Transportation Research Part A 41 (2007) 523–536

www.elsevier.com/locate/tra

The European freight railway system

as a hub-and-spoke network
Seung-Ju Jeong a, Chi-Guhn Lee b, James H. Bookbinder c,*

a
Center for Northeast Asian Transportation, Logistics and Transport Economics, The Korea Transport Institute (KOTI),
Koyang-city, Kyunggi-do 411-701, Republic of Korea
b
Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, Ont., Canada M5S 3G8
c
Department of Management Sciences, University of Waterloo, 200 University Avenue West, Waterloo, Ont., Canada N2L 3G1

Received 23 August 2005; received in revised form 20 September 2006; accepted 21 November 2006

Abstract

This paper addresses a hub-and-spoke network problem for railroad freight, where a central planner is to find transport
routes, frequency of service, length of trains to be used, and transportation volume. Hub-and-spoke networks, often found
in air freight, have not been favoured by railways in the past. Such a structure could be profitable, however, if there exist
concentrated freight flows on some service links. We formulate a linear integer programming model whose objective func-
tion includes not only the typical operational cost, but also cost due to the transit time spent by freight in the network. We
then develop heuristic algorithms to solve large scale instances occurring in rail freight systems in France plus Italy;
Germany; and a 10-country European network. By assuming that every node is equipped with consolidation capability,
we let the final solution naturally reveal potential hub locations, the impact of several of which is studied by sensitivity
analysis.
 2006 Elsevier Ltd. All rights reserved.

Keywords: Hub-and-spoke problem; Rail freight; Network design; Optimization; Logistics; Europe

1. Introduction

With the rise in membership in the EC (European Community), it is logical to expect greater flows of trans-
border freight. The trucking mode is highly convenient for pickups and local deliveries; linehaul is another story.
Road networks are already under intense pressure due to concerns about pollution from the larger number of
shipments. Physical boundaries such as mountains cause environmental worries, and slow speeds in any case.
One effect is the trend toward transshipments, associated with ‘‘City Logistics’’. Thompson and Taniguchi
(2001) and Whiteing et al. (2003) elaborate on this concept: Several carriers will transship to another, with the

*
Corresponding author. Tel.: +1 519 888 4013; fax: +1 519 746 7383.
E-mail address: [email protected] (J.H. Bookbinder).

0965-8564/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tra.2006.11.005
524 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

latter trucking company making all deliveries in the urban core. This idea may hold considerable promise for
goods transported within a given country.
It is clear, however, that only some of the increased freight can be absorbed by trucking; the European
highway network simply does not have the potential for expansion. For the right products, barge transporta-
tion on inland waterways is very useful. The barge mode is more applicable in Europe than in North America,
but its discussion is beyond our scope here.
Extensive rail networks exist in Europe. The International Journal of Transport Management recently pub-
lished a special issue, volume 2 (1), on European Rail Policy. Most of the articles concern rail passenger ser-
vice, but Crozet (2004) compares Great Britain, France and Germany in terms of their infrastructure charges
for access to the railroad system by a third-party operator.
Transportation Research A has also published a special issue, volume 25 (4), on European Rail. Dejax and
Bookbinder (1991) discuss the reliability of special services of SNCF, the French National Railway, that were
designed to compete with medium- or long-distance truck in the JIT (Just-in-Time) market. The rail freight
services in the EC do tend to have a schedule, whereas railroads in the US or Canada would more often wait
to dispatch a train, doing so only when sufficiently many cars have been accumulated.
Our focus in this paper is thus goods transport by rail. The enhanced volumes mentioned above raise the
importance of operating the freight system more efficiently. Hub-and-spoke networks, although better known
for air passenger transport and LTL-trucking, may merit consideration. That is one purpose of the present arti-
cle: We test the viability of several railway networks containing various numbers and locations of potential hubs.
In this paper, we consider rail freight between 10 countries, all but two of which are now in the EC. Our
sources of data on flows and costs are EUFRANET (1999) and SIMIQ (2000); the input parameters on loca-
tions and distances are from readily available maps. We will design various types of hub networks. These will
later be described as having (a) only some countries in the network; (b) more countries involved, but [in both
(a) and (b)] only particular countries can have hubs, or none can; and (c) all countries involved, and any node
can be a hub. Results for these systems will be compared in terms of cost or service, and on each country’s
‘‘contribution to the performance of the overall network’’.
A hub-and-spoke network structure aims to save transportation cost by concentrating volume at the hubs.
Research has been done on the optimal routing of freight and on performance analysis. The hub location
problem has been studied by O’Kelly and Bryan (1998) and by Pirkul and Schilling (1998) and others. A com-
prehensive survey of research on hub-and-spoke networks appears in Campbell et al. (2002). The high com-
plexity of the hub location problem has often required assuming that all the movements of goods should
involve one or more hubs between source and destination.
Aykin (1995) considers a hub-and-spoke strategy for air transportation which we will refer to as a hybrid
network: Material or passenger movements originating from a source can reach a destination without visiting
any hub; can go to a hub and then move directly to destination; or go to a hub, then to a second hub after
being consolidated with other materials (passengers), and moving onward to final destination. Different coef-
ficients for goods movement with and without hubs capture the extra cost due to shipment consolidation at
hubs. Jaillet et al. (1996) address a more general hub-and-spoke problem whose patterns of dispatch include
direct air-service, 1-stop service, 2-stop, . . ., up to unlimited-stop service. Every site is a potential hub, but time
delay at hubs is not considered.
The research discussed so far addresses the problem in the context of air freight. SIMIQ (2000), one of the
few studies that solve the rail hub location problem, treats economies of scale in the form of a piecewise-linear
objective function; this approximates the cost that varies with transportation volume. A heuristic solution is
obtained for a model with 28 nodes selected from the actual railway system in Europe. That network is the
focus of the present paper as well. Racunica and Wynter (2005) also consider a rail hub-and-spoke network,
in a more theoretical treatment emphasizing polyhedral properties.
Another research avenue is the rail service-network design problem, where the frequency of service, the
routes, and hub location are simultaneously found. Assad (1980) determines the type of trains, the number
of each, and service routes so as to minimize the total operating cost. Crainic et al. (1984) and Crainic and Rous-
seau (1986) include additional decisions on service frequency and train makeup. Decomposition or metaheu-
ristics are incorporated in more recent studies by Haghani (1989), Keaton (1992) and Marı́n and Salmerón
(1996a,b).
 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536 525

