You are on page 1of 12

European Journal of Operational Research 175 (2006) 1616–1627

www.elsevier.com/locate/ejor

A time–space network based exact optimization model


for multi-depot bus scheduling
Natalia Kliewer *, Taı̈eb Mellouli, Leena Suhl
Decision Support & Operations Research Lab, University of Paderborn, Warburger Street 100, 33098 Paderborn, Germany

Available online 6 April 2005

Abstract

The vehicle scheduling problem, arising in public transport bus companies, addresses the task of assigning buses to
cover a given set of timetabled trips with consideration of practical requirements, such as multiple depots and vehicle
types as well as depot capacities. An optimal schedule is characterized by minimal fleet size and minimal operational
costs including costs for unloaded trips and waiting time. This paper discusses the multi-depot, multi-vehicle-type
bus scheduling problem (MDVSP), involving multiple depots for vehicles and different vehicle types for timetabled
trips. We use time–space-based instead of connection-based networks for MDVSP modeling. This leads to a crucial size
reduction of the corresponding mathematical models compared to well-known connection-based network flow or set
partitioning models. The proposed modeling approach enables us to solve real-world problem instances with thousands
of scheduled trips by direct application of standard optimization software. To our knowledge, the largest problems that
we solved to optimality could not be solved by any existing exact approach. The presented research results have been
developed in co-operation with the provider of transportation planning software PTV AG. A software component to
support planners in public transport was designed and implemented in context of this co-operation as well.
 2005 Elsevier B.V. All rights reserved.

Keywords: Transportation; Vehicle scheduling; Multi-depot bus scheduling; Time–space network; Decision support systems

1. Introduction transport companies focus on efficient use of re-


sources, especially vehicles and drivers. Operations
1.1. Motivation research based decision support systems play here-
by a crucial role.
Been obliged to act market-oriented instead Service trips are provided as products with
of the traditional monopolistic approach, public fixed departure and arrival times as well as
start and end locations. The transport resour-
*
Corresponding author. ces, operating these trips have to be allocated guar-
E-mail address: kliewer@upb.de (N. Kliewer). anteeing their availability at the right time on

0377-2217/$ - see front matter  2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2005.02.030
N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627 1617

the right place, by minimizing total costs of tion and Lagrangean relaxation for the multi-com-
resources. modity flow formulation (see Löbel (1999)), or an
We consider the scheduling of vehicles under auction algorithm for the quasi-assignment formu-
constraints and objectives arising in urban and lation (see Freling (1997)) of the MDVSP.
suburban public transport. Thus, each timetabled In cases where suboptimal solutions are suffi-
trip can be served by a vehicle belonging to a given cient, other authors propose heuristic methods,
set of vehicle types. Each vehicle has to start and such as schedule first—cluster second approaches
end its work day in one of the given depots (here: (Daduna and Paixão (1995)), or shortest path-
parking garage). based heuristics (DellÕAmico et al. (1993)).
After serving one timetabled (loaded) trip, each Another approximate approach to cope with
bus can serve one of the trips starting later from real-world instances is the artificial limitation of
the station where the vehicle is standing, or it degrees of freedom for vehicle type and depot
can change its location by moving unloaded to assigning. Furthermore, one may consider only a
an another station (deadhead trip) in order to subset of possible deadhead trips (see, e.g., Löbel
serve the next loaded trip starting there. (1999)).
The cost components include fixed costs for From our point of view, the explicit modeling
required vehicles as well as variable operational of all possible connections seems to be inefficient
costs. The variable costs consist of distance-depen- for large scale planning problems, because the
dent travel costs and time-dependent costs for time number of such pairs grows quadratically with
spent outside the depot—the case where a driver is the number of scheduled trips.
obliged to stay with the bus. All cost components
depend on vehicle type. Since the fixed vehicle cost 1.2. Intention
components are usually orders of magnitude high-
er than the operational costs, the optimal solution The objective of this paper is to introduce a new
always involves the minimal number of vehicles. modeling approach by applying a time–space net-
The combinatorial complexity of the multi- work technique which avoids the explosive in-
depot bus scheduling problem (MDVSP1 in the crease of the model size with a growing
following) is determined by numerous possibilities timetable. The new model should enable us to
to assign vehicle type to each trip, to build se- solve multi-depot vehicle scheduling problems of
quences of trips for particular buses, and to assign practical sizes and complexity, arizing in large
buses to certain depots. To represent these se- public transport companies. Thus, we might be
quences of trips, exact modeling approaches able to use standard optimization software (such
known in the literature consider explicitly all pos- as ILOG CPLEX or MOPS (see ILOG (2003)
sible connections—pairs of trips that can be served and Suhl (2003))), and profit from the recent sig-
successively. nificant improvements in optimization software.
In context of this paper, we refer to all models If we would use the specific algorithms developed
that represent connections explicitly as ‘‘connec- for connection based models, it would not be pos-
tion-based’’ models. Because connection-based sible to take advantage of the improvements in
MDVSP models inherently include a large number standard software.
of arcs, such models of practical dimension with
thousands of trips and multiple depots can often 1.3. Results
not be solved with standard optimization software
packages. Known publications describe in detail We implemented the time–space network based
applications of techniques such as column genera- modeling approach as a software component
which has been integrated in commercial software
packages to support planning processes in public
1
MDVSP means in the sense of this paper the MDMVTBSP— transport. This software component generates
the multi-depot, multi-vehicle-type bus scheduling problem. mathematical models for given instances and
1618 N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627

