Sie sind auf Seite 1von 24

A Cy

le Based Optimization Model


for the Cy li Railway Timetabling Problem
Leon Peeters1;2 and Leo Kroon1
1
2

Erasmus University Rotterdam, Rotterdam S hool of Management, PO Box


1738, 3000 DR Rotterdam, The Netherlands
l.peetersfbk.eur.nl.

The paper presents an optimization model for y li railway timetabling


that extends the feasibility model by S hrijver and Steenbeek (1994). We use a
mixed integer non-linear programming formulation for the problem, where the integer variables orrespond to y les in the graph indu ed by the onstraints. Obje tive fun tions are proposed for minimizing passengers' travel time, maximizing the
robustness of the timetable, and minimizing the number of trains needed to operate
the timetable. We show how to approximate the non-linear part of the formulation,
thereby transforming it into a mixed integer linear programming problem. Furthermore, we des ribe prepro essing pro edures that onsiderably redu e the size of the
problem instan es. The usefulness and pra ti al appli ability of the formulation and
the obje tive fun tions is illustrated by several variants of an instan e representing
the Dut h inter ity train network.

Abstra t.

1 Introdu tion
The demand for railway transportation in Europe has been in reasing rapidly
during the past few years, partly in uen ed by the ongoing e orts of European governments to persuade people to use publi transportation. Be ause of the resulting in rease in railway tra , the pro ess of onstru ting
railway timetables is be oming more and more omplex. At the same time,
the introdu tion of ompetition on the European railroads has pushed the
management of railway ompanies to onsider various methods for improving their business planning pro esses, one important planning pro ess being
the onstru tion of timetables. Therefore railway ompanies are interested
in methods for generating several timetables, preferably of high quality, in a
reasonable amount of time, so as to be able to evaluate several alternatives
for their future timetables.
These developments have led to an in reasing interest in the problem
of onstru ting y li railway timetables, as operated in several European
ountries. In su h a y li timetable, train onne tions are operated on a
regular basis, e.g., a train for a ertain destination will leave every hour at 5
minutes past the hour. Re ently, several authors have published work on this
problem, with a rst su essful attempt by S hrijver and Steenbeek (1994),
who studied the problem of onstru ting a feasible timetable, and more re ent

Leon Peeters, Leo Kroon

publi ations by Na htigall (1999), Goverde (1999), and Lindner (2000) on the
optimization of y li railway timetables. These authors all use models that
are based on the Periodi Event S heduling Problem (PESP) introdu ed by
Sera ni and Ukovi h (1989).
This paper proposes some new obje tive fun tions that an be used within
the above des ribed models. These obje tive fun tions aim at minimizing the
passengers' travel time, onstru ting a robust timetable, and minimizing the
number of train ompositions to operate a timetable. Also, any ombination
of these obje tives is possible, with a possible weight fa tor.
A y li railway timetabling problem an be represented by a so- alled
onstraint graph. At the ore of the mixed integer linear program des ribing
the y li railway timetabling problem lies a basis of the y le spa e of this
onstraint graph. In this paper, we also show the onsiderable impa t that
the hoi e for a ertain type of y le basis has on the omputation times.
1.1

Pra ti al ba kground

During the past de ade, the Netherlands' largest passenger railway operator
NS Travelers and the Dut h railway apa ity manager Railned have been
putting a lot of e ort into the development of the automati timetabling
system DONS (short for Designer Of Network S hedules). One of the intelligent modules of DONS, alled CADANS, was developed by S hrijver
and Steenbeek (1994) to assist timetable planners in generating a tentative
timetable based on onstraints dedu ed from the rough layout of the railway
network, the train line system, safety regulations, and quality requirements.
The timetable onstru ted by CADANS is y li with a y le length of one
hour.
The DONS system has been in use at NS Travelers and Railned for some
years now, and CADANS generally performs well, both from a mathemati al
point of view and in the view of the planners that use it on a daily basis. However, the operational use has also indi ated some problems. Firstly,
CADANS is an algorithm based on onstraint satisfa tion, that sear hes
for some timetable that satis es all requirements. After nding a feasible
timetable, provided that one exists, the algorithm terminates without giving information on the quality of the returned timetable, or on what other
timetables may exist that also satisfy the requirements. Although it is possible to lo ally optimize the obtained timetable, su h a lo al optimization
ignores many other, possibly better, timetables. Se ondly, when no feasible
timetable exists, a minimal set of mutually on i ting onstraints is returned,
but pra ti al experien es have shown that this information does not ompletely answer the planners' requirements. In this ase, one would ideally like
to re eive information on how the spe i ed requirements should be adjusted.
A solution to these two problems is o ered by using an optimization approa h, rather then just sear hing for a feasible timetable. Su h an optimization approa h enables one (i ) to sear h for a timetable that is optimal with

Cy le Based Optimization for Railway Timetabling

respe t to the de ned obje tive fun tion, and (ii ) to gain insight into the
ne essary hanges to an infeasible instan e, by allowing a penalized violation
of the onstraints or some subset of them.
1.2

Outline of the paper

The remainder of the paper is outlined as follows. The next se tion gives
an overview of the literature on related railway timetabling problems. Se tion 3 presents the optimization extension to the existing railway timetabling
model, resulting in a mixed integer program. In Se tion 4 we propose some
prepro essing pro edures to redu e the size of instan es, and thus the number
of integer variables. The presented methods were tested on several variants of
the Dut h inter ity train network, the results of whi h are dis ussed in Se tion 5. The nal se tion ontains the on lusions and some ideas for further
resear h.