In the present article, we consider hub-and-spoke service networks for rail freight, in which goods can be
transported through however many hubs as long as doing so reduces overall cost. That cost is defined based
on total time, incorporating delays at hubs, and modelled using cost functions discussed in Appendix A. There
we also summarize a heuristic method (which we term a hybrid approach, HA) developed to solve large-scale
problem instances. Numerical examples from the European railway system, including policy implications, are
presented in Section 3. The final section concludes with a summary of our results. We begin in Section 2 with
the precise hub-and-spoke problem that we address.

2. Problem definition

Let us consider a hub-and-spoke network where (A1:) A central planner decides and supplies transporta-
tion service (type of train, frequency of operation, transportation routing patterns, consolidation volume at
each node). (Designations A1, . . . , A6 indicate our assumptions.) (A2:) Every node is a potential hub with con-
solidation capability. To identify hub locations which minimize total cost, we employ a hybrid network: (A3:)
Rail cars are allowed to move from origin to destination directly or to be consolidated at one or more inter-
mediate hubs.
Materials experience non-negligible delay at each hub due to consolidation, classification of cars, and wait-
ing for the next available train. Computation of that time delay is facilitated by assumptions (A4:) Freight
arrives according to a Poisson process, and (A5:) When a train heading for a destination departs a node,
no freight remains there for the same destination. (It is sufficient that all shipments tendered to the railway
be in full-carload multiples, and train capacity be great enough to clear all goods waiting at a node.)
In more detail, assumptions A2 and A3 allow the model to select the locations of hubs from a set of can-
didates: A node that consolidates a high volume of freight is potentially a good site at which to install a hub.
We also assume (A6:) The transportation network is complete. This is easily imposed by regarding the path
between two nodes as a ‘‘direct’’ arc over which a train can travel, and whose distance is the path length.
A mathematical formulation of our problem is in Appendix A. (Additional details are contained in O’Kelly
and Bryan, 1998 or Campbell et al., 2002.) A node represents a location (shipment origin, destination, or
transshipment point); an arc denotes a segment of track connecting two nodes. A service network is thus a
graph of all feasible directed arcs defined between nodes, while a route is an ordered set of service links form-
ing a connected path from one node to another (see Fig. 1).
In the hub-and-spoke problem of interest for us, we try to simultaneously find good service frequencies and
transportation routes. Freight moves from origin to destination according to that best route, with transit time
dependent on service levels of the associated arcs and nodes. The objective is minimization of total cost, i.e.
both operational and time-delay cost.

3. Choices of European hubs

3.1. Candidate locations

The heuristic solution HA of Pb4 (Table 1), defined on the entire European network, sheds light on the use
of our model to identify potential hub locations: High volumes of freight do move to several sites from which
shipments are redistributed to final destination. In Table 1, the second column exhibits the total volume of
freight moving into the node shown in the first column. The third column indicates freight volume that comes

Fig. 1. Service network and routes: (a) physical network, (b) service network, (c) examples of routes from 1 to 3.
526 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

