day of the planning horizon. Tan and Beas-
ley formulate the initial assignment problem as a 0-1 integer program and solve
the associated LP relaxation obtaining very good integrality results. The ap-
proach generally yields better initial results than does the surrogate median
problem.
Ball [2] has formulated the PVRP as a 0-1 integer program. His formulation
incorporates some ideas from Golden and Wasil’s [ 121 work on computerized
vehicle routing in the soft drink industry. This more general formulation utilizes
delivery-day patterns with a fraction of total demand delivered each day. The
demand for a given day is allowed to vary depending on the pattern. Addition-
ally, a customer pattern variable gipis used to incorporate customer preference
for a given pattern. Computational results using the model are not reported.
In this paper, we present a multiphase approach to the PVRP. The first phase
MULTIPHASE APPROACH TO PERIOD ROUTING PROBLEM 749

involves the generation of a feasible initial assignment of customers to day


combinations. The first phase is a “cluster first” approach. A generalized
network is constructed and used to assign customers to seed points that repre-
sent days of the planning period. Vehicle routing procedures are then used to
develop the routes on each day of the planning period. The generalized network
approach is very efficient and provides very good integrality results. The sec-
ond phase is a “tour improvement” algorithm and is a straightforward adapta-
tion of the Christofides and Beasley [5] interchange heuristic for the surrogate
PTSP. The third phase is a modified interchange heuristic that works directly
on the daily vehicle routing problems rather than on a surrogate problem. A
fourth phase is a mathematical modeling approach employing a 0-1 tour im-
provement model. Each phase tends to require more computational effort than
the previous phase. Computational results relating solution quality to computa-
tional effort are presented in the Computational Results section of this article.

11. A GENERALIZED NETWORK FORMULATION


An initial PVRP solution can be obtained by solving a generalized network
with integer flow restrictions. Satisfying the integer flow restrictions can re-
quire the use of Lagrangian relaxation, branch and bound, or possibly other
computationally intensive methods. Fortunately, the relaxation of integer flow
requirements provides integer or very nearly integer solutions for the test prob-
lems examined in this paper. The approach is analogous to the LP formulation
of Tan and Beasley [17], but uses only one seed point for each day in the
planning horizon and offers the computational efficiency of a network model.
The generalized network approximation to the PVRP is represented by a
tripartite transshipment graph with source nodes (customers), transshipment
nodes (allowable delivery combinations), and sink nodes (days of the planning
horizon). The generalized network is a directed graph with a set of nodes N , a
set of arcs, A , with ordered pairs of nodes represented as (i, j ) , where i denotes
the tail node a n d j denotes the head node of the arc. For each arc there is a cost
per unit flow cjj based on the flow leaving the tail node and a multiplier pi that
alters the flow from tail node to head node. Upper and lower bounds uij and Ijj
are specified for flow on arc (i, j).
To construct the generalized network formulation of the initial assignment of
the PVRP with n customers and T days in the planning horizon, let

A = set of arcs in the network


N, = set of n customer nodes
N2 = set of nodes representing delivery combinations
N3 = set of T seed points representing the days of the planning horizon
qi = demand per day of customer i , i = I , 2, . . . , n
p i = number of required deliveries per planning horizon for customer i ,
i = 1,2,. . . , n
xij = flow from node i to node;, (i,j ) E A
hj = total vehicle capacity on day j.
750 RUSSELL AND GRlBBlN

The mathematical model of the generalized network problem can be stated as

Min c
W E A
cij xij

s.t.

C
(i.j)EA
xij = I , i E N,

xij I1, i E N z , (i, j ) E A

xij 2 0, (i, j ) E A .

The objective function (1) minimizes the total distance of inserting the cus-
tomers on seed-point routes. Equation (2) insures that each customer is as-
signed to one delivery combination. Equation (3) insures that the customer
flows associated with a given delivery-day combination are routed to the appro-
priate day(s). Inequality (4) insures that the total vehicle capacity on a given
day is not exceeded. Inequality (5) enforces an upper bound of 1 on each of the
flows out of the delivery combination transshipment nodes.
The cij cost matrix is generated in a way similar to Fisher and Jaikumar’s
definition of route insertion costs for the VRP. In the PVRP, a single seed point
is generated for each day of the T-day planning period. The seed points are
generated in the manner described by Fisher and Jaikumar [ 101. For customer i,
an insertion cost, dik,is determined by calculating the extra distance traveled
when customer i is inserted on the round-trip route from the depot to seed point
k. The cij are calculated as the sum of the appropriate dik for a given delivery-
day combination. Thus,

Figure 1 illustrates the generalized network representation of the 5Ob test


problem created by Christofides and Beasley [ 5 ] . The (u, c , p) on each arc
represent the upper bound, unit flow cost, and multiplier, respectively. All arcs
have a lower bound of zero. In the problem, customers 1-7 require service on
each day of the 5-day planning period. Thus, no delivery-day decision is re-
quired and the first seven customers are excluded from the network. Customer
8 has a demand of 16, requires service two times per week, and has Monday-
MULTIPHASE APPROACH TO PERIOD ROUTING PROBLEM 751

Possible Days of
Customer Combinations the Week

Demard
SUPPLY
271

271

271

271
I

271

FIG. I . Generalized network representation of 5Ob problem.

