Fast Approaches to Improve the Robustness of a Railway Timetable

Author(s): Matteo Fischetti, Domenico Salvagnin and Arrigo Zanette

Source: Transportation Science , August 2009, Vol. 43, No. 3 (August 2009), pp. 321-335
Published by: INFORMS

Stable URL: https://www.jstor.org/stable/25769455

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/25769455?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Transportation
Science

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 TRANSPORTATION SCIENCE
Vol. 43, No. 3, August 2009, pp. 321-335
doi 10.1287/trsc.l090.0264
issn 0041-16551 eissn 1526-54471091430310321
?2009 INFORMS

Fast Approaches to Improve the Robustness of

a Railway Timetable
Matteo Fischetti
Department of Information Engineering, University of Padova, 1 35131-Padova, Italy,
[email protected]

Domenico Salvagnin, Arrigo Zanette

Department of Pure and Applied Mathematics, University of Padova, 1 35121-Padova, Italy
{[email protected], [email protected]}

The train
fies sometimetabling problem
operational constraints (TTP)some
and maximizes consists of finding
profit function a train
that accounts scheduleof on a railway network that satis
for the efficiency
the infrastructure usage. In practical cases, however, the maximization of the objective function is not enough,
and one calls for a robust solution that is capable of absorbing, as much as possible, delays/disturbances on
the network. In this paper we propose and computationally analyze four different methods to improve the
robustness of a given TTP solution for the aperiodic (noncyclic) case. The approaches combine linear program
ming (LP) and ad hoc stochastic programming/robust optimization techniques. We computationally compare
the effectiveness and practical applicability of the four techniques under investigation on real-world test cases
from the Italian railway company Trenitalia. The outcome is that two of the proposed techniques are very fast
and provide robust solutions of comparable quality with respect to the standard (but very time consuming)
stochastic programming approach.
Key words: timetabling; integer programming; robustness; stochastic programming; robust optimization
History: Received: December 2007; revision received: November 2008; accepted: January 2009. Published
online in Articles in Advance June 29, 2009.

1. Introduction described in Caprara, Fischetti, and Toth (2002). Our

The train timetabling problem (TTP) consists of find approach combines linear programming (LP) with
ing an effective train schedule on a given railway net stochastic programming (SP) and robust optimization
work. The schedule needs to satisfy some operational techniques.
constraints given by capacities of the network and We propose the following three-stage framework as
security measures. Moreover, it is required to exploit a practical tool for improving and testing robust solu
efficiently the resources of the railway infrastructure. tions for the TTP:
In practice, however, the maximization of some Stage (1) Nominal problem solution. The TTP is
objective function is not enough: The solution is also modeled without taking into account robustness and
required to be robust against delays/disturbances solved (not necessarily to optimality) by a standard
along the network. Very often, the robustness of MIP solver or by using ad hoc heuristics.
optimal solutions of the original problem turns out Stage (2) Robustness training. Borrowing an ex
not to be enough for their practical applicability, pression typical of the artificial intelligence field,
whereas easy-to-compute robust solutions tend to be starting from the previous stage solution the model
too conservative, and thus unnecessarily inefficient. is "trained to robustness," typically by exploiting a
As a result, practitioners call for a fast, yet accurate, restricted set of samples (scenarios).
method of finding the most robust timetable whose Stage (3) Robustness validation. The robustness of
efficiency is only slightly less than the theoretical opti the final solution found in stage 2 is evaluated by
mal one. This is typically obtained by adding "buffer using a validation tool, thus allowing for a fair com
times" to the schedule according to certain simple parison of different training methods.
rules (see ?2.2 in Kroon, Dekker, and Vromans 2007). In the present work, we focus mainly on the sec
The purpose of the present paper is to pro ond stage, robustness training. We assume nominal
pose and evaluate new methods to improve the solutions are given in input while, as far as the valida
robustness of a given TTP solution. In particular, tion stage is concerned, we use a simple LP validation
we address the aperiodic (noncyclic) TTP version model.
321

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
322 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

The paper is organized as follows. In ?2 the TTP observed empirically in Kroon, Dekker, and Vromans
literature is quickly reviewed. In ?3 we present the (2007). Very recently, a new notion of robustness,
TTP in detail and give a natural event-based MIP called recoverable robustness, has been proposed in
formulation. Section 4 is a short introduction to the Liebchen et al. (2007a), which integrates the notion of
SP paradigm and to the sampling techniques that robustness and recoverability into a common frame
are used in practice to implement it. In ??5 and 6 work. Applications to integrated timetabling/delay
we present the main building blocks of our solu management in railway systems are described and
tion framework. Extensive computational results are evaluated in Liebchen et al. (2007a, b) and Cicerone
given in ?7, showing that two of the new methods et al. (2007).
we propose are very fast and provide robust solutions The aperiodic TTP is especially relevant on heavy
of comparable quality with respect to the standard traffic, long-distance corridors, where the capacity of
(but very time consuming) stochastic programming the infrastructure is limited because of greater traffic
approach. Finally, some conclusions are drawn in ?8. densities, and competitive pressure among the train
An early version of the present work was pre operators is expected to increase in the near future. In
sented at ATMOS 2007 (Fischetti, Zanette, and the Caprara, Fischetti, and Toth (2002) setting, a train
Salvagnin 2007). operator applies for allocating its trains on the infras
tructure, and specify a profit for the "ideal timetable"
2. Literature Review for which they are asking. Then the infrastructure
manager collects all requests from train operators and
The TTP problem has two main variants: the peri
computes a feasible timetable maximizing the overall
odic and aperiodic versions. The periodic TTP's goal
profit, i.e., the difference between the profits of the
is to design a timetable that is operated cyclically
scheduled trains and a cost-penalty function, which
after a (small) period of time; this is a typical require takes into account the deviations of the final timeta
ment for passenger trains to come up with an easy
to-remember timetable. The first authors who devel bles with respect to the ideal ones (possibly leaving
some trains unscheduled).
oped a model for generating periodic timetables were
Serafini and Ukovich (1989), who introduced a math Different ILP models based on a graph represen
tation of the problem were presented in Caprara,
ematical model called the Periodic Event Scheduling
Problem (PESP). In PESP, a set of repetitive events is
Fischetti, and Toth (2002) and Caprara et al. (2006).
In these papers, the problem is modeled by means of
scheduled under periodic time-window constraints.
a directed acyclic multigraph, in which nodes corre
Consequently, the events are scheduled for one cycle
spond to arrival and departure events from the sta
in such a way that the cycle can be repeated. Most
tions and arise at some discretized time instances, and
models for periodic TTP are based on PESP. A nat
ural LP formulation of PESP is quite weak, due arcs correspond to train stops within a station or to
to a kind of big-M constraints (where M is the train trips. A Lagrangian relaxation method is used to
derive bounds on the optimal solution value, as well
period). An alternative stronger formulation is treated
as to drive a heuristic procedure.
in Nachtigall (1999), Liebchen and Peeters (2002), and
Peeters (2003), among others, and is based on cycle
bases and separation of valid inequalities. 3. The Nominal Model
As to robustness, Kroon, Dekker, and Vromans In this section we describe the specific aperiodic TTP
(2007) describe a stochastic optimization variant of problem we consider, and give a basic event-based
PESP. Their model explicitly takes into account formulation for the "nominal" version where robust
stochastic disturbances of the railway processes, dis ness is not taken into account.
tinguishing between a planned timetable and sev Following Caprara, Fischetti, and Toth (2002), the
eral realizations of the timetable under predetermined aperiodic TTP can be described as follows: Given a
stochastic disturbances. The model can be used to railway network, described as a set of stations con
allocate time supplements and buffer times to the nected by tracks, and an ideal train timetable, find
processes in the planned timetable in such a way an actual train schedule satisfying all the operational
that the average delay of the realizations of the constraints and having a minimum distance from the
trains is minimized. To keep the computation times ideal timetable.
within an acceptable bound, they start with an exist The entities involved in modelling the problem are
ing timetable and fix the precedences among trains. as follows.
They show that a substantial increase in robustness Railway network: A graph N = {3>, 2), where Sf is
can be achieved by taking into account stochastic dis the set of stations and EE is the set of tracks connecting
turbances in the design of the timetable. For the case them.
of one trip serving 10 stations, Liebchen and Stiller Trains: A train corresponds to a simple path on the
(2006) provide a theoretical explanation for the effects railway network N. The set of trains is denoted by T.