Table 1
Heuristic solution HAa for Pb4b (European network as a whole)
Zone Inc Consolc Ratiod Zone Inc Consolc Ratiod
1. Antwerp 5.25 0 0 25. Sachsen 9.01 0.16 1.7
2. Copenhagen 6.82 2.65 38.9 26. Neumunster 6.50 0 0
3. Stuttgart 14.6 0.85 5.8 27. Paris 26.2 13.6 52.1
4. Karlsruhe 62.0 50.8 82.0 28. Reims 3.96 1.46 36.8
5. Freiburg 11.8 4.13 35.1 29. Le Havre 4.46 0 0
6. Tübingen 1.47 0 0 30. Montlouis 6.41 4.54 70.9
7. Munich 51.3 24.4 47.5 31. Lile 7.08 0.85 12.0
8. Landshut 5.51 0.84 15.2 32. Metz 12.7 7.62 60.2
9. Regensburg 14.6 5.60 38.4 33. Strasbourg 5.91 2.56 43.2
10. Nürnberg 30.5 22.7 74.4 34. Poitiers 1.38 0 0
11. Augsberg 11.9 5.13 43.0 35. Bordeaux 2.86 0.22 7.7
12. Berlin 23.7 14.9 62.9 36. Lyon 20.2 12.2 60.6
13. Bremen 40.1 7.81 19.5 37. Montpellier 1.61 0.57 35.2
14. Hamburg 88.5 34.6 39.1 38. Marseilles 9.70 1.98 20.5
15. Frankfurt 25.8 13.5 52.3 39. Milan 14.4 6.96 48.3
16. Braunschweig 8.64 0.11 1.3 40. Verona 5.92 1.05 17.8
17. Hannover 7.52 2.55 33.9 41. Parma 2.25 0 0
18. Oldenburg 3.80 0 0 42. Rotterdam 4.62 0 0
19. Düsseldorf 52.3 29.9 57.2 43. Linz 17.9 3.05 17.0
20. Cologne 19.6 7.35 37.4 44. Steiermark 3.20 0 0
21. Münster 1.99 0 0 45. Norra ellansverige 0.40 0 0
22. Dortmund 30.0 12.70 42.2 46. Basel 3.20 0 0
23. Koblenz 1.11 0.54 48.7 47. Warsaw 1.23 0 0
24. Saarland 3.00 0.59 19.6 48. Prague 1.80 0.13 7.3
Total 695 299 43.0
a
Method described in Appendix A.
b
Problem instances Pb1, . . . , Pb4 are summarized in Table 3 (Appendix A).
c
000, ton/week.
d
Consol/in, %.

to the node, is consolidated, and distributed to another node, and the last column presents the percentage of
freight which stops by the node just for consolidation purposes.
Locations that receive a large volume of freight and demonstrate a high consolidation ratio can be
considered as potential sites for a hub. The model suggested as hubs (hence bold font in Table 1)
Karlsruhe, Munich, Nürnberg, Hamburg, Düsseldorf, Paris, Metz and Lyon. It should be noted that
Germany has already adopted hub-based railroad freight transportation; the five hub sites there each con-
solidate more than 20,000 ton of freight per week. Moreover, Paris and Metz are currently used as hubs
and Lyon has been considered for such use. With the exception of Munich and Hamburg, the remaining
six locations show a consolidation ratio above 50%. Karlsruhe especially seems to be an excellent choice of
hub, due to its large freight volume (more than 60,000 ton/week) and its high consolidation ratio (over
80%).
Economies of scale in transportation operations are found in a solution for the whole European network:
Hub-to-hub links have 5840 ton/week, hub to or from non-hub links carry only 3502 ton, while links between
two non-hubs move just 1869 ton weekly, on average. Naturally, trains with the greater capacity (600 ton) are
intensively used on hub-to-hub links (90%), whereas the 300-ton-capacity trains are employed on 73% of the
hub-to-non-hub links and on 84% of links between two non-hubs. The solution consists of direct shipments
(45%), one-stop routes (38%), and routes with two stops (13%) and three-or-more (4%).

3.2. Synthesizing the hub-and-spoke solution

The problem instances Pb1–Pb4 pertain to networks of different sizes, in particular countries. They have
this in common: Once we specify the freight origins and destinations, any such city can be a hub if it makes
 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536 527