Wednesday, Tuesday-Thursday, and Wednesday-Friday as allowable delivery


combinations. Three transshipment nodes representing the three delivery-day
combinations are linked to source node 8. Customer 50 has a demand of 10 and
requires service once during the 5-day week. Source node 50 is linked directly
to the five sink nodes. Three vehicles, each having a capacity of 160, create a
total capacity of 480 on each of the 5 days of the week. The total demand of the
first seven customers is 209, creating a remaining capacity of 271 for each of the
five sink nodes of the network.
Ten test problems were used to test the performance of the generalized
network approximation and the proposed phase 2, 3, and 4 procedures. The
first seven test problems are described by Christofides and Beasley [ 5 ] . The
eighth problem is the Christofides and Beasley 5-day-period relaxation of the
126 points test problem originally described by Russell and Igo [161. In the
original test problem, a 6-day period was required. The ninth problem is a new
period routing test problem based on a previous refuse collection application
[7]. The tenth problem is also a new period routing application in the fast food
industry. It is introduced to test performance on a large-scale problem. Details
of the test problems are presented in Table 1.
The generalized network formulation provided an integer solution in six out
of the 10 test problems. Table I1 shows the results of applying the generalized
network model to the 10 test problems. One to two fractional values occurred
on the 75a and 75b test problems, but these were resolved manually by round-
ing the fractional values to delivery combinations whose days had slack capac-
ity. It was also possible to round the nine fractional values of the 417 problem to
TABLE I. Test problem details. -I
VI
Eilon et al. 191 Length of Vehicles h)

Problem No. problem period each


no. customers no. (days) day Combination details
;D
5Oa 50 8 2 213 One delivery in period c
v)
50b 50 8 5 3 Demand 5 10 one delivery in period u)
I I S demand I 25 (10100) or (01010) or (00lOl) m
I-
demand 2 26 delivery every day r
75a 75 9 2 5 One delivery in period r,
z
75b 75 9 5 6 Demand 5 15 one delivery in period 0
16 5 demand 5 27 (10100) or (01010) or (00101) D
demand 2 28 delivery every day E
1OOa 100 10 2 4 One delivery in period
m
%
1OOb 100 10 5 5 Demand 5 10 one delivery in period z
I I 5 demand I 25 (10100) or (01010) or (00101)
demand 2 26 delivery every day
1OOd I00 10 5 4 Demand 5 10 one delivery in period
I I 5 demand I25 (10100) or (01010) or (00101)
demand 5 26 (10101) or (0101 I) or (11010)
126 I26 - 5 4 See Russell and Igo [I61 and [5]
163 163 Cook and 5 3 Demand 5 9 one delivery in period
Russell [7] 10 5 demand I19 (10100) or (OlOlO) or (00101)
demand 2 20 (10101) or (01011) or (11010)
vehicle capacity = 140
distance = distance X 10
417 417 New problem 7 9 Customer i, i 5 40
details in (1001000) or (1000100) or (0100100) or (0100010)
Appendix A or (0010010) or (0010001) or (0001001)
Customer i, 41 5 i C. 417
(IOOOOOO) or (OIOOOOO) or (0010000) or (OOOlOOO)
or (0000l00)or (0000010) or (000000l)
(01010) means no delivery on the first day of the period, delivery on the second day, no delivery on the third day, etc.
MULTIPHASE APPROACH TO PERIOD ROUTING PROBLEM 753

TABLE 11. Comparison of the generalized network approximation with LP.


Network No. CPU No. CPU
size fractional time in LP size fractional time in
Problem nodeslarcs values secondsa varlconst values secondsd
50a 52,100 0 >O 100,52 0 2
50b 140.36I 0 2 170,55 0 4
75a 77,150 1 >O 150,77 I 5
75b 193,522 2 3 263,180 2 9
I 00a 102,200 0 I 200,102 0 6
lOOb 257,698 0 4 352,105 0 15
lood 285,782 0 4 352,107 0 10
I26 352,1196 0 5 256,133 I 18
I63 213,896 4 3 785,168 2 48
417 704,3479 9 39 2,919,424 5 429
* 20 M H z Dell 310 personal computer.

