Transportation Research Part E 39 (2003) 325339

www.elsevier.com/locate/tre

Mixed truck delivery systems with both

hub-and-spoke and direct shipment
a,*
Jiyin Liu , Chung-Lun Li b, Chun-Yan Chan a

a
Department of Industrial Engineering and Engineering Management, The Hong Kong University of Science
and Technology, Clear Water Bay, Kowloon, Hong Kong
b
Department of Shipping and Transport Logistics, The Hong Kong Polytechnic University, Hung Hom,
Kowloon, Hong Kong

Abstract

This paper studies a mixed truck delivery system that allows both hub-and-spoke and direct shipment
delivery modes. A heuristic algorithm is developed to determine the mode of delivery for each demand and
to perform vehicle routing in both modes of deliveries. Computational experiments are carried out on a
large set of randomly generated problem instances to compare the mixed system with the pure hub-and-
spoke system and the pure direct shipment system. The experiment results show that the mixed system can
save around 10% total traveling distance on average as compared with either of the two pure systems.

2003 Elsevier Science Ltd. All rights reserved.

Keywords: Truck delivery system; Hub-and-spoke; Direct shipment; Mixed system; Vehicle routing; Heuristic algorithm

1. Introduction

The ow of physical goods from manufacturers to customers is a major focus of logistics

systems. Moreover, product delivery at a reasonable cost has recently become a critical factor in
the survival of emerging e-businesses (Lee and Whang, 2001). Logistics and operations researchers
have done extensive research on the design and operations of local delivery systems in order to
determine the most cost-eective methods of delivery. Because obtaining the optimal solutions to
these problems is extremely dicult, it is essential to develop good heuristics for organizing and
operating delivery systems.

*
Corresponding author. Tel.: +852-2358-7113; fax: +852-2358-0062.
E-mail address: [email protected] (J. Liu).

1366-5545/03/$ - see front matter
2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S1366-5545(03)00005-X
326 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

The design or organization of a local truck delivery system has a critical impact on its per-
formance. There are two common types of delivery system designs: the direct shipment system and
the hub-and-spoke system. In a direct shipment system, each supplier operates independently with
its own eet delivering goods to customers. Each vehicle visits only one customer in a trip. This
method should be utilized when the lead-time requirement is tight, the goods need to be isolated,
or the shipment is large. If these criteria are not satised, then transportation costs can be reduced
by having each delivery vehicle visit several customer locations, provided that the total quantity of
goods to be delivered does not exceed the vehicles capacity. This type of arrangement is called
direct shipment with milk runs. Whenever a milk run is included, a decision on the routing of each
vehicle needs to be made. When there are multiple suppliers in the delivery region, especially when
customers have common suppliers, another type of delivery system can be utilized. In such a
system, goods from all suppliers are collected and consolidated in a central facility, called the hub,
and then redistributed to the customers. If each vehicle visits only one supplier or customer in a
collection or redistribution trip, the system is called a hub-and-spoke system. When the delivery
order sizes are relatively small, a vehicle can visit several stops in a collection trip or a redistri-
bution trip. This delivery network is termed hub-and-spoke with milk runs. We will consider only
the direct shipment with milk runs and hub-and-spoke with milk runs. For simplicity, we will
simply refer to these systems as direct shipment and hub-and-spoke respectively in the following
discussion.
The hub-and-spoke system takes advantage of the economies of scale in vehicle utilization. It
can also improve customer service in terms of delivery frequency. When direct shipment is used,
smaller suppliers need to wait until a sucient amount of goods are ordered to maintain cost
eectiveness in transportation. With the hub-and-spoke system, suppliers can provide a higher
frequency of delivery (improved service quality) by combining the demands or orders of others.
Intuitively, when the customers of each supplier are located very close to the supplier and the
delivery quantity is large enough to justify the shipping of goods with full truckloads, the direct
shipment system is better. Otherwise, the hub-and-spoke system is more appropriate.
In reality, suppliers and customers are located quite randomly and delivery quantities vary
from order to order. The advantage of one of the systems over the other is neither obvious nor
unchanged from day to day. In this situation, a mixed delivery system can be benecial and better
than either of the two pure delivery systems. Such a mixed system can be viewed as a hub-and-
spoke system allowing some orders to be directly shipped whenever benecial. Therefore, in the
mixed system, dierent delivery modes may be used for dierent shipments depending on the
quantity to be shipped and geographical locations of the supplier and the customer (see Fig. 1).
There has been extensive research on the vehicle routing problem (VRP), which is the main
component of a direct shipment system with milk runs. It is a problem of determining routes
through one or more depots and a set of customer locations to minimize the total distance
traveled. A VRP can take various forms based on the constraints and requirements of the network
and the shipment demands, such as vehicle capacity, the delivery time window, line-haul and
back-haul demands, and multiple depots, etc. Bodin and Golden (1981) presented an overview of
dierent types of VRPs. Our analysis is related to one of these VRPs, namely, the capacitated
vehicle routing problem (CVRP), in which each vehicle has a given capacity. Various solution
methods have been developed to solve the CVRP. Well-known solution techniques include sav-
ings heuristics (Clarke and Wright, 1964), the sweep heuristic (Gillett and Miller, 1974), k-opt tour
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 327