economic sense. Here, we break down the problem instances to understand which hubs are most important in
networks of various sizes. Conversely, what is the opportunity cost of allowing no hubs? Or of dis-allowing
hubs in a particular country?
Examples a–j are described in Table 2. The cases b, d, f and j are exactly the problem instances Pb1–Pb4. In
addition, however, we show the implications in those problems of allowing no hubs anywhere (respectively,
Examples a, c, e and g).
Instance Pb4 is further delineated by Examples h (no hubs in Germany) and i (none in France), thereby
permitting the hubs to be ranked in terms of their cost influence. Comparison of cases g and j shows a total
savings in Table 2 of about 30% in cost per week, when no hubs are permitted vs hubs are allowed anywhere.
However, 80% of that benefit (i.e. 25% saving in overall cost) comes when hubs are possible only in France or
in Germany. It is interesting to see that the cost savings in the whole European network, when hubs are pos-
sible in either the German network (Example i in Table 2) or the French network (Example h), is almost as
great as when both the German and French networks are allowed to have hubs (Example j in Table 2). Intrigu-
ingly, benefits from acting as hubs in their own isolated network is a different story; the French network enjoys
significant saving when their locations are permitted to be hubs (Examples a vs b), while only minor savings
result in the German network when hubs are allowed there (cf. Examples c and d).
Let us now study the roles played by particular cities, and how their influence changes when the railway
system is constrained to varying degrees. Figs. 2–5 contain the details of Examples g–j, respectively. Cities
are designated in those maps by the numbering scheme of Table 1.
Consider Example g, the European network with no hubs allowed at any of the 48 locations. Thus, none of
those nodes has a consolidation ratio greater than 50%; inbound volumes exceed 20,000 ton/week at only four
sites (Fig. 2): Munich, Bremen, Hamburg and Düsseldorf. Only non-German hubs are permitted in Example
h, so while the four preceding cities retain the same scale of inbound volumes, that group then expands (Fig. 3)
to include nodes in France, namely Paris, Metz, Strasbourg and Lyon, plus Linz in Austria. Consolidation

Table 2
Ten examples showing details of problem instances Pb1–Pb4
Problem Number Countries included in Example Description Total Hybrid-Algorithm Computation
Instancea of network Freight (HAb) Solution cost time, HAb
nodes (000 ton/ (two train types) (min)
week) (000 Euros)
Pb1 15 France + Italy a Pb1, but no hubs 47.8 4295 0.35
b Pb1, with hubs 3189 0.40
Pb2 24 Germany c Pb2, but no hubs 210 13,798 6.47
d Pb2, with hubs 11,925 8.62
Pb3 33 Europe  (France + Italy) e Pb3, but no hubs 256 19,924 21.3
f Pb3, with hubs 15,462 27.6
Pb4 48 European network as a g Pb4, but no hubs 396 35,854 99.2
whole anywhere
h Pb4; no hubs in 26,729 163
Germany, but hubs
possible elsewhere
i Pb4; no hubs in 25,861 197
France, but hubs
possible elsewhere
j Pb4, with hubs 25,197 149
possible
everywhere
Each sub-network contributes in varying degrees to the overall hub-and-spoke solution (Example j) for the European rail network. Nodes
in each case are the appropriate ones in Table 1.
a
Problem instances Pb1, . . . , Pb4 are further summarized in Table 3 (Appendix A).
b
Method described in Appendix A.
528 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

Fig. 2. Example g: Inbound volumes for the 48 cities (as designated in Table 1), and appropriate consolidation ratio.

ratios exceeding 50% are then attained by those four French nodes, plus other locations (Milan, Rotterdam,
Basel, Prague) whose weekly inbound volumes are smaller.
Moving to Example i, where hubs are now allowed in Germany, Fig. 4 shows that only Düsseldorf of the
four original German cities has a consolidation ratio greater than 50%. Ten nodes in Germany (i.e. six addi-
tional ones) have inbound volumes exceeding 20,000 ton/week, of which four sites (Karlsrühe, Nürnberg,
Berlin and Dortmund) also have consolidation ratios above 50%.
It is interesting as well to observe the effect on other locations of the hub restrictions in Examples g and h,
relative to the final solution of Example j. We first discuss the case of Prague. With hubs allowed anywhere
(Table 1), this city has just 1800 ton of inbound freight, of which 7% is consolidated. By contrast, with no hubs
allowed in Germany, Fig. 3 shows that Prague has more than 10,000 ton inbound each week, and the majority
of that is consolidated. This Example h suggests that Prague is a good approximation to a ‘‘German’’ hub, if
the latter is impossible.
Comparison of Examples j and h is even more striking for the case of Basel: Nothing at all is consolidated of
its small volume in the unconstrained final solution (Table 1), while more than half is consolidated of a weekly
volume that exceeds 10,000 ton in Example h. When hubs are not permitted in Germany, Basel thus behaves
like one (Fig. 3). But because its location also straddles France, Basel can simulate a French hub in Example i
(Fig. 4).
 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536 529

Fig. 3. Example h: Inbound volumes for the 48 cities (as designated in Table 1), and appropriate consolidation ratio.

3.3. Policy implications

In the formulation of our model, we assumed (A2) every node had the capability to consolidate, hence
could be a potential hub. But in the present practice, many busy nodes could not be operated that way. Some
new investment is required; its annual return would be the decrease in system costs when that node is a hub, vs
not.
We have studied, in Examples a–j, systems constrained to various degrees in the locations of hubs. The rea-
sons for those constraints could be political or financial. Eventually, among the feasible systems, it will be nec-
essary to answer the question, Which single hub will have the earliest priority in implementation? Which
choice of two hubs?
The optimal multiple-hub system, Example j, is the one in Table 2 of lowest total cost for Europe overall.
But that system might differ considerably from present operations: Many nodes may have alternative roles, of
varying importance.
Suppose it could be agreed as to the one new hub most important for initial implementation. The sequence
would be as follows. Investment would be made, that hub would become operational, and additional infor-
mation on costs and throughput would be gathered there and for the whole network. This updating of figures
on flows and costs is quite important for the ongoing use of the model. It would be run again, with this first
530 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