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 323

For each train h eT we have an ideal profit irh (the presentation we consider the situation where one
profit of the train if scheduled exactly as in the ideal is required to schedule all the trains. A natural
timetable), a stretch penalty 0h (the train stretch being event-based model in the spirit of the periodic event
defined as the difference between the running times in scheduling problem (PESP) formulation used in the
the actual and ideal timetables), and a shift penalty ah periodic (cyclic) case (Serafini and Ukovich 1989) can
(the train shift being defined as the absolute difference be sketched as follows:
between the departures from the first station in the
actual and ideal timetables). z*=maxheT?p*C
Events: Arrivals and departures of the trains at the
stations. The set of all the events is denoted by E. V/zgT, / = l,...,len(/z)-l (1)
With a small abuse of notation, we will denote by t\
both the zth event of train h and its associated time. |ff-f*|>Afl Vff, r*eAs, VSe^ (2)
We also define
?A: set of all arrival events |t?-*jl>Ad Vtlt*eDs,VSe3> (3)
?D: set of all departure events, ^<r^ O t\<t)
whereas As, Ds, and Es denote the restriction of the
above sets to a particular station S. Each train h is Vff, t)eDs, VSe^ (4)
associated with an ordered sequence of length len(h) Ph = irh-crh\t\-t\\
of departure/arrival events t\ such that tf+1 > t\, the
first and last event of train h being denoted by t\ ~ ?h ((*len(/z) ~ *1 ) ~ (t\en(h) ~ *1 ))
and tfen(hy respectively. In addition, let ff denote the V/zgT (5)
ideal time for event t\.
(Partial) schedule: A time assignment to all the
l<t<u WteE. (6)
events associated with a subset of trains.
Constraints (1) impose a minimum time differ
Objective: Maximize the overall profit of the ence di i+1 between two consecutive events of the
scheduled trains, the profit of train h being com
same train, thus imposing minimum trip duration
puted as
(trains are always supposed to travel at the maximum
ph = 7Th- ah shifth- 6h stretchy allowed speed for the track) and minimum stop time
at the stations.
where Constraints (2)-(3) model the headway times
shifth^\t\-t\\ between two consecutive arrival or departure events
and in the same station (Ad and Afl being the minimum
departure and arrival headway, respectively). Because
stretch, = (ttn(h) - t?) - (fLw - *?) these constraints are nonlinear and we do not know
denote the shift and stretch associated with train h, in advance the order in which events occur at the sta
respectively. Trains with negative profit are intended tions, in our MIP model we introduce a set of binary
to remain unscheduled and do not contribute to the
variables x^'J to be set to 1 iff t\ < along with a
overall profit. big-M coefficient M, so that conditions
Operational constraints include:
Time windows: It is possible to shift an event from |t?-t?l>A
its ideal time only within a given time window.
Headway times: For safety reasons, a minimum can be translated to
time distance between two consecutive arrival/depar
ture events from the same station is imposed. f?-fj>A-Mxf;/
Track capacities: Overtaking between trains is al
lowed only within stations (assumed of infinite t*-t*>A-MxfJ
capacity). h,k i_ y.k, h_-i
As already mentioned, in the present paper we do
not address the solution of the nominal TTP problem Given the linearization of constraints (2)-(3), it is
explicitly, in that a nominal solution is assumed to be easy to translate the track capacity constraints (4) as
provided in input. Nevertheless, we next outline the
structure of an MIP formulation for the nominal TTP
problem, because a relaxed version of it is at the basis
of the LP models used in ??5 and 6. Constraints (5) define the profits of the trains,
Although in the nominal problem one is allowed whereas constraints (6) model the user-defined time
to leave some trains unscheduled, to simplify our windows of each event.

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
324 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

It is important to notice that, although we are inter the second-stage recourse problem corresponding to
ested in integer values (minutes) for the events to first-stage decision x and parameters ((o).
be published in the final timetable, we do not force If O contains a finite number of scenarios {(o1/(o2/
the integrality on variables tj. This has the important ...,ce>|0|} with associated probabilities pk, k e 1,
consequence that after fixing the event precedence 2.. .., |ft|, then the expectation can be evaluated as
variables x, the model becomes a plain linear model. a finite sum, and the two-stage model (with linear
On the other hand, the possible fractional value of the recourse) becomes a standard linear model:
final time variables t need to be handled somehow
in a postprocessing phase to be applied before pub
lishing the timetable. For arrival events, one can just
round the corresponding fractional times to the near
w^mm^x+J^p^r^ rkeYk, Vfc = l,...,|0|, (7)
est integer because there is no problem if an arrival where rk are the recourse variables with linear costs qk,
arises a little earlier (or later) than published. An easy
and Yk is a polyhedron depending on the first-stage
procedure for departure times is instead to simply variables x.
round down all the f-values even if this results in a
Because is often very large, various sampling
slightly infeasible published timetable, so as to guar based approaches have been proposed to estimate
antee that all events arise no earlier than their pub the second-stage objective function. Interior sampling
lished time value. In a sense, this policy amounts methods use samples during the algorithm execution
to using an "infinite" time discretization during the to estimate, from time to time, algorithmic param
optimization phase, the difference between the actual eters such as function values, gradients, and opti
and the published event times being perceived by the mality cuts. Exterior sampling methods, instead, use
travelers as a small (less than one minute) delay. the sample average approximation (SAA) algorithm to
As far as the objective function is concerned, the estimate the optimal objective. We have chosen exte
nonlinear term
rior sampling because it has some advantages over
if?-*?i, interior sampling (Shapiro 2003): ease of numerical
giving the shift sh of train h, can be easily linearized as implementation, good theoretical convergence prop
erties (Verweij et al. 2003), well-developed statisti
cal inference (validation and error analysis, stopping
rules). Furthermore, it is easily amenable to variance
reduction techniques, ideal for parallel computations.