solves them to optimality. We have carried out There are several variations of the bus schedul-
tests on real-life timetables of several public trans- ing problem involving different side constraints or
port companies in Germany, such as Munich. In- numbers of depots and/or of bus types. The
stances with more than 11,000 trips were restrictions and optimization criteria may differ
provided by PTV AG. Concerning groups of vehi- from one problem setting to another.
cle types allowed for certain trips, this problem The multi-depot vehicle scheduling problem in-
decomposes into four disjunct multi-depot prob- volves several depots, so that a vehicle has to re-
lems. The largest of them has more than 7000 turn in the evening to the same depot from
scheduled trips. which it started in the morning.
Multi-vehicle-type VSP considers heteroge-
1.4. Outline neous fleet of vehicles—for example normal bus,
minibus, and kneel bus. For a given trip we define
The paper is organized as follows: Section 2 con- a group of vehicle types this trip can be served by.
tains a definition of MDVSP. Section 3 describes For example, a trip to a bus stop near a hospital
MDVSP models known in literature, having sub- can be served only by handicapped accessible vehi-
stantial similarity in the representation of possible cles such as kneel bus or normal bus with a ramp
connections. In Section 4, a time–space network extension. In a feasible solution each rotation is as-
MDVSP model is introduced, and its advantages signed to exactly one depot and at least to one
compared to existing models are discussed. Section vehicle type.
5 shows computational results on test instances and Furthermore, it is possible to state capacity
MIP-solver performance depending on problem restrictions for depots, which means that a depot
complexity. Section 6 gives conclusions. has only a restricted number of (overnight) park-
ing slots for buses. Other kinds of capacity con-
strains set a limit for the number of available
2. The multi-depot multi-vehicle type scheduling vehicles of certain bus types and for the number
problem of certain type vehicles in a given depot.
It is well known that the single-depot homoge-
We define the vehicle scheduling problem neous fleet vehicle scheduling problem is a polyno-
(VSP), arising in public bus transportation, as mially solvable minimum cost flow problem, while
the task of building an optimal set of rotations the multiple depot and the heterogeneous fleet ver-
(vehicle schedule), such that each trip of a given sions and their combinations imply that the opti-
timetable is covered by exactly one rotation. For mization problem becomes NP-hard (see Bertossi
each trip the timetable specifies a departure time et al. (1987)).
and an arrival time with start and end stations The complexity of instances of the MDVSP de-
respectively. pends on various factors, such as:
Within a bus tour consisting of several (loaded)
service trips chained with each other, the use of • the number of timetabled trips,
deadhead trips (unloaded trips between two end • the number of depots, or more precisely, the
stations) often provides an improvement in order average number of depot-vehicle type combina-
to serve all trips of a given timetable by a mini- tions per timetabled trip,
mum number of buses. Thus a work day for a • the number of possible unloaded trips, which
given bus is defined as a sequence of trips, dead- can vary depending on the completeness of the
heads, waiting times at stations (parking stops) distance matrix for stop points.
and pull-out/pull-in trips from/to the assigned de-
pot. Since deadhead trips mean an additional cost Fig. 1 shows a simple example for a timetable
factor, minimization of this cost and minimization with three trips and a station topology with dis-
of waiting time cost are important optimization tances expressed in time units. Stations k3 and k4
goals. are depots. Trips a and c can be served by buses
N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627 1619

