Improving Punctuality and

Transfer Reliability by
Railway Timetable Optimization

TRAIL Research School, Delft, December 1999

Author:
Drs. Rob M.P. Goverde
Transportation Planning and Traffic Engineering Section,
Faculty of Civil Engineering and Geo Sciences, Delft University of Technology

Thesis supervisors:
Prof.Dr.-Ing. Ingo A. Hansen
Transportation Planning and Traffic Engineering Section,
Faculty of Civil Engineering and Geo Sciences, Delft University of Technology

Prof.Dr. Geert Jan Olsder

Department of Technical Mathematics,
Faculty of Information Technology and Systems, Delft University of Technology
 Contents

Abstract

1 Introduction 1

2 Modelling Timetable Constraints 4

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Time Window Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Cycle Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Rolling Stock Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Railway Timetable Optimization 16

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 The Periodic Railway Timetable Optimization Problem . . . . . . . . . . . . . 16
3.3 Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Preprocessing: Integer Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4 The Network Synchronization Problem 20

4.1 The Optimization Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Tight Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 Reliable Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.4 Transfer Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

5 Conclusions 28

Acknowledgements 28

References 28
Abstract

The Dutch railway network is operated close to capacity with the current safety
system. This leaves little space for control by process operators by which a delayed
train can cause severe delay propagation. The NS currently invest a large amount
in punctuality improvement of the railway operations. This paper shows how the
effect of delays can already be incorporated in the timetable design process to
obtain robust and optimized timetables for a punctual and reliable operation.
A detailed modelling approach is presented to compute periodic network timeta-
bles with an optimal distribution of buffer times at those points where they are
needed the most without being too conservative. The variables are the buffer
times and the proposed objective is to minimize a weighted sum of individual
costs for each buffer time. Here, the weights define the relative importance of the
buffer time costs. The following cost functions are proposed and proved to be
convex for general running time distributions: buffer time (fast travel times for
through passengers and small operating costs); expected delay due to buffer time
failure of rolling-stock and crew connections (increasing punctuality) and train
circulations (stability of train circulations); and expected transfer waiting time
(trade-off between small transfer times and transfer reliability).
The constraints on a feasible timetable are modelled conveniently by time window
constraints concerning interactions between two periodic events (arrivals and de-
partures). This is compatible with the timetable design system DONS of the NS.
For the optimization problem these constraints are reformulated using the buffer
times as variables rather than the periodic event times. The resulting cycle con-
straints state that the oriented sum of process and buffer times on fundamental
cycles of the constraint graph is a multiple of the cycle time. The buffer times
are not periodic by which they can be used straightforwardly in the objective
functions. The periodic time windows are thus replaced by periodic cycles.
Theoretical relevance
The periodic timetable optimization problem extends the associated feasibility
problem, the Periodic Event Scheduling Problem (PESP), and improves several
problem issues. The problem is formulated as an optimization problem with con-
vex cost function and buffer times as variables rather than the periodic events
which is most common in literature. Solving the resulting large-scale convex
mixed-integer programming problem requires a synthesis of various techniques
from linear algebra, network flows, integer programming, and convex program-
ming.
Societal relevance
An optimal robust timetable contributes to improve performance and attractivity
of the railway operation, increase service quality to passengers, and decrease costs
to railway operators and passengers due to delays. Moreover, a punctual and re-
liable railway operation may attract more potential passengers who currently use
other modes (cars). This reduces the congestion problem on the Dutch highways
and the associated environmental pollution.
1 Introduction

Punctuality is a major concern to all transportation companies, especially to an

underlying transportation network with lots of interdependent services and low
service frequencies (up to 6 services per hour). The level of punctuality largely
influences the reliability of connections, including crew, vehicle, and transfer con-
nections. On the other hand scheduled connections are a main source for sec-
ondary (or knock-on) delays (for example if a connecting service waits for a de-
layed feeder service to secure a transfer connection) and hence on its turn affects
punctuality. An optimal coordination or synchronization of connected services
is thus important to improve punctuality whereas at the same time reliability
of connections is obtained. Both service punctuality and connection reliability
are main indicators for the performance and attractiveness of a transportation
network. This paper concentrates on railway networks as they represent a par-
ticularly interesting case.
Railway networks have an additional difficulty due to the safety and signalling
systems which increase the possibilities of conflicting movements. Moreover, the
train service network of the Netherlands Railways (NS) is highly interconnected
and is served near to its capacity (for the present safety and signalling system).
Additionally, the railway timetable in the Netherlands is an integrated periodic
timetable, i.e., all train services run in regular intervals. As a consequence, the
railway planners are faced with an extreme hard and time-consuming task to
design a feasible timetable that satisfies all constraints as imposed by for instance
the line system, required connections and infrastructure limitations. Therefore,
specific attention to the synchronization of connected services is currently very
modest.
A feasible timetable is by definition realizable (conflict-free) for punctual op-
eration. However, perturbations during operation result in variations of train
running and dwell times. Perturbations are for instance varying train charac-
teristics, varying driving behaviour, fluctuations in dwell times depending on the
number of alighting and boarding passengers, and extreme weather conditions. A
timetable has to be robust to small variations in running times which is achieved
by adding running time margins to the minimum running times. This way the
(scheduled) running times can compensate for most of the variations. The more
severe perturbations result in exceeding the scheduled running times, which give
primary delays. Moreover, once a train is delayed it may hinder other trains or
propagate the delay to connecting trains, resulting in secondary delays. The sec-
ondary delays are indirectly the result of the robustness of the timetable design.
A robust timetable design aims at minimizing secondary delays during operation
and thereby improving punctuality and reliability.
A railway timetable consists of the scheduled arrival and departure times of all
train services at each station. Implicitly this also gives the dwell times for each
line and the transfer times between lines at transfer stations where lines meet.
From an operational point of view the timetable also fixes the connection times
for rolling stock and/or crew connections, and headways between trains (at e.g.

TRAIL Research School, December 1999 1

 conflict points). All these process times consist of two components: a necessary
minimum process time (e.g., minimum dwell time for alighting and boarding; min-
imum transfer time for alighting, walking to the departure platform and boarding;
and minimum arrival headway between two arriving trains at a station) and a
possible buffer time or recovery time which improves robustness with respect to
delays.
We are concerned with a periodic railway timetable. Obviously, this imposes
severe restrictions on the timetable. In combination with the network interde-
pendencies of all train services, it is no surprise that there is an amount of slack
in a timetable necessary to fit all process times in the regular interval pattern.
The resulting buffer times lead to larger process times which at first glance is
annoying. However, buffer times may be very desirable to improve robustness
of the processes with respect to variations in process times. The challenge is to
distribute the buffer times over the service network such that they are advanta-
geous during railway operation. This is the periodic railway timtable optimization
problem.
The development of the advanced design system DONS (Designer of Network
schedules) [12, 25] is a great step forward to computing a feasible timetable that
satisfies all constraints, or if no such timetable exists, giving information about
conflicting constraints. DONS also features a timetable optimization option that
can be used after having computed a feasible timetable. The optimization has
then be understood with respect to the computed train (sequence) orders. So
the train orders remain fixed which obviously results in local optimal solutions,
i.e., better (feasible) timetables might be obtained if train orders are allowed
to be swapped. Moreover, more advanced cost functions have to be defined
to (sub)optimize the timetable with respect to operation reliability. Another
improvement would be to obtain (nontrivial) multiple feasible solutions (if there
are any). Currently, the computations stop when a feasible solution has been
found. The existence of multiple solutions with perhaps more convenient train
orders then remains unknown. This paper gives a solution to the above problems
and thereby contributes to the further development of the DONS features. The
problem of computing multiple feasible solutions to the system of time window
constraints is also the subject of a parallel research project [24].
The optimization of railway network timetables has had some attention in the
literature. The problem is typically modelled as a (linear) mixed integer pro-
gram (MIP). Weigand [28, 29] reports a graphical oriented modelling approach.
The MIP-formulation is based on cycles in a graph representing all stop and
transfer connections. Infrastructure constraints are not dealt with explicitly.
Nachtigall [19, 20] formalizes this approach and proves that the problem is N P-
complete. An alternative MIP-formulation is based on constraints imposed by
interrelated train pairs resulting in time window constraints [15, 16, 18]. Here
also infrastructure constraints (minimum headway between a train pair) are in-
corporated. The corresponding feasibility problem has been formalized by Ser-
afini and Ukovich [26] and is known as the Periodic Event Scheduling Problem
(PESP). They proved that PESP is N P-complete, and hence also the associ-
ated optimization problem is N P-complete. The PESP-formulation is also used

