Transportation Research Part C: Xiaojie Luan, Francesco Corman, Lingyun Meng

Transportation Research Part C 80 (2017) 148–174
Contents lists available at ScienceDirect
Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc
Non-discriminatory train dispatching in a rail transport market

with multiple competing and collaborative train operating
companies
Xiaojie Luan a, Francesco Corman a, Lingyun Meng b,⇑
a
Section Transport Engineering and Logistics, Department of Maritime and Transport Technology, Faculty of Mechanical, Maritime and Materials Engineering,
Delft University of Technology, Mekelweg 2, 2628 CD Delft, The Netherlands
b
State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, No. 3 ShangYuanCun, HaiDian District, Beijing 100044, China
a r t i c l e i n f o a b s t r a c t
Article history: Train dispatching is vital for the punctuality of train services, which is critical for a train
Received 18 October 2016 operating company (TOC) to maintain its competitiveness. Due to the introduction of com-
Received in revised form 6 March 2017 petition in the railway transport market, the issue of discrimination is attracting more and
Accepted 20 April 2017
more attention. This paper focuses on delivering non-discriminatory train dispatching
Available online 5 May 2017
solutions while multiple TOCs are competing in a rail transport market, and investigating
impacting factors of the inequity of train dispatching solutions. A mixed integer linear pro-
Keywords:
gramming (MILP) model is first proposed, in which the inequity of competitors (i.e., trains
Train dispatching
Equity
and TOCs) is formalized by a set of constraints. In order to provide a more flexible frame-
Train Operating Company (TOC) work, a model is further reformulated where the inequity of competitors is formalized as
Mixed-integer linear programming the maximum individual deviation of competitors’ delay cost from average delay cost in
the objective function. Complex infrastructure capacity constraints are considered and
modelled through a big M-based approach. The proposed models are solved by a standard
MILP solver. A set of comprehensive experiments is conducted on a real-world dataset
adapted from the Dutch railway network to test the efficiency, effectiveness, and applica-
bility of the proposed models, as well as determine the trade-off between train delays and
delay equity.
Ó 2017 Elsevier Ltd. All rights reserved.
1. Introduction
Providing punctual and reliable services is a main goal of train operating companies (TOCs) in order to maintain and fur-
ther improve their competitiveness in the rapidly changing multimodal transport market. As tactical plans, train timetables
are typically computed offline, months before operations to specify a physical network route and detailed arrival and depar-
ture times at passing stations for each train. While a planned train timetable is put into operation, unavoidable stochastic
perturbations (e.g., bad weather, extra passenger flow, and capacity breakdowns) may influence the scheduled train running
and dwelling times, thus causing primary delays to normal train operations. Due to the high interdependency between trains
for the available capacity, primary delays may further result in snowball effect on other trains with consecutive delays in a
rail network. The key task of train dispatching is to take proper measures which can recover the impacted schedules from
⇑ Corresponding author.
E-mail addresses: x.luan@tudelft.nl (X. Luan), f.corman@tudelft.nl (F. Corman), lymeng@bjtu.edu.cn (L. Meng).
http://dx.doi.org/10.1016/j.trc.2017.04.011
0968-090X/Ó 2017 Elsevier Ltd. All rights reserved.
X. Luan et al. / Transportation Research Part C 80 (2017) 148–174 149
perturbations and further reduce potential negative consequences. Ineffective train dispatching could significantly down-
grade the punctuality of train services and the overall system performance.
Railways have developed as vertically-integrated (state-owned) organizations, which have been the most common struc-
ture for the rail sector in most countries, with responsibility for both the railway infrastructure facilities and train operations
(Kurosaki, 2008). Since the 1990s, rail policy regulations in Europe have fostered competition into the rail transport market.
This led to a vertical separation between infrastructure management and train operations, the progressive opening up to the
market for new operating companies, and the rules regarding the allocation of slots and the pricing of infrastructure use,
administered by an independent regulator (Nash and Rivera-Trujillo, 2004). Directive 91/440/EC (Commission of the
European Communities, 1991) is one of such policies, which forced separation of concerns in the railway transport field,
by specifying the roles of Infrastructure Manager (IM) and Railway Undertakings (RUs) or Train Operating Companies (TOCs).
The former is in charge of making infrastructure available for both tactical train timetabling and operational train dispatch-
ing, and the latter has economic interests to strive for increasing ridership. Such policies consider competition among TOCs
as a key element to achieve efficient operations. Nevertheless, situations of quasi-monopoly are common, which may result
in discriminatory treatment among different TOCs, in both tactical train timetabling and operational train dispatching. Sim-
ilar situations exist in China, where passenger trains are generally put in a high priority in using tracks than freight trains.
This is a rather standard allocation approach, but it seriously affects the interests of freight TOCs and downgrade the effi-
ciency of the whole system, particularly during perturbations. To protect the legitimate rights and interests of TOCs and keep
an orderly market, providing non-discriminatory access to rail infrastructure for TOCs is of great importance, in both plan-
ning and operational control levels.
The competitive interaction, concerning equity among multiple TOCs, has been studied so far mostly from a policy and
financing point of view. Those are offline issues addressed during design and strategic planning, including for instance
the equitable allocation of timetable slots. As requested in Directive 2001/14/EC (Commission of the European
Communities, 2001), the access to the rail infrastructure for all TOCs should be provided in a fair and non-discriminatory
manner. This requirement is reflected in the timetable planning process, which follows a sequence of applications of TOCs
for infrastructure capacity, scheduling the requested applications, coordination of the conflicting requests, (if conflicts still
exist, then) declaring the infrastructure congested, and employing non-discriminatory priority criteria to allocate the con-
gested infrastructure. However, the rules for access and use of the infrastructure during real-time traffic management mainly
focus on restoring the normal situation and do not require a special focus on non-discriminatory actions. Additionally, penal-
ties may be charged for the actions that disrupt the normal operation, compensation may be granted for the TOC which suf-
fer from disruption, and TOCs may be rewarded for better than planned performance. During online operations, the available
capacity can be reduced by delays and delay propagation, which may result in infeasibility of the planned train timetables.