4.1. The Sample Average Approximation Method

4. The Stochastic Programming The SAA consists of approximating the mean of the
Paradigm random variable Q(x, ?(<*>)) with the sample mean of
Having stated the problem as an MIP, a well-known
[Q{x, ?k)), Q{x, l{a>2)),..., Q(x, ((oN))} independent
tool to find robust solutions is the (two-stage) stochas
and identically distributed (i.i.d.) samples from the
tic programming approach; see Birge and Louveaux distribution of Q(x,i;(a))). If Q(x/^(a))) has finite
(1997) and Ruszczynski and Shapiro (2003) for an mean and variance, the sample mean QN(x,g((oi)) =
introduction to the SP methodology. In SP, the set
(l/N)?fiiQ(*, *s an unbiased estimator of the
of constraints is decomposed in structural constraints, actual mean
which represent the deterministic part of the model,
and control constraints, which have a stochastic nature
E[QN(x,t(a>i))] = E[Q(x,Z(o>))],
and whose coefficients depend on the particular sce
nario. Roughly speaking, the approach allows one and it satisfies the following central limit theorem:
to take decisions in the first stage by ignoring the
stochastic part of the model, but enforces some costly Vn[qn(^^k))-e[Q(^^))]]
recourse action when indeterminacy will eventually
occur. Thus, a natural optimization objective for this
two-stage strategy is to minimize the total expected
cost where denotes convergence in distribution, and
mm{crx + E[Q(x,^))]}, 7V(0, a2) is a normal random variable with zero mean
xeX
and variance a2 = var Q(x, ?(<w)).
where x denotes the first-stage decision whose fea The SAA approximation of (7) reads:
sibility set is X, w g ft denotes a scenario that is
unknown when the first-stage decision x has to be
made, and Q{x, %{(o)) represents the optimal value of <=^{cTx+^T,tir*}> rkeYk,Vk = l,...,N. (8)
This content downloaded from
81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 325

Under mild assumptions it was proved that the opti numbers between 0 and 1, but typically they are
mal value of SAA problem (8) converges with proba treated as random variables.
bility 1 to w* (the optimal value of stochastic problem) When using random sampling, the outcome can be
as N tends to infinity (see Verweij et al. 2003). affected by a large variance, making it difficult to inter
Also, it is possible to use SAA to estimate the pret. Therefore, we decided to use the Latin Hyper
optimality gap by deriving lower and upper bounds cube (LH) variance reduction technique when sam
on w*. These bounds will be used to quantify the con pling from each distribution of P(l;)/ P[7 and 8h.
fidence intervals of the optimal solution of the stochas Other techniques such as, e.g., importance sampling
tic model (see ?7, Figure 9). Indeed, an upper bound (Clark 1961), can in principle fit our simulation set
(UB) on w* is ting as well, but they are quite involved. On the con
trary, LH sampling is of general applicability and
cTx + E[Q(i, f (a>))] = cTx + E[QN(x, f fa))], (9) comes with a straightforward implementation. The
standard approach to get a sample from a particular
where x is a given feasible, yet suboptimal, first-stage
probability distribution is to apply the inverse of the
decision vector. The expectation in the right-hand side
desired cumulative distribution function (CDF) to a
of (9), on its own, can be estimated as the mean of N'
sample drawn from a uniform distribution. The pro
N (say) independent SSA Q]N(x, ?(<*> )), obtaining
cess is then repeated until the required number of sam
ples N is collected. Using LH sampling instead, we
ub = 4 EQn(*-(io) first subdivide the [0,1] interval in N equal subin
tervals, and from each of them we draw a sample
from a uniform distribution spanning the subinterval.
o-?2 = varQN (*,?(?,))
Then the obtained sample vector is inverted through
the CDF and randomly permuted. For more theoret
= (N,_\)N, E<&<*'- ub). (id ical insights on LH sampling, the interested reader is
referred to Loh (1996).
It is easy to show (see Linderoth, Shapiro, and Wright LH sampling proved to be quite effective in our ap
2006) that a lower bound (LB) on w* is given by E[k;^]. plication. Figure 1 shows the reduction in variance o~
Again, we can compute this expectation by sampling: when sampling from an exponential distribution with
or without LH sampling. In our computational testing,

LB = ^IX' (12)
we observed an even larger variance reduction (one
order of magnitude or more).

5. Validation Model
^ = Var ^ = (N'-l)N' ;|(^ " ^ (13) An important component in our framework is robust
ness validation. Validation is often carried out inside
4.2. Sampling
the model itself, as is the case when an SP approach
Sampling of delays has been carried out using the
is used. However, we decided to implement an exter
following per-line model. A line EE is defined as a nal simulation-based validation module that is inde
sequence of stations operated by trains during the
pendent from the optimization model itself, so that it
24 hours. Each line section (the path between two con
can be of general applicability and allows one to com
secutive stations i and ;) can have a certain probability
pare solutions coming from different methods. The
P(i ;) to be affected by delay. Also, each time interval
module is required to simulate the reaction of the
[Z, k] in the 24-hour time horizon can have a certain
railway system to the occurrence of delays by intro
probability of delay, say P[//Jtj. Then each single train
h arrives at the last station with a cumulative random ducing small adjustments to the planned timetable
received as an input parameter. Validation is a major
delay Sh. The actual delay incurred by train h operat
topic on its own. Indeed, the set of actions the rail
ing on section (i, j) in time interval [Z, k] is computed
way system can make to react to disruptions is quite
using the following formula:
large?see, for example, Hofman et al. (2006)?and the
decision-making process is often complicated by strict
8lj)([l,k]) = 8hP[lk] P{Uj) , (14)
L{i,j)e%L(i,j) real-time requirements and complex business rules.
Validation can be carried out by optimization meth
where we normalize sections delay probabilities to ods, as proposed in Liebchen et al. (2007a, b) and
distribute the cumulative delay 8TP[l k] incurred by Cicerone et al. (2007). However, the complexity of such
train T operating on line EE through each line section models grows rapidly as soon as we allow complex
Note that P(l;) and P[hk] could be deterministic decisions to be made. Thus, simulation methods are

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
326 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