Fig. 1. A mixed delivery system with both direct shipment and hub-and-spoke deliveries.

improvement methods (Lin, 1965), etc. For recent surveys on these solution techniques, see La-
porte (1992), Fisher (1995), and Laporte et al. (2000).
In order to increase the eciency of delivery systems, some researchers have studied the design
and operation of hub-and-spoke systems, in which the hub location is a critical decision. For
example, OKelly (1987), Campbell (1996), Abdinnour-Helm and Venkataramanan (1998), and
Pirkul and Schilling (1998) solved the location-allocation problem that determines locations of
hubs and the assignment of nodes to each hub. OKelly and Bryan (1998) considered the above
problem with economies of scale taken into account, where the marginal cost decreases with ow
volume. All these papers used air passenger ow data to illustrate their methods. But the models
and algorithms in these studies are general and may be applied in air, truck or telecommunication
networks. Recently, Sasaki et al. (1999) proposed solution algorithms to solve multi-hub location
problems in the airline industry.
The mixed delivery system has received less attention than the two pure systems. Aykin (1995a)
studied the location-routing problem. The problem was to nd the hub locations and at the same
time to determine the delivery mode for each demand. Aykin (1995b) proposed a simulated
annealing procedure to solve the problem with an initial solution generated using a greedy
interchange heuristic. The interchange was based on the savings estimate calculated for each
hub-node pair if that node would be served by that hub. Hall (1987) developed similar models and
used the EOQ concept to determine which delivery mode a demand should be assigned to with a
predetermined hub location. However, all these models for mixed delivery systems assumed that
328 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

each trip only involved one origin or one destination through the hub(s). These models were built
for applications in air transportation networks. They considered only the assignment of demands
to particular hub(s) without dealing with the issue of routing the vehicle in each trip. To the best
of our knowledge, there has been no previous work on a mixed system with milk runs.
In this study, we propose a heuristic for scheduling vehicles in a mixed truck delivery system
and evaluate through extensive computational experiments the traveling cost (distance) savings of
the mixed system as compared with the traveling costs of the pure systems. For simplicity, we
ignore the xed cost of operating the hub. Furthermore, we assume that any variable cost of
operating the hub is included in the transportation cost of entering and leaving the hub (see
Section 2 for further discussion).
We assume that homogeneous vehicles are used. We further assume that the required delivery
quantity from any supplier to any customer does not exceed the capacity of one vehicle. A cus-
tomer order from a supplier cannot be split into two trips in the direct shipment system. Since
demands are aggregated in the hub-and-spoke system, the total supply from one supplier or the
total demand by one customer may be larger than a truckload. Therefore, it is inevitable for us to
allow splitting of shipments into several vehicles.
This paper is organized as follows: Section 2 presents the model of our mixed delivery system.
The heuristic procedure for scheduling vehicles in the delivery system is given in detail in Section
3. Section 4 reports the computational results and analyzes the relationship between the benets
of the mixed delivery system and the problem parameters. Finally, some concluding remarks are
provided in Section 5.

2. Model description

In our study, the delivery system is dened on an undirected network G V ; E. The vertex set
is V fu0 g [ Vs [ Vc , where u0 is the given hub location, Vs fu1 ; u2 ; . . . ; um g is the set of sup-
pliers, and Vc fum1 ; um2 ; . . . ; umn g is the set of customers. Associated with network G is a
shortest distance matrix with elements tij being the shortest distance from ui to uj , for i; j 0;
1; . . . ; m n. This shortest distance matrix is symmetric and satises the triangle inequality, that
is, tij tji and tij tjk P tik for any i; j; k 0; 1; . . . ; m n. We let
D fui ; uj ji 1; 2; . . . ; m; j m 1; m 2; . . . ; m ng