obtain an integer solution. Rounding will often produce a feasible solution since
real-period routing problems typically have some amount of “inherent slack”
in total daily vehicle capacity. Without slack, the bin packing or loading of
discrete entities into fixed-capacity vehicles can be infeasible. Specifying total
daily capacity as 97-98% of actual capacity also facilitates the rounding pro-
cess.
A more sophisticated approach to the rounding problem can be achieved by
applying a 0-1 integer model. A small 0-1 model can be developed to assign the
customers having fractional values to seed points with remaining capacity.
Since the number of fractional values is usually very small, the 0-1 model
should be efficient to solve. Although a 0-1 rounding model was not used in
solving the test problems, it should provide a solution at least as good as
manual rounding.
The ninth test problem was the only problem whose solution could not be
obtained through manual rounding. A few of the customers in this problem
have a demand quantity that is large relative to vehicle capacity, (i.e., 105 vs.
140). Rounding a fractional variable whose demand is large relative to capacity
is not always feasible. To obtain a feasible solution, the three customers with
the largest demand were “preassigned” to their least costly delivery combina-
tion. The network was then solved for the remaining 160 customers and a
solution amenable to rounding was obtained.
Table I1 shows the generalized network integrality and efficiency properties
relative to the LP approach of Tan and Beasley [17]. On the first eight test
problems, the integrality properties are almost identical with the generalized
network having one less fractional variable. However, on the larger two test
problems, the LP approximation exhibits slightly better integrality properties.
The generalized network model, however, is significantly more efficient, partic-
ularly for the large-scale problems. Solutions were obtained using the GNO
(Generalized Network Optimizer) software package [ 1 I] for IBM compatible
personal computers. The LP solution was derived using the LINDO/PC soft-
754 RUSSELL AND GRlBBlN

ware [ 141. Either approximation generates solution5 superior to other “cluster-


first” approaches. The solution quality ranges from 0 to 12.2% greater than the
total routing distance of the best-known solutions. The mean error was 6.49%.

111. INTERCHANGE HEURISTICS


In this section, we describe heuristics that can be used to improve an initial
feasible solution to the PVRP. The first heuristic (phase 2) is a straightforward
adaptation of the interchange heuristic of Christofides and Beasley [ 5 ] . This
heuristic solves a surrogate period traveling salesman problem or PTSP. Our
modification to the basic heuristic is the substitution of the more robust Lin and
Kernighan TSP algorithm for the more efficient 2-opt heuristic used by Christo-
fides and Beasley. The second heuristic (phase 3) is an extension of the inter-
change heuristic that solves the actual PVRP rather than the surrogate PTSP.
In applying the interchange heuristics, it is assumed that an initial feasible
allocation of customers to days of the period T has been made. Thus, for each
day, a traveling salesman tour (phase 2) or a vehicle routing solution (phase 3)
can be generated and a total route distance for all days can be calculated.
The interchange heuristics attempt to alter customer delivery combination
assignments in order to reduce total route distance. The heuristics proceed by
choosing a subset of customers and totally enumerating the delivery combina-
tion assignments to seek an improved solution. A subset U is generated for
each customer on each day during the period T. Each subset consists of a
specific customer i, the ( L = 0, 1, or 2) customers that follow customer i on the
tour, and two other customers that are physically near to customer i but have
few delivery days in common. (More specific details of the subset U construc-
tion are described by Christofides and Beasley.)
Let the family of all subsets U be denoted by %, and let d* be the total route
distance of the current solution. The interchange heuristics operate by tempo-
rarily deleting all customers in subset U from the current routes. The route
distance d l of the reduced routes is then calculated. For each possible combina-
tion of day-delivery assignments of the customers in U , the additional costs d2
of inserting the customers onto the reduced routes is calculated. If d , + d2 < d*
and route capacities are not violated, then an improved solution has been
found.
The phase 2 PTSP heuristic terminates when no improvement has been found
after examining all I/ E OU. If an improvement has been found, the heuristic
applies a TSP algorithm to the daily tours and recalculates the family OU to begin
another iteration. The phase 3 interchange heuristic is similar to the phase 2
heuristic except that the actual vehicle routes on a given day are considered
rather than a single tour surrogate. Given subset U , the phase 3 heuristic must
consider a greater variety of possible insertion points since each day typically
has multiple vehicle routes. The individual vehicle capacities also impose
tighter capacity constraints than the overall capacity of the surrogate “one
giant tour.”
To reduce the combinatorics of the evaluation of each subset U (and, conse-
MULTIPHASE APPROACH TO PERIOD ROUTING PROBLEM 755

quently, the computational effort), a bin packing type of heuristic was used to
insert each point in subset U on a route associated with the delivery-day combi-
nations being considered. The points in subset U are sequentially assigned to
the most cost-effective routes in the order of largest demand quantity first. No
attempt was made to optimize the simultaneous assignment of all (three or four)
points in I/ to routes associated with the delivery-day combination under con-
sideration. The generation of an efficient procedure to evaluate the simulta-
neous assignment of the points in U could yield further improvements and
could be the subject of future research.