Without LHS With LHS

Figure 1 Reduction of Variance a with LH When Approximating, Through Sampling, the Exponential Probability Distribution Function (Solid Line)

often used to empirically measure the robustness of a value difj is updated by adding the (possibly zero)
given timetable?see, for example, Barber et al. (2008). extra time S? from the current scenario a). Weights re
For the purpose of the present paper, we decided to appearing in objective function (15) are related to
implement a simple LP-based validation tool based on the relative importance of the events, and typically
the following simplified assumptions: depend on the number of passengers affected.
?Limited adjustability in response to delays Constraints (17) are nonanticipatory constraints stat
with respect to the given timetable: In this paper, ing the obvious condition that one is not allowed
timetabling robustness is not concerned with major to anticipate any event with respect to its pub
disruptions (which have to be handled by the real-time lished value in the timetable. Because these values are
control system and require human intervention), but known, these constraints act as simple lower bounds
is a way to control delay propagation, i.e., a robust on the decision variables. Instead, we impose no upper
timetable has to favor delay compensation without bounds because we allow for an unlimited stretch of
heavy human action. As a consequence, at validation the timetable to recover from delays, i.e., a feasible
time no train cancellation is allowed, and event prece timetable is always achievable.
dences are fixed with respect to the planned timetable. The objective function is to minimize the "cumula
?Speed of validation: The validation tool should be tive delay" on the whole network.
able to quickly analyze the behavior of the timetable Given a feasible solution, the validation tool keeps
under many different scenarios. testing it against a large set of scenarios, one at a
Given these guidelines, we designed a validation
Single Delay time, gathering statistical information on the value of
model that analyzes a single delay scenario w at a the objective function and yielding a concise figure
time. Because all precedences are fixed according to (the average cumulative delay) of the robustness of the
the input solution to be evaluated, constraints (l)-(3) timetable.
all simplify to linear inequalities of the form:

6. Finding Robust Solutions

In this section we present three different approaches to
where dif] can be a minimum trip time, a minimum cope with robustness. We introduced two simplifying
rest, or headway time. We will denote with the set hypotheses: (1) all input trains have to be scheduled;
of ordered pairs (/, /) for which a constraint of type (2) all event precedences are fixed according to a given
(1) can be written. The problem of adjusting the given input "nominal" solution. These strong assumptions
timetable t under a certain delay scenario co can thus were made to obtain tractable (LP) models.
be rephrased as the following simple linear program
ming model with decision variables t": 6.1. A Fat Stochastic Model
Our first attempt to solve the robust version of the TTP
minjeET.w^-tj) (15) is to use a standard scenario-based SP formulation.
The model can be outlined as
(16)
min
t?>ti V/eE. (17)
Constraints (16) correspond to the linear inequalities
just explained, in which the nominal right-hand side
Y,ph>(l-a)z*
heT
(18)

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 327

t?-tf>ditj + S?fj V(i,j) &, V(oea (19) As to objective function, it involves a weighted
sum of the recourse variables. Finding meaningful
t?>ti VieE,Va>eSl (20)
values for weights wifj turns out to be very impor
ti-tjZd^ V(i,/)e3> (21) tant. Indeed, we will show in ?7 how to define these
l<t<u. (22) weights so as to produce solutions whose robustness is
comparable with that obtainable by solving the (much
The structure of the model is similar to that used more time consuming) fat model.
in the validation tool, but takes into account several
scenarios at the same time. Moreover, the nominal 6.3. Light Robustness
A different way to produce robust solutions is to use
timetable values tj are now viewed as first-stage deci
sion variables to be optimized?their optimal value the Light Robustness (LR) approach proposed recently
will define the final timetable to be published. The by Fischetti and Monaci (2008). This method is based
model is composed by the original one and a copy of it on the consideration that, roughly speaking, robust
with a modified right-hand side for each scenario. The ness is about putting enough slack on the constraints
original variables and the correspondent second-stage of the problem. In its simpler version, the LR counter
copies in each scenario are linked through nonantici part of the LP model
patory constraints.
min{cTx: Ax < b, x > 0}
The objective is to minimize the cumulative delay
over all events and scenarios. The original objec reads
tive function J^Ph *s taken into account through con
straint (18), where a > 0 is a trade-off parameter and min f(y) (29)
z* is the objective value of the reference solution.
For realistic instances and number of scenarios this Ax-y<b-p (30)
model becomes very time consuming (if not impossi cTx<{l + a)z* (31)
ble) to solve?hence, we called it "fat." On the other
hand, also in view of its similarity with the validation
x > 0 (32)
model, it plays the role of a kind of "perfect model" in 0<y</3, (33)
terms of achieved robustness; hence, it has been used
for benchmark purposes. where j8z is a parameter giving the des
level (or slack) on constraint i, and
6.2. A Slim Stochastic Model sion variable giving the corresponding u
Because the computing time required by the full The objective is to minimize a given fu
stochastic model is quite large, we present an alter y variables (typically, a linear or quadra
native model that is simpler, yet meaningful for our Moreover, (31) gives a bound (controll
problem. In particular, we propose the following efficiency loss because of the increased
recourse-based formulation: the solution, where z* is the value of th
solution.
min(i,j)e&,
? wifjs^
(oeCl
(23) In our TTP model, a typical constraint reads

2>?>(l-a)z* (24)
heT
and its LR counterpart is simply
ti-tj+sr.jZdi.j + S^ V(i,j)e9,V<oe(l (25)
ti-tj + y^zd^ + b^ %,;>0,
s?/>0 V(i,j)e?,V(oeil (26)
where Az ; is the required protection-level parameter.
ti-tj>difj V(i,/)e3> (27)
l<t<u. (28)
7. Computational Results
In this model we have just one copy of the origi We carried out tests on four single-line medium-size
nal variables, plus the recourse variables s?;, which TTP instances provided by the Italian railway com
account for the unabsorbed extra times 8?. with pany, Trenitalia. Data refers to unidirectional traffic on
respect to the minimum train trip times. It is worth different corridors.
noting that the above "slim" model is inherently An almost-optimal heuristic solution for each of
smaller than the fat one. Moreover, one can drop all these instances was computed by using the algorithm
the constraints of type (25) with 8? = 0, a situation described in Caprara, Fischetti, and Toth (2002). The
that occurs very frequently in practice because most algorithm is a Lagrangian heuristic based on the com
extra times in a given scenario are zero. putation of paths on a time-expanded network, whose

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
328 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

Table 1 Nominal Solution Characteristics buffer of one minute, we can absorb an exponentially
distributed disturbance of mean 1.44 minutes half of
Instance No. of No. of
Stations Scheduled trains the times.