2 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 in the DONS-system [22, 23, 25]. A related modelling approach is the schedule
synchronization problem in which the line schedules (running and dwell times)
are fixed. The problem is to determine the optimal departure times of all lines
at their starting terminal station such that the costs of the resulting transfer
times at stations where the lines meet are minimized. This optimization problem
can be formulated as a binary integer program (BIP) [13] or an integer program
(IP) [2]. However, infrastructure constraints can not be handled trivially by these
formulations. Domschke [7] proved that the problem is N P-complete. Note that
this problem can also be modelled as an MIP using the time window (or PESP)
constraints as it is a special case for fixed dwell times. Then also infrastructure
constraints can be incorporated.
Some literature is more or less devoted to reliability and performance measures
of transportation services. Hallowell and Harker [10] studied expected delays due
to interference at meet/pass points on a partially double-tracked railway line.
Carey [3] and Ferreira and Higgins [8] consider the minimization of train run-
ning times and expected arrival delays at a single railway line including knock-on
delays. Carey (and Kwieciński) [6, 5] and Higgins and Kozan [11] have general-
ized these models to networks of various lines. Carey [4] moreover studied the
behavioural response of actual running times if the scheduled running time is in-
creased to minimize arrival delays. Knoppers and Muller [14] studied the expected
passenger transfer waiting time including the probability of missing a connection.
Finally, Meng [17] developed a model to determine the required average dwell and
transfer buffer times at a railway station for random (Gamma-distributed) inter-
arrivals of trains, incorporating the number of trains and connections, the number
of platform tracks, average secondary delays due to connections and conflicting
train paths, and the risk of exceeding platform capacity.
The paper is organized as follows. Chapter 2 shows how to model the timetable
constraints and get a constraint system of the buffer times. Chapter 3 defines
the general periodic railway timetable optimization problem and considers its
computational complexity. The network synchronization problem is addressed in
Chapter 4 where various convenient cost functions for the individual buffer times
are derived to obtain either tight connections, reliable connections, or optimal
transfer connections, whatever is appropriate.

TRAIL Research School, December 1999 3

2 Modelling Timetable Constraints

2.1 Introduction

The train service network in the Netherlands is operated according to a regular

interval or periodic timetable, which is also popular in train service networks in
most densely populated countries in Europe, for long-distance railway networks
as in Germany, and other public transportation systems. A periodic timetable
is convenient for passengers but imposes severe restrictions on the timetable.
All arrival and departure times must fit a regular interval pattern. Additional
difficulties in finding a railway timetable are the network interdependencies and
infrastructure limitations.
A (periodic) network timetable is a feasible solution to a (periodic) constraint
system that specifies all interactions between the train services in the network.
We assume that a line system is given in advance and that the process times are
deterministic and given as well. Here a line system is defined as a set of lines, the
train type of each line, and the connections between the lines at stations where
the lines meet. A line is a sequence of train services on a particular route between
two terminal stations, and the train type (for instance express train or local train)
determines the intermediate served stations on the route. We define a timetable
point as a node in the service network where interactions between trains are
observed. Examples are transfer stations, level-crossings, and emerging sections.
Then our input data contains

• train running times between timetable points including dwell times at in-
termediate stops;

• minimum dwell times at transfer stations;

• minimum turn-around times at terminal stations;

• minimum transfer times between train pairs at stations where lines meet;

• minimum crew and rolling stock connection times between train pairs at
stations where a crew transfer or train coupling/decoupling is required;

• minimum headways between train pairs at timetable points.

Additionally maximum process times may be given to prevent too large dwell
times, transfer times, etc. The crew schedules are usually planned after the
timetable has been established. However, some crew connections may be known
in advance when a timetable design is based on an existing timetable, with some
adjustments due to for instance completed infrastructure projects. Typical crew
connections are for example endpoints of lines. An extension of the planning
process may include a feedback to the timetable design after crew schedules and
rolling stock circulations have been obtained. Then slight adjustments of the

4 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 timetable design may increase the robustness of the crew connections. Note that
a crew connection does not have to occur at each timetable period.
A timetable is expressed in arrival and departure times of all services at each
station. In the sequel it is more convenient to refer to an event which can be
both an arrival or a departure. Generally, an event is the start or end of an
activity. So a train departure from a station is an event that activates a train run
to the next station, and an arrival at a station is the end of a train run from a
preceding station. A periodic event is an event that repeats at a regular interval
with cycle time T . For instance, a departure in an hourly timetable repeats every
T = 60 min. An event time is the occurrence time of the event, e.g. the departure
time.
If running times between stations are assumed deterministic (for planning pur-
poses) then it suffices to consider departure events only: an arrival time can be
given as the sum of a departure time at a preceding station and the train run-
ning time between the stations. This is convenient if large service networks are
considered as this gives a reduction to half the number of variables. The vari-
able τi (mod T ) denotes the (departure) event time of the periodic (departure)
event i. Below we express all timetable constraints in departure events. We also
somewhat sloppy refer to i as both the departure and the activated train service.

2.2 Time Window Constraints

Connections at transfer stations impose synchronization constraints on the train

services. Also the track layout and the safety and signalling system of a railway
network imposes limitations on train movements. For instance, trains can not
cross a level-crossing at the same time, and a safety distance has to be respected
for two trains running on joint track sections.
A train service is defined as a train trip between two stations. All timetable
constraints can then be defined as an interaction (e.g. a connection) between two
train services. For example, train 1 must depart no sooner than 2 to 5 min after
the arrival of train 2 to allow a transfer. Such a constraint is called a time window
constraint. Note that an interaction has a ‘direction’, i.e., train 1 waits for train
2, but the reverse does not have to hold unless also an interaction is defined from
train 2 to train 1. An interaction of two services occurs for example at a stop, a
transfer, or a conflict point (mutual track section). Each interaction produces a
minimum process time and a buffer time. So a buffer time can for instance be a
dwell buffer time, a transfer buffer time, or a headway buffer time.
Time window constraints specify that the time difference between two periodic
event times is restricted to a particular time window. Let τi and τj be two event
times. Then a time window constraint is

τj − τi ≡ lij + rij ( mod T ) and rij ∈ [0, rijmax]. (1)

Here, lij is the minimum time interval between event i and j, and rij is an addi-
tional nonnegative buffer time that is bounded from above by a given maximum

TRAIL Research School, December 1999 5

 buffer time rijmax ∈ [0, T ]. Equation (1) has to be evaluated modulo the cycle time
T which means that an integral multiple of T can be subtracted where necessary.
Without loss of generality we can write (1) as the MIP constraint
τj − τi = lij + rij + zij T and rij ∈ [0, rijmax ]
or
lij ≤ τj − τi − zij T ≤ lij + rijmax (2)
where zij ∈ ZZ.
All timetable constraints can be modelled conveniently as time window con-
straints, see also Schrijver and Steenbeek [25]. The constraints can be classified
in synchronization constraints, infrastructure constraints, and additional market
constraints, which are considered sequentially below.
Synchronization constraints specify that a train service j has to wait for a train
service i from a preceding station to guarantee a connection. A connection can
either be a stop or transfer, and we thus distinguish between stop constraints and
transfer constraints. Let tri be the running time of train service i from the preced-
ing station, and tmin
ij be the minimum dwell/transfer time to the connecting train
j. Then lij = ti + tmin
r
ij is the minimum time after the departure of train service
i from its preceding station before the connecting train service j can depart. If
rijmax is the maximum (dwell or transfer) buffer time then the synchronization
constraint is given as
τj − τi = tri + tmin
ij + rij + zij T,

where rij ∈ [0, rijmax ] is the dwell/transfer buffer time, see Figure 1.
Crew and rolling stock connections give special synchronization constraints with
respect to the logistics of crew and rolling stock schedules. First, a turn at a
terminal station is a special stop where the minimum turn-around time does not
just depend on alighting and boarding time, but also includes for instance the
necessary time for cleaning the coaches and turning the locomotive. Second,
a crew connection is a special transfer in which the crew of the feeder train is
scheduled to transfer to the connecting train. In this case the connection can not
be cancelled unless a reserve crew can be allocated to the connecting train. Third,
a rolling stock connection is a connection in which trains are coupled/decoupled.
Infrastructure constraints describe headway restrictions due to the safety and
signalling system and infastructural limitations. Consider two trains services i
and j that use a common track section some distance from the (distinct) departure
stations. Let tri be the running time of train service i to reach this point from its
departure at the preceding station, and analog for trj . Assume that a minimum
headway hij has to be respected from train service i to j, and hji if this sequence
order is reversed. Then we have
hij ≤ (τj + trj ) − (τi + tri ) − zij T ≤ T − hji.
Rewriting in the form (2) gives the infrastructure constraint
hij + tri − trj ≤ τj − τi − zij T ≤ T − hji + tri − trj . (3)

6 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 distance
transfer station
time 0
ti

tir

tijmin
rij
rijmax
tj

Figure 1: Synchronization constraint: the time window between the departure

time τi of the feeder train and the departure time τj of the connecting train at
the next station consists of the running time tri , the minimum transfer time tmin
ij ,
and the transfer buffer time rij