IV. A 0-1 TOUR IMPROVEMENT MODEL


The phases 1-3 heuristics are a practical means to develop an effective
solution to the PVRP. However, they do not consider all possible delivery
combinations. In this section, we present a 0-1 integer model that can be used
to improve a given feasible solution. To describe the model, let

rn = number of vehicle tours during the T-day period


R; = set of delivery combinations not chosen by customer i
q, = demand of customer i
1 if customer i is assigned to

{
x,, = combination r
0 otherwise
1 if delivery combination r involves delivery on

{
arL= tour k
0 otherwise
1 if delivery combination r does not involve

- {
ark = delivery on tour k
0 otherwise
dir = net distance improvement from assigning customer i to combination r
DL= Set of customers currently assigned to tour k , k = 1, 2, . . . , rn
DL= Set of customers currently not assigned to tour k , k = 1, 2, . . . , m
ex = excess vehicle capacity on tour k , k = 1, 2, . . . , rn

The djr are calculated by determining the best insertion location of customer i
on the best vehicle tour on each day that is associated with delivery combina-
tion r. Finally, to calculate arkand ark for a given customer, the best tour on
each day associated with delivery combination r must be predetermined.
The model can be stated as

Maximize, d;, xir


i = l rER,

s.t.
756 RUSSELL AND GRlBBlN

xir = 0 or 1 , i = 1, . . . ,n; r E R;

The objective function (6) maximizes the net route distance reduction by
reassigning customers to alternate delivery combinations. Constraint (7) in-
sures that each customer is reassigned to at most one new delivery combina-
tion. Constraint (8) states that the total customer load being reassigned to tour k
less the total customer load being removed from tour k does not exceed the
vehicle capacity.
In some problems, it is necessary to add constraints to prevent “double
switching.” Constraints (7) and (8) do not prevent customer i from being
switched from route w to route z to be near customerj and simultaneously
switching customerj from route z to route w.This double-switching possibility
can be foreseen in the calculation of the djr,i = 1, 2 , . . . , n ; that is, it is
possible to generate a report that shows the insertion points associated with
each dir.Thus, it is possible to foresee the possibility that customer i would like
to be adjacent to customerjon route z and vice versa. The double switching can
be prevented by adding the constraint

where r and s are the appropriate delivery combinations for customer i andj,
respectively.

V. COMPUTATIONAL RESULTS
The interchange heuristics were programmed in FORTRAN IV and compiled
under MICROWAY NDP FORTRAN-386 Version 3.02 and run on a 20 MHz
IBM compatible personal computer with a math coprocessor. The 0-1 tour
improvement model was solved using the LINDO/PC software package.
Table 111 shows the results of the heuristics in solving the 10 test problems.
The total routing distance is reported together with the computation time re-
quired for phases 2 and 3. As expected, the phase 2 interchange heuristic
develops solutions comparable to those reported by Christofides and Beasley.
Note, however, that the initial solutions are different, the TSP heuristics are
different, and Christofides and Beasley’s approach uses multiple passes with L
= 0, I , and 2 , whereas the phase 2 approach in this paper utilizes passes only
for L = 0 and 1. The phase 2 or PTSP heuristic provides solutions whose mean
improvement is approximately 3.53% better than the phase 1 generalized net-
work solution. The phases 3 and 4 heuristics average 5.75% and 5.95% less
total distance than the initial network solution. The solution to the 50b problem
is 6.1% better than the previous best solution. The graph of the 50b solution is
shown in Figure 2 .
TABLE 111. Results for phases 1-4. %
rn
Phase I CPU CPU Phase 4 B
general Phase 2 time in Phase 3 time in 0- I Christofides and Tan and W
Problem network PTSP minutes" PVRP minutes model Beasley PTSP Beasley LP 2
5Oa 537.3 537.3 (0.05) 537.3 (0.05) 537.3 547.4 - 0
b
70b 1,521.o 1,430.0 (0.19) 1.355.6 (0.39) 1,355.4 1,443.I 1,481.3 0
75a 868.5 868.5 (0.03) 867.8 (0.08) 867.8 843.9 - I
75b 2,220.0 2,184.7 (0.35) 2,150.0 (0.72) 2,141.3 2,187.3 2,192.5 -I
lOOa 833.7 833.6 (0.05) 833.6 (0.33) 833.6 839.2 - 0
lOOb 2,270.0 2,188.6 (0.54) 2.108.5 (1 33) 2,108.3 2,153.3 2,281.8 W
m
lOOd 1,805.0 1,719.4 (0.79) I H8.8 (1.23) 1,638.5 1,674.0 1,833.7 E
126 877.0 838.8 (1.81) 826.2 (5.09) 820.3 847.3 878.5 0
I63 1,464.4 1,365.4 (3.75) I ,3 14.9 (19.57) I ,3 12.0 - - 0
417 3,820.0 3,678.3 (23.53) 3,641.9 (9.12) 3,638. I - - ;
II

Total 16,216.9 15,644.6 I 5,284.6 I 5 ,252.6 --I


Mean %
-z
improvement - 3.53 5.75 5.95 D
W
a 20 MHz Dell 310 personal computer. ;
II

s
r
z
758 RUSSELL AND GRlBBlN

43

29

’= v7
28 48 24

(4

A4

(4
FIG. 2 . Problem 50b routes: days 1 (a), 2 (b), 3 (c), 4 (d), and 5 (el.