27 127 As for the validation model, weights Wj appearing

BZVR
BrBO 48 68 in objective function (15) are assumed to be equal to 1,
MUVR 48 48 i.e., all events are considered equally important.
PDBO 17 33 The results are shown in Table 2, whereas graphical
representations are given in Figures 2 and 3.
According to the figures, slim2 always yields a very
computing time was in the order of minutes on a tight approximation of fat, whereas sliml is often
Pentium IV, 2.4 GHz PC. The corresponding solutions poorer. As for LR, it usually produces good results that
were used as the input nominal solutions to freeze the are only slightly worse than slim2, mainly in the most
event precedences and to select the trains to schedule. realistic cases where the trade-off parameter a is small.
Solution characteristics are given in Table 1. As for computing times (reported in Table 2), the fat
We implemented our framework in C++ and car model is one order of magnitude slower than sliml
ried out our tests on a AMD Athlon64 X2 4200+ com and slim2, although it uses only 50 scenarios instead of
puter with 4 GB of RAM. ILOG CPLEX 10.1 (ILOG Inc. 400. LR is much faster than any other method?more
2007) was used as MIP solver. than two orders of magnitude w.r.t the fat stochas
According to the sampling model described in ?4.2, tic models. Therefore, LR qualifies as the method of
we generated an extra time Sh((o) corresponding to choice for addressing large-scale real cases, because it
each train h and to each scenario a), drawing them guarantees good levels of robustness and requires very
from an exponential distribution with mean /i = 5%. short computing times.
For lack of more detailed data from the Italian rail
We also tried a variation of the slim2 (and LR) objec
ways operator about the actual distribution of delays tive function. The variation is motivated by observa
in line sections, we assume a proportional distribution tions in Kroon, Dekker, and Vromans (2007) about the
of delays along line segments. Accordingly, probabili optimal distribution of buffers on a single corridor.
ties P(z y) in (14) are proportional to the length of train There it was observed that buffers that are placed too
segments, barring a small additive white Gaussian early risk being left unused, because the probability
noise (standard deviation a = 0.01, i.e., a random of facing any delay at this early position is too small.
adjustment of l%-2%), and probabilities P[l k] are As a consequence, it might be worthwhile to lower
deterministically set to 1. the weights wifj arising in the early sections of the
Given this setting, the first test we performed was line. Figure 4 plots different parametric variants of the
aimed at comparing four different training methods slim2 objective function. All of them obey a common
for each reference solution, with different values of the formula, namely,
trade-off parameter a, namely, 1%, 5%, 10%, and 20%.
We compared the following alternative methods: {l-e-ki){\en(h)-i),
?fat: fat stochastic model (50 scenarios only)
?sliml: slim stochastic model with uniform objec parametrized in A (A = oo gives the original slim2
tive function?all weights equal (400 scenarios) weighing scheme). Table 3 reports the percentage
?slim2: slim stochastic model with enhanced objec improvements with respect to case A = oo for slim2
tive function (400 scenarios), where events arising ear and LR, respectively. It turns out that the new objec
lier in each train sequence receive a larger weight tive function typically produces slightly worse results
in the objective function. More specifically, if the zth for LR, whereas slim2 takes advantage of it for large
event of train h is followed by k events, its weight values of A. In any case, the improvement is not sub
in (23) is set to k + 1. The idea behind this weigh stantial (up to 3%-4%).
ing policy is that unabsorbed disturbances s? in a One might also wonder what the effect is of the
train sequence are likely to propagate to the next ones, input nominal solution to the subsequent robust
so the first ones in the line are the most important to ness improving phase. To answer this question, we
minimize. performed the following experiment. We took the
?LR: Light robustness model with objective func scheduled trains in the heuristic nominal solutions of
tion as in slim2 (using the sliml objective function Caprara, Fischetti, and Toth (2002) used in the previ
produces significantly worse results). Protection-level ous experiments, and we constructed the MIP model
parameters are set to A = ? \l In \, where /x is the mean described in ?3, where the choice of precedences is
of the exponential distribution. This is the protection left open. Then we collected a series of heuristic nom
level required to absorb a delay drawn from such a inal solutions of increasing efficiency for that model.
distribution with probability \. For example, setting a This was obtained by running the MIP solver with a

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 329

Table 2 Comparison of Different Training Methods with Respect to Computing Time, Percentage WAD, and Validation Function (Cumulative Delay
in Minutes), for Different Lines and Trade-Off a

Fat Sliml Slim2 LR

a (%) Line Delay WAD (%) Time (s) Delay WAD (%) Time (s) Delay WAD (%) Time (s) Delay WAD (%) Time (s)
0 BZVR 16,149 9,667 16,316 532 16,294 994 16,286 2.27
0 BrBO 12,156 384 12,238 128 12,214 173 12,216 0.49
0 MUVR 18,182 377 18,879 88 18,240 117 18,707 0.43
0 PDBO 3,141 257 3,144 52 3,139 63 3,137 0.25
Tot: 49,628 ? 10,685 50,577 800 49,887 1,347 50,346 3.44
1 BZVR 14,399 16.4 10,265 15,325 45 549 14,787 17 1,087 14,662 18 2.13
1 BrBO 11,423 21.6 351 11,646 42 134 11,472 21 156 11,499 23 0.48
1 MUVR 17,808 12.9 391 18,721 37 96 17,903 12 120 18,386 8 0.48
1 PDBO 2,907 15.6 250 3,026 51 57 2,954 11 60 2,954 13 0.27
Tot: 46,537 66.5 11,257 48,718 175 836 47,116 61 1,423 47,501 62 3.36
5 BZVR 11,345 15.9 9,003 12,663 48 601 11,588 19 982 12,220 22 1.99
5 BrBO 9,782 18.9 357 11,000 50 146 9,842 22 164. 10,021 23 0.51
5 MUVR 16,502 14.5 385 18,106 41 86 16,574 13 107 17,003 11 0.45
5 PDBO 2,412 14.7 223 2,610 44 49 2,508 20 57 2,521 19 0.28
Tot: 40,041 64 9,968 44,379 183 882 40,512 74 1,310 41,765 75 3.23
10 BZVR 9,142 21.4 9,650 10,862 50 596 9,469 24 979 10,532 33 2.01
10 BrBO 8,496 19.1 387 10,179 51 132 8,552 20 157 8,842 23 0.51
10 MUVR 15,153 14.7 343 17,163 49 84 15,315 15 114 15,710 13 0.43
10 PDBO 1,971 19.9 229 2,244 49 50 2,062 27 55 2,314 37 0.25
Tot: 34,762 75.1 10,609 40,448 199 862 35,398 86 1,305 37,398 106 3.2
20 BZVR 6,210 28.5 9,072 7,986 50 538 6,643 31 1,019 8,707 52 2.04
20 BrBO 6,664 22.1 375 8,672 53 127 6,763 23 153 7,410 30 0.52
20 MUVR 13,004 17.1 384 15,708 52 91 13,180 18 116 13,576 19 0.42
20 PDBO 1,357 28.4 230 1,653 49 55 1,486 34 60 1,736 53 0.28
Tot: 27,235 96.1 10,061 34,019 204811 28,072 106 1,348 31,429 1543.26
40 BZVR 3,389 35.4 10,486 4,707 50 578 3,931 37 998 5,241 51 2.31
40 BrBO 4,491 27.7 410 6,212 52 130 4,544 29 166 6,221 52 0.53
40 MUVR 10,289 21.8 376 13,613 52 95 10,592 25 108 11,479 34 0.45
40 PDBO 676 37.1 262 879 49 55 776 41 57 1,010 52 0.28
Tot: 18,845 122 11,534 25,411203 858 19,843 132 1,329 23,951 189 3.57

five-minute time limit and by storing all the incum (For all instances, the optimality gap after one hour
bent solutions produced during the run. Moreover, was fewer than 4%.) Then we compared the robust
we ran the solver with a one-hour time limit so as to ness achieved by our fat model when starting from
produce an almost optimal solution of value, say, z^{. these solutions, by allowing for a relative efficiency