So here lij = hij + tri − trj and rijmax = T − hij − hji , see Figure 2. The associated
headway buffer time at the junction is thus rij ∈ [0, T − hij − hji ]. Also the
headway buffer time between train service j and the next train service i is here
determined as
rji = rijmax − rij (4)
since the sum of the minimum headways and headway buffer times of a train pair
must equal the cycle time T .
Infrastructure constraints also arise with respect to simultaneous arrivals and
departures at transfer stations, due to limitations on available train paths at
the station track layout. Moreover, train services that use the same route to
the next station are restricted by the safety distance of trains on these common
track sections. This results in minimum arrival headways between arriving train
pairs and minimum departure headways between departing trains. The associ-
ated constraints are also given as (3), where for the arrival headway constraints
the running times tri and trj correspond to the running times from the preceding
stations to the mutual station or to overtaking points at an intermediate stop,
and for the departure headway constraints tri = trj = 0, see Figure 2.
Finally, additional market constraints are restrictions with respect to market re-
quirements. This models for instance relative restrictions on the departure times
of two train services to force semi-regular interdeparture times. For instance, if
a line has twice the frequency of the overall cycle time then this line is modelled
by two sequences of train services. The frequency constraint may then specify
that two train services have a semi-regular interdeparture time. Suppose that

TRAIL Research School, December 1999 7

 distance distance
conflict point 0
time 0 time
h ji
ti ti
hij
tir
tj rij
tj
h ji
hij t rj
rij
r ji

T T

Figure 2: Infrastructure constraints: conflict point (left) and departure headway

(right). The sum of the minimum headways in both directions (hij and hji) and
the headway buffer times rij and rji equals the cycle time T

train services i and j must be separated by half the cycle time with a tolerance
of δ so that twice in a period a train service departs in a certain direction in a
semi-regular fashion. Then the frequency constraint is
1 1
T − δ ≤ τj − τi − zij T ≤ T + δ.
2 2
Here lij = 12 T − δ and rijmax = 2δ. The interpretation of the buffer time is
in this case obsolete. It can now be interpreted as a tolerance where rijmax is the
absolute tolerance. Market constraints may also specify exact departure times for
a train service. This models prefixed departure times of for instance international
(high speed) trains. For this a fictitious train service 0 is defined that has a
predetermined departure time τ0 = 0. Assume that a train service j has to
depart at exactly τj = τ then

τ ≤ τj − τ0 − z0j T ≤ τ. (5)

The maximum buffer time is 0 min and so r0j = 0 as well.

The time window constraint system can also be given in matrix notation. Let n
be the total number of events, and m be the number of time window constraints.
Number all constraints from 1 to m. Suppose that the kth constraint is τj − τi =
lij + rij + zij T . Then define the matrix M ∈ ZZn×m as (M)jk = 1, (M)ik = −1
and (M)lk = 0 for all l ∈ {1, . . . , n}\{i, j}. So a constraint corresponds with a
column of two nonzeros. Let τ ∈ [0, T )n be the vector of all event times, and
define vectors l ∈ IRm and z ∈ ZZm according to their constraint (arc) number, so
e.g. lk := lij . Furthermore, denote the vector of buffer times as x ∈ IRm+ defined

8 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 [44,54]
1 [10,36]
2

]
47
[52,62] [2,28] 8] [2,28] [28,38]

2,

8]
[4
,4

4
3

3,
[4

[4
[17,43]
3 4
[43,53]

Figure 3: Constraint graph

as xk := rij , and equivalently x̄ ∈ [0, T ]m is defined as x̄k := rijmax . Then the

constraint system is
MT τ − T z = l + x
(6)
x ∈ [0, x̄], z ∈ ZZm .

Note that we use the transpose matrix M T rather than directly defining a matrix
N = M T . This is motivated by the interpretation of the matrix M as the node-arc
incidence matrix of a directed graph, the so-called constraint graph.
The constraint graph is a directed graph (or digraph) G = (V, E), where V is the
set of n nodes and E is the set of m arcs, see Figure 3. The nodes i ∈ V correspond
to the periodic events τi and for each pair ij in a time window constraint there
is a directed arc ek = (i, j) ∈ E from i to j. Each arc has an associated lower
and upper capacity given as lij and uij = lij + rijmax , respectively. Note that the
constraint graph can have multiple arcs between two adjacent nodes, if several
time window constraints are defined for one pair of event times. The columns of
M define the arcs of the constraint graph: an arc ek starts at node i for which
(M)ik = −1 and terminates at node j for which (M)jk = 1.
In the sequel we assume that the constraint graph is strongly connected, i.e.,
there is a directed path between any two nodes. Otherwise, the problem can
be decomposed into subproblems for which the constraint graphs are strongly
connected. These subproblems can then be solved separately, and the solutions
can afterwards be linked to one overall timetable. Note that the constraint graph
contains all interactions between train services, including infrastructure restric-
tions. Second, we assume that the constraint graph does not contain loops, i.e.,
cycles of one arc. Note that a loop can not be defined by a node-arc incidence
matrix. A loop would imply a relationship between an event to itself. This occurs
for instance if a train circulates on a route without interactions to other trains
but for one station. The loop constraint then specifies that the sum of the (cir-
culation) running time and the dwell time at the one station where interactions
occur has to be a multiple of the cycle time. Clearly, this constraint can be solved
independently and does not influence the remaining constraint system.
The time window constraint system (6) contains both the event times τ and the
buffer times x. Note that these variables are redundant: if we know the event
times then we also know the resulting buffer times, and vice versa. Since we are
interested in optimizing the buffer times it is desirable to use these buffer times

TRAIL Research School, December 1999 9

 as decision variables. Note that since each buffer time xk ∈ [0, T ], these variables
do not suffer from modular arithmetic and can be used as variables in objective
functions without causing discontinuities and thereby disturbing convexity. The
next section derives an equivalent constraint system in the variables xk (k =
1, . . . , m) only.

2.3 Cycle Constraints

Consider a simple railway network of two stations and trains circulating between
the stations. Then the sum of running times and dwell times on this circulation
must be a multiple of the cycle time since after returning at its origin station the
train has to wait for its periodic departure time, which is exactly a multiple of an
hour (or more generally, the cycle time) after its former departure, before starting
a new circulation. So the sum of running times, minimum dwell times, and dwell
buffer times has to be a multiple of the cycle time. Since the running times and
minimum dwell times are given, this results in a constraint with respect to the
buffer times. This is an example of a cycle constraint.
The above example can be generalized to circuits in a complex railway network
with lots of interactions, represented by the constraint graph. From graph theory
it is known that any circuit can be obtained by a linear combination of funda-
mental cycles [1]. A cycle is a sequence of adjacent arcs (a path) where the initial
and final node is the same. The arcs on a cycle may have different directions. So
it is not necessary that a cycle is an ordered sequence of successive processes (a
circuit). Thus, more abstract, on each cycle in the constraint graph the oriented
sum of the process times (running times, minimum dwell times, minimum trans-
fer times, and minimum headway) and buffer times must be an integral multiple
of the cycle time T . It is sufficient to consider only fundamental cycles of a cycle
basis from which all other cycles (and all circuits) can be obtained. The (finite)
number of fundamental cycles in the cycle basis, the so-called cyclomatic number,
is ν = m − n + 1, where m and n are the number of arcs and nodes in the graph,
respectively. This can be clarified as follows: a spanning tree of a digraph G is
a subgraph that contains all nodes of G but no cycles. The number of arcs of a
spanning tree is n − 1 because any additional arc gives a cycle. The number of
nontree arcs is therefore m − (n − 1) = ν. Each nontree arc generates a funda-
mental cycle (together with the path on the spanning tree from the tail of the
nontree arc to its head). The resulting ν cycles form a cycle basis. Note that a
cycle basis is not unique.
The cycle constraint system defines all interactions between services given in
terms of buffer times and taking into account the network structure. The peri-
odicity of the event times is replaced by the periodicity of circulation times on
cycles. Thus, the system of ν = m − n + 1 cycle constraints is equivalent with
the system of m time window constraints (all interactions). Note that less cycle
constraints are necessary to model all interactions as a result of incorporating
the network structure. This implies that the number of integer variables in the
constraint system is reduced by n − 1.