The 50a, 75a, and lOOa problems consist of only 2 days in the planning
period. Thus, the number of possible combinations is quite small and the final
solution is highly dependent upon the seed-point selection and quality of the
initial solution. The phase 1 network solution to the 50a and lOOa problems is
better than any period routing solution previously reported. However, the 75a
MULTIPHASE APPROACH TO PERIOD ROUTING PROBLEM 759

initial solution is not as good as the Christofides and Beasley solution and
phases 2-4 were not able to achieve a significant improvement on this problem.
The other seven test problems with planning horizons of 5-7 days are more
representative of real-world applications.
The 0-1 model was run with an upper bound of 10,OOO iterations on the
LINDO branch and bound process. Only the smaller 50a, 50b, and 75a prob-
lems solved to optimality. The remaining phase 4 solutions represent the best
solution found before termination. On the large-scale 417 point problem, the 0-
1 solution was approximated by using only the 80 variables that were nonzero
in the solution of the LP relaxation.
The solution times of the phases 2 and 3 heuristics are reported in parenthe-
ses adjacent to the solution value. The solution times depend not only on the
number of interchanges performed per iteration, but also on the number of
iterations performed before no improvements could be found. The computation
times do not reflect the time required to obtain the vehicle routing solution that
is required at the end of phase 2 or the beginning of phase 3. The final vehicle
routing solutions were obtained using the MTOUR method [15].
The larger solution time of the 126 and 417 test problems is partly attributed
to the greater number of combination possibilities. In the other eight test prob-
lems, there are one, three, or five delivery combinations for each customer. In
the 126 and 417 test problems, there are up to seven delivery combinations
depending on the customer.
The solution times appear to be competitive with the PTSP approach of
Christofides and Beasley, although it is difficult to compare personal computer
and CDC 7600 mainframe computation times. In solving five of the test prob-
lems, phases 2 and 3 required a total of 3.68 and 8.76 personal computer
minutes, respectively. On the same five test problems, the PTSP approach of
Christofides and Beasley required a total of 3.19 minutes on a CDC 7600 main-
frame.

VI. CONCLUSION
The multiphase approach to the period routing problem generated solutions
that are better than the previous best-known solutions for nine of the 10 test
problems. The phase 1 generalized network approximation is very efficient and
demonstrates good integrality properties. The total route distances were 6.49%
more than the best-known solutions. The efficiency of the network approach
suggests that it could be used to solve large-scale problems in a personal com-
puter or mainframe environment.
The FTSP surrogate approach of Christofides and Beasley is currently the
most effective method for making initial improvements to a feasible period
routing solution. In the test problems, a PTSP approach was able to improve
the initial solutions by a mean of 3.53%. Further improvements require a
method that addresses the actual PVRP rather than the surrogate FTSP. In
phase 3, the modified interchange heuristic generated solutions that on the
average reduced the test problem route distances by another 2.22%. Addition-
APPENDIX A: DETAILS OF 417-POINT PROBLEM