be the set of all suppliercustomer pairs. Associated with each suppliercustomer pair, ui ; uj 2
D, is a nonnegative demand parameter, qij , which indicates that a quantity of qij is required to be
transported from supplier ui to customer uj . The goods are transported by homogeneous vehicles
with capacity Q, and we assume that there is an innite supply of vehicles. We further assume that
qij 6 Q for all ui ; uj 2 D. The objective is to determine the vehicle routes, some of which will serve
the customers directly from a supplier while others will be connected to the hub, to minimize the
total travel distance of the vehicles. In our model, one of the decisions is to partition the set D into
subsets Dd and Dh , where the demand in Dd will be satised via direct shipments (with milk runs)
and the demand in Dh will be served by hub-and-spoke deliveries (with milk runs). For i 1;
2; . . . ; m and j m 1; m 2; . . . ; m n, let
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 329


qij ; if ui ; uj 2 Dd ;
qdij
0; otherwise;


qij ; if ui ; uj 2 Dh ;
qhij
0; otherwise;
X
mn
q~hi qhij ;
jm1

X
m
q~hj qhij :
i1

Note that in our model, the distance between any two points can also be interpreted as the cost
of traveling between those two points. Here, all xed costs (e.g., the xed cost of operating the
hub) are ignored. Furthermore, if there is a variable cost, j, for handling each unit of shipment at
the hub, then our model can still be used for solving the mixed hub-and-spoke delivery problem
by adding j=2 to the length of every arc incident to the hub.
For any given set Dd , the optimal direct shipment is obtained as follows: For each supplier, ui
(i 1; 2; . . . ; m), we solve a CVRP with the depot located at ui to satisfy all the demand, qdij , in-
curred by each customer, uj , that satises ui ; uj 2 Dd .
For any given set Dh , the optimal hub-and-spoke shipment is obtained as follows: To determine
the optimal collection routes, we solve a CVRP with the depot located at u0 to pick up the goods,
q~h1 ; q~h2 ; . . . ; q~hm , from suppliers, u1 ; u2 ; . . . ; um , respectively. To determine the optimal delivery routes,
we solve a CVRP with the depot located at u0 to satisfy the demand q~hm1 ; q~hm2 ; . . . ; q~hmn of cus-
tomers um1 ; um2 ; . . . ; umn , respectively.
Hence, our problem is to nd a partition fDd ; Dh g of D such that (i) the demand qdij is satised
by direct shipment for every suppliercustomer pair, ui ; uj 2 Dd , where the delivery arrange-
ments are determined by solving a CVRP for each supplier, and (ii) the demand qhij is to be
satised by hub-and-spoke deliveries for every suppliercustomer pair, ui ; uj 2 Dh , where the
delivery arrangements are determined by solving two CVRPs, one for the collection of goods from
the suppliers and one for the delivery of goods to the customers (see Fig. 1). We will call this
model the mixed hub-and-spoke and direct shipment delivery problem, or simply the mixed de-
livery problem.
Two other problems related to our mixed delivery problem can be described as follows: (i)
Instead of determining the optimal partition fDd ; Dh g of D, suppose that Dd D and Dh are
given. Then, our decision is to determine the optimal direct shipment arrangements to satisfy all
the customers demands. We will refer to this problem as the pure direct shipment problem. (ii)
Instead of determining the optimal partition fDd ; Dh g of D, suppose that Dh D and Dd are
given. Then, our decision is to determine the optimal hub-and-spoke deliveries to satisfy all the
customers demands. We will refer to this problem as the pure hub-and-spoke problem. Clearly, for
any given set of data, the optimal objective function value of the mixed delivery problem must be
no larger than that of either pure problem. In Section 4, we will study the benets of the mixed
delivery model as compared with the two pure delivery systems.
Note that both the mixed delivery problem and the two pure delivery problems involve solving
the CVRP as a subproblem. However, the CVRP belongs to the class of NP -hard problems, which
330 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

indicates that the existence of an ecient algorithm to solve the problem optimally is unlikely (see,
for example, Christodes, 1985). Hence, we solve our CVRP subproblems using the well-known
ClarkeWright savings heuristic (Clarke and Wright, 1964). In solving each CVRP, those vertices
with zero quantity will not be considered, i.e., they need not be visited by any vehicle.

3. Solution procedure

In this section, we present a heuristic algorithm to nd a near-optimal solution to the mixed

delivery system. In this heuristic, we rst obtain the solution to the pure direct shipment problem
and the solution to the pure hub-and-spoke delivery problem. The better of them is then taken as
the initial solution of our improvement procedure that searches for improvements in the solution.
Thus, the solution generated by the heuristic will be guaranteed to be no worse than the solutions
of the pure delivery systems obtained by the ClarkeWright heuristic. The heuristic can be de-
scribed as follows.