10 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 A cycle constraint system can be obtained by using the cycle matrix representa-
tion of the constraint graph. The cycle (-arc) matrix Γ is defined as the ν × m
matrix where each row is a vector representation of a fundamental cycle. That is,
(Γ)ij = 1 if arc j is a forward arc on cycle i, (Γ)ij = −1 if arc j is a backward arc
on cycle i, and (Γ)ij = 0 otherwise. The direction of an arc on a cycle depends
on the orientation of the cycle. Using the cycle matrix the constraint system
becomes
Γx − T z = b
(7)
0 ≤ x ≤ x̄, z ∈ ZZν ,
where Γ ∈ IRν×m is the cycle matrix, b ∈ IRm is the vector of the negative oriented
sum of minimum process times of the constraint graph/time window constraints,
x̄ ∈ [0, T ]m is the vector of maximum buffer times, and T ∈ IN is the cycle time.
Note that the right-hand side can be computed as b = −Γl, where l = (l1 , . . . , lm )T
are the minimum process times of the constraint graph/time window constraints.
If a solution x to (7) has been computed then we can also simply compute the
associated event time vector τ by solving M T τ = l + x for τ , and subsequently
set τ := τ ( mod T ). Since l + x is now a given vector M T τ = l + x can be
computed efficiently using standard numerical algorithms such as LU decompo-
sition or Gauss-Jordan elimination. Moreover, note that we in fact only need
n − 1 independent rows of M T to completely determine τ up to a constant (one
degree of freedom). The event time vector τ is then completely determined by
fixing one event time. Note that if absolute constraints (5) are imposed then τ
is uniquely determined by fixing the fictitious event time τ0 = 0. On the other
hand, if we have found a solution τ to the time window constraint system (6) then
the associated buffer time vector is simply obtained as x = M T τ − l ( mod T ).
This shows that the time window constraint system (6) and the cycle constraint
system (7) are equivalent.
Above we gave a graph-theoretical construction of a cycle basis based on a span-
ning tree in the graph. This gives an algorithm to compute the cycle matrix.
Below we present an alternative algorithm for computing the cycle matrix based
on linear algebra.
Recall from linear algebra that a null space of a matrix M is the space spanned
by vectors x such that Mx = 0. A null space can be computed efficiently by
Gaussian elimination with back-substitution (or by Gauss-Jordan elimination)
in O(m3 ) time for general matrices, see e.g. Strang [27]. The following theorem
states that this algorithm can also be applied to compute a cycle matrix from the
incidence matrix. Note that since the incidence matrix is very sparse (m = 2n),
the efficiency of the algorithm is better than O(m3 ).

Theorem 1 Let G be a directed graph and M the associated (node-arc) incidence

matrix. Then

1. a cycle basis of G is also a basis of the null space of M,

2. if a basis of the null space of M is given by vectors in {0, ±1}m−n+1 then it

is also a cycle basis,

TRAIL Research School, December 1999 11

 3. the matrix of which the rows are the basis vectors of the null space of M
computed by Gaussian elimination with back-substitution, is a cycle matrix.

Proof. Consider the incidence matrix M in more detail. The rows correspond to
nodes and the columns to arcs in the associated digraph. At each row a 1 implies
that the arc is an incoming arc of the node and a −1 implies an outgoing arc.
On the other hand, each column in M can be viewed as a vector representation
of an arc, where the entry −1 denotes the tail and the entry 1 denotes the head.
Now, observe that on a cycle the number of incoming arcs over all nodes equals
the number of outgoing arcs. Thus, a cycle can be found as a linear combination
of all columns (arcs) giving the zero vector, that is, by x ∈ {0, ±1}ν solving
Mx = 0. These vectors x are just elements of the null space of M. On the other
hand, the range of an incidence matrix M is the space spanned by the vectors
corresponding to the arcs in the graph, i.e., all possible combinations Mx of the
columns of M. A basis of the range is a set of independent vectors spanning this
vector space. This corresponds to a spanning tree in the graph. The dimension
of the range of a node-arc incidence matrix equals its rank, which is n − 1. The
dimension of the null space then is m − (n − 1) = m − n + 1. Recall that a cycle
basis also contains ν fundamental (independent) cycles. It follows that a cycle
basis is also a basis of the null space of M. This proves the first statement of
the theorem. Reversely, since a cycle is represented as a vector with entries in
{0, ±1} this is an additional requirement for a basis of the null space to be a cycle
basis, which proves the second statement.
Finally, we prove the third statement. Assume that we compute ν linearly inde-
pendent solutions x to Mx = 0 by Gaussian elimination with back-substitution
(choosing unit vectors for the free variables in the back-substitution process).
Then x ∈ {0, ±1}m. This can be proven as follows. The independent columns
correspond to a spanning tree on G. Taking one free variable xk = 1 at a time
and the remaining free variables zero, and solving for the independent variables
by back-substitution, gives the free variable (nontree arc) in terms of the inde-
pendent variables which corresponds exactly to the associated path from the tail,
across the spanning tree, to the head and closing the cycle. Or equivalently, the
nontree arc combined with the reversed path on the spanning tree is zero. The
tree arcs are traversed only once in the reversed direction and so x = {0, ±1}m .2

2.4 Rolling Stock Constraints

The amount of buffer time to be scheduled in the timetable may be bounded

by a maximum amount of rolling stock. This imposes additional constraints to
the timetable constraint system. These constraints can be formulated as linear
(in-)equalities of the buffer times and do not contain additional integer variables.
An alternative way to include rolling stock constraints is to express them in terms
of the integral multipliers z. In this way the rolling stock constraints restrict the
feasible region of the integral variables. This results in a smaller search space and
can thus decrease the computational effort necessary to find the optimal solution.

12 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 Assume that trains circulate on fixed sequences of lines exclusively. An example
is the popular case that trains are assigned to two lines corresponding to running
a specific route in both directions. In this case a train arriving at its terminating
station turns around to run in its opposite direction and is not assigned to a
different line. Also more complex routings are possible covering a certain sequence
of different lines that give a circuit in the service network. Such a circuit is called
the circulation of a train.
A rolling stock constraint is modelled as follows. Let w̄i be a prescribed maximum
number of trains on circulation i. Then the rolling stock constraint is given as
m
X
uik xk + di ≤ w̄i T,
k=1

where T is the cycle time, m is the number of connections/arcs in the constraint

graph, xk is the buffer time of stop k on the circulation, di is the minimum
circulation time of circulation i (the sum of running times and minimum dwell
and turn-around times), and uik = 1 if circulation i contains stop k and uik = 0
otherwise. Note that a train circulation is a circuit.
These constraints can also be written in vector notation. Let the number of
circulations be denoted as υ and define the matrix U ∈ {0, 1}υ×m as (U)ij = uij .
Then the rolling stock constraints are given as

Ux + d ≤ T w̄. (8)

Here the inequality is defined componentwise. The vector d is given as d = Ul,

where l is the vector of minimum process times (or lower arc capacities in the
constraint graph). The rolling stock constraints can now be taken into account
by adding (8) to the periodic constraint system (7).
An alternative way to include the rolling stock constraints in the constraint system
is to express the maximum number of trains of all circulations in the integral
multipliers z. This restricts the feasible region of the integral variables and the
constraint system (7) is not expanded. The rolling stock constraints are now
implicitly present in the feasible region of the integral variables.
Recall that any circuit can be written as a combination of fundamental cycles.
The rows of the matrix U are the vectors representing the circulations, i.e., U =
(u1 , . . . , uυ )T , where ui = (ui1 , . . . , uim). The cycle basis consists of the rows of
the cycle matrix. A circulation i represented by the row vector ui can thus be
written as
ui = ci Γ, i = 1, . . . , υ, (9)
where ci = (ci1 , . . . , ciν ) is a row vector representing the scalar multiples of the
rows of the cycle matrix Γ (the fundamental cycles) which generate the circulation
i. Note that (9) can be written in matrix notation as U = CΓ, where C =
(c1 , . . . , cυ )T ∈ ZZυ×ν is an unknown matrix. This matrix C can be found by
solving (9) (or in standard form uTi = ΓT cTi ) for all i = 1, . . . , υ using Gaussian
elimination.

TRAIL Research School, December 1999 13

 Now the feasible set of integral variables z can be restricted as follows.
Ux + d ≤ T w̄ ⇔
CΓx + d ≤ T w̄ ⇔
C(b + T z) + d ≤ T w̄ ⇔ (by (7))
Cz ≤ w̄ − (Cb + d)/T.
This results in the following constraint on the feasible region:
1
Cz ≤ bw̄ − (Cb + d)c.
T
In the same vein, a given minimum number wi of trains on all circulations i results
in the following constraint on the feasible region of the integral variables:
1
Cz ≥ dw − (Cb + d)e.
T
Note that a minimum number of trains on a circulation can be obtained if all
buffer times are zero. By (8) this gives w ≥ dd/T e. So if no explicit minimum
number of trains is given than this lower bound can be used to restrict the feasible
integer region. Note that it is now easily checked whether or not a given minimum
number of trains on a circulation is feasible by comparing it with this lower bound.
If the number of trains w on all circulations are fixed in advance then the feasible
region is restricted by
Cz = w − (Cb + d)/T
Of course also combinations of the above three rolling stock circulation constraints
are possible.
A different kind of a rolling stock constraint is obtained when the maximum
amount of trains on the entire service network is given. In this case the allocation
of trains to routes is still free. The optimization problem then also determines
the optimal number of trains taking the given upper bound into account. Let
v̄ ∈ IN be the maximum number of trains. In the same notation as introduced
above, the rolling stock constraint becomes
υ X
X m
uik (xk + lk ) ≤ v̄T. (10)
i=1 k=1