Fig. 4. Example i: Inbound volumes for the 48 cities (as designated in Table 1), and appropriate consolidation ratio.

node constrained to be a hub, to see the model’s prediction of the best successor hub to be implemented. Again
the pattern would be investment in that hub; operation; collection of data on the now-current system; use of
that data in continuing model runs.
When aided as above by an optimization model, the progression of capital investments can be quite system-
atic. As in any model-based analysis, a user would wonder about the extent to which the input data on costs
and volumes are ‘‘representative’’. This is a major reason for the gathering of data next year and the year after.
Sensitivity analyses concerning flows and costs are facilitated, and of course, the following step in the series of
hub investments will be based on the latest system results.
Many of the preceding considerations hold as well for a private-sector logistics decision such as sequential
improvement of a company’s distribution system. For example, any study may be limited by its data and
model assumptions. Additional factors in the public sector, for our specific problem, include the following:

• If the first-choice hub is in Germany, say, how do we distinguish between transportation solely within that
country, vs flows originating or ending outside Germany? Distinguish between German investment, vs
investment by the European Community?
• To what degree is the EC policy on rail transport expected to remain constant, or change over time?
 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536 531

Fig. 5. Example j: Inbound volumes for the 48 cities (as designated in Table 1), and appropriate consolidation ratio.

Directives of the European Community have taken the following positions on transportation by rail. The
state-run rail operator in each country is to have no advantage over new entrants. Those new carriers, whether
operators of intermodel transport or international groupings of rail companies, are to be given access at a fair
price to the infrastructure capacity in each nation of the EC.
There is thus an obvious tension between national regulatory authorities and the policy makers of the EU.
But the required neutrality in pricing and in deciding access to any country’s rail system is to be guaranteed by
a clear division: The railway network is to be run by separate infrastructure managers, under the supervision of
a proposed EC rail regulator, independently of any rail operators. The goal is that each national network be
opened up to international freight transit.
Those types of goods movement are highlighted in this paper. It was not necessary to ask, which of several
competing rail carriers will haul the freight? That is naturally of concern to a particular railway. But assuming
the preceding arrangements guarantee competitive access to each rail operator, determination of the mini-
mum-cost network flows can proceed as in our analysis.
Model (2)–(7) suggests certain choices of hubs are favourable to the system overall. Above, we outlined how
to determine the implementation sequence to progress from a network of few hubs to one with near-optimal
numbers and locations of hubs. The benefits of such a system extend to the EC as a whole. A European frame-
work thus seems necessary for the infrastructure investments required.
532 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

4. Conclusions

In this paper, we have studied the hub-and-spoke network transportation problem, where service frequency
and routes of trains are to be determined so as to minimize total cost. The objective function includes tradi-
tional expenses for operational items (train cost, mileage charge), but also transit-time cost. The latter is pro-
portional to the total time that freight has to spend in the network and includes the sum of transportation
time, processing time at hubs, and the delay or wait until the next available train.
We formulated a linear 0–1 programming model and developed heuristic algorithms FI and HA (see Appendix
A) to solve realistically-sized instances. Our heuristics perform almost as well as the exhaustive algorithm based
on the work of Crainic and collaborators (see Tables 4 and 5). Our model was employed to identify potential loca-
tions for hubs. Study of the actual European railway network revealed interesting patterns of freight movement.
Our main transportation results are contained in Tables 1 and 2, and may be summarized as follows. Cities
in France and Germany are the most important hubs of the European rail network. Use of one or more hubs
decreases total costs by 26%, 14%, 22% and up to 26%, in instances Pb1, . . . , Pb4, respectively. In Pb4, for
example, 65% of goods (by weight) or 55% of shipments go through hubs to their destination. (Problem
instances are outlined in Tables 2 and 3.)
In light of these benefits, a reader may wonder why hub-and-spoke rail networks are not employed more
widely, e.g. in North America. Railway networks in Europe are much more ‘‘complete’’ than in the US and
Canada. Even with the trend in North America to fewer, larger railroads, there is nothing like the single,
‘‘national-flag’’ carrier within the EC countries. Networks in Canada or the US may thus not have been sub-
ject to the same broad scrutiny (rationalization, or ‘‘filling in’’) as in Europe.
Consider the following example. London, Ontario is a major city that is 120 miles west of Toronto and
132 miles east of the American city of Detroit. Shipments from London to Detroit used to go direct, on
the track of the CN Railway. The Goderich and Exeter shortline railroad today owns that London–Detroit
portion, so a shipment between this origin and destination now follows the route London ! Toronto !
Detroit (where each link belongs to the CN network). Although Toronto is an obvious hub, the goods have
to travel east to go west, a total distance of 353 miles vs 132 miles previously.
This inconvenience is caused by the layout of the track. The carrier benefits in cost savings from a hub, but
the consignee may suffer poor customer service, unless the elapsed time can be decreased due to additional
dispatches permitted by a hub’s greater volume. Successful use of railway hubs in Germany and France
may be due to a more condensed geography; a rail network of greater ‘‘completeness’’ (as noted above); or
the larger quantity of freight available at a railroad hub, in countries where truck is less prevalent than in
North America.