Heuristic H:
Step 1. Solve the pure direct shipment problem. This is done as follows: For i 1; 2; . . . ; m,
solve a CVRP using the ClarkeWright heuristic with the depot located at ui to serve the n
customers with demands qdi;m1 ; qdi;m2 ; . . . ; qdi;mn . Let the solution value (i.e., the total travel dis-
tance) be Z d .
Step 2. Solve the pure hub-and-spoke delivery problem. This is done as follows: Solve a CVRP
using the ClarkeWright heuristic with the depot located at u0 to collect the goods from the m
suppliers with supplies q~h1 ; q~h2 ; . . . ; q~hm . Next, solve a CVRP using the ClarkeWright heuristic with
the depot located at u0 to serve the n customers with demands q~hm1 ; q~hm2 ; . . . ; q~hmn . Let the solution
value (i.e., the total travel distance of all collection and distribution trips) be Z h .
Step 3. If Z d 6 Z h , then let the pure direct shipment solution be the current solution, put the
direct shipment delivery in sending mode, put the hub-and-spoke delivery in receiving mode,
and let Dd D, Dh . Otherwise, let the pure hub-and-spoke delivery solution be the current
solution, put the hub-and-spoke delivery in sending mode, put the direct shipment delivery in
receiving mode, and let Dh D, Dd . Let Z minfZ d ; Z h g be the value (i.e., the total travel
distance) of the current solution. Set Z m Z, which is the value of the best solution obtained so
far.
Step 4. Consider the current solution.
Case (i): If the direct shipment delivery is in sending mode, then for every ui ; uj 2 Dd ,
compute vdij , which is an estimate of the improvement in the solution value if the suppliercus-
tomer pair ui ; uj is transferred from Dd to Dh . Transfer all those pairs with positive vdij from direct
shipment delivery to hub-and-spoke delivery, i.e., set
Dd Dd n fui ; uj jvdij > 0g and Dh Dh [ fui ; uj jvdij > 0g;
Case (ii): If the hub-and-spoke delivery is in sending mode, then for every ui ; uj 2 Dh ,
compute vhij , which is an estimate of the improvement in the solution value if the suppliercus-
tomer pair ui ; uj is transferred from Dh to Dd . Transfer all those pairs with positive vhij from hub-
and-spoke delivery to direct shipment delivery, i.e., set
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 331

Dh Dh n fui ; uj jvhij > 0g and Dd Dd [ fui ; uj jvhij > 0g:

(The estimates vdij and vhij will be discussed in detail later.)

Step 5. Solve the mixed delivery problem with demand partition fDd ; Dh g. This includes solving
a subproblem with direct shipment for the suppliercustomer pairs in Dd and a subproblem with
hub-and-spoke delivery for the suppliercustomer pairs in Dh . Let Z 0 be the solution value of the
mixed system.
Step 6. If Z 0 < Z (i.e., the new solution is better than the previous one), then let the new solution
be the current solution and set Z Z 0 , otherwise, interchange the sending and receiving modes of
the direct shipment delivery and the hub-and-spoke delivery. Now, Z equals the value of the
current solution. If Z < Z m , then the new solution becomes the best solution obtained so far, and
we set Z m Z. If the best solution has not been improved for N consecutive iterations or the
current solution has not been improved for N 0 consecutive iterations, then stop; otherwise, go to
Step 4.

Note that Steps 13 of Heuristic H determine an initial solution, while Steps 46 form a so-
lution improvement procedure. The solution values of the mixed delivery problem, the pure direct
shipment problem, and the pure hub-and-spoke problem generated by the algorithm are Z m , Z d ,
and Z h , respectively.
In Case (i) of Step 4, the quantity vdij is an estimate of the improvement in the solution value if
demand qij is transferred from direct shipment to hub-and-spoke delivery, i.e., if the supplier
customer pair ui ; uj is moved from Dd to Dh . Similarly, in Case (ii) of Step 4, vhij is an estimate of
the improvement in the solution value if demand qij is transferred from hub-and-spoke delivery to
direct shipment.
Since vdij or vhij are updated at every iteration, it is possible that some demands transferred from
one delivery mode to the other are transferred back later. In the following, we discuss how the
estimates vdij and vhij are obtained. Since they are updated frequently, the computational time spent
on determining them signicantly aects the eciency of Heuristic H. Therefore, we suggest a
simple formula for determining these estimates. We let