2 Literature Review
Several resear hers have studied problems in the eld of railway timetabling
or railway s heduling. Here, we only give an overview of the literature that
is related to our resear h: network based y li railway timetabling for the
types of railway networks that one nds in several European ountries, i.e.
generally one tra k for ea h dire tion of tra , and high frequen y tra
that is inter-related by many onne tions.
The model that lies at the basis of the y li railway timetabling model
was developed by Sera ni and Ukovi h (1989). Their paper des ribes the Periodi Event S heduling Problem (PESP), that onsiders the s heduling of a
set of periodi ally re urring events under periodi time window restri tions
on pairs of these events. An obje tive fun tion was not taken into a ount;
their main interest was in nding feasible periodi s hedules. They proved
that the general PESP is NP- omplete, and proposed a Bran h&Bound pro edure for nding feasible solutions. They also presented some appli ations
in job shop s heduling, transportation s heduling and tra light s heduling.
Hassin (1996) des ribes an optimization model for the Network Syn hronization Problem (NSP), a mathemati al model that provides an optimization
formulation for PESP, but does not take into a ount onstraints as hard
rules. Instead, forbidden time window values are highly penalized.
S hrijver and Steenbeek (1994) solve a PESP formulation for the Dut h
railway timetabling problem with a onstraint propagation algorithm. The
basis of this algorithm is enumeration: whenever an arrival or departure time
has been xed, the onsequen es for all other times are propagated. Moreover, some lever bran hing and ba ktra king tri ks were invented to speed
up the solution pro ess. Hurkens (1996) uses the same formulation to test a

Leon Peeters, Leo Kroon

polyhedral solution method. He on ludes that the method works, but omputation times in rease rapidly with the instan e sizes. Odijk (1996, 1997)
uses a utting plane algorithm to generate timetables. The obje tive of his
resear h is to qui kly generate a family of timetables, in order to evaluate
possible infrastru ture lay-outs around and within stations. His results are
promising, but only for rather small problems, i.e. a station and a small network immediately surrounding it. Na htigall (1999) studied optimization for
y li railway timetabling, using as obje tive minimal waiting times, minimal
travel times, and a multi- riteria optimization. He developed two lasses of
fa et de ning inequalities for the problem, and uses these in a utting plane
method, whi h is able to solve medium sized real life problems. We refer to
Na htigall (1999) for a more detailed overview of y li railway timetabling
problems. The onstru tion of minimum ost timetables was studied by Lindner (2000). His model ombines the assignment of train types and train units
to train lines with the timetabling problem.

3 The Cy li Railway Timetabling Problem


Very brie y stated, the Cy li Railway Timetabling Problem (CRTP) onsists
of nding y li arrival and departure times for all trains at their orresponding relevant stations in a railway system, taking into a ount aspe ts su h as
safety, servi e levels, onne ting trains, operating e ien y, et . In this paper
we will mainly fo us on passenger train timetabling, although it is possible
to handle argo trains more or less in the same way.
This se tion gives a des ription of a model for the CRTP. We start with
an overview of the assumptions, and ontinue with a des ription of the onstraints in our model. Next, a y le based formulation for these onstraints
is presented. Se tion 3.4 des ribes the obje tive fun tion, that may be nonlinear. A method for linearizing the obje tive fun tion is proposed in Se tion
3.5. The nal subse tion des ribes a heuristi for nding a good y le basis.
3.1

Assumptions

Sin e the 1930's, a y li timetable for passenger trains has been in use in
the Netherlands. This means that trains leave (almost) every hour of the
day at the same time from a ertain station for a ertain dire tion (with
a possible deviation of a few minutes). Be ause of the y li hara ter of
the passenger servi e timetable, one usually also reserves y li time paths
for argo trains; these are assigned to spe i trains at some later point in
time. In order to onstru t a y li timetable, one only needs to onstru t a
y li timetable for one hour, alled an hourly pattern. This hourly pattern
forms the basis for the a tual timetable, where adjustments are made for
rush hours, weekends, holidays, spe ial trains, et . For onstru ting su h an
hourly pattern, we assume the following information to be given:

Cy le Based Optimization for Railway Timetabling

We assume the infrastru ture to onsist of nodes (stations,


jun tions, bridges, rossings et .) and tra ks that onne t these nodes.
It is possible that there are multiple parallel tra ks between two nodes;
in that ase, however, we assume that ea h train has been assigned to
one of these tra ks a priori. We do not onsider the infrastru ture within
stations in detail.
Train lines: The information on trains is given in the form of train lines. A
train line is a dire t train onne tion between an origin and a destination
station that is operated with a ertain frequen y (number of trains per
y le time). Furthermore, a train line has a ertain type (inter ity, interregional, lo al, argo) that determines what intermediate stations the
trains of that line will serve. Finally, we are also given the xed trip times
between every onse utive pair of nodes on the train line.
Constraints: The timetable should satisfy three types of onstraints: safety
onstraints, marketing onstraints and trip time onstraints. The safety
onstraints are used to ensure that trains will not ollide, requiring that
there should be a ertain time bu er between possibly on i ting train
movements. Marketing onstraints are used to model any other requirements that the timetable should have in order to either make it as attra tive as possible for passengers, or to be ost e e tive. With respe t
to the passenger-friendliness one should imagine transfer possibilities to
onne ting trains, xed departure times for ertain trains, and syn hronization of the departure times of trains heading for the same dire tion.
Regarding the ost-e e tiveness there are matters like short turning times
for trains at destination stations, so as to use the rolling sto k as e iently
as possible. The trip time onstraints, nally, relate the departure and
arrival times of a train at subsequent stations.

Infrastru ture:

3.2

The Basi Model

As mentioned, we want to onstru t a y li pattern whi h onsists of arrival


and departure times for every train at ea h traversed node in the railway
network. To model this, we introdu e the on ept of an event. An event is
a ombination of (1) either an arrival or a departure, (2) a train, and (3) a
node in the railway network. The set of events, denoted by N , an be derived
from the infrastru ture and from the train line data. In order to onstru t a
timetable, we then need to determine the time instant at whi h ea h event
takes pla e. These time instants should respe t the onstraints that were
des ribed in the previous se tion, whi h all turn out to be onstraints on
pairs of events, as will be explained in more detail below. The set of pairs of
events for whi h we have a onstraint is denoted by A. The model uses the
following sets, de ision variables, and parameters:

Leon Peeters, Leo Kroon

N
A  (N  N )
vi 2 f0; : : :; T
pij 2 Z
T
[lij ; uij

the set of events that need to be s heduled.


the set of event-pairs for whi h a onstraint exists.
1g the integer time instant within period T at whi h
event i 2 N takes pla e.
the integer variable modeling the y li nature of the
onstraint relating events i and j .
the y le time of the timetable, whi h for our purposes will usually equal 60 minutes.
the time window relating the events i and j , meaning
that event j should take pla e between lij and uij
minutes after event j takes pla e.