TOCs only look at maximizing their interests and suffer from negative effects of delays, leading to passenger dissatisfaction,
refund, and penalties. The problem we have is then how to allocate this (reduced) capacity among competing TOCs without
favoring any of them, i.e., how to provide non-discriminatory access to the limited capacity for the competing TOCs. In fact,
few online (i.e., in relation with real operations) approaches are known to address this problem. Most existing studies on
train dispatching focus on minimizing the negative impacts of perturbations and pay little attention to discrimination (which
corresponds to delay inequity) among competing TOCs while generating dispatching solutions (see the review paper by
Cacchiani et al., 2014). This brings about the motivation of this paper, i.e., delivering non-discriminatory train dispatching
solutions in order to protect the rights and interests of TOCs during real-time train dispatching, and filling the research
gap in the literature.
In this paper, we focus on generating non-discriminatory train dispatching solutions (or achieving an acceptable degree of
equity while dispatching), and exploring the aspects related to delay equity. We address the problem of dispatching trains in
a non-discriminatory way: this means that we use an optimization approach to explicitly consider delay equity among mul-
tiple competing TOCs or trains, in addition to minimizing average (consecutive) train delay time. We consider delay equity as
the degree of homogeneity of the delays faced by different trains, or trains of different TOCs. An inequitable (or discrimina-
tory) situation occurs when some trains or some TOCs, face much larger delays than other trains or TOCs.
The remainder of this paper is organized as follows. Section 2 provides a detailed literature review on relevant studies,
e.g., network access and competition policies, equitable capacity allocation in train timetabling, real-time train dispatching
without considering delay equity, and equitable traffic control in other transport modes. In Section 3, mathematical models
are proposed, including a model (P1) representing equity in constraints, a model (P2) representing equity in objective func-
tion, a model (P3) without considering equity as a benchmark, and a model (P4) considering consecutive delay equity only.
Section 4 presents a detailed description of experimental settings, followed by the analyses of the experimental results in
Section 5, which quantify the trade-off between train delays and delay equity and the key determinants of delay equity.
Finally, conclusions and suggestions for further research are given in Section 6.
2. Literature review
This section presents a detailed review on relevant policies and scientific studies. We first investigate the policies con-
cerning network access and competition in the railway transport field in Section 2.1, followed by discussing the studies
on the equitable capacity allocation (offline train timetabling) in Section 2.2. Then, in Section 2.3, we review the studies
150 X. Luan et al. / Transportation Research Part C 80 (2017) 148–174
on the real-time train dispatching problem, where delay equity is not explicitly considered. We further focus on the studies
addressing equitable control problem in other transport modes, i.e., air traffic and road traffic, in Section 2.4. Those would
give us inspiration for our optimization approaches.
2.1. Network access and competition policies
Railways were once the dominant means of land transport, and have developed as vertically-integrated organizations,
typically a state-owned firm in most countries, with responsibility for both the railway infrastructure facilities and train
services (UNESCAP, 2003). Such state-owned (vertically-integrated) railways are not necessarily technically inefficient or
lacking in investment funds, and the provision of rail transport services is typically regarded as a classic example of nat-
ural monopoly (Kurosaki, 2008). In situations of monopoly status of state-owned railways, the issue of discrimination in
capacity allocation between different train categories or train companies is common and neglected by railway managers
and regulators (for instance in China as discussed in Wu and Nash, 2000). However, the state-owned railways can com-
pensate relatively easily financial deficit, and in general proved to have little incentive to be cost-effective or to respond
flexibly to changes in customers’ demand. Moreover, strategic political goals interfering with daily operations have some-
times led to hardly quantifiable performance targets and poor adaptation to customer demands. As a result, railways in
many countries have seen their market share decline sharply over the past thirty years. In order to maintain and further
improve the competitiveness of railways in the rapidly changing multi-modal transport market, many countries have
introduced reforms designed to improve the operational and financial performance of national railways, e.g., turning
loss-making state-owned railways into stand-alone companies operated on commercial principles (UNESCAP, 2003;
Kurosaki, 2008).
The European legislator considers deregulation and competition as effective ways to reduce cost, optimize use of the
infrastructure, and improve quality of public transport services. Directive 91/440/EC (Commission of the European
Communities, 1991) introduced separation of concerns between IM and TOCs. The IM holds a monopoly in the supply of
access to its network and has the duty of providing fair and non-discriminatory access to the available infrastructure capac-
ity, as required in Commission of the European Communities (2001). The TOCs are companies that compete to offer services
to customers. The access to the railway transport market differs by county, as investigated in Kurosaki (2008). In order to
gain the access to the market in Europe, a TOC must hold an operating license issued by the Member State where the
TOC is established (according to Directive 95/18/EC, Commission of the European Communities, 1995), and has to pay the
access charges for enabling IM to recover the costs of maintaining and renewing its infrastructure. A detailed discussion
of the legal environment of the European railway market can be found in Klabes (2010).