vdij pijd chij ; 1

where pijd is an estimate of the savings if demand qij is removed from the current direct shipment
requirements and chij is an estimate of the cost increase if demand qij is added to the predicted hub-
and-spoke delivery requirements. However, without resolving the modied CVRP for the hub-
and-spoke mode in the mixed problem, it is not easy to predict to which route qij should be added.
To avoid resolving the CVRP, we let chij be the cost savings of taking out qij from the pure hub-
and-spoke problem. This estimate can be calculated from the result of Step 2 and used whenever
needed in the iterations.
Similarly, we let

vhij pijh cdij ; 2

where pijh is an estimate of the savings if demand qij is removed from the current hub-and-spoke
delivery requirements and cdij is an estimate of the cost increase if demand qij is added to the
332 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

predicted direct shipment requirements. Again, cdij here is set to the cost savings of taking out qij
from the pure direct shipment problem. It can be calculated from the result of Step 1.
To show the calculation of pijd and pijh , we dene the following notation for the current solution:

ni number of direct shipment routes from supplier ui (i 1; 2; . . . ; m);

sri length of the rth direct shipment route from supplier ui (i 1; 2; . . . ; m; r 1; 2; . . . ; ni );
srij savings in travel distance if we remove customer uj and its demand from the rth direct
shipment route of supplier ui (i 1; 2; . . . ; m; j m 1; m 2; . . . ; m n; r 1; 2; . . . ; ni );
ncol number of collection routes from the hub;
ndel number of delivery routes from the hub;
ncol;r quantity handled by the rth collection route in the hub-and-spoke system (r 1; 2; . . . ;
ncol );
ndel;r quantity handled by the rth delivery route in the hub-and-spoke system (r 1; 2; . . . ; ndel );
scol;r length of the rth collection route (r 1; 2; . . . ; ncol );
sdel;r length of the rth delivery route (r 1; 2; . . . ; ndel );
Rcol
i set of collection routes containing the goods of supplier ui (i 1; 2; . . . ; m);
Rdel
j set of delivery routes containing the goods of customer uj (j m 1; m 2; . . . ; m n).

We rst consider the cost savings when the demand, qij , is removed from the current direct
shipments. Let qi denote the direct shipment route that handles the customer demand, qij . Let
Ddqi fui ; uj jdemand qij is handled by route qig:
Then, we set
qi qi
sij  si
pijd P qi
: 3
k s:t:ui ;uk 2Ddqi sik

To understand the rationale behind this formula, consider the removal of demand qij from the
qi
route. When this demand is removed, the length of the route decreases by sij . However, when
two demands qij , qik are removed from the same route, the decrease in route length is not nec-
qi qi
essarily equal to sij sik if customers uj and uk are adjacent to each other (see Fig. 2). In the
extreme case, if all demands
P in the route are removed, then the reduction in route length should
qi qi
equal si instead of k s:t:ui ;uk 2Dd sik . Hence, for the removal of each customer, uj , from the
qi
qi
route, we approximate the cost savings by an amount proportional to sij , with proportionality
qi P qi
constant si = k s:t:ui ;uk 2Dd sik .
qi
Next, we consider the cost reduction when the demand, qij , is removed from the current hub-
and-spoke system. Note that when the customer demand, qij , is removed from a collection route in
the hub-and-spoke system, the length of that route remains unchanged unless qij is the only de-
mand from supplier ui to be collected on that route. Thus, we need P a fair assessment on the at-
tractiveness of removing demand qij from the route. Note that r2Rcol scol;r is the total P length of
i col;r
collection routes carrying products of supplier ui in the hub-and-spoke system, P and r2Rcol n is
i
the total demand handled by those routes. In other words, if the demand r2Rcol ncol;r is removed
i
from
P thecol;r hub-and-spoke system, the reduction in the total length of collection routes would be
r2Rcol s . Hence, we estimate the reduction in the total length of collection routes due to the
i
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 333

Fig. 2. The situation when demands qij and qik are removed from the route: (i) original route, (ii) if demand qij is
removed, (iii) if demand qik is removed, (iv) if both demands qij , qik are removed.
P P
removal of demand qij as qij r2Rcol scol;r = r2Rcol ncol;r . Similarly,
P we estimate
P the reduction in the
i i
total length of delivery routes due to the removal of qij as qij r2Rdel sdel;r = r2Rdel ndel;r . Therefore,
j j
we set
0P P 1
col;r del;r
col s r2Rdel s
r2R
pijh qij  @ P i col;r P j del;r A: 4
r2Rcol ni r2Rdel n j

Finally, we consider the cost increase when inserting demand qij into the existing direct shipment
requirements and the cost increase when inserting demand qij into the existing hub-and-spoke
delivery requirements. According to earlier discussions, we set the estimates of these quantities to
chij \pijd in the pure direct shipment problem" 5
and
chij \pijh in the pure hub-and-spoke problem": 6
Corresponding to the notations used for the current mixed solution, notations for the two pure
b col , R
delivery problems are dened as n^i , s^ri , ^srij , n^col , n^del , n^col;r , n^del;r , s^col;r , s^del;r , R b del , and D
bd ,
i j qi
respectively.
334 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