Line BrBO Line PDBO

5 10 1 5 10
Efficiency loss (%) Efficiency loss

Figure 2 Comparison of Different Training M

Notes. On the x-axis there is the efficiency loss (a

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
330 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

xio4 LineMUVR

Efficiency loss (%)

xio4 LineBZVR
1.8 i-1-1-1

0 Len(/7)
/

Figure 4 Alternative Weighing Functions for slim2 and LR, Giving

Weight wri as a Function of Position / in the Line

solutions have a loss of efficiency ranging from 5.5%

to 0.4% with respect to zref. Rows correspond to the
different thresholds a used (1%, 5%, 10%, and 20%).
Efficiency loss (%) The table entries then give the percentage increase in
robustness (as measured by the validation tool) with
Figure 3 Comparison of Different Training Models Applied to the
Best Reference Solution forrespect
Each to robustness
Instance measured when starting from the
Notes. On the x-axis there
almost optimal solution of value zref. For example, for
is the efficiency loss (a), whereas the y-axis
reproduces the confidence BrBO, if we allow for a 10%
intervals of the validation figure efficiency loss with respect
(run with
500 scenarios). to zref and start from a noininal solution that is already
4.5% worse, we lose 13.9% in terms of robustness
achievable through the fat training method. Missing
loss a with respect to zref. The left-hand side part of entries correspond to infeasible cases.
Table 4 gives the outcome of the experiment for two As expected, starting from a worse nominal solu
instances (BrBO and MUVR). Columns correspond to tion reduces the degree of freedom in the subsequent
the 10 best solutions obtained within the five-minute
training phase, leading to a robustness loss. This nega
time limit, sorted from left to right by increasing tive effect could in principle be counterbalanced by the
efficiency. For example, for instance BrBO, the 10 different precedence structure of the solution, in the

Table 3 Percentage Robustness Improvement with Respect to A = oo for the Different Weighing Functions Plotted in
Figure 4
A
BZVR BrBO MUVR PDBO
a= 0.01 0.05 0.10 0.20 0.01 0.05 0.10 0.20 0.01 0.05 0.10 0.20 0.01 0.05 0.10 0.20
Slim2
3 -1.9 2.6 1.7 3.5 0 -2.5 -2.3 -0.1 -1.2 -0.8 -1.2 -5.3 -2.3 0.2 -0.1 -1.7
6 -0.7 2.6 1.6 3 0.4 -1.3 0.6 2.8 -0.8 -0.9 0.2 -1.5 -1.8 0.2 1 3.6
15 0 3.7 1.7 4.9 1.1 1.2 3.4 3.8 -0.6 -0.3 0 1 -0.3 0.7 1.8 1.4
30 0.4 3.6 0.1 3.5 0.8 1.8 2.2 3.7 0 0.2 0.3 0.7 0.3 -0.2 1.1 1.8
LR
3 -0.3 -0.2 1.1 -2.2 -0.1 0.2 -0.5 -0.5 0.1 -0.4 1.2 0.2 -0.7 -0.7 -2.2 0.2
6 -0.2 -0.8 2 -2 0.1 0.6 0.5 -0.7 0.4 0.2 -0.1 1.3 -1.5 -0.3 0.2 -1.7
15 -0.3 -0.5 1.7 -1.1 0.1 0.3 0.4 -0.5 -0.2 0.7 1.3 0.6 -1.3 0 -0.7 1.3
30 0.1 -0.1 1.2 -0.7 0.3 0.2 0 -0.7 -0.4 0.4 0.9 -0.8 -0.4 -0.8 -0.4 -0.9
Note. A negative value corresponds to a worse robustness.

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 331

Table 4 Effects of Nominal Input Solution on Robustness

BrBO

LR-MIP

Fat vs. Fat vs. LR

Eff(%)= -5.5 -4.5 -3.9 -2.7 -2.2 -1.7 -1.3 -1.2 -0.8 -0.4 0.0 0.0 0.0
a (%)
1 ________ _4.1 -2.7 0.0 -0.4 -0.1
5 ? -20.3 -18.2 -8.1 -7.4 -4.4 -2.8 -3.1 -1.6 -2.2 0.0 1.7 4.1
10 -23.9 -13.9 -15.2 -5.2 -5.6 -2.9 -1.9 -2.8 -1.4 -2.6 0.0 -1.0 4.3
20 -22.2 -11.9 -14.9 -4.6 -4.5 -3.1 -2.1 -2.8 -2.4 -3.0 0.0 -11.8 2.7
MUVR

LR-MIP

Fat vs. Fat vs. LR

Eff(%)= -27.4 -14.9 -9.9 -9.2 -7.6 -6.8 -2.7 -1.6 -1.6 -1.3 0.0 0.0 0.0
a (%)
1 __________ o.O -0.6 0.0
5 ______ _3.7 -1.7 -1.2 -1.7 0.0 -1.2 -0.1
10 ? ? -19.2 -16.5 -12.6 -10.1 -1.6 -0.8 -0.3 0.2 0.0 -1.4 1.1
20 ? -25.5 -13.1 -12.7 -9.1 -8.6 -2.1 -0.9 0.1 -0.8 0.0 -4.3 1.4