d1(k3) Stop points and depots


Timetable

Trip From To Dep. Arr. d1 D2 1 time unit


1
time time k1 k2
a k1 k2 1 2 + +
b k2 k1 3 4 - +
c k2 k1 5 6 + +
d2(k4)

Fig. 1. An example for a timetable with three trips and distance matrix for 2 stop points and 2 depots.

from depots d1 and d2, whereas trip b can only be coefficient matrix of the linear programming
served from depot d2. relaxation.
In the following, we discuss state-of-the-art For the multi-depot case, the network is multi-
solution techniques for the MDVSP as well as plied, so that there is a network layer for each
present a new model for solving them with stan- depot. Mathematical formulation of the model
dard optimization software. Differences between contains additional restrictions, called cover con-
connection-based and time–space-network-based straints, guaranteeing that each scheduled trip is
models will be illustrated on the example from carried out by exactly one vehicle belonging to
Fig. 1. one of the required vehicle types. Because of the
explicit integrality constraints (all flow variables
must be equal to 0 or 1), the optimization model
3. Existing connection-based MDVSP models is no more a minimum cost flow model and is
now difficult to solve.
Traditionally, mathematical optimization mod- Fig. 2 illustrates an application of the traditional
els for public mass transit are based on a network modeling approach to the example from Fig. 1.
where trips and depots are represented by nodes Each trip to be scheduled is represented by a depar-
(trip-nodes and depot-nodes, respectively), and ture node and an arrival node, together with an arc
possible connections between trips given by arcs joining them. Other arcs join arrival nodes with
(see, e.g., Daduna and Paixão (1995)). The rota- later departure nodes, representing the possibility
tion of a vehicle is thus generated as a flow of of a single vehicle undertaking the two correspond-
one unit starting in the depot and returning back ing trips in succession. Further arcs represent bus
into it. The nodes connected by a flow represent movements to and from depot nodes. Each arc
the scheduled trips, carried out by the correspond- has an associated cost factor (deadhead cost for
ing vehicle. arcs joining arrival nodes to later departure nodes,
The single-depot case of the VSP can be mod- vehicle cost for pull-out and/or pull-in arcs), and
eled as a minimum cost flow problem. Vehicles capacity (one for service trips, depot capacities
represent the flow units circulating in this network. for circular arcs). The network involves two net-
Each arc is provided with the cost of its corre- work layers: one layer for each depot.
sponding activity—stand time cost, deadhead or A flow solution is feasible only if the flow value
fixed vehicle cost. The objective is to minimize is equal to one on exactly one of all service trip
the total flow costs over all arcs. All flow units in- arcs representing a certain service trip in different
volved in the single-depot problem are of the same layers (see for example Fig. 2: unbroken arcs a in
type, so that the incoming and outgoing flow of d1 and d2 layers). The optimal flow (feasible flow
each trip-node has to be equal to one. The mathe- with minimum total costs) for this network now
matical model consists of the minimum cost flow determines the optimal vehicle schedule for the ori-
problem with flow conservation constraints for ginal timetable. A circulation flow down to the
each network node. No integrality constraints right gives an example for feasible and fleet-mini-
are needed because of the unimodularity of the mal solution.
1620 N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627

Service trip arcs


a Trip compatibility arcs
Pull-in/pull-out arcs
depot d1
Circulation flow arc
c d1-Layer

depot d2 b

d2-Layer

Fig. 2. The example from Fig. 1 modeled with the connection-based approach.