This constraint implies that the sum of the minimum circulation times and buffer
times on all circulations must not be larger than the maximum number of trains
to be allocated multiplied by the cycle time. If a minimum amount of rolling
stock is desired then the analog to (10) holds with a ≥ inequality, and for a fixed
number of rolling stock the analog to (10) holds with equality.
Note that these rolling stock constraints may result in infeasibility if the number
of trains is not enough to operate the service network according to a desired cycle
time T . This is easily checked by considering inequality (10) with all xk = 0. The
minimum necessary number of trains is thus at least
& υ m '
uik lk
XX
v≥ .
i=1 k=1 T

14 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 Again the rolling stock constraint can also be expressed in the integral variables.
In the same vein as before we then obtain for a given maximum amount of rolling
stock   
υ Xν  υ Xν υ X m 
X  1 X X 
cij zj ≤ v̄ −  (cij bj ) + (uik lk ) . (11)
i=1 j=1 T i=1 j=1 i=1 k=1

This constraint gives an upper bound to a linear combination of the integral

multipliers zj associated with the jth cycle constraints. For a given minimum
amount of rolling stock the analog to (11) holds with a ≥ inequality (and rounding
to the nearest integer above), and for a fixed amount of rolling stock the analog to
(11) holds with equality. Note that the right-hand side of (11) is a fixed integer.

TRAIL Research School, December 1999 15

3 Railway Timetable Optimization

3.1 Introduction

The railway timetable optimization problem can be viewed as computing a fea-

sible network timetable with an optimal allocation of buffer times. Buffer times
are the variables in this optimization problem. The objectives differ between the
various buffer times. Dwell buffer times should be as small as possible to ob-
tain small travel times. Transfer buffer times can be used to guarantee reliable
connections and so some buffer time has a positive impact. Turn-around buffer
times should guarantee the stability of train circulations. Buffer times in rolling
stock and crew connection times are convenient for improved punctuality. Head-
way buffer times reduce the probability of conflicts and are desired as large as
possible.
This chapter presents the general periodic railway optimization problem based on
the cycle constraints as derived in the previous chapter. These constraints agree
with the DONS interface, that is, the generated files of variables and constraints
used by CADANS (the solver of DONS) to compute a feasible timetable are also
sufficient to obtain the variables and constraints in the optimization problem
presented in this paper. So the optimization problem is compatible with DONS.

3.2 The Periodic Railway Timetable Optimization Problem

Let m be the total number of buffer times (or interactions), and let xk (k =
1, . . . , m) be the buffer time with respect to an interaction k. Then the periodic
railway timetable optimization problem can be given as

min F (x1 , . . . , xm )
m
X
γik xk − T zi = bi i = 1, . . . , ν (12)
k=1
0 ≤ xk ≤ x̄k k = 1, . . . , m
zi ∈ ZZ i = 1, . . . , ν.

Here, F : IRm → IR is a cost function of the buffer times xk (k = 1, . . . , m), x̄k ∈

[0, T ] (k = 1, . . . , m) are the maximum buffer times, T is the cycle time (or period
length), −bi is the oriented sum of minimum process times on a fundamental cycle
i of the constraint graph, ν is the number of fundamental cycles in the constraint
graph, zi (i = 1, . . . , ν) are integers, and γik = 1 if interaction k corresponds to
a forward arc on cycle i, γik = −1 if interaction k corresponds with a backward
arc on cycle i, and γik = 0 otherwise. Note that the (cycle) constraints state that
the oriented sum of process times and buffer times on a fundamental cycle of the
constraint graph equals an integral multiple of the cycle time T .

16 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 The optimization problem can also be formulated in vector notation as

min F (x)
Γx − T z = b (13)
0 ≤ x ≤ x̄, z ∈ ZZν ,

where Γ ∈ IRν×m is the cycle matrix, b ∈ IRm is given as b = −Γl where l ∈

[0, ∞)m is the vector of the minimum process times of the constraint graph/time
window constraints, x̄ ∈ [0, T ]m is the vector of maximum buffer times, and
T ∈ IN is the cycle time.
Additionally, the integers can be constrained by a convex set Ω ⊆ ZZν defined by
constraints on the amount of rolling stock, see Section 2.4. Also integer lower
and upper bounds can be computed, see Section 3.4. These integer bounds can
also be incorporated in Ω.
The function F : IRm → IR models the ‘cost’ of any allocation of buffer times over
the timetable. From a computational point of view the cost function F has to be
convex implying that any local optimal solution is also the global optimum. For
nonconvex F the problem is practically unsolvable. A convenient cost function is
a (weighted) sum of the relative costs of all buffer times, i.e.,
m
X
F (x) = ck Fk (xk ) (14)
k=1

where ck ≥ 0 are positive real numbers, and the functions Fk : IR → IR are

smooth convex functions of a single buffer time. The weights ck give the relative
importance of the individual cost contributions and thereby tune the priorities
of the buffer times. This particular form of F is a convex separable function
of the buffer times. Note that a linear cost function cT x is a special case of
(14). This separable formulation is convenient for modelling the individual cost
contributions of all buffer times. In the next chapter several appropriate cost
functions are presented representing the buffer time costs to passengers and/or
railway operators.
Another possible formulation of cost functions is F (x) = maxk=1,...,m [Fk (xk )],
resulting in a minimax problem. The objective is here to minimize the maximum
of all buffer time costs. For example, if Fk (xk ) = xk (k = 1, . . . , m) then the
objective is to minimize the maximal buffer time over all buffer times. However,
for our purposes this objective is not felt appropriate as the individual buffer times
vary in interpretation. Only for special cases of uniform variables this objective
may be interesting. For instance, if we are only interested in minimizing the
transfer times over the network. Also the priorities of the individual buffer times
can not be modelled in a minimax problem.

3.3 Computational Complexity

The minimization problem (12) is a convex mixed-integer programming problem,

with linear mixed-integer constraints and a convex (separable) cost function. Note

TRAIL Research School, December 1999 17

 that although the arrival/departure times are periodic events, the buffer times
xk ∈ [0, T ] are not periodic. Therefore the cost function is not affected by pe-
riodicity. If the optimization problem would have been formulated directly with
respect to the time window constraints and the cost function as a function of the
periodic arrival/departure times then the cost function would have been noncon-
vex by the modular arithmetic (arrival and departure times have to be evaluated
modulo the cycle time).
For fixed integers z (and F convex in x), the optimization problem (12) becomes
a convex programming problem, and if F is linear in x then the problem is a
linear programming problem. These problems can be solved efficiently.
However, the integer variables z in the mixed-integer constraints make the prob-
lem difficult to solve. In fact, the optimization problem (12) is N P-complete
which follows from the N P-completeness of the underlying feasibility problem of
finding solutions to the cycle constraint system (or the time window constraint
system). For small networks the railway optimization problem can be solved
exactly by a branch-and-bound algorithm [21] using convex programming relax-
ations.
For large-scale networks more effort is necessary to solve the problem. Both for
solving the convex relaxation problem and the mixed-integer part. In particular
the structure of the timetable constraint system (the timetable polyhedra) can
be examined for preprocessing, and to tighten the feasible region by which the
branch-and-bound procedure becomes more effective. Nachtigall [20] examined
the timetable polyhedron and found two classes of facet defining valid inequalities.
Moreover he developed two polynomial separation algorithms which can be used
to generate strong cutting planes.

3.4 Preprocessing: Integer Bounds

Solving the constraint cycle system is complicated by the integrality condition on

the vector z. However, lower and upper bounds on the integral variables zi can
be obtained by exploiting the network structure. This reduces the search space
considerably.
Recall that x ∈ [0, x̄]. Then a lower bound on a cycle is obtained if all backward
arcs have a maximum buffer time and the forward arcs have zero buffer time.
Likewise, an upper bound is given if all forward arcs have a maximum buffer
time and the backward arcs have zero buffer time. Consider a cycle i in the
+ +
constraint graph. Define γik = max(0, γik ), i.e., γik = 1 if arc k is a forward arc
+ −
on cycle i, and γik = 0 otherwise. Equivalently, define γik = min(0, γik ), i.e.,
− −
γik = −1 if arc k is a backward arc on cycle i, and γik = 0 otherwise. Then
m
X m
X
− +
γik x̄k ≤ bi + zi T ≤ γik x̄k .
j=k k=1

Let the ν × m matrices Γ+ and Γ− be defined accordingly as (Γ)+ +

ij = γij , and

18 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 (Γ)− −
ij = γij . Then we obtain the lower bound vector

1
z := d (Γ− x̄ − b)e (15)
T
and upper bound vector
1
z := b (Γ+ x̄ − b)c. (16)
T
Here dae is the nearest integer above a (the ceiling operator) and bac is the near-
est integer below a (the floor operator), defined componentwise. An additional
constraint to the cycle constraint system (7) is thus

z ≤ z ≤ z. (17)

Note that the choice of the cycle basis influences the lower and upper bounds on
the integral vector z, and thus also the solution algorithm efficiency.
The above derived integer bounds depend on the decision variable upper bounds
x̄. These upper bounds (maximum buffer times) may be taken equal to the cycle
time T . In this way the overall optimal solution results. Note that the upper
bounds have become somewhat superfluous as large buffer times are penalized by
the cost functions. However, removing the buffer time upper bounds magnifies
the gaps between the lower and upper bounds on z given by (15) and (16). In
a branch-and-bound scheme this implies larger computation times. On the other
hand, tight upper bounds x̄ on buffer times to avoid large waiting times may
result in infeasibility of the constraint system.