Appendix A. Mathematical formulation and solution procedure

A.1. Model formulation

Let us introduce the following notation:

S(r) set of service links that constitute route r
P set of train types differentiated by capacity
cTp cost of using train of type p per km and ton
dpr total distance serviced by train of type p on route r
cD cost of time delay of freight per unit time
s average time taken to complete operation at a single hub
nr the number of hubs involved in route r
ch cost of switching and consolidation at a hub per ton of freight
H number of periods in a planning horizon
qpj service frequency of train of type p on link j
mp the average speed of train of type p
wm volume of freight m
 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536 533

The cost of waiting for the next train can be modelled as delay in a queueing system. Following Daganzo
(1986) or Keaton (1992), and based on assumptions A4 and A5, we employ the average waiting time over a
period of H as H/(2qpj), when link j is serviced by a train of type p at frequency qpj. The total waiting cost is
thus (cD Æ H Æ wm)/(2qpj).
When route r transports freight m from origin to destination, the associated cost is
( )
X X X 1 X d pr
R T D h D D
cmr ¼ wm cp  d pr þ c  s  nr þ c  nr þ c  H þc ð1Þ
p2P p2P j2SðrÞ
2qpj p2P
vp

The routing cost given in Eq. (1) contains five terms, beginning with the sum of railway operating costs over
the entire route r. Next is the time-delay cost of sorting, handling, etc. at all hubs involved in that route. The
third term is the total consolidation cost (proportional to the volume consolidated), while the fourth repre-
sents waiting for the next available service. The last term is the cost proportional to the length of travel along
route r.
The waiting cost in Eq. (1) is non-linear in qpj, the service frequency. In an effort to linearize the total cost
function, we introduce binary variables xmr = 1, if route r is used for load m, and 0 otherwise; and ypfj = 1, if a
train of type p is used at frequency f on link j, and 0 otherwise. Note that decision variable ypfj indicates the
availability of a specific service level (or frequency) on a particular link by a train of a certain type, whereas
decision variable xmr shows whether route r is in use to transport freight m. The latter binary variable is
defined for every possible combination of physical route and service level, for each link on that route.
The average waiting time, determined a priori from the service frequencies along the given route, is there-
fore treated as constant. With a known routing cost, it only remains to decide which route r should transport a
given load. The downside of such linearization is a dramatic increase in the number of decision variables or the
number of distinct routes. The cost cspfj of using a train of type p, whether loaded or empty, at the frequency qpj
on service link j is cSpfj ¼ qpj  cUp  lj , where cUp is the cost of using a train of type p over a unit distance, lj is the
length of service link j, and fpj is the service frequency of a train of that type on the given link.
Let

M set of shipments (each is defined by its origin, its destination and volume)
R(m) set of routes that can be used to transport freight shipment m
R(m, j) {r 2R(m)jj 2 S(r)}
J set of all service links
F(p,j) set of feasible service frequencies for trains of type p on link j
apfj transportation capacity of link j when a train of type p operates on that link at service frequency f

R(m, j) is thus the set of routes that can be used to transport the load m and which involve service link j. The
hub-and-spoke network transportation problem can now be written as
X X XX X
min cRmr  xmr þ cSpfj  y pfj ð2Þ
m2M r2RðmÞ j2J p2P f 2F ðp;jÞ
X
s:t: xmr ¼ 1 8m 2 M ð3Þ
r2RðmÞ
X X X X
wm xmr 6 apfj  y pfj 8j 2 J ð4Þ
m2M r2Rðm;jÞ p2P f 2F ðp;jÞ
X X
y pfj 6 1 8j 2 J ð5Þ
p2P f 2F ðp;jÞ

xmr 2 f0; 1g; 8r 2 RðmÞ 8m 2 M ð6Þ

y pfj 2 f0; 1g; 8f 2 F ðp; jÞ 8j 2 J ð7Þ
Constraints (3) ensure that only a single route should be used to serve a particular shipment. Transportation
capacity is guaranteed by inequality (4) to be sufficient to move all freight. Constraints (5) connect the two
534 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

binary variables (service level decision and route decision), forcing the use of just one service type on each link,
while constraints (6) and (7) denote non-negativity and integrality of the decision variables.