By introducing competition to the railway transport market, some discriminatory situations among competing TOCs have
unavoidably arisen, as analyzed in many policy studies (e.g., IBM Global Business Services, 2011; Van de Velde et al., 2012)
and considered as possible outcome in many other situations. Wong et al. (2002) discussed the lessons learned and the appli-
cability of such an approach to the Chinese high speed railway market. The review of the relevant policies (e.g., Commission
of the European Communities, 2001, 2013) points out the urgent need to decrease monopolies, which cause anti-competitive
practices and tangible advantages for incumbent operators (i.e., those companies deriving directly from the state-owned rail-
way companies operating as monopolists before the Directive 91/440/EC). In the Access Rights Policy (2016) provided by
Network Rail (the Infrastructure Manager of most of the rail network in England), discrimination is considered as an impor-
tant factor in making decisions. Current situations of discrimination in capacity allocation should be reduced, in both plan-
ning and operational control levels.
We report here the most relevant studies, which address competition from an online train dispatching point of view.
According to the study on the effects of competition in railway market (Lalive and Schmutzler, 2008), the quality of the train
operations in regional lines might not strongly depend on whether competing TOCs are operating them, rather than a single
TOC. In fact, delays caused by railway operators appear to be unrelated to the introduction of competition by lines tendered
to other TOCs. For those lines, there is no evidence that competition has had an adverse effect on service quality. On the other
hand, when TOCs compete for the same infrastructure capacity, competition plays a major role. Alexandersson et al. (2012)
reported that in Sweden, online capacity allocation between competing TOCs is a major issue. Online traffic control is com-
monly managed by priorities, i.e., some train categories are more highly ranked than others. When TOCs in the same train
category or having the same priority compete with each other, the rules of thumb currently used for manual handling will
not be sufficient, due to the intricate interdependencies between trains.
The traffic control based on priorities across trains of multiple classes is typically done neglecting equity and non-
discrimination. For instance, in China, trains are dispatched according to their classes, which are categorized by train speed
and customers served. The descending sequence of trains priority goes from High-Speed Electric Multiple Unites (EMU) train
(G), Intercity EMU train (C), EMU train (D), direct express train (Z), express train (T), fast train (K), accommodation train to
freight train. Such a priority-rule measure is quite inequitable for the trains with lower priority, and finally they are more
delayed. A relatively improved situation occurs in Canada, where Rail Traffic Control is advised prior to leaving equipment
(train) on a siding track (i.e., let the train on the siding track go first), as reported in Canadian Rail Operating Rules (2015).
Such a rule (a kind of First-In-First-Out rule, FIFO) neither explicitly considers equity in train operations, nor results in a dis-
criminatory situation (as the situation in China discussed above).
2.2. Equitable allocation of capacity in timetabling
A substantial amount of studies deals with offline capacity allocation, i.e., equitable allocation of resources among com-
petitors in the train timetabling stage. We next discuss the studies where equity is a concern while allocating capacity
offline.
An auction-based allocation mechanism for railway capacity was considered in many studies, in order to establish a fair
and non-discriminatory access to a railway network. In this setting, TOCs compete for the use of a shared railway infrastruc-
ture by placing bids for trains that they intend to run. Such a mechanism is desirable from an economic point of view,
because it can be argued that it leads to the most efficient use of the capacity. The main motivation and argumentation
of that idea can be found in Borndorfer et al. (2006). Harrod (2013) discussed the problem of pricing the train paths for ‘‘open
access” railway networks in the U.S. market. An approach based on bidding and auctioning for time slot allocation was
described, in which equity is related to the possibility of handling all railway traffic in a transparent manner. As stated by
the author, ‘‘Looking back at the history of the Interstate Commerce Commission in the United States, it would appear that
the long and arduous investigations of cost allocation was in essence a pursuit of fairness.”. Schlechte (2011) used the same
basic assumption that optimization approaches considering all stakeholders provide more equitable allocation than an incre-
mental or the current human assignment. The idea is that the competing TOCs can bid for any imaginable use of the infras-
tructure. Possible conflicts will be resolved in favor of the party with the higher willingness to pay.
Karsu and Morton (2015) reviewed the operational research literature on inequity averse optimization and focused on the
cases where there is a trade-off between efficiency and equity. The operational research approaches that incorporate equity
concerns alongside other concerns (mostly efficiency) were discussed in detail, for different problem types. Xu et al. (2014)
considered equity measure as the ratio between the maximum delay encountered by a train and the total planned time with-
out delays. Genetic algorithms were used to solve the resulting problem for a small artificial railway line. In the urban sub-
way traffic, Wu et al. (2015) proposed a timetable synchronization optimization model to equitably optimize passengers’
waiting time over all transfer stations, with an aim of improving the worst transfer by adjusting the departure, running,
and dwell times for all directions.
The approaches based on auctions and those based on scheduling are two common ways to allocate capacity with some
consideration of equity. Those latter appear to be more applicable in case a solution is required in a very short computation
time, as it is the case in the real-time train dispatching problem.
2.3. Real-time train dispatching without considering delay equity
The train dispatching problem has been well studied in the past decades; interested readers might refer to the surveys in
Pacciarelli (2013), Corman and Meng (2014), Cacchiani et al. (2014), and Fang et al. (2015). We give a brief review on recent
studies, particularly focusing on train dispatching models with multiple (conflicting) objectives.