a of 10% with respect to zref, the MIP version of LR

sense that a less-efficient solution could involve prece
dence patterns, leading to improved robustness. How is able to improve by only 4.3% over the simple lin
ear LR. Furthermore, the second-last column of Table 4
ever, our experiments seem to indicate that the prece
reports, for comparison sake, the percentage differ
dence structure of the solutions plays only a secondary
role. This supports the viability of our approach, ence between the solution robustness obtained by the
MIP LR and the robustness obtained by using fat on
where only the most efficient nominal solution avail
able is "trained" for robustness. the same almost-optimal solution zref. Results show
that the new scheme produces only marginal robust
To better quantify the effect of fixing all precedences
when improving robustness of the nominal solution, ness improvements with respect to the simple linear
we performed a second experiment consisting of solv LR. This confirms that for the cases in our testbed,
ing the MIP version of the LR model where all prece the precedence structure of the solutions is not really
dences are left unfixed. Note that this is only viableimportant, efficiency being the key figure in determin
for LR, because the other models are too large to ingbethe maximum achievable robustness. However,
attacked by a general-purpose MIP solver. As in our may no longer be the case for more-complex net
this
previous experiments, we considered a loss of work effi topologies.
ciency a ranging from 1% to 20% with respect A to simple, yet often used in practice, policy to
the almost-optimal solution value zref. The solution enforce robustness in a timetable is to allocate a buffer
of value zref was also used to provide a first incum that is just proportional to the train duration. Figure 5
bent to the MIP solver. In these runs the MIP solver gives the results of this simple policy on a sample
performed quite well, in that the optimality gap afterinstance, where we first compute the maximum total
one hour of computing time was less than 1% for all amount of buffer we can allocate for a given efficiency
instances. (Note, however, that the model does loss, not and then distribute it proportionally along the
line. According to the figure, the proportional buffer
take into account the possibility of leaving some trains
unscheduled.) allocation policy and sliml behave quite similarly. This
The results of our second experiment are reported is not surprising, because model sliml actually favors
in the right-hand side part of Table 4. For PDBO and a proportional buffer allocation?this is confirmed in
BZVR, the MIP model did not find any better solution the other instances as well (not shown in the figure).
than the incumbent, so these cases are not reported On inthe other hand, much better results are obtained by
the table. The last column of Table 4 reports the perapplying more clever optimization methods, showing
centage robustness improvement of the MIP LR model the practical relevance of the optimization approaches.
described above, over the linear LR model described Although the validation output gives a reliable mea
in ?6.3. For example, for case BrBO with a thresholdof how robust a solution is against delays, other
sure

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
332 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

x104 LineMUVR Line BrBO

2.0 p?i-1-1 60 I-=f-1-1?

Efficiency loss (%)

Figure 5 Comparison of a Simple "Proportional" Buffer Allocation

Strategy Against the Proposed Methods
Note. The percentages shown are the total amount of buffer it was possible
to allocate within a given trade-off.

figures exist that somehow summarize the "static"

structure of a solution. These figures are useful to get Efficiency loss (%)
insights into the structure of the solutions obtained
with different training methods. In particular, we used
Figure 6 Comparison of Different Training Models from the WA
the weighted average distance (WAD) of the Point allocated
of View (WAD Is Given Within Its Confidence
Intervals)
buffer from the starting point. The WAD of the single
train h is calculated as
WAD of an entire line is calculated as the mean of
1 \en(h)-l (fh ,ih/9 all the WADs of the trains of the line. The reader is
WAD, - 1 V s'*'+i^+i + *'J/z cu\
2*1=1 Si,i+l 1=1 rlen(fc) Tl referred to Kroon, Dekker, and Vromans (2007) for a
more detailed discussion.
A to
where s, i+1 is the amount of buffer allocated from t{ comparison of the various WADs is reported in
ti+1. The WAD is a number between 0 and 1 that meaTable 2 and illustrated in Figures 6 and 7. It can be
seen that there is a significant correlation between the
sures how the buffers are distributed along the train
degree
trip. For example, a value of 0.5 means that the same of approximation of the various WADs with
amount of buffers were allocated in the first halfrespect
and to "perfect WAD" (WADfat) and the robust
ness of the solution?as computed by the validation
in the second half of the trip; values smaller or bigger
than 0.5 relate to a shift in buffers distribution towards
tool and reported in Figure 2 and 3. In Figures 6
the beginning or the end of the trip, respectively.and
The7, sliml WAD is almost always 50%, meaning a

Line MUVR Line BZVR

Efficiency loss (%) Efficiency loss (%)

Figure 7 Comparison of Different Training Models from the WAD Point of View (WAD Is Given Within Its Confidence Inter

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 333

iiniform allocation of buffers. On the other hand, the Again, note that the poorest method (sliml) produces
other methods tend to allocate buffers earlier in the an almost uniform distribution of the buffers, whereas
line, resulting in a lower value of the WAD. More the best ones tend to allocate them earlier. This con
over, as the allowed efficiency loss increases (x axis), firms the findings reported in Kroon, Dekker, and Vro
the WAD increases as well, meaning that uniform allo mans (2007).
cation becomes a good choice. We can also note that Finally, given the intrinsic approximation of the
LR behaves better for small efficiency losses. Indeed, stochastic methods because of the evaluation of the
LR uses a fixed buffer /3 to deal with disturbances. expectation, we have computed lower and upper
When the problem is less constrained in efficiency, bounds on the optimal solutions of the stochastic mod
these buffers can become too small, and the LP solver els, as described in ?4. A typical plot obtained for
will start to distribute the buffer excess in a some
the slim stochastic model is reported in Figure 9,
how unpredictable way, so as to meet the increased showing very narrow estimation gaps. Similar results
degree of freedom, thus degrading the performance are obtained with the other models, except that fat
of the method. For example, this is the case of lines behaves a little worse due the reduced number of
BZVR and PDBO. Moreover, BZVR and PDBO are scenarios.
more congested than the other two instances, which
also explains the better performance of the uniform
allocation strategy. 8. Conclusions and Future Work
Figure 8 reports how the buffers are distributed In this paper we have described a three-stage frame
along the line. The figure is obtained by normalizing work as a practical tool for building and testing robust
each line by the length of the corridor, and averag solutions for the train timetabling problem, we mainly
ing the buffers allocated in each normalized line sec focused on robustness improvement of a given nomi
tion. The averages are then normalized by the total nal solution. Robustness was then validated in terms
amount of allocated buffer, so that the area of each of the total cumulative delay, computed by solving an
chart approximately sums up to 1. For example, sliml LP model.
allocates buffers almost uniformly along the line?the We examined different robustness-improving
particular structure of the timetable being responsi models. The best-performing, in terms of validated
ble for local fluctuations. It is clear that slim2 pro cumulative delay, is a "fat" stochastic reformulation
duces a very tight approximation of fat, whereas sliml of the nominal TTP problem. However, the solution
does not. It is worth noting that LR uses a smoother of this model turned out to be very hard (if not
allocation of buffers, whereas sliml yields a better impossible) for practical instances. A "slim" version
approximation of their oscillations, but misses the performed much better/provided that a clever objec
global allocation policy. In this respect, slim2 per tive function was used. The fastest method, light
forms quite well instead. This is due to the fact that robustness (LR), proved to be quite accurate when
LR does not directly exploit the scenario informa dealing with a reasonable robustness-efficiency trade
tion; thus, it has to cope with very little information. off, allowing for a fast solution of large instances.
On the whole, light robustness qualifies as a suitable
Buffer allocation curves in MUVR (efficiency loss 10%)
4.0 r tool for addressing large-scale real scenarios, and
-Fat can even be embedded in the nominal solver to find
.Slim 1
3.5 h -Slim 2 optimized train-precedence patterns leading to more
- - LR robust timetables.
3.0 h Future direction of research should address the
important topics below.
In the present paper, we quantified (for the LR
model) the gain in terms of robustness resulting from
relaxing the requirement that all precedences in the
nominal solution must be preserved. It would be inter
esting to extend this analysis to the (much more diffi
cult to solve) slim2 model.
We performed our computations on real-world uni
directional corridors operated by the Italian railways
0 _,_
operator; it would be interesting to address more com
o 10 20 30 40 50 60 70 80
Trip length (%)
plex network topologies.
Finally, in our study we used a simplified LP-based
Figure 8 Comparison of Different Training Models from the validation tool to estimate the cumulative delay in
Allocated-Buffer Point of View a set of random scenarios. An interesting research