A.2. Solution procedure

Our heuristic method is composed of two sub-problems: Frequency improvement (FI) is employed to
change service frequency on a link in a current solution, while RI (route improvement) determines those routes
to serve all freight movements at given frequencies.
The heuristic starts with an effort to find routes, with service frequencies initialized at values just high
enough to serve all freight. (A frequency this high will not exclude any node as a potential hub location.)
Once a set of routes is found, FI reduces those frequencies if such a decrease results in lower total cost.
To this end we compute optimal service frequency for each link, and find a link whose frequency adjustment
leads to the greatest improvement in cost. Now that a service frequency differs from the previous solution,
RI modifies the routes by solving shortest path problems for affected freight. This solution-improvement
process will continue until reduction in total cost due to lower service frequencies is less than a pre-specified
threshold.
The basic idea is similar to Crainic et al. (1984) and Crainic and Rousseau (1986), but while their service
frequencies can be adjusted by only ±1, we allow adjustments in magnitude of any integer (e.g., frequency can
be adjusted from 2 to 4, from 5 to 1, and so forth). Our heuristic thus achieves more rapid computation at the
expense of higher total cost, in general. Their approach checks a greater number of feasible solutions, which is
why we call theirs the ‘‘Exhaustive Algorithm’’ (this should not be taken to mean that the latter is a complete
enumeration).
Since an optimal routing problem has to be solved whenever a service frequency is adjusted, an efficient
frequency update is crucial. We noted that FI does take a bigger step in search space, and so may reduce
the quality of solution obtained. This leads to a hybrid approach (denoted HA in our tables of results),
whereby FI and the exhaustive algorithm are used in turn.
Given the service frequencies on each link, i.e. once transportation capacity there is set according to the
given frequencies, the objective-function service-cost is constant as well, not being affected by the routing deci-
sion in the sub-problem. Therefore, that sub-problem is to find a set of optimal routes, subject to an upper
bound on transportation service.
A decrease in service frequency on link j leads to increased routing cost on that link, due to lengthened wait-
ing times before the next available freight service. Changes in frequency thus necessitate re-optimization of the
routing problem, which is aided by the following observations. When service frequency decreases between iter-
ations, we have greater cost per unit shipment on the link. Therefore, all shortest paths that did not use link j
should still not, whereas the paths that used this link may not be shortest any more. We need recompute only
those shortest paths that employed this link, thus requiring little calculation.
An exactly opposite situation occurs when a frequency increases. The service cost thus becomes smaller,
hence all shortest paths involving link j continue to be optimal, but those not using this link may now have
to. Although it may seem to require much numerical work to update all shortest paths not using the given
link, we expect very few iterations of the algorithm to experience increasing service frequency. That is con-
firmed by computational experience.

A.3. Example numerical results

The heuristic algorithm of Section A.2 is now applied to four test problems (Table 3). Those examples are
based on the actual European rail system. The problem designations Pb1, . . . , Pb4 stand respectively for the
networks in France plus Italy; Germany; Germany plus surrounding countries (Belgium, Netherlands, Swit-
zerland), i.e. the European network minus (France + Italy); and the whole European network. Freight infor-
mation (origin–destination pairs, volume) and cost parameters are from two projects done for the European
Union (EUFRANET, 1999; SIMIQ, 2000). Our algorithm was implemented in C++ on a desktop computer
with a Pentium III (565 MHz) and 64 M RAM. We considered two sizes of trains, with respective capacity
300 ton and 600 ton. When available, those with higher capacity are used between hubs.
 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536 535

As expected, the exhaustive algorithm (EA) of Crainic et al. (1984, 1986) finds better solutions than our
optimal frequency algorithm or hybrid heuristic. The optimal frequency algorithm combines our frequency
improvement (FI) method with the route decision procedure of this appendix. The exhaustive algorithm, how-
ever, takes more computational time. Numerical results are reported in Table 4, when only a single length of
train is used, and in Table 5 when two train types are available. (In our tables, HA denotes the hybrid
approach as above.)
For a single train type, the final cost found by FI is higher than that of the exhaustive algorithm by 4–6% on
average, while the final cost found by HA exceeds that of EA on average by less than 1%. HA actually finds a
better solution than EA in the case of problem instance Pb3.
Let us turn now to the case of two sizes of trains (Table 5). The mean additional cost, when FI is employed
rather than EA, is 1–2%; the corresponding figure for HA is less than 1% above that of EA. FI in fact finds a
better solution than EA in the case of network Pb1. We also note that EA cannot solve the largest problem,
Pb4, for two train types. (We ran EA for 24 h before stopping the search.) It is also observed that using two
different train lengths allows 4–8% savings, when a given problem instance is compared for the same
algorithm.
The parameter b, for the case of problem instances Pb2 and Pb3, has more impact on computational time
than on solution quality. When b increases from 0.1 to 7.0, the gap between solutions by the two algorithms
slowly improves from 1.5% to 0.3%, while computation time increases from 1.1 min to over 16 min (for Pb2)
and from 4.3 min to 71 min (Pb3).