TRAIL Research School, December 1999 19

4 The Network Synchronization Problem

4.1 The Optimization Problem

The aim of this paper is to find an optimal network synchronization. So we are

concerned with buffer times between train connections. These include buffer times
at stops, transfers, rolling stock connections, crew connections, and turns. The
resulting optimization problem will be referred to as the network synchronization
problem.
We consider the following optimization model as obtained in the two previous
chapters:
m
X
min ck Fk (xk )
k=1
m
X
γik xk − T zi = bi i = 1, . . . , ν (18)
k=1
0 ≤ xk ≤ x̄k k = 1, . . . , m
zi ∈ ZZ i = 1, . . . , ν.

where ck ≥ 0 are positive real numbers, and the functions Fk : IR+ → IR are
smooth convex functions of a single buffer time xk ≥ 0. Additionally, the integer
vector z may be constrained by a feasible region Ω defined by the rolling stock
constraints (Section 2.4) and the integer bounds (Section 3.4). In this chapter we
will present some appropriate cost functions Fk for the network synchronization
problem.
We are concerned with an optimal allocation of buffer times over the network.
Therefore, it is convenient to express cost in terms of time (minutes). Of course
other cost units are possible as for instance the cost equivalent in cash. The
following cost functions are proposed:

• Buffer time: minimizing dwell buffer times to obtain small travel times.
• Expected departure delay: maximizing buffer time reliability of logistics
connections (crew connections, rolling stock connections, and turns of trains
at terminal stations (stability of train circulation)).
• Expected transfer waiting time: optimizing a trade-off between a small
transfer buffer time and a high transfer reliability.

The cost function F in the optimization problem is a (weighted) sum of these

cost functions. The weights give the relative importance of the individual cost
contributions. This tunes the priorities of circulation stability, crew and rolling
stock connections, transfers, and travel times. The resulting network synchro-
nization problem aims at finding a timetable with small travel times and transfer
waiting times, and a punctual and reliable railway operation.

20 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 Also other choices of cost functions are possible representing the specific objectives
of the problem at hand. Recall that we are concerned with connections only
and so we did not consider cost functions for headway buffer times. Of course,
appropriate cost functions can be derived that penalize small headway buffer
times between a pair of trains in both directions. The subsequent sections consider
the individual cost functions for the network synchronization problem.

4.2 Tight Connections

An obvious measure for the quality of a timetable is the total (synchronization)

buffer time, or the total passenger waiting time resulting from these buffer times.
Optimizing this measure results in a linear cost function to the network synchro-
nization problem: the weighted sum of buffer times. So the linear cost function
is m X
F (x) = ck xk , (19)
k=1

where ck ≥ 0 is a constant weight reflecting the priority of dwell/transfer connec-

tion k, for instance an estimate of the (relative) passenger flow, and ck = 0 for
headway buffer times.
The solution to the linear network synchronization problem gives an optimal
timetable for punctual operation. However, with respect to arrival delays the
resulting timetable may be far from optimal. The cost function (19) tends to
give tight transfer times for high priority connections. In case of arrival delays
this implies that either the connecting train is frequently delayed as well or, in
the case of transfer connections, large amount of transferring passengers miss
their connection. The linear cost function for all buffer times should therefore be
considered only if the arrival times have a large punctuality percentage.
However, for dwell buffer times xk the linear cost function Fk (xk ) = ck xk is
convenient as it minimizes passenger travel times and train circulation times.
Note that train running times include running time margins to neutralize the
stochastic driving behaviour. Dwell buffer times also compensate arrival delays
but this is just a (positive) side-effect. In principle the dwell times should be as
small as possible and possible variation in running times should be compensated
by running time margins.
An interesting observation is that the linear network synchronization problem
(where F (x) = cT x) has integral optimal buffer times as a result of a special
property of the cycle matrix and the node-arc incidence matrix (total unimodu-
larity).

Theorem 2 Consider the (linear) network synchronization problem (18) with

linear cost function (19). If all parameters T , bk , x̄k (k = 1, . . . , m) are integral
then the linear network synchronization problem has integral solutions, i.e., the
optimal buffer times and corresponding event times are also integers.

Proof. Note that b = Γl is integral if l is, since then b is just a linear combination

TRAIL Research School, December 1999 21

 of integral vectors. Moreover, since z ∈ ZZν and T ∈ ZZ also zT ∈ ZZν . From inte-
ger programming theory it is known that for a totally unimodular (TU) matrix A
the solution (if any) of the linear programming problem {min x|Ax ≤ b, x ∈ IRm +}
is integral [21]. So in particular the relaxation problem of an MIP has integral
solutions, and thus also the MIP itself. Now, the cycle time is TU which can
be proven from its construction of which all operations are closed under total
unimodularity. Note that the proof also holds for the network synchronization
problem with time window constraints. The matrix in this case is just the trans-
pose of the incidence matrix which is totaly unimodular. 2
Note that it is convenient that all arrival and departure times in a timetable are
given in minutes (integers). Theorem 2 then states that this is automatically
obtained by the optimal solution of the linear network synchronization problem.

4.3 Reliable Connections

A major cause of (secundary) delays is represented by crew connections in which

a driver or conductors have to change trains, and by rolling-stock connections
for, e.g., coupling of coaches. These connections are hard, that is, a connecting
service can only depart after the processes of the connection have been completed.
For these connections the reliability of the buffer time between the arriving and
departing services is hence very important. A third category to which buffer
time reliability is of major concern is the stability of train circulations. The
desire of a punctual initial departure time of a new train circulation has to be
assured by a buffer time in the turn-around time at the terminating station.
Note that running time margins and dwell buffer times during the circulation
also compensate for possible delays at the cost of increased (passenger) travel
times. A relative large buffer in the turn-around time does not influence travel
times but assures a punctual departure at the start of a new circulation. In
general, a reliable connection has to be assured for logistics constraints.
A measure for the reliability of a buffer time is the expected departure delay of the
connecting service due to failure of the buffer time. The expected departure delay
is formally given as follows. Consider a connection k = (i, j) at a railway station
from a train service i to a subsequent train service j. Let tri be the scheduled
running time of service i from the preceding station, and xk ≥ 0 be the connection
buffer time. Furthermore, let Ti be a random variable denoting the stochastic
running time of service
R
i, with density function gi and associated distribution
function Gi (t) = 0t gi (τ )dτ , see Figure 4. Common distributions are the shifted
Weibull, lognormal, Erlang, or Gamma distributions, but also empirical density
functions may be used. Denote the stochastic departure delay of train service
j as Yj . Then the probability that train j is delayed is P(Yj > 0) = P(Ti >
tri + xk ). Thus, the expected departure delay is E[Yj ] = E[(Ti − tri − xk )+ ] with
(·)+ = max(0, ·), that is
Z ∞
E[Yj ] = (τ − tri − xk )gi (τ )dτ
tr +xk
Z i∞
= (p − xk )+ gi(p + tri )dp.
0

22 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 g i (t )

buffer time
failure risk

0 tir tir + xk t

Figure 4: Buffer time failure of running time density function

Note that the integration variable p corresponds to the arrival delay of service i.
So instead of the running time distribution, also the distribution of the induced
stochastic arrival delay Pi of service i is sufficient. Note that the probability that
train service i arrives on-time is given as P(Ti ≤ tri ) = Gi (tri ). So typically, a
random arrival delay Pi has a density function
(
Gi (tri ) if p = 0
fi (p) = r
gi (p + ti ) if p > 0,

where gi , Gi , and tri are interpreted as above.

Above we silently assumed that the arrival delay of a feeder train is the result
of a stochastic running time only. Thus, a possible departure delay at the start
of the train run is neglected. The underlying idea is that at each connection the
buffer time is computed to cope with the primary delays as good as possible.
But this also implies that the secondary delays are controlled as good as possible.
Therefore, this approach can be interpreted as the minimization of a first-order
approximation of the delay propagation with respect to primary delays.
Instead of looking at the expected departure delay we can more generally look at
the expectation that an event is δ minutes late, with δ ≥ 0. The probability that
train j departs δ minutes late is given as P(Yj ≥ δ) = P(Ti ≥ tri + xk + δ) and
the expectation of a departure delay more than δ minutes is E[(Ti − tri − xk − δ)],
i.e.,
Z ∞
E[Yj − δ] = (τ − tri − xk − δ)gi (τ )dτ
tr +xk +δ
Z i∞
= (p − xk − δ)+ gi (p + tri )dp.
0

The expected departure delay is now a special case for δ = 0. The following
theorem states that the expectation E[Yj − δ] = E[Pi − xk − δ] is convex in the
buffer time xk for all δ ≥ 0.