We can clearly identify three different ap- ing (see Hane et al. (1995) as an example), because
proaches among the existing modeling techniques: they are advantageous in modeling possible con-
nections between arriving and departing flights.
• path-oriented—leading to set partitioning for- In a time–space network, connections within a
mulation (see Ribeiro and Soumis (1994)), location are realized by using a time line that con-
• arc-oriented—leading to multi-commodity flow nects all possible landing and takeoff events within
formulation (compare e.g. Forbes et al. (1994)), the location. Thus, there is no need to explicitly
and model connections for each feasible pair of events
• combinations of these two approaches (see Car- within a location.
paneto et al. (1989)). Time–space network models were not used for
bus scheduling problems until now, because, com-
In all these modeling approaches, the possible pared to airline scheduling where deadheading is
trip connections are considered explicitly and the generally not allowed, bus scheduling permits
number of such connections, corresponding to unrestricted deadheading. Thus the advantages
the number of integer variables, grows quadrati- given by TSN remained negligible, because of
cally in dependence of the number of loaded trips. too many deadhead arcs.
Therefore, models with several thousand sched- However, we show in this section that a new
uled trips become too large to be solved directly modeling technique leads to a drastic reduc-
by standard optimization tools in a reasonable tion of arcs, simultaneously allowing all possi-
time. Various techniques to reduce the number of ble deadheads, so that we may exploit the
possible connections have been proposed in the lit- advantages of TSN models for bus scheduling
erature. Some approaches discard arcs with too problems.
long waiting times, others generate arcs applying
the column generation idea to the network flow
4.1. Modeling of connections within the
representation.
time–space network

4. A new time–space network flow model for The new TSN model is based on aggregation of
MDVSP possible connection arcs so that it is possible to
carry several connections on one single arc simul-
Time–space network (TSN) models have been taneously. Basically, we have to take into account
proposed for routing problems in airline schedul- three types of potential connections:
N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627 1621

• connections within one station, start station to the end event on the time line of
• connections over a depot with pull in/pull out the end station.
trips to/from the depot, and In contrast to connection-based models from
• connections through a directly connecting trip the literature, the connections within one station
(deadhead). must not be modeled explicitly by connection arcs,
but can be implicitly represented by the flow
through the timeline. The compatible trip pairs
4.1.1. Connections within one station (a, b) and (a, c) from Fig. 2 can now be connected
A connection within one station means the by waiting arcs (Fig. 3: dotted arcs) in the time line
(standard) case where the successor trip starts at of station k2.
the same station where the predecessor trip ends.
We build a time line for each station, as well as 4.1.2. Connections over a depot with pull
for depots. A time line contains all arrivals and in/pull out trips to/from the depot
departures at the given location at their corre- In analogy to stations we build a time line for
sponding points of time. The network contains a each depot, although there may not be scheduled
trip arc (Fig. 3: unbroken arcs) for each scheduled trips starting or ending directly in a depot. To each
trip—from the start event on the time line of the scheduled trip i we introduce, if necessary, arcs for

Fig. 3. The example from Fig. 1 modeled with new time–space network approach.
1622 N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627

potential pull-out and pull-in trips from/to each Thus, a crucial modeling technique is aggrega-
depot directly before/after carrying out i (with tion of possible matches—directly connecting trips
associated deadhead costs) (Fig. 3: punctured between scheduled trips. Fig. 4 shows the equiva-
arcs). Because it is more favorable for buses to lence to the connection-based approach relating to
stand at a depot than at other stations, we place the feasible solutions set of VSP. The aggregation
a higher cost for waiting arcs outside the depots, of possible trip connections is carried out in three
therefore avoiding long waiting times outside the steps, as explained below:
depots. First stage aggregation
For each arriving scheduled trip i at station k
4.1.3. Connections through a directly connecting we determine the first trip compatible with i at
trip (deadhead) each other station l(k 5 l). We call this trip first-
Basically it is possible to connect each pair of match (i, l). We introduce only those deadheading
compatible trips through deadheading. However, arcs into the model, which correspond to the first
as the number of possible deadhead trips can be matches. Thus, the number of arcs is reduced sig-
extremely high, a direct modeling of all of them nificantly compared to the original situation. Nev-
in large networks implies a problem size which ertheless, all possible connections remain feasible.
cannot be handled by state-of-the-art solvers. Each scheduled trip j compatible to a scheduled

1. Stage with all possible matches


Space

Time
Station k

Matches
Station l

2. Stage with waiting arcs in station l

Station k

First Matches
Station l

Station k

Latest First Matches


Station l

Fig. 4. Two-stage aggregation of possible connections.