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
 Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
334 Transportation Science 43(3), pp. 321-335, ?2009 INFORMS

Line BrBO Line PDBO

12,000 2,500 r

E 10,000
03
<D
"D
CD 8,000

? 6,000
O

4,000 5 10 5 10
Efficiency loss (%) Efficiency loss (%)

x104 Line MuVR Line BZVR

5 10 5 10
Efficiency loss (%) Efficiency loss (%)

Figure 9 Confidence Intervals of Upper and Lower Bounds of

Caprara, A., M.
topic would be to measure theMonaci, P.actual
Toth, P. Guida. 2006. A Lagrangian requir
price
heuristic algorithm
recover a delayed timetable by using the for a real-world train timetabling problem.
same st
Discrete Appl. Math. 154(5) 738-753.
gies used in real-world delay management.
Cicerone, S., G. D'Angelo, G. Di Stefano, D. Frigioni, A. Navarra.
2007. On the interaction between robust timetable planning
Acknowledgments and delay management. Technical Report ARRIVAL-TR-0116,
This work was supported by ARRIVAL theproject. Future and Emerging
nologies unit of the EC (1ST
Clark, C. E.priority),
1961. Importance sampling inunder Contract
Monte Carlo analyses.
021235-2 (Project "ARRIVAL"), and by MIUR, Rome
Oper. Res. 9(5) 603-620.
(PRIN 2006 project "Models and algorithms for robust
Fischetti, M, M. Monaci. 2008. Light robustness. Technical Report
work optimization"). The authors thank
ARRIVAL-TR-0119, ARRIVAL project. Paolo Toth, A
Caprara, and Valentina Cacchiani for providing th
Fischetti, M., A. Zanette, D. Salvagnin. 2007. Fast approaches
nal TTP solutions used in to their
robust railwaycomputational
timetabling. C. Liebchen, R. K. Ahuja, study
also thank three anonymous referees
J. A. Mesa, for
eds. ATMOS 2007?7th their
Workshop on Algorith constr
comments. mic Approaches for Transportation Modeling, Optimization, and
Systems, Internationales Begegnungs- und Forschungszentrum
fur Informatik (IBFI), Schloss Dagstuhl, Germany.
References Hofman, M., L. Madsen, J. J. Groth, J. Clausen, J. Larsen. 2006.
Barber, R, S. Cicerone, D. Delling, G. D. Stefano, M. Fischetti, Robustness and recovery in train scheduling?A case study
from DSB S-tog a/s. R. Jacob, M. Muller-Hannemann, eds.
L. Kroon, D. Salvagnin, P. Tormos, C. Weber, A. Zanette.
2008. New frameworks for the interaction between robust ATMOS 2006?6th Workshop on Algorithmic Methods and Models
and online timetable planning, and for monitoring the status for Optimization of Railways, Internationales Begegnungs- und
quo of the system. Technical Report ARRIVAL-D3.4, ARRIVAL Forschungszentrum fur Informatik (IBFI), Schloss Dagstuhl,
project. Germany.
ILOG Inc. 2007. ILOG CPLEX 10.1 User's Manual.
Birge, J. R., F. Louveaux. 1997. Introduction to Stochastic Program
ming, Springer Series in Operations Research and Financial Engi Kroon, L., R. Dekker, M. Vromans. 2007. Cyclic railway timetabling:
neering, 1st ed. Springer, New York. A stochastic optimization approach. Algorithmic Methods
Caprara, A., M. Fischetti, P. Toth. 2002. Modeling and solving the for Railway Optimization, Lecture Notes in Computer Science.
train timetabling problem. Oper. Res. 50(5) 851-861. Springer, Berlin /Heidelberg, 41-66.

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms
Fischetti, Salvagnin, and Zanette: Fast Approaches to Improve the Robustness of a Railway Timetable
Transportation Science 43(3), pp. 321-335, ?2009 INFORMS 335

Liebchen, C, L. W. Peeters. 2002. On cyclic timetabling and cycles Nachtigall, K. 1999. Periodic network optimization and fixed inter
in graphs. Technical Report 761-2002, Mathematical Institute, val timetables. Habilitation thesis, Deutsches Zentrum fur Luft
TU Berlin. und Raumfahrt, Braunschweig, Germany.
Liebchen, C, S. Stiller. 2006. Delay resistant timetabling. Technical Peeters, L. W. P. 2003. Cyclic railway timetable optimization. Ph.D.
Report ARRIVAL-TR-0056, ARRIVAL project. thesis, Erasmus University, Rotterdam, The Netherlands.
Liebchen, C, M. Lubbecke, R. H. Mohring, S. Stiller. 2007a. Ruszczynski, A., A. Shapiro, eds. 2003. Stochastic Programming
Recoverable robustness. Technical Report ARRIVAL-TR-0066, (Handbooks in Operations Research and Management Series). Else
ARRTVAL-project. vier Publishing Company, Amsterdam.
Liebchen, C, M. Schachtebeck, A. Schobel, S. Stiller, A. Prigge. Serafini, P., W. Ukovich. 1989. A mathematical model for periodic
2007b. Computing delay resistant railway timetables. Technical scheduling problems. SIJDM: SIAM J. Discrete Math. 2 550-581.
Report ARRIVAL-TR-0071, ARRIVAL project, October. Shapiro, A. 2003. Monte Carlo sampling approach to stochastic
Linderoth, J., A. Shapiro, S. Wright. 2006. The empirical behavior programming. ESAIM: Proc. 13 65-73.
of sampling methods for stochastic programming. Ann. Oper. Verweij, B., S. Ahmed, A. J. Kleywegt, G. Nemhauser, A. Shapiro.
Res. 142(1) 215-241. 2003. The sample average approximation method applied to
Loh, W. L. 1996. On Latin hypercube sampling. Ann. Statist. 24(5) stochastic routing problems: A computational study. Comput.
2058-2080. Appl. Optim. 24(2-3) 289-333.

This content downloaded from

81.79.176.106 on Wed, 27 Mar 2024 20:15:23 +00:00
All use subject to https://about.jstor.org/terms