TRAIL Research School, December 1999 23

 Theorem 3 Let g : [0, ∞) → [0, ∞) be a density function, and t, δ ∈ [0, ∞).
Then Z ∞
(p − x − δ)+ g(p + t)dp (20)
0

is a convex function of x on the domain [0, ∞).

Proof. Recall that max(0, x) is a convex function. Denote the integral function
(20) as h(x). Let α ∈ [0, 1]. Then we have to prove that h(αx1 + (1 − α)x2 ) ≤
αh(x1 ) + (1 − α)h(x2 ). We have
Z ∞
h(αx1 + (1 − α)x2 )) = (p − αx1 − (1 − α)x2 − δ)+ g(p + t)dp
0
Z ∞
= (α(p − x1 − δ) + (1 − α)(p − x2 − δ))+ g(p + t)dp
0
Z ∞
≤ [α(p − x1 − δ)+ + (1 − α)(p − x2 − δ)+ ]g(p + t)dp
0
Z ∞ Z ∞
= α (p − x1 − δ)+ g(p + t)dp + (1 − α) (p − x2 − δ)+ g(p + t)dp
0 0
= αh(x1 ) + (1 − α)h(x2 ).

2
This cost function can also be applied to optimize running time margins. So let
the running time also consist of a constant minimum running time and a variable
running time margin that has to be determined in the optimization. Then the
running time margins are also decision variables in the optimization problem.
Now the running time margin and the dwell buffer time (or turn-around buffer
time, etc.) are jointly optimized. An appropriate cost function for the running
time margin is then for instance the expectation of arriving 2 min late, and for
the dwell buffer time just the buffer time. Then the optimization problem aims
at optimizing the trade-off between a reliable running time (margin) and a small
dwell (buffer) time. Note that more flexibility of the process times represented
by the variable (buffer time/margin) part, also reduces the chance of infeasible
constraints.
A further extension of the optimization model is to include knock-on delays along
the route of a train. Carey and Kwieciński proved that the resulting expected
delays along the route are convex in the successive buffer times/margins. How-
ever, the evaluation of these cost functions may take a lot of computation time.
So here is a trade-off between a more accurate model and tractability of the op-
timization problem. A better approach may be to use these stochastic models in
a post-evaluation of the solution to the optimization model.

4.4 Transfer Connections

The effect of arrival delays on transfer reliability can be dealt with by incorpo-
rating the risk and significance of missing connections in the objective function.
This can be accomplished by considering the mean or expected transfer waiting
time. Note that a missed connection results in large waiting times. This section

24 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 derives a model for the expected transfer waiting time based on a modified ver-
sion of the approach in Knoppers and Muller [14]. Knoppers and Muller derive a
periodic function of the transfer waiting time, whereas here we derive a smooth
convex function of the transfer buffer time, see also Goverde [9].
Consider a connection k = (i, j) from a train service i to j. The transfer waiting
time is a function of the arrival delay pi of the feeder service i. If the feeder service
arrives on time or within the transfer buffer time xk then the transferring pas-
sengers have to wait for the scheduled departure time. For intermediate arrival
delays of the feeder service the connecting service j may wait to secure the con-
nection, and departs right after the arrival of the transferring passengers. If the
feeder service has an arrival delay for which securing the connection would result
in exceeding the maximum admissible train waiting time, the synchronization
control margin s̄j , then the connection is cancelled and the transferring passen-
gers miss the connection. The passengers then have to wait on the next train
of the connecting line. Assuming that the connecting train departs on-time, the
transfer waiting time can be given as


 xk − pi if 0 ≤ pi ≤ xk
wk (pi ) = xk + ωj − pi if xk + s̄j < pi ≤ xk + ωj (21)


0 otherwise,

see Figure 5. Here xk is the transfer buffer time of connection k = (i, j), ωj is the
interdeparture time (the reciprocal of the frequency) of the connecting service j,
and s̄j < ωj is the synchronization control margin of train j.
The above derivation of the transfer waiting time assumes that transferring pas-
sengers who miss the connection are able to catch the next train of the connecting
line and also do not leave the station in a train of an alternative line. Moreover
it is assumed that the connecting train and its successor depart on time. These
assumptions are reasonable from a scheduling point of view and result in a con-
siderable reduction of complexity.
To obtain an expression for the expected transfer waiting time the distribution
of the arrival delays is required. In the sequel it is assumed that early arrivals

wk

0 xk xk + s j xk + h j
pi

Figure 5: Transfer waiting time

TRAIL Research School, December 1999 25

 do not contribute to transfer waiting time costs, i.e., early arrivals are treated as
on-time arrivals.
Let gi be the density function
Rt
of the stochastic running time Ti of the feeder train
service i and Gi (t) = 0 gi (τ )dτ the corresponding distribution function. Let tri
be the scheduled running time. The probability that the train arrives on-time is
simply P(Ti ≤ tri ) = Gi (tri ). The delay density function is therefore
(
Gi (tri ) if p = 0
fi (p) = r
gi (p + ti ) if p > 0.

The following theorem gives a sufficient condition for which the resulting expected
transfer waiting time is a convex function.

Theorem 4 Consider a transfer connection k = (i, j). Let ωi > s̄j ≥ 0, the
function wk (p) defined as in (21), and P ≥ 0 be a stochastic variable with a
nonincreasing differentiable density function fi with a possible dicontinuity at
p = 0. Then the expectation
Z ∞
E[wk (p)] = xk P(p = 0) + wk (p)fi (p)dp
0+

is a convex function of xk ≥ 0.

Proof. Recall that 0+ = lim↓0 is necessary because of a possible discontinuity

of fi at 0. Let w̄k (xk ) = E[wk (P )]. Below it is shown by construction that the
second derivative of w̄k exists and is nonnegative which proves convexity. Let Fi
be the distribution function associated with fi . Note that fi is integrable as it is
a density function and so Fi exists. Moreover note that Fi (0) = P(p = 0). Using
partial integration the expectation of wk (P ) can be rewritten as
Z xk Z xk +hj
w̄k (xk ) = xk P(p = 0) + (xk − p)fi (p)dp + (xk + hj − p)fi (p)dp
0+ xk +s̄j
 Z xk 
= xk P(p = 0) + lim [(xk − p)Fi (p)]x k + Fi (p)dp
↓0 
Z xk +hj
+ [(xk + hj − p)Fi (p)]xxkk +h
+s̄j +
j
Fi (p)dp
xk +s̄j
Z xk
= xk P(p = 0) − xk Fi (0) + Fi (p)dp
0+
Z xk +hj
−(hj − s̄j )Fi (xk + s̄j ) + Fi (p)dp
xk +s̄j
Z xk Z xk +hj
= Fi (p)dp − (hj − s̄j )Fi (xk + s̄j ) + Fi (p)dp.
0+ xk +s̄j

The first derivative of w̄k can be computed using the fundamental theorem of
d Rx
calculus, i.e., dx a F (p)dp = F (x) for any fixed a ≤ x. Hence,

w̄k0 (xk ) = Fi (xk ) − (hj − s̄j )fi (xk + s̄j ) + Fi (xk + hj ) − Fi (xk + s̄j ).

26 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 15

expected transfer waiting time (min)

s ranging from 0 to 5
j

q = 0.7, h = 30, λ = 0.5

j j
10

sj =0

sj =5
0
0 5 10 15
transfer buffer time (min)

Figure 6: Expected transfer waiting time

By the assumption of fi being nonincreasing, we obtain for the second derivative

of w̄k

w̄k00(xk ) = fi (xk ) − fi (xk + s̄j ) − (hj − s̄j ) fi0 (xk + s̄j ) + fi (xk + hj )
| {z } | {z }| {z } | {z }
≥0 >0 ≤0 ≥0
≥ 0.

2
Typically, railway planners (unconsciously) schedule the running time after the
top of its density function by which most trains (e.g. 87%) arrive on-time, see
Figure 4. The expected transfer waiting time is then a convex function by The-
orem 4. Examples of convenient distributions are the Weibull and Erlang dis-
tribution, or any empirical density function with a decreasing tail, see Figure 4.
The expected transfer waiting time can be computed numerically. Goverde [9]
considers a negative-exponential density for the arrival delays with an additional
punctuality rate. This can be viewed as approximating the tail of the running
time density by a negative exponential density. For this case an analytical ex-
pression for the expected transfer waiting time can be computed [9], which is
convenient in the overall optimization problem.
Figure 6 shows a typical expected transfer waiting time function of the transfer
buffer time for some values of the synchronization control margin. Note that
the function is convex. For large buffer times (larger than 8 minutes) the influ-
ence of the synchronization control margin is negligible: any arrival delay can be
compensated by the buffer time. On the other hand, the effect of the synchroniza-
tion control margin is considerable for small buffer times. If the synchronization
control margin is large (about 5 minutes and larger) then the minimum of the
expected transfer waiting time is obtained for zero transfer buffer time. The syn-
chronization control can then compensate all (mainly occurring) arrival delays.