N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627 1623

trip i can now be reached in the model network However, the number of first matches is limited
over first-match (i, start-station (j))—possibly via by the number of bus stations multiplied by the
waiting arcs at start-station (j). number of timetabled trips. Since the number of
Second stage aggregation stations is always smaller than the number of trips,
At the second stage, the number of arcs is fur- we notice a crucial model reduction compared to
ther reduced. We aggregate the set of first matches existing approaches. The last two aggregation
in a smaller set of latest first matches (see Fig. 4). stages yield an additional strengthening of the
The latest first matches can be determined in the model size.
following way: Let S be the set of incoming trips Table 1 illustrates the impact of the three-stage
i at station k, having the same first match jS at sta- aggregation process for timetables of some Ger-
tion l. Let iL be the latest incoming trip in S. Then man public transport companies. We define here
the latest first match for each element in S is the matches as possible connections between compati-
first match (iL, k). By removing all arcs corre- ble trips at different stations. The table shows
sponding to first matches but being no latest first explicitly the number of matches, first matches,
matches, we reduce the network significantly, but and latest first matches. Furthermore, the right-
do not loose any possible connections. Every most column denotes the number of latest first
scheduled trip j compatible to scheduled trip i matches that are cheaper than traveling over de-
can now be reached over the corresponding latest pot, thus leaving out the direct connection trips
first match to the start station of j, and possibly that are dominated by trips over depot. The largest
via waiting arcs in one or both connection lines. reduction arises for urban timetables with short
Third stage aggregation distances between stations (see T2 to T4). Only
There are still some redundancies in the net- 1% of the matches remain in the reduced set of lat-
work after the elimination all possible connec- est first matches. Even for the regional timetable
tions being no latest first matches. For deadhead (T1), having long distances between stations, we
trips, having enough time for temporarily return achieve a reduction of 97%.
to depot, we evaluate costs arising for such a
temporary return and compare them to costs of 4.2. Construction of the network
directly connecting with a deadhead trip. In
case of positive cost savings by a depot-return we Because we have to assure that each bus returns
eliminate this specific latest first match in our net- back to its original depot at the end of a day, we
work, guaranteeing the connection still by depot- need to distinguish between buses from different
return. home depots. Furthermore, certain trips have to
be taken over by buses of a group of given bus
4.1.4. Model reduction: Computational results types, so that we need to distinguish between bus
The total number of matches, which determines types as well. Therefore, we construct one network
the number of variables in connection-based net- layer for each depot-vehicle-type-combination,
work models for MDVSP, depends quadratically called depot for simplicity, containing the follow-
on the number of trips given by the timetable. ing arcs:

Table 1
Scale of problem reduction by the new formulation
Timetable Scheduled trips Stations Matches First matches Latest first matches Lfm not-over-depot
T1 682 50 195,618 19,457 6292 6262 (3%)
T2 2047 21 1,143,868 42,863 17,791 11,885 (1%)
T3 3054 49 3,906,045 81,265 52,898 48,572 (1%)
T4 7068 124 20,555,784 513,393 246,580 213,433 (1%)
T1: regional bus company; T2–T4: urban.
1624 N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627