x Y Q x Y Q X Y Q
Depot 330.37 850.34 0 31 348.06 869.73 I65 62 323.91 862.50 350
I 347.70 866.17 250 32 347.30 877.01 200 63 323.71 863.02 300
2 348.17 876.69 I50 33 345.54 869.8 I I75 64 342.62 862.05 340
3 336.67 858.33 300 34 328.46 866.38 I90 65 313.09 858.09 327
4 323.90 870.08 224 35 304.10 872.87 160 66 312.45 854.26 295
5 332.33 875.95 210 36 304.26 872.24 220 67 328.41 85 1.70 350
6 338.40 872.79 250 37 344.80 853.47 I80 68 306.98 88 I .66 290
7 305.23 879.03 220 38 347.06 853.66 240 69 306.67 881 .oo 245
8 341.82 867.99 300 39 347.69 849.67 I33 70 306.12 882.06 290
9 318.26 866.27 I80 40 342.49 887. I9 215 71 338.16 857.60 245
10 329.90 858.65 I90 41 312.45 854.26 200 72 308.94 877.77 27 I
11 318.79 851.51 I 60 42 324.70 850. I3 250 73 3 12.83 864.52 I75
12 335.42 865.29 270 43 346.78 865.68 300 74 345.05 877.29 290
13 340.09 860.66 210 44 348. I7 876.69 I80 75 3 10.90 870.66 280
14 337.59 870.54 255 45 347.25 865.78 320 76 304.05 876.82 255
15 335. I I 868. I3 228 46 347.15 866.27 250 77 334.30 861.01 250
16 333.73 866.09 220 47 347.25 865.78 250 78 313.09 857.13 360
17 3 10.00 872.58 I65 48 347.15 866.27 230 79 325.20 862.03 270
18 344.46 857.15 215 49 347.25 865.78 410 80 343.44 863.1 I 215
19 312.81 862.56 I55 50 348. I7 876.69 205 81 336.55 868.09 300
20 335.63 868.14 200 51 336.69 858.38 340 82 346.89 866.91 290
21 344.42 869.42 210 52 33 I .99 875.32 325 83 325. I 1 878.49 195
22 343.22 865.01 220 53 323.30 862.71 215 84 332.75 863.54 33 1
23 348.06 869.73 280 54 323.91 862.50 I95 85 334.44 869.74 250
24 314.31 869.26 190 55 347.54 876.75 275 86 334.65 869. I2 315
25 335.46 867.57 240 56 326.08 854.50 280 87 324.09 862.78 380
26 331.48 853.64 265 57 33 1.99 875.32 330 88 336.21 866.17 355
27 341.71 861.55 230 58 318.06 859.4 I 250 89 315.28 878.83 345
28 344.83 858.59 255 59 3 13.92 856.82 345 90 332.01 867.88 280
29 331.99 875.32 225 60 324.64 864.58 315 91 345.54 869.81 335
30 339.53 860.20 170 61 323.61 863.37 330 92 325.19 858.87 330
X Y Q x Y Q X Y Q
93 329.42 871.33 282 I26 323.63 862.96 310 159 301.67 858.00 I60
94 319.19 877.42 310 I27 336.69 858.38 255 160 307.03 848.42 300
95 306.74 878.92 215 I28 336.69 858.38 366 161 304.44 850.44 300
96 336.87 878.31 330 I29 328.28 857.53 335 162 301.56 856.47 185
97 341.38 879.86 290 I30 317.17 862.62 235 163 301.56 856.47 22 1
98 333.86 866.09 300 131 323.33 862.83 360 164 306.07 866.03 215
99 334.11 868.98 295 132 326.41 853.78 325 165 307.60 865.66 430
100 336.67 866.67 259 133 310.06 863.24 219 166 304.41 865.76 350
101 335.33 867.67 360 I34 339.20 878.00 360 167 301.67 858.00 100
102 308.11 880.89 260 I35 321.38 857.10 265 168 303.68 872.00 245
I03 347.32 869.73 268 I36 330.98 867.54 270 I69 297.26 849.93 250
104 310.24 874.94 285 I37 349.78 857.63 240 170 309.60 855.14 315
105 345.54 869.81 380 138 330.75 868.44 380 171 303.68 872.00 300
106 314.17 860.72 220 I39 313.63 853.34 275 I72 304. I7 872.33 265
I07 326.08 854.50 226 I40 317.21 858.16 250 173 332.45 842.53 240
108 335.78 867.59 290 141 326.31 873.53 260 I74 326.24 836.56 165
I09 346.57 860.24 330 I42 339.28 856. I 1 260 I75 330.37 850.34 220
I10 338. I6 857.60 I75 I43 346.57 860.24 250 I76 320.72 842.27 171
111 304.10 882.76 330 I44 328.57 851.99 250 I77 319.65 837.82 155
I12 306.47 881.52 237 I45 350.00 910.00 250 178 325.53 838.89 190
1 I3 342.29 876.2 1 445 I46 307.75 852.35 230 I79 324.77 849.47 325
1 I4 306.96 880.96 245 147 305.17 864.77 460 I80 324.57 837.33 160
115 323.61 863.37 350 I48 303.64 871.76 265 181 322.50 834.67 215
I I6 339.84 859.5 I 265 149 307. I9 861.15 315 182 333.64 847.87 I65
I I7 310.38 858.50 305 I 50 306.46 870.75 290 I83 335.81 850.79 150
118 33 1.67 863.24 220 151 303.68 872.00 225 I84 313.77 849.42 240
I I9 335.60 868.00 33 1 52 308.87 856.62 400 185 336.77 840.03 I45
120 315.58 852.49 257 53 304. I8 865.45 225 I86 339.86 833.79 325
121 335.84 863.91 360 54 304.26 872.24 290 I87 339.30 845.35 200
I22 339.16 864,46 380 5.5 306.07 866.03 465 188 326.24 836.56 315
I23 312.50 854.33 360 56 302.54 856.47 185 189 330.52 84 1.65 190
I24 341.52 856.78 277 57 301.56 856.47 I80 I90 336.67 842.83 200
I25 323.63 862.96 400 58 297.96 852.92 220 191 323.88 833.51 305
APPENDIX A: (continued)
~