A sequential optimization setup was used in Corman et al. (2011a), which divided trains in different categories and then
found the optimal schedule for each class, given the division in classes. This was inspired by the priority scheduling of prac-
tice, when the lower ranked classes need to be scheduled in such a way to not decrease the performance for the higher
ranked classes. Meng and Zhou (2014) investigated the benefits of simultaneous train rerouting and rescheduling compared
to sequential approaches in general rail networks. Network-wide cumulative flow variables were used to implicitly model
capacity constraints, which enabled an easy problem decomposition mechanism. The decomposed sub-problems were then
solved by an adapted time-dependent least cost path algorithm.
As equity and efficiency of schedules are often conflicting objectives, we then review papers that address multiple objec-
tives. Ghoseiri et al. (2004) addressed the problem of finding the Pareto optimal solutions concerning fuel consumption and
passenger travel time. Mazzarello and Ottaviani (2007) introduced a Traffic Management System (TMS) for real-time traffic
control, by applying an alternative graph based model with objectives of improving train punctuality and saving energy.
Corman et al. (2009) considered two conflicting objectives of energy efficiency and delay minimization, in which the first
objective was used as a hard constraint. By using green-wave operations, trains face no unnecessary/unplanned decelera-
tions or stop, while delays might increase slightly. Another kind of bi-criteria optimization models arises when delays of
trains and passengers’ satisfaction are jointly considered. This issue is dealt in a unified manner by so-called delay manage-
ment approaches. Schachtebeck and Schöbel (2010) added the priority decisions to the integer programming formulation of
the delay management problem and were hence able to deal with the capacitated case. In their model, connections were
decided to be maintained or dropped by minimizing the number of missed connections, while minimizing the sum of all
delays of all events. Hence, the objective function was an approximation of the sum of all delays over all passengers.
Corman et al. (2012) instead considered explicitly the two objectives. They defined a heuristic search which was able to
determine the Pareto frontier of the bi-objective problem of reducing delays and keeping as many passengers’ connections
as possible, by repeatedly solving a real-time train dispatching problem with side constraints. Fuchsberger (2012) addressed
an equivalent problem by a time-indexed formulation. In his approach, the conflicting objectives of reliability (scheduling all
trains), punctuality (reducing delays) and keeping as many passengers’ connections as possible were treated together as a
weighted sum. Albrecht et al. (2013) proposed a meta-heuristic algorithm so-called Problem Space Search (PSS) for simul-
taneous train rescheduling and track maintenance. The considered objectives were to minimize both train delays and main-
tenance delays. Boccia et al. (2013) described two heuristic approaches to solve the problem published by the Institute for
Operations Research and the Management Sciences (INFORMS) Railway Applications Section (RAS) in 2012, based on a mixed
integer linear programming formulation. The multiple objectives were weighted sum of the minimum train delays, devia-
tions from the required timetable, and train travel times on the unpreferred tracks.
Table 1 summarizes the studies on train dispatching problem discussed above, classified by the objective function, math-
ematical formulation (model structure), and solution algorithm. As listed in Table 1, the equity issue is not explicitly consid-
ered in most existing studies.
2.4. Equitable control of air traffic and road traffic
Pellegrini and Rodriguez (2013) analyzed in detail the similarities between railway and air transport modes in the critical
battle for improving efficiency. The key issues are the strategic interaction of competitors for capacity allocation and the dif-
ficulty of the real-time control. Air traffic controllers are in charge of movement safety on air segments, while railway dis-
patchers are controlling traffic in a saturated infrastructure. For both situations, safety critical tasks are fulfilled by a safety
system or mechanism. Only quality of the traffic control is at stake, and the worst consequence is a large scale delay prop-
agation. A stream of studies focuses on the current fairness concept in air traffic control. The FCFS (or FIFO, or Ration-by-
delay) rule gives relatively equitable decisions, while better operations could be achieved if delays are to be spread equally
over as many operations as possible.
Manley and Sherry (2010) introduced a number of metrics concerning the interaction of passenger delay, fuel burn, and
equity. Current regulations achieve high equity at the cost of a reduced throughput; equity and delay are in general conflict-
ing objectives.
In Vossen et al. (2003), two problems were solved in cascade, first the unconstrained problem of finding an equitable allo-
cation, and then improving its performance with a limited deviation from the equitable allocation determined before. A fol-
low up work of Glover and Ball (2013) introduced stochasticity in the model, to find solutions that achieve higher levels of
equity. Kuhn (2013) addressed performance and equity as once, determining efficiently the Pareto front for those two con-
flicting objectives.
Zhong (2012) also considered equity in air traffic operations, defining a bi-criteria optimization model to offload demand
from a congested airspace. The Pareto frontier of efficiency and equity was generated to allow decision makers identifying
the best trade-off solutions, based on a system view. Equity was considered as a set of additional side constraints. Lagrangian
relaxation was further used to relax those latter constraints, yielding a decomposition in a series of single-flight scheduling
problems.
Equity in air traffic operations was also considered by Del Pozo y de Poza et al. (2009), providing definitions and metrics
for equitable air traffic control, combining geometric and arithmetic mean of the delay of the different operations. Kim and
Hansen (2013) investigated role of sharing information in achieving equitable and collaborative resource allocation for air
traffic flow control. A model was proposed by considering public and private information. Sharing such private information
can achieve a clear benefit in terms of efficiency.
Hoffman and Davidson (2003) pointed out that equity is a prerequisite for achieving efficient management of distur-
bances. Equity is achieved when the welfare of each user of the air traffic network is increased to the maximum extent pos-
sible, given limited resources, after taking proper account of individual claims and circumstances. They also used a two-stage
approach, similar to Vossen et al. (2003), which first found an equitable allocation and then increased its efficiency.