All onstraints that were mentioned in the previous se tion, safety onstraints, marketing onstraints, and trip time onstraints, an be represented
by onstraints on pairs of events. Su h a onstraint on the event-pair (i; j ) 2 A
is written in the following way:

vj vi 2 [lij ; uij T :
(1)
This onstraint states that event j should take pla e between lij and uij
minutes after event i takes pla e. The notation [:T means that the onstraint
is y li with y le time T , i.e. ea h onstraint should be taken modulo T .
We assume that 0  uij lij < T , sin e else the onstraint would have no
meaning in a y li setting with y le time T . Multiple onstraints may exist
for the event pair (i; j ), these will be ignored to keep the notation simple. A
representative set of example onstraints for the CRTP an be found in the
appendix.
We model the y li nature of the onstraints by the integer de ision
variable pij . The p-variables will be used to subtra t, when ne essary, an
integer multiple of the y le time T from the time window [lij ; uij . The
feasible region for the railway timetabling problem an be des ribed as
8
>
<

vj vi + Tpij 2 [lij ; uij 8(i; j ) 2 A >


=
V = > v 2 Zn 0  vi  T 1
8i 2 N > (2)
;
:
p 2 Zm
where n is the number of events and m the number of onstraints.
The onstraints of the timetabling problem an also be represented by
a dire ted onstraint graph G = (N; A), with a node i for every event that
needs to be s heduled, and an ar (i; j ) whenever there exists a onstraint
vj vi 2 [lij ; uij T . The ar (i; j ) will be asso iated with the time window
[lij ; uij . Figure 1 shows an ar in the onstraint graph G.
If a onvex obje tive fun tion is used, the variables vi in (2)

Theorem 1.

an be relaxed to real values.

Cy le Based Optimization for Railway Timetabling

If the integer variables pij are xed, we are left with an integer program with time window onstraints on the v -variables only. The onstraint
matrix for this integer program is the node-ar in iden e matrix for the onstraint graph G. This means that all verti es of the polyhedron V are integral,
and we an drop the requirement that v should be integer.


Proof

Sera ni and Ukovi h (1989), Odijk (1997) and Na htigall (1999) showed
that the Periodi Event S heduling Problem (PESP), whi h onsiders the
problem of nding a feasible solution to V , is NP- omplete.
[lij , uij]T

Ar ( ) in the onstraint graph

Fig. 1.

3.3

i; j

A Cy le Based Formulation

The previous se tion introdu ed y li onstraints involving an integer variable to mathemati ally model the CRTP. These onstraints prove to be very
powerful for learly expressing timetable requirements by relations between
pairs of event times. For solving the problem, we shall however use a di erent
model. Instead of studying the feasible region (2), we shall use a formulation
based on y les in the onstraint graph G, as proposed by Na htigall (1999)
and S hrijver (1999).
The variables vi , orresponding to the nodes in the graph G, are also
known as potentials, see e.g. Ro kafellar (1984). Correspondingly, an ar variable xij = vj vi is a so- alled tension. It is well-known that a ve tor x is a
feasible tension if and only if the oriented sum of tensions xij along any y le
in the onstraint graph G equals zero. It is however su ient to onsider a
basis C of the y le spa e. For a y le , let ij be 1 for forward ar s, {1 for
ba kward ar s, and 0 for ar s that are not in the y le. Then x is a feasible
tension if and only if
X

2A

(i;j )

ij xij = 0

8 2 C :

Rewriting V in terms of tensions xij gives


8
>
>
>
<

X = >x 2 R m
>
>
:


P

(i;j ) A


ij
ij


m


8 2 C
2 ij xij = 0
l  x + Tpij  uij 8(i; j ) 2 A
p2Z

9
>
>
>
=
>
>
>
;

(3)

Leon Peeters, Leo Kroon

Note the strong relation between potentials vi and tensions xij . Given potentials vi , tensions are straightforwardly al ulated. Conversely, given tensions
xij , a mat hing set of potentials an be onstru ted by xing one of them,
say v0 = 0, al ulating all other potentials by the formula vj = vi + xij , and
afterwards taking all potentials modulo T .
Sin e we are onsidering a y li problem with y le time T , there is no
di eren e between studying values x or values x + Tp. Therefore the above
set an be formulated as
8
>
>
>
<

X = > x 2 Rm
>
>
:


P

(i;j ) A


ij
ij


m

P


2 ij xij = T i;j 2A ij pij 8 2 C
l  x  uij
8(i; j ) 2 A
p2Z
(

9
>
>
>
=
>
>
>
;

(4)

The rst ondition in this expression states that the oriented sum of tensions
along ea h y le should equal a multiple of the y le time T . This an also
be expressed using a new integer variable q for ea h 2 C , giving
8
>
>
>
<

X = >x 2 R m
>
>
:


P


(i;j ) A ij


ij
ij


m
n+1

2 xij = Tq 8 2 C
l  x  uij
8(i; j ) 2 A
q2Z

9
>
>
>
=

(5)

>
>
>
;

where m n +1 is the number of y les in a y le basis. Note that a y le basis


may be onstru ted by starting with a spanning tree of G, and iteratively
reating a y le by adding one of the m (n 1) non-tree ar s.
Finally, we an obtain bounds on the q -variables by substituting the
bounds on the x-variables. This gives
X
=1
(i;j ): ij

lij

= 1
(i;j ): ij

uij  Tq 

X
=1
(i;j ): ij

Be ause of the integrality of q , we an divide by


the following bounds l  q  u
2

l
u

=
=

6
6
6
6
6
6
4

uij

= 1
(i;j ): ij

13
X

=1
(i;j ): ij

lij

(6)

T and round, whi h yields

lij :

=1
(i;j ): ij

uij

X
= 1
(i;j ): ij

uij A77

X
= 1
(i;j ): ij

lij

17
7
7
A5

(7a)
(7b)

Cy le Based Optimization for Railway Timetabling

Putting all the above together, the feasible region of the CRTP is des ribed
by
8

9
P
x = Tq
>
>
>

>

8

2
C
ij

>
>
(i;j )2A ij
>

>
>
>
<

X = >x 2
>
>
>
>
>
:

Rm




lij  xij  uij


l  q  u
q 2 Zm n
+1

8(i; j ) 2 A
8 2 C

>
>
=
>
>
>
>
>
>
;

(8)

The model based on the feasible region X turns out to solve mu h faster
than the model based on V . The remainder of the paper fo uses on the
model des ribed by (8). For experien es with the original model, we refer to
an earlier paper (Peeters, 1999).
3.4

Obje tive Fun tions

Clearly, railway ompanies are interested in a high quality timetable. Several


aspe ts an in uen e the quality of a timetable. One ould think of passenger
satisfa tion (e.g. the available time to transfer between trains), operating
e ien y (e.g. running trains at the highest possible speeds so that traveling
times are minimal and the rolling sto k usage is maximal), and robustness
(e.g. there should be su ient bu er time in the timetable to absorb delays).
All of these desired hara teristi s an be in luded in one spe i timetable
only to a ertain extent; some of them are on i ting, in whi h ase a trade-o
between them has to be made.
The rst part of the obje tive fun tion is aimed at dire ting ertain tensions towards ertain values. Let Ao denote the subset of tensions that are
in orporated into the obje tive fun tion. The obje tive fun tion will be the
sum over fun tions of the tensions in Ao . We de ne a fun tion fij (xij ) for
every (i; j ) 2 Ao to represent the preferred values of the tension xij . The
fun tions fij will be linear or quadrati . In Figure 2 these two fun tions are
illustrated. In Figure 2.I the value in the middle of the window [lij ; uij is
preferred, in Figure 2.II the value lij is preferred. Both preferred values are
assigned a fun tion value of zero. Pra ti ally, one an think of fun tions that
re e t a preferen e for the middle of a frequen y window or the middle of a
safety window (see example onstraints 4 and 5 in Appendix A), whi h would
result in a fun tion resembling Figure 2.I. Another possibility is a preferen e
for minimal dwell and onne tion times (see example onstraints 2 and 3 in
Appendix A), whi h gives Figure 2.II. Linear fun tions fij an be used for situations in whi h ea h minute of deviation from the preferred value is equally
important. Quadrati fun tions fij are onvenient for expressing passengers'
per eptions of waiting: it is well known that ea h additional minute of waiting
is valued heavier than the previous one. Moreover, quadrati fun tions an
be used to reate a di eren e between solutions that would yield the same
solution value when using linear fun tions. With the fun tions fij , the rst

10

Leon Peeters, Leo Kroon


fij

fij

uij

lij
I

Fig. 2.

T
dij

uij

lij

dij

II

Linear and quadrati fun tions ij ( ij )


f

part of the obje tive fun tion is expressed as


Minimize

2Ao

(i;j )

fij (xij ):

Minimizing the number of trains needed to operate a timetable will be the


se ond obje tive. Consider Figure 3, whi h illustrates the outward and return
trip for a train, and the so- alled turn-around onstraints that tie these two
trips together. These last onstraints atta h a time-window to the time that
a train spends at a terminal station before exe uting its return trip. Su h a
time-window is needed to lean the train, to shunt it, or to repla e the rew.
The task- y le [outward trip!turnaround waiting!return trip!turnaround
waiting is alled a rolling sto k ir ulation y le, or just ir ulation y le.
The q -variable for a ir ulation y le orresponds exa tly to the number of
hours it takes a single train to exe ute one ir ulation, assuming a planning
y le time of one hour. Therefore, q equals the number of trains that will be
needed to exe ute the part of the timetable orresponding to the ir ulation
y le, sin e every hour one train has to leave the origin node. Let Ct be the
set of all su h ir ulation y les in the onstraint graph G. Minimizing the
number of trains needed to operate a timetable an then be expressed as
Minimize

2Ct

q :

Note that more ompli ated ir ulation stru tures an be dealt with similarly,
e.g. ir ulation stru tures in whi h a train exe utes several other trips before
returning to the origin station of its initial outward trip.

Cy le Based Optimization for Railway Timetabling

Rolling sto k ir ulation y le in

Fig. 3.

11

The following mixed integer non-linear program then de nes the CRTP:
Minimize
subje t to

2Ao

(i;j )

(i;j )

fij (xij ) +

ij xij = Tq

l  q  u
lij  xij  uij
x 2 Rm ; q 2 Zm

2Ct

q
8 2 C
8 2 C
8(i; j ) 2 A

(9)

n+1

Here, and are weights for the two obje tives. Of ourse, one an also
asso iate weights with ertain train y les, or in orporate weights in the
fun tions fij . Note that in this formulation, the set of train y les Ct must
be part of the y le basis C . One an also onstru t a y le basis that does
not ontain the train y les expli itly, and write the q -variables for the train
y les as linear ombinations of q -variables for y les in the y le basis.
The mixed integer non-linear program (9) an be transformed
to a mixed integer linear program by repla ing all quadrati fun tions ij ( ij )
0 that are exa t for integer values of ij . The
by linear approximations ij
resulting mixed integer linear program will have exa tly the same optimal
solution and solution value as (9).
Theorem 2.

f x
x

On e the integer ve tor q is xed, the onstraint matrix for X be omes


a y le-ar in iden e matrix, and is therefore totally unimodular. This means
that the polyhedron represented by X has integral verti es only. The only
non-linear part of (9) are quadrati fun tions fij . If we approximate these by
linear fun tions fij0 that are exa t for integer values of xij , the solution and
solution value of the approximated mixed integer linear program will be the
same as the solution and solution value of (9).


Proof

3.5

Linearizing Quadrati Obje tive Fun tions

We an use Theorem 2 to linearize quadrati obje tive fun tions fij , and
thereby formulate the CRTP as a pure mixed integer program. Consider

12

Leon Peeters, Leo Kroon

the quadrati fun tion in Figure 2.II. For all integer values xij 2 [lij ; uij ,
the quadrati fun tion fij (xij ) = x2ij passes through the following points
(xij ; fij (xij ))
(lij + ; 2 )
= 0; : : :; uij lij :
The line through the onse utive points
(lij + ; 2 ) and (lij + + 1; ( + 1)2 )

is des ribed by

y = a()xij + b()

with

8
<

a()
b()

= 0; : : :; uij lij

= 0; : : :; uij lij

1:

= 2 + 1;

( + 1) (2 + 1)lij :
We will linearize the quadrati fun tion fij by introdu ing an auxiliary variable fij0 . These variables are bounded from below as follows
fij0  a()xij + b()
2 ij ;
(10)
with ij = [0; : : :; uij lij 1. For the quadrati fun tion in Figure 2.I both
:

the right and left half an be approximated similarly by linear equations.