TRAIL Research School, December 1999 27

5 Conclusions

The computation of a periodic railway timetable can be modelled conveniently

as an optimization problem, where the buffer times are the decision variables
and the cost function is convex, by taking the network structure of the (periodic
event time window) constraints in account. The resulting MIP constraint system
of the buffer times is compatible with the DONS system of the NS, whereby the
optimization problem is suitable to extend the DONS features.
An appropriate convex cost function of the buffer times can be derived represent-
ing the individual required performance of a tight connection, a reliable connec-
tion, or a transfer connection, and evaluating the priorities between the individual
buffer time performance costs. The convex cost function guarantees that a local
optimal solution gives the global optimum as well.
Using the network structure of the constraints in the problem formulation signifi-
cantly reduces the required number of integer variables as opposed to the conven-
tional periodic event (time window) constraint system. The network structure is
also used in CADANS to solve the time window constraint system [25]. However,
modelling the network structure explicitly gives opportunities to restrict the feasi-
ble integer region by integer bounds, rolling stock constraints, and strong cutting
planes. This is advantageous for solving large-scale problems.
Another progress is made by using cost functions to penalize large buffer times at
stops, connections, and turns, making upper bounds on these buffer times super-
fluous. This considerably increases the opportunity to find feasible solutions, and
thus implicitly solves the problem of finding conflicting constraints. Moreover,
also running time margins can be included as decision variables in the optimiza-
tion problem which improves the feasibility issue as well as timetable efficiency
even more.
Current research concentrates on efficient solution algorithms to solve the pre-
sented optimization problems. Moreover, the optimization approach will be vali-
dated by a case study. Experiments with small networks give promising results.
The proposed optimization model contributes to design efficient and robust rail-
way timetables to improve performance and attractivity of railway operations,
increase service quality to passengers, and decrease costs to railway operators
and passengers due to delays.

Acknowledgements

This publication is a result of the research programme Seamless Multimodal Mo-

bility (SMM), carried out within the TRAIL Research School for Transport, In-
frastructure and Logistics, and financed by Delft University of Technology.

28 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization

 References
[1] Berge, C., Ghouila-Houri, A., Programming, Games and Transportation Networks,
Wiley, New York, 1965.

[2] Bookbinder, J.H., Désilets, A., “Transfer Optimization in a Transit Network”,

Transportation Science, Vol. 26, No. 2, pp. 106-118, 1992.

[3] Carey, M., “Reliability of Interconnected Scheduled Services”, European Journal

of Operational Research, Vol. 79, pp. 51-72, 1994.

[4] Carey, M., “Optimizing Scheduled Times, Allowing for Behavioural Response”,
Transportation Research B, Vol. 32, No. 5, pp. 329-342, 1998.

[5] Carey, M., “Ex Ante Heuristic Measures of Schedule Relibility”, Transportation
Research B, Vol. 33, pp. 473-494, 1999.

[6] Carey, M., Kwieciński, A., “Properties of Expected Costs and Performance Mea-
sures in Stochastic Models of Scheduled Transport”, European Journal of opera-
tional Research, Vol. 83, pp. 182-199, 1995.

[7] Domschke, W., “Schedule Synchronization for Public Transit Networks”, OR Spek-
trum, Vol. 11, pp. 17-24, 1989.

[8] Ferreira, L., Higgins, A., “Modelling Reliability of Train Arrival Times”, Journal
of Transportation Engineering, pp. 414-420, 1996.

[9] Goverde, R.M.P., “Optimal Transfer Times in Railway Timetables”, Paper pre-
sented at the 6th Meeting of the EURO Working Group on Transportation,
Gothenburg, September, 1998, to appear in Transportation Research B.

[10] Hallowell, S.F., Harker, P.T., “Predicting On-Time Line-Haul Performance in

Scheduled Railroad Operations”, Transportation Science, Vol. 30, No. 4, pp. 364-
378, 1996.

[11] Higgins, A., Kozan, E., “Modelling Train Delays in Urban Networks”, Transporta-
tion Science, Vol. 32, No. 4, pp. 346-357, 1998.

[12] Hooghiemstra, J.S., “An Automatic Timetable Generator for the Dutch Rail-
way Network”, In: Murthy, T.K.S., Mellitt, B., Brebbia, C.A. (eds.), Comput-
ers in Railways IV, Vol. 2, pp. 109-116, Computational Mechanics Publications,
Southampton, 1994.

[13] Klemt, W.-D., Stemme, W., “Schedule Synchronization for Public Transit Net-
works”, In: Daduna, J.R., Wren, A. (eds.), Computer-Aided Transit Scheduling,
Proceedings of the Fourth International Workshop on Computer-Aided Schedul-
ing of Public Transport, Lecture Notes in Economics and Mathematical Systems,
No. 308, pp. 327-335, Springer-Verlag, Berlin, 1988.

[14] Knoppers, P., Muller, Th., “Optimized Transfer Opportunities in Public Trans-
port”, Transportation Science, Vol. 29, No. 1, pp. 101-105, 1995.

TRAIL Research School, December 1999 29

 [15] Krista, M., Verfahren zur Fahrplanoptimierung dargestelt am Beispiel der Syn-
chronzeiten (in German), PhD Thesis, Schriftenreihe des Instituts für Eisenbah-
nwesen und Verkehrssicherung, Heft 56, Technische Universität Braunschweig,
Braunschweig, 1996.

[16] Krista, M., “Fahrplanoptimierung durch Minimierung der Synchronzeiten” (in

German), Signal + Draht, Vol. 91, No. 1+2, pp. 34-38, 1999.

[17] Meng, Y., Bemessung von Pufferzeiten in Anschlüssen von Reisezügen (in Ger-
man), PhD Thesis, Veröffentlichen des Verkehrswissenschaftlichen Instituts der
Rheinischen-Westfälischen Technischen Hochschule Aachen, Vol. 46, Aachen, 1991.

[18] Migom, A., Valaert, W., “The Optimization of a Rhytmic Inter-city Train Net-
work”, Rail International, Vol. 12, pp. 71-91, 1981.

[19] Nachtigall, K., “Periodic Network Optimization with Different Arc Frequencies”,
Discrete Applied Mathematics, Vol. 69, pp. 1-17, 1996.

[20] Nachtigall, K., Cutting Planes for a Polyhedron Associated with a Periodic
Network, Internal Report IB 112-96/17, Institut für Flugführung, Deutsche
Forschungsanstalt für Luft- und Raumfahrt (DLR), Braunschweig, 1997.

[21] Nemhauser, G.L., Wolsey, L.A., Integer and Combinatorial Optimization, Wiley,
New York, 1988.

[22] Odijk, M.A., “A Constraint Generation Algorithm for the Construction of Periodic
Railway Timetables”, Transportation Research B, Vol. 30, No. 6, pp. 455-464,
1996.

[23] Odijk, M.A., Railway Timetable Generation, PhD Thesis, Department of Technical
Mathematics and Informatics, Delft University of Technology, Delft Univeristy
Press, Delft, 1997.

[24] Peeters, L.W.P., “Computer Aided Timetabling for Railway Systems: Research
Proposal”, Proceedings of the 4th TRAIL Annual Congress 1998, Part 2, The
Netherlands TRAIL Research School, Delft, 1998.

[25] Schrijver, A., Steenbeek, A., Dienstregelingontwikkeling voor Railned: rapport

CADANS 1.0 (in Dutch), Report, CWI, Amsterdam, 1994.

[26] Serafini, P., Ukovich, W., “A Mathematical Model for Periodic Event Scheduling
Problems”, SIAM Journal on Discrete Mathematics, Vol. 2, No. 4, pp. 550-581,
1989.

[27] Strang, G., Linear Algebra and Its Applications, 3rd Edition, Harcourt Brace
Jovanovich, Orlando, 1988.

[28] Weigand, W., Graphentheoretisches Verfahren zur Fahrplangestaltung in Trans-

portnetzen unter berücksichtigung von Pufferzeiten mittels interaktiver Reken-
technik (in German), PhD Thesis, Technische Universität Braunschweig, Braun-
schweig, 1981.

[29] Weigand, W., “The Man-Machine Dialogue and Timetable Planning”, Rail Inter-
national, Vol. 14, pp. 8-25, 1983.

30 Improving Punctuality and Transfer Reliability by Railway Timetable Optimization