• a loaded trip arc for each timetabled trip which In case of minimizing the number of buses be-
can be operated from this depot, fore considering operational cost components as
• deadhead arcs, according to latest first matches, an objective, we should choose the cost factors
which are cheaper than temporarily return to properly: The costs on a circulation arc must be
the depot, and large enough in comparison to the operational
• pull-in/pull-out arcs to the depot after/before costs (costs for deadheads) in order to induce the
each loaded trip. effect of minimizing the number of buses.
Fig. 3 shows the complete time–space network
For each station, including depots, we build a for the MDVSP—example from Fig. 1. Down to
connection line representing all possible arrival the right a feasible and fleet-minimal solution is
and departure events at the given station—after given by some arcs. Flow values on these arcs
pre-sorting of arcs according to arrival/departure are equal to one; missing arcs from the above net-
time respectively. work have a null-flow in this solution.
One node of this network connects a group of The resulting flow model contains one network
possible arrivals to the following group of possible layer for each depot (as defined above), where 0/1-
departures (see also Fig. 3). In this way all stations, variables on trip arcs and integral flow variables
including depots, are represented as ordered sets of on other arcs are defined. In addition to the flow
(connection) nodes, linked together by waiting balance equations, it is required by cover con-
arcs. On each layer, there is one circulation flow straints that for each scheduled trip the sum over
arc connecting the last node of the depot line with flow variables on the corresponding trip arcs of
its first node. This arc is provided with fixed cost all network layers is equal to one.
for corresponding vehicle type and represents vehi- The optimal solution of the mathematical
cles parking over night in depot. model describes flow values in this network with
There are different ways to express the capacity minimal total costs.
of a depot. Our approach incorporates the follow-
ing specific modeling approaches: 4.3. Decomposition of the aggregated flows

• We set flow upper bounds on circulation flow It is an important characteristic of the time–
arcs for capacities of depot-vehicle type space network formulation that due to the aggre-
combinations. gation of possible connections, any feasible flow,
• The number of parking slots in certain depots including also an optimal flow, represents a bundle
generates additional restrictions in the mathe- or a class of vehicle schedules. All of them have
matical model formulation constraining the minimal total costs but different distributions of
sum of flow values on corresponding circulation waiting times. With the help of a suitable flow
flow arcs. decomposition procedure, we may extract a vehicle
• For a limited number of vehicles of certain schedule with an optimal flow and desired charac-
types we also constrain the sum of correspond- teristics. A large number of possible flow decom-
ing flow values. position algorithms may be constructed to
• If it is required to use a fixed number of buses of decompose a given optimal flow. For example,
certain type or of certain depot, we set a lower the first-in-first-out (FIFO) strategy connects the
bound equal to the upper bound (variable fixing bus that arrived first with the trip that leaves first,
in mathematical model) on the corresponding whereas the last-in-first-out (LIFO) strategy con-
arc. nects the bus that arrived last with the trip that
leaves first, both strategies having specific typical
Waiting arcs together with arcs for unloaded properties.
trips are also provided with corresponding opera- Fig. 5 demonstrates the number of choices in
tional costs. Upper bounds on the loaded trip-arcs the case that the flow of the given connection arc
are equal to one. has the value of three units in the optimal solution.
N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627 1625