From Eqs. (1), (3), (4) and (6), we have

0P P 1
col;r del;r
qi qi s
^ b s
^
sij  si B r2bR col r2 R del
C
vdij P qij @ P i col;r P j del;r A:
qi
k s:t:ui ;uk 2Ddqi sik n^ n^
r2b
R col
i
r2b
R del
j

From Eqs. (2)(5), we have

0P P 1
col;r del;r
col
r2Ri s del
r2Rj s qi qi
^sij  s^i
h @
vij qij  P P A P :
r2Rcol ncol;r r2Rdel ndel;r qi
bd ^sik
i j k s:t:ui ;uk 2 D qi

4. Computational experiments

In this section, we assess via computational experiments the savings on the total travel distance
of vehicles due to the use of the mixed delivery system, in comparison with the use of a pure
delivery system. In the implementation of Heuristic H, the termination condition is set as N 7
and N 0 5. In general, the termination condition should set to balance the solution quality and
the computational eort. The above setting is based on test run results, which indicate that more
iterations will hardly make further improvement. The experiments are carried out on a variety of
dierent problem settings. Results are analyzed in terms of the relative improvements and the
number of cases improved. Additional experiments are also carried out on the impact of including
milk runs in the mixed system.

4.1. Problem instance generation

All the problem instances are dened within a square of unit length, which may be considered
as a scaled version of practical delivery regions. For each instance, the m n supplier and cus-
tomer locations are uniformly distributed in the square area. Euclidean distance is taken as the
travel distance between any two of these locations. There is a delivery order for each customer
from each of the suppliers. The delivery demand of customer j from supplier i, qij , is randomly
generated from Uniforma; b, where a and b are prespeciedPnumbers. The hub location is de-
termined by solving the gravity problem that minimizes mn ~i t0i (see Francis and White,
i1 q
1974, [p. 170]). The vehicle capacity, Q, is set to 10 units, and the number of suppliers, m, is
set to 5.
To represent a wide range of situations, the number of customers and the parameters of the
demand distribution are set to vary at several levels. The number of customers, n, is set to 10, 15,
20, and 25. The values of a and b, which determine the mean and coecient of variation of the
demand distribution, are set according to Table 1. We only use those combinations of a and b that
satisfy the condition of b 6 10, since, by our assumption, all demands should not be larger than
the vehicle capacity. There are 72 such combinations. Taking into account the four dierent
values of n, there are altogether 288 dierent settings of problem parameters. For each of these
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 335

Table 1
Combinations of a and b used in the computational experiments
Mean Coecient of variation
0 0.058 0.115 0.173 0.231 0.289 0.346 0.404 0.462 0.520 0.577
1 b 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0
a 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0
2 b 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8 4.0
a 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0
3 b 3.0 3.3 3.6 3.9 4.2 4.5 4.8 5.1 5.4 5.7 6.0
a 3.0 2.7 2.4 2.1 1.8 1.5 1.2 0.9 0.6 0.3 0.0
4 b 4.0 4.4 4.8 5.2 5.6 6.0 6.4 6.8 7.2 7.6 8.0
a 4.0 3.6 3.2 2.8 2.4 2.0 1.6 1.2 0.8 0.4 0.0
5 b 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 10.0
a 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0
6 b 6.0 6.6 7.2 7.8 8.4 9.0 9.6
a 6.0 5.4 4.8 4.2 3.6 3.0 2.4
7 b 7.0 7.7 8.4 9.1 9.8
a 7.0 6.3 5.6 4.9 4.2
8 b 8.0 8.8 9.6
a 8.0 7.2 6.4
9 b 9.0 9.9
a 9.0 8.1

settings, 40 problem instances are randomly generated. Therefore, a total of 11,520 problem in-
stances are generated and used in the experiments.

4.2. Experiment results

For each of the above problem instances, we apply Heuristic H to obtain Z d , Z h , and Z m . The
relative savings from adopting the mixed system as compared with the pure systems,
minfZ d ; Z h g Z m = minfZ d ; Z h g, can then be calculated. For any group of the problem instances,
we can analyze the savings using the following two performance measures:

The average relative savings;

The proportion of instances with positive savings.