The above approximation yields the following mixed integer programming
formulation for the CRTP
Minimize

subje t to

2Ao

(i;j )

(i;j )

fij0 +

2Ct

ij xij = Tq

l  q  u
lij  xij  uij
fij0  a()xij + b()
x 2 Rm ; q 2 Zm n

q
8 2 C
8 2 C
8(i; j ) 2 A
8(i; j ) 2 Ao ; 2 ij

(11)

+1

Note that, in order to apply Theorem 2, the quadrati fun tions fij must
be approximated su h that the approximation is exa t for every integer value
in the time window [lij ; uij . For wide time windows, this will result in many
linear inequalities that bound the variables fij0 from below. This in turn may
lead to very large linear programming relaxations in a Bran h&Bound pro edure. A ompromise might then be to approximate the quadrati fun tions
su h that the approximation is exa t only for = 0; : : :; max , and to approximate to remainder of fij by the last linear inequality

fij0  (2

max

+ 1)xij

max

(max + 1)

(2max + 1)lij :

Cy le Based Optimization for Railway Timetabling


3.6

13

Cal ulating a Cy le Basis

Consider the following quantity, the so- alled width of a y le basis C

W (C ) =

2C

( u

l + 1) :

(12)

The width of a y le basis is an indi ation of the size of the solution spa e of
the CRTP, sin e it ounts the possible values of the ve tor q . It is therefore
lear that a good y le basis would be one with small W (C ). However, onstru ting a y le basis that minimizes W (C ) is hard be ause of the rounding
in al ulating u and l . We use a heuristi proposed by Na htigall (1999) to
onstru t a y le basis with small width. First, ompute a spanning tree that
is minimal with respe t to the quantities wij = uij lij + 1, the width of the
time windows. Then, iteratively add a non-tree ar , whi h generates a y le.
Ea h of the wide non-tree ar s will therefore only appear in one y le. Moreover, the part of a y le that is ontained in the tree will have a relatively
small width. Therefore, the total width of the y le basis will hopefully be
small. A similar pro edure was used by Sera ni and Ukovi h (1989) for their
algorithm to solve PESPs.
If we want to in lude rolling sto k ir ulation y les into the obje tive
fun tion (see Se tion 3.4), we pro eed as follows in onstru ting a so- alled
ir ulation y le basis. For all ar s orresponding to trip time onstraints,
dwell time onstraints, and turn-around onstraints, we set wij = 0. For
all other ar s, wij = uij lij + 1 as before, and we iteratively add a nontree ar as before. This way, all rolling sto k ir ulation y les have width
zero, and will therefore be in luded in the y le basis. After all ir ulation
y les have been added to the y le basis, the pro edure omplements the
y le basis with small width y les as in the previous pro edure. Clearly, the
ir ulation y le basis may be mu h wider than the y le basis onstru ted
by the heuristi , espe ially for trains that stop frequently, resulting in many
dwell time onstraints, and wide turn-around windows.

4 Prepro essing
The instan es that we use are obtained from the de ision support system
DONS. The size of instan es an be redu ed drasti ally by applying prepro essing. By deleting ar s from G, we may also delete some y les, thereby
redu ing the number of integer variables, whi h an onsiderably improve
the solution pro ess.
The DONS instan es typi ally ontain
many parallel onstraints, e.g. safety onstraints for trains that belong to
the same train series, and that are therefore already separated by their
frequen y onstraints.

Removing parallel onstraints.

14

Leon Peeters, Leo Kroon


Nodes with degree one in the onstraint
graph an be deleted: they are no part of any y le, and the tension that
orresponds to the deleted ar an be al ulated in a post-pro essing
phase. Nodes with degree two an be ontra ted, i.e. their two adja ent
ar s are merged into a single ar , the window of whi h is the sum of the
previous two windows. Consider the following example

Nodes with degree one or two.

vj vi 2 [lij ; uij T ;
vk vj 2 [ljk ; ujk T :

Here, node j is assumed to have degree two, so that the ar s (i; j ) and
(j; k ) an be merged. This ontra ted ar then represents the onstraint

vk vi 2 [lij + ljk ; uij + ujk T :

A ontra ted ar an be expanded during post-pro essing to al ulate


the tensions for the two original ar s. A ontra tion an however only be
exe uted if (uij + ujk ) mod T  (lij + ljk ). Otherwise, the ontra ted
ar represents two disjun t time windows. Consider the example

vj vi 2 [20; 40 ;
vk vj 2 [20; 40 ;
60

60

resulting in the merged onstraint

vk vi 2 [40; 80

60

= [0; 2060 [ [40; 5960 :

The two disjun t windows may be modeled by two new parallel ar s, but
we may then just as well not ontra t the two original ar s.
Removing subsequent safety onstraints. Consider the situation of two
trains, 1 and 2, running from station A to station B, and using the same
tra k. Let the trip times along tra k AB be r1 and r2 for train 1 and 2
respe tively, and denote the departure times from station A by v1 and v2 .
The arrival times at station B will then be (v1 + r1 ) and (v2 + r2 ). Upon
leaving and entering a station, there should be a bu er time between the
two trains (see also example 5 in Appendix A). Pra ti ally, this bu er
time is often the same for both entering and leaving a station. Supposing
the identi al bu er times equal h, we get the following two onstraints

v v 2 [h; T hT
(v + r ) (v + r ) 2 [h; T hT
2

for leaving station A,


for entering station B,

whi h an be rewritten as

v
v

2
2

v 2 [h; T hT
v 2 [h; T hT + (r
1

r)
2

for leaving station A,


for entering station B.