Table 2 summarizes the sizes of the mathemat-


ical models, objective values and optimization
3 flow units
times for 25 test instances based on three timeta-
bles. All solution times are shown in seconds of
CPU time on a 2.1 GHz processor using ILOG
CPLEX 8.0 as a standard optimization package.
Except for the larges problem instances, we also
Fig. 5. Degree of freedom in flow decomposition.
received acceptable run times with the standard
optimizer MOPS (Suhl (2000)).
There are 3! = 6 variations to connect arrivals to The number of depots means actually depot-
the departures. vehicle-type combinations, as mentioned above,
and thus corresponds to the terminology in Löbel
(1999). The given number of depots may be differ-
5. Implementation and optimization runs ent from the average depot group size—for such
cases we additionally provide the average depot
Based on the description given in the previous group size in parenthesis (last 10 instances). The
section, our implementation of the time–space net- entry ‘‘3 (2.5)’’ means, that for a timetable-trip
work based solution procedure to solve a given on the average 2.5 different depot-assignments
MDVSP involves the following steps: are possible, although there are three depots in
the network.
• Read the data and construct the layers of a Note that the number of trip matches in the
time–space network. corresponding column gives a lower bound for
• Create a mathematical model for min cost flow the number of variables in the existing
with additional cover constraint for each connection-based multi-commodity flow for-
unloaded trip. mulation.
• Optimize the mixed-integer programming With the proposed modeling approach, we were
model with standard software, and compute able to solve quite large problems in a reasonable
an optimal flow representing a class of aggre- amount of time without special tuning of the used
gated schedules. optimization software. Despite the fact that this
• Decompose this flow to rotations of one con- problem is NP-hard, most of the solution time is
crete optimal vehicle schedule, according to spent solving the first LP relaxation. Comparing
LIFO, FIFO or any intermediate strategy. various LP solvers leads to the conclusion that
the barrier optimizer is the best choice to solve
In the context of a co-operation with the soft- large problems with many degrees of freedom for
ware developer PTV AG, we designed and imple- each loaded trip.
mented a software component which supports This time–space network based model has a
public transport planners in constructing schedules very good MIP behavior; the IP-gap is infinitesi-
for buses. PTV AG provided us test instances of mal small or null, and almost all variables have
several public transport companies. The largest integer values in the optimal (basis) solution of
timetable (T4 in Table 1: City of Munich) the LP-relaxation. The reason for this lies in the
contains: model properties: because of the aggregation of
possible connections, the mathematical model
• 11.018 scheduled trips, tends to use one general integer variable instead
• 33 vehicle types in 8 groups, of several binary variables. Thus, we practically
• 24 depots, and shift the corresponding decisions to the postpro-
• results in 4 disjunctive MDVS problems; largest cessing phase, where we construct an optimal vehi-
problem has 7.068 scheduled trips and 124 end cle schedule from the optimal network flow via
stations. flow decomposition.
1626
Table 2
Computational results for timestables T2, T3 and T4
Instance Trip matches arcs Model size TS-MCF Obj./1000 MIP optimization time Best LP solver

N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627


(in seconds)
Trips Stop Depots* Trip First matches Latest first Rows · Cols · Nzs Network Dual Barrier Best time
points matches TS-MCF matches
MCF TS-MCF(d)
1 20,555,784 246,580 213,433 20,575 · 248,021 · 503,110 1,014,213 12 12 207 12 Dual = Network
2 41,111,568 493,160 432,737 34,072 · 501,903 · 1,017,942 1,014,110 136 150 1,313 136 Network
3 61,667,352 739,740 616,432 47,574 · 720,181 · 1,461,566 1,014,050 7079 4608 3435 3435 Barrier
4 82,223,136 986,320 825,624 60,690 · 963,175 · 1,954,622 1,013,999 24,762 10,030 5863 5863 Barrier
7.068 124 5 102,778,920 1,232,900 1,039,137 73,866 · 1,210,542 · 2,456,424 1,013,913 53,143 49,394 11,887 11,887 Barrier
1 3,906,045 52,898 48,572 9597 · 64,231 · 131,516 508,223 2 1 22 2 Dual
2 7,812,090 105,796 98,057 16,060 · 129,295 · 264,698 508,103 24 20 158 20 Dual
3 11,718,135 158,694 136,290 22,611 · 183,195 · 375,552 508,056 246 136 420 36 Dual
4 15,624,180 211,592 187,299 29,076 · 249,775 · 511,766 508,049 805 482 988 482 Dual
5 19,530,225 264,490 233,316 35,267 · 310,703 · 636,686 508,002 1106 646 1510 646 Dual
6 23,436,270 317,388 273,126 41,370 · 365,185 · 748,694 507,939 1115 858 1510 858 Dual
7 27,342,315 370,286 322,513 47,766 · 429,933 · 881,244 507,922 1448 1237 1826 1237 Dual
8 31,248,360 423,184 363,118 54,096 · 485,834 · 996,100 507,870 2039 1334 2568 1334 Dual
9 35,154,405 476,082 409,191 60,416 · 547,138 · 1,121,762 507,829 2617 2893 4126 2617 Network
3.054 49 10 39,060,450 528,980 450,937 66,747 · 604,151 · 1,238,842 507,822 3264 4423 3136 3136 Barrier
3 (2.5) 1,143,868 17,791 11,885 8798 · 27,599 · 58,485 421,582 4 4 15 4 Dual
6 (5) 2,115,896 34,955 23,118 14,997 · 53,249 · 113,039 421,545 37 40 67 37 Dual
9 (7.5) 3,397,601 52,864 40,813 21,939 · 87,128 · 184,111 421,544 148 106 146 106 Dual
12 (10) 4,754,636 70,882 56,101 29,031 · 118,768 · 250,675 421,537 371 271 244 244 Barrier
15 (12.5) 6,153,772 90,390 72,007 36,235 · 151,369 · 319,161 421,533 972 669 356 356 Barrier
18 (15) 7,473,783 108,458 85,873 43,232 · 181,409 · 382,525 421,531 1566 1179 824 824 Barrier
21 (17.5) 9,005,549 128,764 101,448 50,647 · 214,110 · 451,211 421,525 3385 1652 598 598 Barrier
24 (20) 10,468,848 148,143 119,162 57,841 · 248,580 · 523,435 421,521 2798 3099 568 568 Barrier
27 (22.5) 12,014,270 168,591 138,008 65,223 · 284,539 · 598,637 421,517 3708 3551 785 785 Barrier
2.047 21 30 (25) 13,558,845 188,927 157,526 72,636 · 321,198 · 675,239 421,516 6069 4689 905 905 Barrier
N. Kliewer et al. / European Journal of Operational Research 175 (2006) 1616–1627 1627