X Y Q X Y Q X Y Q
I92 338.83 845.33 206 223 330.35 839.40 21 1 254 343.00 838.33 325
I93 338.83 845.33 220 224 328.79 843.28 245 255 341.01 845.19 265
194 335.73 844.16 250 225 337.00 846.67 250 256 343.53 829.28 260
195 336.67 847.17 I30 226 338.84 845.31 I79 257 328.49 836.77 245
196 336.67 843.67 176 227 336.67 843.67 250 258 321.67 831.33 230
197 338.21 841.17 168 228 333.64 847.87 260 259 342.94 838.25 274
198 338.53 840.22 200 229 335.24 843.56 242 260 342.41 844.81 240
199 338.06 837.59 340 230 320.72 842.27 155 26 1 332.94 839.61 315
200 338.00 841.67 204 23 1 336.67 840.00 I40 262 337.35 827.46 290
20 1 3 17.74 844.41 210 232 337. I I 84 I .06 178 263 330.93 820.18 370
202 339.40 842.07 150 233 338.83 845.33 405 264 340.74 842.96 423
203 342.76 85 1.94 200 234 339.33 845.33 250 265 345.98 837.62 280
204 338.35 843.16 200 235 337.50 847.47 280 266 326.24 836.56 300
205 332.45 842.53 245 236 339.40 842.07 250 267 333.33 845S O 365
206 338.48 844.31 250 237 339.33 842.17 220 268 342.76 85 I .94 160
207 336.67 842.83 250 238 338.00 841.67 200 269 326.58 837.54 300
208 338.21 841.17 300 239 325.00 828.00 215 270 340.66 846.72 350
209 338.41 841.02 220 240 330.86 832.30 260 27 1 342.29 838.84 455
210 339.33 842.17 150 24 I 334.36 8 19.67 240 272 323.02 840.24 220
21 1 342.04 834.55 285 242 334.63 824.99 270 273 325.42 828.92 230
212 338.06 841.71 180 243 334.67 820.17 170 274 339.51 839.89 260
213 337.24 842.83 166 244 334.38 819.41 289 275 328.49 836.77 340
214 338.08 846.34 200 245 328.41 836.20 275 276 340.66 846.72 265
215 340.00 833.83 205 246 328.10 836.76 230 277 329.98 824.05 300
216 341.23 840.03 340 247 328.10 836.76 220 278 310.42 848.78 280
217 338.14 843.58 190 248 337.94 846.59 275 279 340.98 835.68 400
218 328.53 846.06 260 249 350.19 851.82 240 280 337.09 85 1.47 275
219 337.05 842.42 150 250 334.34 820.15 250 28 I 334.22 845.21 128
220 337.55 843.87 205 25 I 334.38 819.41 245 282 317.70 847.94 250
22 1 324.77 849.47 200 252 347.67 847.97 250 283 344.68 844.28 285
222 336.69 838.58 215 253 334.63 819.86 232 284 335.42 842.33 290
-
X Y Q X Y Q X Y Q
285 325.58 842.36 250 318 334.68 819.92 250 35 1 317.04 891.26
286 333.25 833.89 315 319 299.46 901.21 I80 352 312.31 904.59
287 346.84 844.72 390 320 325.00 92 I .33 250 353 339.88 884.79
288 341.69 847.99 250 32 I 299.46 901.21 250 354 303.98 890.80
289 337.26 843.36 230 322 307.81 898.57 220 355 324.50 886.50
290 338.06 841.71 200 323 302.00 932.50 250 356 303.98 890.80
29 1 336. I I 838.52 250 324 304.58 910.70 250 357 342.49 887. I9
292 339.94 850.04 240 325 302.50 932.58 164 358 305.32 896.67
293 335.73 844.16 200 326 304.09 908.99 100 359 341.92 887.22
294 335.37 852.53 275 327 302.10 932.06 200 360 323.20 891.61
295 342.80 83 1.09 275 328 302.00 932.33 200 36 1 323.11 901.79
2% 315.67 841.67 250 329 299.88 902.43 I70 362 303.98 890.80
297 348.75 839.60 200 330 302.79 897.73 I 40 363 354.44 846.00
298 349.50 853.00 165 33 1 335.71 884.20 335 364 351.49 848.67
299 345.32 839.84 250 332 334.1 1 910.08 295 365 353.53 862.05
300 338.61 846.77 300 333 324.62 886.60 235 366 356.13 870.41
30 I 341.36 828.26 I95 334 317.04 891.26 181 367 35 1.49 848.67
302 348.78 833.99 255 335 313.02 892.99 250 368 354.84 864.52
303 328.75 85 1.78 250 336 303.83 885.29 313 369 354.44 846.00
304 328.87 848.30 190 337 303.83 887.00 290 370 355.00 850.17
305 334.68 819.92 290 338 303.92 886.34 260 37 I 350.30 851.85
306 334.34 820. I5 375 339 304.16 899. I6 300 372 350.28 852.33
307 340.24 843.44 250 340 304.16 889. I6 245 373 356.05 844.64
308 337.33 849.13 300 34 I 3 16.77 886.48 380 374 351.49 848.67
309 345.12 835.26 I80 342 315.68 913.66 245 375 350.09 852.29
310 33 1.75 848.74 265 343 346. I9 899.67 255 376 352.63 85 1.76
31 I 345.32 839.84 220 344 304.40 885.46 245 377 35 I .49 848.67
312 335.97 834.68 200 345 324.21 901.34 I65 378 351.51 865.76
313 343.69 849.38 250 346 349.04 885.67 300 379 351.99 870.29
314 345.05 849.49 322 347 342.00 905.73 190 380 350.21 857. I I
315 324.77 849.47 295 348 323.41 908.69 235 38 I 352.26 845.28
316 324.77 849.47 430 349 334.63 888.18 240 382 357.93 842.74
317 333.52 816.96 250 350 334.00 884. I7 410 383 350.07 852.03
D
c
v)
v)
m
r
r
APPENDIX A: (continued) P
2
0
x Y Q x Y Q x Y Q G)
384 356.55 888.75 245 396 351.17 899.67 245 407 356.05 851.90 185 z03
385 354.50 867.95 370 397 350.57 898.92 245 408 352.59 889.99 220
386 355.21 843.59 375 398 360.37 893.83 265 409 351.17 899.67 250 E
387 350.09 852.84 310 399 357.34 843.43 200 410 351.25 853.24 160 z
388 356.51 883.91 315 400 353.71 853.89 205 411 354.44 846.00 250
389 352.40 858.31 265 401 350.07 852.03 140 412 354.44 846.00 280
390 352.25 882.38 250 402 351.86 861.09 375 413 351.36 900.08 lo00
391 356.13 870.41 430 403 351.06 899.59 138 414 350.08 852.08 250
392 355.93 892.60 305 404 350.44 896.63 235 415 341.82 867.99 67
393 351.06 899.59 215 405 359.19 887.60 205 416 345.54 869.81 119
394 351.52 899.81 235 406 350.74 850.63 200 417 345.54 869.81 125
395 351.17 899.67 150
Truck capacity = 2000; Euclidean distance metric
MULTIPHASE APPROACH TO PERIOD ROUTING PROBLEM 765