Cy le Based Optimization for Railway Timetabling

15

If r1 = r2 , learly one of the onstraints an be removed. If r1 6= r2 , then


both onstraints an be repla ed by the following onstraint

v 2 [h; T hT \ [h + (r

r ); T h + (r
2

r )T :

Consider the
ase where several trains, not ne essarily belonging to the same series,
use a ertain segment of railway tra k. Suppose that these trains do not
stop at ea h station they pass. This means that we an ontra t their trip
times over the segment, under the ondition that the safety onstraints
are also adjusted. Su h a situation is illustrated in Figure 4, where the

Shrinking systems of trip time and safety onstraints.

r1 2

r2 3

r1 2+r2 3
3

2
6

s2 5

s1 4

s3 6

1 3

32
6

s3 6

s1 4'

s1 4''
6

r4 5
Fig. 4.

56

r5 6

4 6

r4 5+r5 6

Shrinking a system of trip time and safety onstraints

nodes (1,2,3) orrespond to one train, and the nodes (4,5,6) to the other.
The windows for all safety onstraints are assumed to be [h; T h, and
are denoted by s14 ; s25 ; s36 . The ontra tion of the equality onstraints,
respe ting all safety bu ers, an be done by moving the safety onstraint
between the heads of two ar s to the tails of the ar s. For the trip ar s (1,2)
and (4,5) this means that s14 should be interse ted with s25 + r12 r45 ,
giving
s014 = s14 \ (s25 + r12 r45 ):
Note that this an also be seen as strengthening the window s14 by onsidering the path 1-2-5-4. Repeating this pro edure gives

s00

14

= s014 \ (s36 + r12 + r23

45

r ):
56

Espe ially for large subgraphs of this stru ture, involving several trains
and segments, the shrinking pro edure onsiderably redu es the number
of ar s and, more importantly, the number of y les in the onstraint
graph.

5 Computational Results
The model was tested on the Dut h inter ity network for 1997/1998, as obtained from NS Travelers. We will rst des ribe the instan e in more detail,
then dis uss the used obje tive fun tions, and on lude with an overview of
the omputational results.

16
5.1

Leon Peeters, Leo Kroon


The IC97 Instan e

The IC97 instan e ontains all the inter ity trains in the Netherlands in the
hourly pattern for the 1997/1998 timetable. The instan e onsists of 50 stations and 25 train lines. The y le time is one hour, i.e. T = 60. Although all
trains have frequen y one, many routes are visited by multiple train lines, and
frequen y onstraints between these di erent lines are de ned so as to have
a half hour servi e on the majority of the routes. Conne tions are de ned
su h that it is possible to travel with a good onne tion between any two
inter ity stations. The instan e also ontains many onne ting and dis onne ting events between train lines, to model the situation where two trains
are onne ted to travel a ommon part of their trips as one train. The initial onstraint graph, as obtained from the DONS system, onsists of 1475
nodes and 3394 ar s. After applying the prepro essing pro edures of Se tion
4, the size of the onstraint graph is redu ed to 217 nodes and 586 ar s.
Consequently, a y le basis for the onstraint graph will ontain 370 y les.

init
1
2
3
4

onne tion dwell turn-around


[2 560 [1 360 [10 4060
[2 560 [1 560 [10 4060
[2 860 [1 860 [10 4060
[2 1260 [1 1160 [10 4060
[2 560 [1 560 [10 5960
Variants of the IC97 instan e
;

Table 1.

We onsider four variants of the instan e, see Table 1, that di er in the


width of the time windows for the three types of market onstraints: onne tion onstraints, dwell time onstraints, and turn-around onstraints. Instan e init is the initial instan e, that turned out to be infeasible. In the
rst three variants, the widths of the onne tion and dwell time windows in rease. The purpose of in reasing the width of these time windows is two-fold.
Firstly, relaxing the market onstraints' time windows is what we propose as
a possible solution for dealing with infeasible instan es. It is however not
lear beforehand how large su h a relaxation should be. Studying these three
variants shows the results of small, middle and large relaxations. Se ondly,
it allows us to study what happens when the instan es be ome harder, sin e
the solution spa e in reases. This issue is of ourse also related to the rst
purpose. The fourth variant onsiders wide turn-around time windows. Note
that the lower bounds for the time windows remain the same for all variants.
We assume that the lower bounds represent hard operational onstraints that
may not be relaxed, i.e. they represent the absolute minimum time needed

Cy le Based Optimization for Railway Timetabling

17

for the asso iated a tivity. For all these variants the frequen y onstraints
are perfe t; their time windows onsist of the singleton 30. We will also study
variants, indi ated by the letter F, for whi h the frequen y time window is
relaxed to [25; 3560. For all variants, the headway time equals three minutes.
5.2

Obje tive Fun tions

The following obje tive fun tions were onsidered in our omputational experiments:
L
Q
Ri
T
TL
TQ

minimizing the sum of onne tion and dwell times by a


linear obje tive fun tion,
minimizing the sum of onne tion and dwell times by a
quadrati obje tive fun tion,
maximizing robustness with max = i,
minimizing the number of trains needed to operate the
timetable,
ombination of C and L,
ombination of C and Q.

For the obje tive fun tions L, Q, and R, the weight parameters are = 1; =
0. For L and Q, the set Ao onsists of all onne tion and dwell time onstraints, that are penalized linearly and quadrati ally, respe tively. These
obje tive fun tions aim at minimizing the travel time of passengers, sin e
ea h minute above the absolute lower bound of a onne tion or dwell time
window is time that a trains spends waiting at a station, and therefore adds
to the total travel time of the passengers in it. For R, the set Ao onsists of
the safety onstraints, and the deviation from the middle of the safety time
windows is penalized quadrati ally. The idea here is to push trains apart as
mu h as possible, so that it be omes less likely that they on i t with ea h
other in ase of delays. These onstraints typi ally have very wide time windows, e.g. for the IC97 instan e [3; 5760 . We will therefore onsider three
variants, that di er in max , the number of inequalities that are used to approximate the quadrati obje tive fun tion. For the obje tive fun tion T the
weight parameters are set to = 0; = 1, and a rolling sto k ir ulation
y le basis is used. Finally, for TL and TQ, the weight parameters are set to
= 0; = 10; 000.
5.3

Computational Results

We used the mixed integer programming solver of CPLEX 6.6 (ILOG CPLEX
website, 2000) to solve problem (11). The tables below ontain the omputation times in se onds (time), the number of Bran h&Bound nodes (nodes),
the number of iterations required to solve the linear programming relaxations