In air traffic control, there are many optimization models formulating equity from different points of view, e.g., the pro-
portion of the delayed flights, the total delay time/cost, the delay time/cost per passenger, etc., by using different represen-
tations, e.g., variance and absolute value. However, they have similar formulations, i.e., keeping the individual values in a
small deviation from the average value, as listed in Table 14 of Appendix A.
The mathematical formulations concerning equity in this paper are inspired on those proposed in air traffic control. Rail-
way optimization models have substantially different constraints than air traffic and thus requiring studying which
approach would best suit the characteristic dynamics of railways. For instance, the train timetable has to be adhered more
strictly, trains have larger separation time, railway operations are more complicated (like overtaking), and capacity along the
lines is much less. We consider equity in both soft and hard ways (i.e., considering equity in objective function and in con-
straints respectively) by using improved representations. These come from direct inspiration of the relevant approaches
reported in Table 14 of Appendix A. For instance, when representing TOC’s equity as a hard constraint, two parameters
are used to make the descriptions more general and applicable (detailed in Section 3.3), instead of using a single deviation
upper bound K.
Furthermore, the issue of equity is also of concern to researchers in road traffic. Some authors dealt with the problem of
exploring the impact of existing strategies over equity (e.g., Ahmed et al., 2008), evaluating the equity in road resources dis-
tribution (e.g., Litman, 2002), or designing transportation networks with consideration of equity (e.g., Santos et al., 2008). In
the context of congestion pricing, optimal pricing models were proposed with social or spatial equity constraints (e.g., Yang
and Zhang, 2002; Yin and Yang, 2004), and a modeling framework was developed to design a more equitable pricing and
tradable credit schemes (e.g., Wu et al., 2012), in order to alleviate congestion or improve social benefit on multi-modal net-
works. Those studies are mostly from points of view of policy, planning, and design, with an aim of suggesting better ways to
incorporate fairness in transportation decisions. Since few approaches are proposed to address the online traffic manage-
ment problem in this field, these studies are not elaborated in this paper.
Table 1
Multiple (conflicting) objectives and mathematical formulations for train dispatching.
Publications Objective(s) Degree of Model Solution algorithm

delay equity structure
Train Unexpected Train Train/passenger Missed Energy Maintenance Delay
X. Luan et al. / Transportation Research Part C 80 (2017) 148–174

delay running time cancellation travel time connection consumption delay equity
Ghoseiri et al. (2004) U U [] MINLP H (Pareto front
based)
Mazzarello and U U [] MILP based H (AMCC)
Ottaviani (2007) on AG
Corman et al. (2009) U U [] MILP based BB, H (FIFO)
on AG
Schachtebeck and U U [] IP EX, H (FSFS, FSFS-
Schöbel (2010) FIX, etc.)
Corman et al. (2012) U U [] MILP based BB, H (Pareto front
on AG based)
Fuchsberger (2012) U U U [] IP EX, H
Albrecht et al. (2013) U U [] – PSS
Boccia et al. (2013) U U U [] MILP EX, H
Corman et al. (2011a) U [–] MILP based BB, H (priority rule
on AG based)
Meng and Zhou U [] IP based on EX, LR, H (priority
(2014) CF rule based)
Luan et al. (this work) U U [+] MILP based EX
on big-M
Symbol descriptions for Table 1: Mixed integer nonlinear programming (MILP); Mixed integer linear programming (MILP); Heuristics (H); Alternative graph (AG); Cumulative flow variable (CF); Branch-and-
bound (BB); First-In-First-Out (FIFO), called First-Come-First-Served (FCFS) or Ration-by-delay as well; First-Scheduled-First-Served (FSFS); FSFS with fixed priority (FSFS-FIX); Exact method (EX) - commercial
solver; Lagrangian relaxation (LR); Problem Space Search (PSS); Avoid Maximum Current Cmax (AMCC); [–] represents negative equity approach, i.e., explicitly leading to an inequitable situation; [+] denotes
positive equity approach, i.e., explicitly considering equity; [] is neutral equity approach, i.e., neither explicitly considering equity, nor resulting in an inequitable situation.
153
2.5. Statement of contribution
In summary, in spite of the rich body of existing train dispatching studies, there is a significant gap in the literature, when
referring to the computation of optimal online train schedules with consideration of delay equity. Previous studies in control
of train operations include negative equity approaches, which are actually discriminatory (e.g., Corman et al., 2011a). Those
include all kind of priority rules which differentiate trains based on their classes, for instance that a freight train should be
held at a signal to allow a faster passenger train to go first. Approaches, which do not explicitly consider classes, do not lead
to such discriminatory situations, including the FCFS rule and the vast majority of the optimization approaches reported in
Section 2.2. Those approaches are not discriminatory, in the sense that they do not specify a hard ranking of train classes, but
all traffic cannot have a systematically guaranteed equity, by using those approaches. In other terms, they achieve equity
only in a statistical sense, i.e., averaging over all possible situations of delays and traffic, assuming they all have equal prob-
ability, and then the resulting average output will be non-discriminatory. For each and every realization and case, instead,
they might actually provide discriminatory solutions which favor a particular train rather than another one. In case of sys-
tematic effects (related to the planned operations, the train services, the delays faced, and the demand), which cannot be
ruled out so easily, the resulting solution will be discriminatory, as we will quantify.