ally, t h e phase 3 approach can a d d r e s s the multiple routes problem discussed


b y Ball [2] in which s o m e vehicles service more than o n e route p e r d a y .
T h e 0-1 integer model improved t h e phase 3 solution by 0-0.72% with a
mean o f 0.21%. Given t h e computational effort, a fourth phase utilizing t h e 0-1
model is probably not cost effective for most applications.

We want to thank the referees for their helpful comments and suggestions for improving
the paper.

REFERENCES
A. Assad, B. Golden, R. Dahl, and M. Dror, Design of an inventory/routing
system for a large propane distribution firm (C. Gooding, Ed.) Proceedings o j t k r
1982 Soiitlienst T l M S Conjerence, Myrtle Beach, 3 15-320.
M . Ball, Allocation/routing: Models and algorithms. Vehic/e Routing: Methods
and Studies (B. Golden and A. Assad, Eds.), North-Holland, Amsterdam (1988).
W. Bell, et al., Improving the distribution of industrial gases with an on-line
computerized routing and scheduling optimizer. Interfiices 13 (1983) 4-23.
E. Beltrami and L. Bodin, Networks and vehicle routing for municipal waste
collection. Networks 4 (1974) 65-94.
N . Christofides and J. E. Beasley, The period routing problem. Nc~ticw4s14
(1984) 237-256.
G . Clarke and J . Wright, Scheduling of vehicles from a central depot to a number
of delivery points. Opertitions Res. 12(4) (1964) 568-581.
T. Cook and R . A . Russell, A simulation and statistical analysis of stochastic
vehicle routing with timing constraints. Dec. Sci. 9 (1978) 673-687.
M. Dror and M. Ball, Inventory/routing: Reduction from an annual to a short-
period problem. Ntiud Re.s. Logist. Q . 34 (1987) 891-905.
S. Eilon. C. D. T. Watson-Candy, and N. Christofides, Distrihirlion Mtintige-
ment: Mathematicril Modelling rind Practicwl Antilysis. Griffin, London ( 197 I ).
M . L. Fisher and R. Jaikumar, A generalized assignment heuristic for vehicle
routing. Networks 11(2) (1981) 109-124.
GNO/PC, Generalized Network Optimization. Copyright 1987. Computer Aided
Planning and Scheduling, Atlanta, GA.
B. Golden and E. Wasil. Computerized vehicle routing in the soft drink industry.
Operations Res. 35 (1987) 6-17.
S. Lin and B. W. Kernighan, An effective heuristic algorithm for the traveling
salesman problem. Operations Res. 21 (1973) 498-516.
LINDO/PC, Linear Interactive and Discrete Optimizer. Copyright 1985, LlNDO
Systems, Inc., Chicago, IL.
R. A. Russell, An effective heuristic for the M-tour traveling salesman problem
with some side conditions. Operutions R P S .25 (1977) 517-524.
R. A. Russell and W. Igo, An assignment routing problem. Networks 9(1) (1979)
1-17.
C. C. R. Tan and J. E. Beasley, A heuristic algorithm for the period vehicle
routing problem. Omega 12(5)(1984) 497-504.

Received June 1989


Accepted April 1991

You might also like