For all the problem instances used in the experiments, the overall average relative savings is
4.1% with a maximum of 24.1%; the proportion of instances with positive savings is 69.4%. This
indicates that the mixed delivery system operated using our heuristic creates savings in most cases
with dierent parameter settings and the savings are substantial. This comparison is made with
the better of the two pure delivery systems. Note that since the pure systems can be considered as
special cases of the mixed system, the performance of the mixed system is at least the same as the
better of the two pure systems. When compared with either of the pure systems, the savings is even
336 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

more signicant. The overall average relative savings and the proportion of instances with positive
savings are 9.9% and 73.4%, respectively, when compared with the pure direct shipment system.
The gures are 11.5% and 84.1%, respectively, when compared with the pure hub-and-spoke
system.
If the mixed system cannot be implemented, e.g., if the hub cost is too high, there is no in-
formation system support, or suppliers are unwilling to cooperate, then we need to identify the
best alternative. To compare the two pure systems based on the experimental results, we dene the
relative change in total travel distance of the pure hub-and-spoke system with the pure direct
shipment system as a reference: Z h Z d =Z d . Again, for any group of problem instances, the
performance measures below can be calculated and evaluated.

The average relative change in total travel distance (T );

The proportion of instances when pure direct shipment is better than pure hub-and-spoke de-
livery (B).

The overall average T is only 1.3%. The overall B is around 61%. This means that the overall
dierence between the performances, in terms of the total distance traveled, of the two pure
systems is minimal though pure direct shipment performs slightly better. This result also indicates
that the problem instances used in this study do not particularly favor either pure delivery system.

4.3. Further discussion and analysis

To analyze the impact of the problem parameters on the savings of the mixed system, we use a
regression model to relate the average relative savings to the following factors: the number of
customers (n), the demand mean (mean), and the coecient of variation of the demand (cv), and
their interactions, n  mean, n  cv, and mean  cv. Each parameter combination and its cor-
responding average relative savings are taken as a data entry for the regression model. The sample
size is 288. The resulting coecient of determination for the model R2 39:1%. Hypothesis
testing is conducted to verify whether the coecient of each factor in the regression model is
signicantly dierent from zero. The estimated coecients of the factors in the model and their P
values in the hypothesis testing are summarized in Table 2. From the P values, we can see that all
factors except n  mean and mean  cv are insignicant. The most signicant factor is mean  cv.

Table 2
Coecients of the factors in the regression model and their P values
Factor Coecient P -value
Constant 0.03501 0.020
n )0.0013336 0.094
Mean 0.001085 0.668
cv 0.00393 0.904
n  mean 0.0003065 0.021
n  cv )0.002142 0.171
Mean  cv 0.014708 0.001
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 337

This indicates that the distribution of the demand quantities signicantly aects the relative
savings of the mixed system.
Note that milk runs are included in our study. The proposed heuristic algorithm not only
decides the mode of delivery for each demand, but also gives vehicle routing results for the de-
liveries. If milk runs are not allowed, then the problem will become only to decide the delivery
mode of each demand and therefore will be much easier to solve. Such a problem can be for-
mulated as the following integer programming model:
X
m X
mn X
m X
mn
Minimize tij xij ti0 yi t0j zj
i1 jm1 i1 jm1

X
mn
subject to qij 1 xij 6 Qyi ; i 1; 2; . . . ; m
jm1
X
m
qij 1 xij 6 Qzj ; j m 1; m 2; . . . ; m n
i1

xij 2 f0; 1g; yi and zj are integers;

where variable xij indicates in which mode demand qij should be delivered (equal to 1 if by direct
shipment; 0 otherwise), yi is the number of trips needed between supplier i and the hub, and zj is
the number of trips needed between the hub and customer j.
To sense how milk runs aect the total traveling distance, we take various problem instances
(n 10, 20; mean 3, 6; cv 0, cvmax , where cvmax is the highest value of cv shown in Table 1)
and optimize the mixed system without milk runs using this model. The total distance traveled in
the system is on average 17.75% longer than that in the system with milk runs. This explains why
most practical delivery systems allow milk runs.
When milk runs are not allowed, the distance of the pure systems will also increase. The results
can also be obtained using the above model by xing all the x values to 1 for the pure direct
shipment system and to 0 for the pure hub-and-spoke system. If all three systems do not include
milk runs, our test on the selected problem instances shows that the mixed system can save 13.28%
of traveling distance compared with the better pure system, and save 24.76% and 24.22% com-
pared with the pure direct shipment system and pure hub-and-spoke system, respectively. All
these savings are higher than in the situation that includes milk runs.