This paper focuses on the train dispatching problem with consideration of both train delays and delay equity, which are
well studied separately in previous studies. It makes a step forward to ensure that traffic is non-discriminatorily controlled,
and each solution computed for any input determination has a satisfactory degree of equity (which can be accepted by all
interested parties). Thus, incorporating delay equity into dispatching models represents one important contribution of this
research. Furthermore, as reported in Karsu and Morton (2015), which reviewed inequity averse optimization in operational
research, even when equity is considered while optimizing, sensitivity analysis (e.g., which controls the value of performance
indicators) is not discussed in most of the studies. In this paper, comprehensive experiments are conducted to study the sen-
sitivity of the solutions to a variety of factors, i.e., what impacts on and how to improve the equity of train dispatching solu-
tions, in order to further provide suggestions for decision making. We are able to provide an evaluation of how equity and
system performance might conflict with each other.
3. Mathematical formulation
Before formulating the non-discriminatory train dispatching problem, we first explain the terms concerning a physical
railway network (e.g., track, station, and segment), as well as the terms used in the following formulations (e.g., node and
link).
3.1. Explanations on relevant terms
In Fig. 1, a simple network is illustrated for instance, at macroscopic level (i.e., presenting railway network, station, single/
double-track segment, etc.), microscopic level (i.e., presenting track, signal, switch, platform, etc.), and modeling level con-
sidered in this paper respectively. The gray portion of the network in Fig. 1(a) is re-plotted in Fig. 1(b) with microscopic view,
and Fig. 1(b) can be further described as shown in Fig. 1(c) by the terms of node and link, which is used in our formulations.
The explanations on relevant terms concerning physical railway network are given as follows.
(i) Tracks are the roadways of a railway system. A track consists of the rails, ties, plates between rails and ties, fasteners,
ballast, etc. Main tracks can be used for regular train movements, except for train stopping. Siding tracks are the tracks
other than main tracks, which can be used for regular train movements, e.g., train stopping, passing, and overtaking.
Fig. 1(b) gives examples for main track and siding track.
(ii) Station is a railway facility where trains may stop to load or unload passengers or goods. A station has main track(s)
and siding track(s) to facilitate trains’ passing or overtaking.
(iii) Segment is the track between two stations, which is divided into block sections for the purpose of safe train separa-
tion. A block section (in a fixed block system) is a section of track, which trains cannot enter when it is blocked
(reserved or occupied) by other trains. A segment may consist of one track (single-track), two tracks (double-track)
or more tracks. Fig. 1(b) also presents single-track segment and double-track segment for instance.
In order to model railway facilities, two concepts are introduced in this paper, i.e., node and link, as shown in Fig. 1(c).
(i) A node represents a beginning/ending point of a block section. Additionally, it can also be viewed as a relevant point of
railway network, corresponding to a main/siding track in station, or a point of merging/diverging of tracks.
(ii) A link corresponds to a block section where only one train is allowed at any time. Nodes are connected in pairs on
links, and a link is directed from a starting node i to an ending node j. The default of link capacity in this paper is
one at any given time. An extension of the model by considering interlocking constraints could be implemented by
refining the concept of cells to match the track circuits on segments and inbound/outbound routes in an interlocking
area of a station. The cell occupation would then correspond to the route-lock sectional-release principles.
sta_B
r ailway netwo rk
(macroscopic)
station
(a) single-track segment
sta_A sta_C double-track segment
sta_D
sta_B single-track
r ailway netwo rk
sta_C
(microscopic)
main track
main track
(b)
double-track siding track

sta_D
sta_B node
bi-dire
sta_C link
modelling
(c)
i j
sta_D
Fig. 1. A simple railway network to illustrate the concepts of node and link.
In the optimization models, following inputs are considered:
(1) A railway network with stations and segments. Stations are simplified to a number of main/siding track(s) and further
modelled as a single link or a set of links. For a double-track railway segment between two stations, each track is mod-
elled as a sequence of directional links (i.e., directional block sections), and for a single-track railway segment, the only
track between two stations is modelled as bi-directional links (i.e., bi-directional block sections), as illustrated in the
red1 dotted line of Fig. 1(c). The turn around capacity for shunting at stations is assumed to be large enough.
(2) A set of trains which belong to a number of TOCs, with their classes (i.e., intercity, local, and freight), origins, destina-
tions, earliest departure times, planned arrival times, and free flow running times over links. Note that the free flow
running times of trains over the links are computed based on the planned speed profile, which can be in general dif-
ferent by train. A fixed speed profile model is used in this paper, as common in most train dispatching models.
(3) A set of TOCs with their delay costs of trains per unit time. Different TOCs may operate different classes of trains, i.e.,
intercity, local, and freight trains.
(4) Safety headway times between trains, which depend on the train length, speed, and route chosen. They are used to
separate the moving trains.
(5) Primary delays of trains at their origins, which are randomly generated by a 3-parameter Weibull distribution. The
formal representation of generating primary delays will be detailed in Section 4.2.
The models proposed in this paper result in choosing a route (a sequence of nodes and links) by which the train traverses
the railway network, and determining the corresponding passing times and train orders at each node. In general, the train
travel time over a link will be longer, if the train is held by the signaling system (before a red signal), or by a dispatching
action to resolve a conflict. This extra time is regarded as the consecutive delay time. Note that the granularity of time is
one second.
3.2. Notation
Tables 2–4 respectively list the general subscripts, input parameters, and decision variables of the models. R indicates the
set of real numbers, N denotes the set of integers, and subsets Nþ and N0 contain positive and non-negative integers respec-
tively, i.e., Nþ ¼ f1; 2; 3; . . .g and N0 ¼ f0; 1; 2; . . .g.
1
For interpretation of color in Figs. 1–5 and 8–12, the reader is referred to the web version of this article.
Table 2
General subscripts.