18

Leon Peeters, Leo Kroon

(iterations), and the optimal obje tive value (zopt ) for ea h of the variants. All
omputations were done on a Pentium III 667 MHz with 128 Mb of memory.
Ea h line in the tables represents a ombination of an obje tive fun tion
and a variant of the instan e. For example, L1 means obje tive fun tion
L for variant 1, Q33 means obje tive Q with max = 3 for variant 3, and
CL4 obje tive CL for variant 4. Whenever an F appears, the frequen y time
windows have been set to [25; 3560 instead of the default [3060 .
Table 2 ompares linear and quadrati obje tive fun tions. It is lear
that linear obje tive fun tions result in faster omputation times, due to
the smaller size of the linear programming relaxations. However, a loser
inspe tion of the optimal solutions shows that quadrati obje tive fun tions
yield solutions in whi h the deviations from the lower bound of the onne tion
and dwell time windows are smaller. A ni e ompromise is to approximate
the quadrati fun tion only partially exa tly, as is shown by omparing Q3,
Q33 , and Q37 . Computation times de rease, while the approximation error is
small. Finally, it an be on luded that the omputation times are a eptable
for pra ti al use.

time (se ) nodes iterations opt


L1
11.99 4564 17589 104
L2
15.16 6502 21799 87
L3
38.62 12442 29513 87
LF1
15.91 2686 17866 93
LF2
55.59 7844 95899 83
LF3
229.34 37127 292591 83
Q1
57.04 6146 62488 233
Q2
71.34 6131 80531 225
Q3
291.11 16968 189701 225
Q33
105.43 13557 126376 199
Q37
180.29 13360 145608 225
QF1
83.21 5020 82013 199
QF2
140.01 6484 135220 196
QF3
438.05 18228 228780 196
Computational results for the IC97 instan e: minimizing passenger travel
z

Table 2.

time.

The results for maximizing the robustness of a timetable are pla ed in


Table 3. From these results it is lear that restri ting the approximation of the

Cy le Based Optimization for Railway Timetabling

19

quadrati fun tions fij to low values of max is essential for omputation times
to stay within reasonable limits. For R18 , i.e. max = 8, the IC97 instan e
ould not be solved within four hours. However, from a pra ti al point of view
this restri ted approximation is no problem. Indeed, with a headway time of
three minutes, approximating exa tly up to max = 3 results in a timetable
for whi h the approximation is exa t up to twi e the minimal required safety
bu er time, whi h is more than adequate for pra ti al purposes.

time (se ) nodes iterations opt


R13
44.95 4530 56844 31
R15
599.12 42187 618784 159
R18
*
Computational results for the IC97 instan e: maximizing robustness.
*11% gap after 4 hrs of omputing.
z

Table 3.

Table 4 shows the results for minimizing the number of trains needed
to operate a timetable, possibly in ombination with minimizing passengers'
travel time. In all solutions, 65 trains are needed to operate the timetable.
Note that the a tual number of trains needed is larger, sin e we did not
in lude international trains and some other spe ial ases into the obje tive
fun tion. The omputation times are again very a eptable, although a lot
of extra time is needed for TL1 and TQ1 in omparison to L1 and Q1. This is
be ause the y le bases are mu h wider when we are minimizing the number
of trains, sin e they must ontain the quite wide ir ulation y les. Also,
omputation times in rease when relaxing the turn-around time windows
from variant 1 to variant 4, but this also results in a de rease in the total
passenger travel time.

time (se ) nodes iterations opt


T1
30.33 5593 26413 65000
TL1
83.12 13165 51711 65127
TQ1
81.9
6578 83438 65295
T4
61.75 12641 81644 65000
TL4
52.02 10440 35892 65091
TQ4
53.18 3994 61014 65219
Computational results for the IC97 instan e: minimizing train usage
ombined with minimizing passenger waiting time.
z

Table 4.

20

Leon Peeters, Leo Kroon

Finally, Table 5 illustrates the importan e of hoosing a good y le basis.


The rst three lines of the table are the same as the orresponding lines
in Table 2, the last three lines represent the same obje tive fun tion and
variant as the rst three, only based on the mu h wider train ir ulation
y le bases (the subs ript stands for ir ulation). From the table it is lear
that omputation times in rease drasti ally when using the train y le bases.
Quite remarkable is the fa tor four di eren e in omputation time between
Q1 in Table 5, and TQ1 in Table 4. Both onsider variant 1 and obje tive
Q, but the latter one also in ludes obje tive T. Apparently, the in lusion of
minimization of the number of trains in the obje tive leads to a mu h faster
solution pro ess.

time (se ) nodes iterations opt


Q1
57.04 6146 62488 233
Q2
71.34 6131 80531 225
Q3
291.11 16968 189701 225
Q1
342.38 28655 281537 233
Q2
741.43 52153 589643 225
Q3
968.18 45306 532134 225
Computational results for the IC97 instan e: omparing y le bases.
z

Table 5.

The infeasibility of the initial variant init is found within a fra tion of a
se ond in CPLEX's presolve phase, regardless of the used obje tive fun tion.
The tables show that the penalized relaxation of the onstraints o ers a quite
adequate tool for reating a timetable that minimally violates the initial
onstraints. With respe t to the aspe t of omputing an optimal timetable,
it is interesting to remark that for most variants of the IC97 instan e, up
to four or ve integer feasible solutions are found during the Bran h&Bound
pro ess, where the value of the rst found integer solution an be about 3
to 4 times worse than the optimal integer solution. Clearly, a lot might be
missed when one just sear hes for a feasible solution.