Several practical side constraints that could not References


be addressed in this paper have been successfully
included into the software component, such as: Bertossi, I., Carraresi, P., Gallo, G., 1987. On some matching
problems arising in vehicle scheduling models. Networks 17,
outsourcing of the parts of timetables to private
271–281.
bus companies, returns to different depots, differ- Carpaneto, G., DellÕAmico, M., Fischetti, M., Toth, P., 1989. A
ent types of fleet size limitations, and so on. branch and bound algorithm for the multiple depot vehicle
scheduling problem. Networks 19, 531–548.
Daduna, J.R., Paixão, J.M.P., 1995. Vehicle scheduling for
public mass transit—an overview. In: Daduna, J.R. et al.
6. Conclusion (Eds.), Computer-aided Transit Scheduling, LNEMS, vol.
430. Springer, Berlin, pp. 76–90.
We propose a new application of the time–space DellÕAmico, M., Fischetti, M., Toth, P., 1993. Heuristic
network flow model for the vehicle scheduling algorithms for the multiple depot vehicle scheduling prob-
lems. Management Science 39, 115–125.
problem arising in public bus transport networks.
Forbes, M.A., Holt, J.N., Watts, A.M., 1994. An exact
The model size has been substantially reduced algorithm for multi-depot bus scheduling. European Jour-
through aggregation of incoming and outgoing nal of Operational Research 72, 115–124.
arcs within each station, without loss of generality. Freling, R., (1997). Models and Techniques for Integrating
Thus, we were able to solve very large practical in- Vehicle and Crew Scheduling. Ph.D. thesis, Tinbergen
stances to optimality through direct application of Institute, Erasmus University Rotterdam.
Hane, C., Barnhart, C., Johnson, E.L., Marsten, R.E., Nem-
standard optimization software. Many of these in- hauser, G.L., Sigismondi, G., 1995. The fleet assignment
stances were not solvable with any existing ap- problem: Solving a large integer program. Mathematical
proach according the current state-of-the-art. Programming 70 (2), 211–232.
The main advantage of the time–space network ILOG, 2003. Cplex v8.0 UserÕs Manual. ILOG, Gentilly,
model is the reduction of the number of variables France, 2002.
Löbel, A., 1999. Solving large-scale multiple-depot vehicle
in the exact optimization model. Inherently, there scheduling problems. Computer-Aided Transit Scheduling
is no straightforward way to explicitly model costs 471, 193–220.
of a given trip-to-trip connection within the aggre- Ribeiro, C., Soumis, F., 1994. A column generation approach
gated model. In bus scheduling case with linear to the multiple-depot vehicle scheduling problem. Opera-
cost structures the individual connections of trips tions Research 42 (1), 41–52.
Suhl, U.H., 2000. MOPS—Mathematical OPtimization System.
or flights are not necessary, and therefore the OR News 8, 11–16.
time–space network model is to be preferred to Suhl, U.H., 2003. MOPS—Mathematical Optimization System.
the traditional connection based models. Available from: <http://www.mops-optimizer.com>.

You might also like