Symbol Description
i, j, k Node index, i; j; k 2 V, V is the set of nodes
e Link index, generated by two adjacent nodes i and j, e ¼ ði; jÞ 2 E, E is the set of links
f Train index, f 2 F, F is the set of trains, jFj is the total number of trains
u TOC index, u 2 U, U is the set of TOCs
M A sufficiently large positive number
Table 3
Input parameters.
Symbol Description
Fu Set of trains belonging to TOC u, F u # F, jF u j is the total number of trains belonging to TOC u
Ef Set of links train f may use, Ef # E
#f ði; jÞ 2 Nþ , free flow running time of train f to drive through link ði; jÞ
ef 2 N0 , planned departure time of train f at its origin node
dprm
f
2 N0 , primary departure delay time of train f at its origin node
rf 2 Nþ , planned arrival time of train f at its destination node
bf 2 R, delay cost of train f (per unit time)
wmin
f ði; jÞ 2 N0 , the minimum dwell time for train f on link ði; jÞ
g f ði; jÞ 2 N0 , setup time of train f on link ði; jÞ between start of link occupancy and train arrival (before a train operation)
hf ði; jÞ 2 N0 , clearance time of train f on link ði; jÞ between end of train departure and link release (after a train operation)
of Origin node of train f
sf Destination node of train f
lu 2 N0 , the maximum acceptable threshold for delay cost of TOC u
cu 2 R, threshold for the maximum allowed deviation of TOC u between delays associated to different TOCs
hf 2 N0 , the maximum tolerable delay time for train f
Table 4
Decision variables.
Symbol Description
af ði; jÞ 2 N0 , arrival time of train f on link ði; jÞ
df ði; jÞ 2 N0 , departure time of train f on link ði; jÞ
xf ði; jÞ 2 f0; 1g, xf ði; jÞ ¼ 1, if train f occupies link ði; jÞ at some time, and otherwise xf ði; jÞ ¼ 0
yf ;f 0 ði; jÞ 2 f0; 1g, yf ;f 0 ði; jÞ ¼ 1, if train f’ arrives at link ði; jÞ after train f, and otherwise yf ;f 0 ði; jÞ ¼ 0
sf ði; jÞ 2 N0 , travel time of train f on link ði; jÞ
ddstn
f 2 N, total delay time of train f at its destination
ddstnCsc
f 2 N, consecutive delay time of train f
Four types of variables are used to formalize the rerouting and rescheduling decisions: route selection variables x, train
order variables y, arrival time variables a, and departure time variables d. Specifically, xf ði; jÞ captures the routing decisions in
a rail network, yf ;f 0 ði; jÞ describes the detailed train orders, and a pair of variables af ði; jÞ and df ði; jÞ are introduced to represent
both temporal and spatial resource consumption of trains. The travel time sf ði; jÞ is then a consequence of the interaction of
all those variables for all trains in the network, as well as the train delay time, which is denoted as ddstn
f .
3.3. Mathematical model
A mathematical model (P1), which formalizes delay equity of competitors by a set of constraints, is first presented. The
objective is to minimize the average train delay costs while respecting all operational and safety requirements, as formulated
in Eq. (1). The total train delays are considered, including primary delays at origin station and consecutive delays encoun-
tered by resolving train conflicts.
P
f 2F ðbf ddstn
f Þ
ðP1Þ Z ðP1Þ ¼ min ð1Þ
jFj
Subject to
Group I: Route choice constraints

X
xf ðof ; jÞ ¼ 1 8f 2 F ð2Þ
j:ðof ;jÞ2Ef
X X
xf ði; jÞ ¼ xf ðj; kÞ 8f 2 F; j 2 V n fof ; sf g ð3Þ
i:ði;jÞ2Ef k:ðj;kÞ2Ef
X
xf ði; sf Þ ¼ 1 8f 2 F ð4Þ
j:ði;sf Þ2Ef
Group II: Time-space network constraints
af ðof ; jÞ P ðef þ dprm

f Þ xf ðof ; jÞ 8f 2 F; ðof ; jÞ 2 Ef ð5Þ
df ði; jÞ P af ði; jÞ 8f 2 F; ði; jÞ 2 Ef ð6Þ
df ði; jÞ ¼ af ðj; kÞ 8f 2 F; j 2 V n fof ; sf g; ði; jÞ 2 Ef ; ðj; kÞ 2 Ef ð7Þ

Group III: Running time and dwell time constraints
sf ði; jÞ ¼ df ði; jÞ af ði; jÞ 8f 2 F; ði; jÞ 2 Ef ð8Þ

h i
sf ði; jÞ P wmin
f ði; jÞ þ #f ði; jÞ xf ði; jÞ 8f 2 F; ði; jÞ 2 Ef ð9Þ
Group IV: Mapping constraints between train orders and link usage
0 0
xf ði; jÞ þ xf 0 ði; jÞ 1 6 yf ;f 0 ði; jÞ þ yf 0 ;f ði; jÞ 6 3 xf ði; jÞ xf 0 ði; jÞ 8f 2 F; f 2 F; f – f ; ði; jÞ 2 Ef \ Ef 0 ð10Þ
0 0
yf ;f 0 ði; jÞ 6 xf ði; jÞ 8f 2 F; f 2 F; f – f ; ði; jÞ 2 Ef \ Ef 0 ð11Þ
0 0
yf ;f 0 ði; jÞ 6 xf 0 ði; jÞ 8f 2 F; f 2 F; f – f ; ði; jÞ 2 Ef \ Ef 0 ð12Þ
Group V: Capacity constraints

0 0
af 0 ði; jÞ g f 0 ði; jÞ þ ½3 xf ði; jÞ xf 0 ði; jÞ yf ;f 0 ði; jÞ M P df ði; jÞ þ hf ði; jÞ 8f 2 F; f 2 F; f – f ; ði; jÞ 2 Ef \ Ef 0 ð13Þ
0 0
af 0 ðj; iÞ g f 0 ðj; iÞ þ ½3 xf ði; jÞ xf 0 ði; jÞ yf ;f 0 ði; jÞ M P df ði; jÞ þ hf ði; jÞ 8f 2 F; f 2 F; f –f ; ði; jÞ 2 Ef ; ðj; iÞ 2 Ef 0 ð14Þ
Group VI: Train delays and delay equity
ddstn
f ¼ df ði; sf Þ rf 8f 2 F; ði; sf Þ 2 Ef ð15Þ
P P
f 2F u ðbf ddstn
f Þ f 2F ðbf ddstn
f Þ lu
6 ð1 þ cu Þ þ 8u 2 U ð16Þ
jF u j jFj jF u j
ddstn
f 6 hf 8f 2 F ð17Þ
Constraints (2)–(4) in Group I ensure the consistency of trains’ movement in the network, at origin, intermediate and des-