5. Conclusions

In this paper, we studied a mixed truck delivery system that allows both direct shipment and
hub-and-spoke deliveries. A heuristic was developed to determine the mode of delivery for each
demand and to perform vehicle routing in both modes of delivery. Computational experiments
were carried out to compare the mixed delivery system with the pure delivery systems. From the
computational results, we can conclude that the mixed system is more eective than both pure
systems. The delivery plan produced using the heuristic for the mixed system saves about 4% in
total distance on average compared with the best of the pure systems. The savings is about 10% on
average if compared with any one of the pure systems. Analysis was also done on the impacts of
338 J. Liu et al. / Transportation Research Part E 39 (2003) 325339

problem parameters on the relative savings of the mixed system. The results showed that the
demand distribution aects the relative performance of the mixed system most signicantly. The
eect of including milk runs in the systems was also discussed. Note that in our computational
study, the xed cost of operating the hub was ignored. Thus, to evaluate the overall cost of the
systems, one should compare the transportation costs of the systems obtained from this study
with the costs of operating the hub in order to make a fair comparison. In fact, the traveling cost
savings of adopting the mixed system can be viewed as an upper bound on the hub cost in order to
make the system more cost eective than pure direct shipment.
This study is limited to the problem with one hub and homogeneous vehicles. However, it can
serve as a basis for further research in a number of directions. The heuristic proposed here is based
on local search. A direct extension is to try other types of methods such as genetic algorithm, tabu
search, etc., to search for better solutions. The problem studied here may be extended to more
complex situations, for instance, to consider multiple hubs, heterogeneous vehicles, or delivery
time window constraints. In particular, the problem with multiple hubs involves many new fea-
tures and is more dicult to solve. Examples of new decisions in the multi-hub system include
which hub to use for a demand and how to arrange the transportation among the hubs. The hub-
and-spoke system in such a problem is more complicated.

Acknowledgements

The authors would like to thank Professor Wayne K. Talley and two anonymous referees for
their helpful comments and suggestions.

References

Abdinnour-Helm, S., Venkataramanan, M.A., 1998. Solution approaches to hub location problems. Annals of
Operations Research 78, 3150.
Aykin, T., 1995a. The hub location and routing problem. European Journal of Operational Research 83, 200219.
Aykin, T., 1995b. Networking policies for hub-and-spoke systems with application to the air transportation system.
Transportation Science 29, 201221.
Bodin, L., Golden, B., 1981. Classication in vehicle routing and scheduling. Networks 11, 97108.
Campbell, J.F., 1996. Hub location and the p-hub median problem. Operations Research 44, 923935.
Christodes, N., 1985. Vehicle routing. In: Lawler, E.L., Lenstra, J.K., Rinnooy Kan, A.H.G., Shmoys, D.B. (Eds.),
The Traveling Salesman Problem. Wiley, Chichester.
Clarke, G., Wright, J.W., 1964. Scheduling of vehicles from a central depot to a number of delivery points. Operations
Research 12, 568581.
Fisher, M.L., 1995. Vehicle routing. In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.),
Handbooks in Operations Research and Management Science, Vol. 8. Network Routing, North-Holland,
Amsterdam.
Francis, R.L., White, J.A., 1974. Facility Layout and Location: An Analytical Approach. Prentice-Hall, Englewood
Clis.
Gillett, B., Miller, L., 1974. A heuristic algorithm for the vehicle dispatch problem. Operations Research 22, 340349.
Hall, R.W., 1987. Direct versus terminal freight routing on a network with concave cost. Transportation Research B
21B, 287298.
Laporte, G., 1992. The vehicle routing problem: an overview of exact and approximate algorithms. European Journal
of Operational Research 59, 345358.
 J. Liu et al. / Transportation Research Part E 39 (2003) 325339 339

Laporte, G., Gendreau, M., Potvin, J.-Y., Semet, F., 2000. Classical and modern heuristics for the vehicle routing
problem. International Transactions in Operational Research 7, 285300.
Lee, H.L., Whang, S., 2001. Winning the last mile of e-commerce. MIT Shan Management Review 42 (4), 5462.
Lin, S., 1965. Computer solutions of the traveling salesman problem. Bell System Technical Journal 44, 22452269.
OKelly, M.E., 1987. A quadratic integer program for the location of interacting hub facilities. European Journal of
Operational Research 32, 393404.
OKelly, M.E., Bryan, D.L., 1998. Hub location with ow economies of scale. Transportation Research B 32, 605616.
Pirkul, H., Schilling, D.A., 1998. An ecient procedure for designing single allocation hub and spoke systems.
Management Science 44, S235242.
Sasaki, M., Suzuki, A., Drezner, Z., 1999. On the selection of hub airports for an airline hub-and-spoke system.
Computers and Operations Research 26, 14111422.