Table 3
Data sets
Problem instance Number of nodes Number of service links Total freight (tons/week)
Pb1 15 210 47,808
Pb2 24 552 210,264
Pb3 33 1056 255,554
Pb4 48 2256 396,002a
a
Total freight for Pb4 is Table 1’s [sum of freight in  sum of freight consolidated], to avoid double counting.

Table 4
Comparison of results when one type of train is used
Inst. Solution (euros 000) Gapb Computation time (s)
EA FI HAa FI HA EA FI HA
Pb1 3471 3623 3478 0.0440 0.0022 12 0 8
Pb2 12,480 13,217 12,483 0.0591 0.0003 369 3 212
Pb3 16,235 17,005 16,230 0.0475 0.0003 2279 9 956
Pb4 26,252 27,595 26,319 0.0512 0.0026 27,469 120 5464
a
b = 5.0.
b
Gap = (solution of algorithm FI or HA)/(solution of exhaustive algorithm)  1.

Table 5
Numerical results (two train types)
Inst. Solution (euros 000) Gapb Computation time (s)
a
EA FI HA FI HA EA FI HA
Pb1 3184 3155 3189 0.0092 0.0014 38 1 24
Pb2 11,858 12,091 11,925 0.0197 0.0057 1460 12 517
Pb3 15,364 15,635 15,462 0.0177 0.0064 8365 102 1658
c c
Pb4 25,257 25,197 – – 660 8958
a
b = 1.5.
b
Gap = (solution of algorithm FI or HA)/(solution of exhaustive algorithm)  1.
c
Cases not solved in 24 h.
536 S.-J. Jeong et al. / Transportation Research Part A 41 (2007) 523–536

References

Assad, A., 1980. Modelling of rail networks: toward a routing/makeup model. Transportation Research B 14, 101–114.
Aykin, T., 1995. Networking policies for hub-and-spoke system with application to the air transportation system. Transportation Science
29, 201–221.
Campbell, J.F., Ernst, A.T., Krishnamoorthy, M., 2002. Hub location problems. In: Drezner, Z., Hamacher, H.W. (Eds.), Facility
Location: Applications and Theory. Springer-Verlag, Heidelberg.
Crainic, T., Rousseau, J., 1986. Multicommodity, multimode freight transportation: a general modelling and algorithmic framework for
the service network design problem. Transportation Research B 20, 225–242.
Crainic, T., Ferland, J., Rousseau, J., 1984. A tactical planning model for rail freight transportation. Transportation Science 18, 165–184.
Crozet, Y., 2004. European railway infrastructure: towards a convergence of infrastructure charging? International Journal of Transport
Management 2 (1), 5–15.
Daganzo, C., 1986. Static blocking at rail yards: sorting implications and tracking requirements. Transportation Science 20, 189–199.
Dejax, P., Bookbinder, J.H., 1991. Goods transportation by the French national railway (SNCF): the measurement and marketing of
reliability. Transportation Research A 25 (4), 219–225.
EUFRANET, 1999. European Freight Railway Network. Supply model, alternative solutions and network load: Part 1. Technical report
for the European Commission, 131 pp. coordinated by INRETS (Institut National de Recherche sur les Transports et leur Securité).
Haghani, A., 1989. Formulation and solution of a combined train routing and makeup, and empty car distribution model. Transportation
Research B 23, 433–452.
Jaillet, P., Song, G., Yu, G., 1996. Airline network design and hub location problems. Location Science 4, 195–212.
Keaton, M., 1992. Designing railroad operating plans: a dual adjustment method for implementing Lagrangian relaxation. Transportation
Science 26, 263–279.
Marı́n, A., Salmerón, J., 1996a. Tactical design of rail freight networks, part I: exact and heuristic methods. European Journal of
Operational Research 90, 26–44.
Marı́n, A., Salmerón, J., 1996b. Tactical design of rail freight networks, part II: local search methods with statistical analysis. European
Journal of Operational Research 94, 43–53.
O’Kelly, M., Bryan, D., 1998. Hub location with flow economies of scale. Transportation Research B 32, 605–616.
Pirkul, H., Schilling, D., 1998. An efficient procedure for designing single allocation hub and spoke systems. Management Science 44,
S235–S242.
Racunica, I., Wynter, L., 2005. Optimal location of intermodal freight hubs. Transportation Research B 39 (5), 453–477.
SIMIQ, 2000. Intermodal Quality (IQ). Technical report for the European Commission, coordinated by INRETS, 67 pp.
Thompson, R.G., Taniguchi, E., 2001. City logistics and freight transport. In: Brewer, A.M., Button, K.J., Hensher, D.A. (Eds.),
Handbook of Logistics and Supply-Chain Management. Pergamon, pp. 393–405.
Whiteing, T., Browne, M., Allen, J., 2003. City logistics: the continuing search for sustainable solutions. In: Waters, D. (Ed.), Global
Logistics and Distribution Planning, 4th ed. Kogan Page, London, pp. 308–320.