6 Con lusions and Further Resear h


The y le based model for the CRTP seems to work quite well on this sample
of test instan es. Solving the model yields timetables that have a onsiderably better obje tive value than other, merely feasible, solutions that might
be found by CADANS. However, CADANS nds a feasible timetable (for the
relaxed variants) within a omputation time varying from 5 to 15 se onds,
whi h is faster than our optimization model needs. Moreover, the presented

Cy le Based Optimization for Railway Timetabling

21

model provides an easy and e e tive on ept for dealing with infeasible instan es. An important fa tor to make the method work well is the appli ation
of the prepro essing pro edures des ribed in Se tion 4. Without these pro edures, omputation times an easily in rease to up to several hours.
Obviously, the IC97 instan e and its variants presented here do not give a
full view of the performan e of the model on general CRTP's. A rst dire tion
for further resear h is therefore to test the model on other, more ompli ated
instan es. A se ond dire tion of resear h is to derive more general ases for
whi h shrinking of the onstraint graph is possible. One idea is to also ontra t dwell time onstraints, viewing the result as a train with a variable,
instead of a xed, trip time. Under ertain onditions, whi h are des ribed
in Kroon and Peeters (1999), safety onstraints involving only the departure
and arrival surrounding su h a variable trip time are ne essary and su ient
to meet all safety restri tions along the route. Other dire tions of resear h
in lude the investigation of methods to onstru t a good, maybe even optimal y le basis, and the in orporation of the lasses of valid inequalities by
Na htigall (1999) into the model. Finally, it is not our intention to start from
s rat h in onstru ting a timetable, without using the very well performing
CADANS algorithm. We therefore want to investigate how to use the knowledge presented in this paper to improve the pra ti al use of CADANS and
DONS.

A knowledgements
The authors would like to thank both referees for their useful omments. This
resear h is sponsored by Railned and NS Travelers as a part of the DONS
resear h and development proje t.

Bibliography
R.M.P. Goverde. Improving pun tuality and transfer reliability by railway
timetable optimization. In P.H.L. Bovy, editor, Pro eedings TRAIL 5th
Annual Congress, volume 2, Delft, The Netherlands, 1999.
R. Hassin. A ow algorithm for network syn hronization.
sear h, 44(4):570{579, 1996.

Operations Re-

C. Hurkens. Een polyhedrale aanpak van treinrooster problemen (A polyhedral approa h to railway timetabling problems). Unpublished. In Dut h,
1996.
ILOG CPLEX website. http://www.ilog. om/produ ts/ plex/, 2000.
L.G. Kroon and L.W.P. Peeters. A variable running time model for y li
railway timetabling. Erasm management report series 28-1999, Erasm,
Rotterdam, The Netherlands, 1999.
T. Lindner. Train S hedule Optimization in Publi Rail Transport. PhD
thesis, Te hni al University Brauns hweig, Brauns hweig, Germany, 2000.
K. Na htigall. Periodi Network Optimization and Fixed Interval Timetables.
PhD thesis, Deuts hes Zentrum fur Luft- und Raumfahrt, Brauns hweig,
Germany, 1999.
M.A. Odijk. A onstraint generation algorithm for the onstru tion of periodi railway timetables. Transportation Resear h, 30(6):455{464, 1996.
M.A. Odijk. Railway Timetable Generation. PhD thesis, Delft University of
Te hnology, Delft, The Netherlands, 1997.
L.W.P. Peeters. An optimization appro h to railway timetabling. In P.H.L.
Bovy, editor, Pro eedings TRAIL 5th Annual Congress, volume 1, Delft,
The Netherlands, 1999.
R.T. Ro kafellar. Network Flows
& Sons, New York, 1984.

and Monotropi Optimization.

John Wiley

A. S hrijver. Timetabling for the Dut h railways. Presented at MAPSP '99,


Renesse, The Netherlands, 1999.
A. S hrijver and A. Steenbeek. Dienstregelingontwikkeling voor Railned
(Timetable onstru tion for Railned). Te hni al report, CWI, Center for
Mathemati s and Computer S ien e, Amsterdam, The Netherlands, 1994.
In Dut h.
P. Sera ni and W. Ukovi h. A mathemati al model for periodi event s heduling problems. SIAM Journal of Dis rete Mathemati s, 2(4):550{581, 1989.

Cy le Based Optimization for Railway Timetabling

23

A Example Constraints
This appendix presents some representative examples of pra ti al railway
timetable onstraints. To improve the larity of these examples, we will use
the variable d for departure times and a for arrival times. The examples below
are illustrated in Figure 5.
Rotterdam

Utrecht
train A

[30]T

[3,57]T

[1,3]T

[2,5]T

train D

Amsterdam
train C
[28,32]T
train B

Fig. 5.

Illustration of example onstraints

Example 1. Trip time onstraint.

Inter ity train A runs from Rotterdam to Utre ht, whi h takes 30 minutes.
We have
aA;Utre ht dA;Rotterdam = [30T :
Example 2. Dwell time onstraint.

Suppose that upon arrival in Utre ht, train A has to stop between 1 and 3
minutes for the boarding and alighting of passengers. This gives

dA;Utre ht aA;Utre ht 2 [1; 3T :


Example 3. Conne tion onstraint.

Next, suppose that train B leaves from Utre ht for Amsterdam, and that
some passengers of train A want to transfer to train B in Utre ht. Train
B should leave between 2 and 5 minutes after train A arrives, so that the
passengers have enough time to transfer. We then have

dB;Utre ht aA;Utre ht 2 [2; 5T :


Example 4. Frequen y onstraint.

Now let train B belong to some train series with frequen y 2 (twi e per hour),
and suppose we want to have a exibility of (plus or minus) 2 minutes in this
frequen y. Let train C be the other train in this series. Sin e twi e per hour
means that a train should leave every 30 minutes, and taking into a ount
the 2 minutes exibility, we obtain

dC;Utre ht dB;Utre ht 2 [28; 32T :

24

Leon Peeters, Leo Kroon

Example 5. Safety onstraint.

Finally, suppose that train D is a Stoptrain, running from Rotterdam to


Utre ht, and using the same tra k as train A. If train A leaves Rotterdam,
then train D should not leave Rotterdam in the following 3 minutes, and vi e
versa.This yields the following two onstraints

dD;Rotterdam dA;Rotterdam 2= [0; 3)T and dA;Rotterdam dD;Rotterdam 2= [0; 3)T ;


whi h together give

dD;Rotterdam dA;Rotterdam 2= ( 3; 3)T :


Be ause of the y li nature of the timetable, this is the same as

dD;Rotterdam dA;Rotterdam 2 [3; 57T


sin e the omplement of the interval ( 3; 3)T is [3; 3T , and [3;
[3; 57T , assuming that the y le time equals 60 minutes.

3T =

Das könnte Ihnen auch gefallen