tination nodes respectively. Constraint (5) guarantees that trains do not leave their origins before the earliest departure time,
i.e., the sum of the planned departure time and the primary delay time. Constraint (6) forces train’s transition within a link,
i.e., train departure time is later than arrival time on a certain link. Constraint (7) guarantees df ði; jÞ ¼ af ðj; kÞ, if two adjacent
links ði; jÞ and ðj; kÞ are consecutively used by train f. Constraint (8) defines the travel time of train f on link ði; jÞ, denoted as
sf ði; jÞ. Constraint (9) enforces the required free flow running time, as well as the minimum dwell times at stations. It is
worth noting that the minimum dwell time is a required time to complete the processes of passengers boarding/alighting,
goods loading/unloading, etc. Constraint (10) links train order variables yf ;f 0 ði; jÞ and link usage variables xf ði; jÞ. Specifically,
constraint (10) makes sure that, if and only if both trains f and f0 traverse on link ði; jÞ, i.e., xf ði; jÞ ¼ xf 0 ði; jÞ ¼ 1, then both
inequalities reduce to yf ;f 0 ði; jÞ þ yf 0 ;f ði; jÞ ¼ 1. This equality further indicates that, either train f0 arrives at link ði; jÞ after train
f or train f arrives at link ði; jÞ after train f0 . Constraints (11) and (12) further ensure that any yf ;f 0 ði; jÞ and yf 0 ;f ði; jÞ are always
less than xf ði; jÞ and xf 0 ði; jÞ. Constraints (13) and (14) explicitly ensure the link capacity requirement by setting a safety head-
way which is computed by g f ði; jÞ and hf ði; jÞ between the departure of a preceding train and the arrival of a successive train,
if those two trains are running on the same link.
In Eq. (15), total delay time of each train is calculated according to its planned arrival time rf . Recall that the total delay
time is the sum of the primary delay time at origin station and the consecutive delays encountered by solving train conflicts.
The consecutive delay time is later formulated in Eq. (23). Three parameters lu ; cu , and hf are used in constraints (16) and
(17), to ensure that any TOC or train cannot incur too large deviations in delays, compared with other TOCs or trains. The
representations of equity in constraints (16) and (17) are inspired from the typical air traffic control. Parameter cu is given as
a percentage, e.g., cu ¼ 10% means that TOC u can take 10% extra delays at most than other TOCs. Parameter lu indicates a
typical punctuality threshold, which can be used to solve a problem on how to equitably allocate a very small delay (e.g., only
ten seconds) to TOCs. For instance, lu ¼ 10 implies that if the delay cost of TOC u is not greater than ten seconds, TOC u is
always satisfied, even other TOCs have no delay costs. In other words, the delay equity constraint (16) will play a role, when
and only when the delay cost of TOC u is greater than ten seconds. For the equity of trains, constraint (17) simply requires
that the delay time of train f cannot be greater than hf . It should be noted that the input parameter hf should not be too small
to avoid model infeasibility.
In order to formalize the delay equity along different dimensions and provide a more flexible framework for numerical
analysis, we next propose another model representing delay equity of competitors in the objective function. We denote
the model (P2) including the objective function (18), constraints (2)-(15) and (19)–(22). Additional parameters and variables
used by model (P2) are shown in Table 5.
Three weights ka ; kb , and kc are respectively used to balance the importance among average delay costs, equity of com-
peting TOCs, and equity of trains. Equity of TOC is the deviation of the delays, measured at the aggregated level of TOC with-
out considering individual train delays, which can be regarded as ‘‘macroscopic” equity. Instead, equity of train refers to the
single deviation of delay at the level of each train, which can be viewed as ‘‘microscopic” equity. The deviation of TOC’s and
train’s delay cost, denoted as wu and /f respectively, are considered as measures of equity in model (P2). The maximum delay
cost deviation variables W and U are used to measure the quality of the overall system from the equity point of view.
" P #
f 2F ðbf ddstn
f Þ
ðP2Þ Z ðP2Þ ¼ min ka þ kb W þ kc U ð18Þ
jFj
In Eq. (18), the overall objective is to minimize train delay costs and delay inequity of competitors (i.e., TOCs and trains)
with respecting all operational and safety requirements. Regarding the equity, we consider the equity of a train for a set of
dispatching actions as the biggest (positive) difference between the average delay cost and the delay cost of each train. Sim-
ilarly, the equity of a TOC is measured as the biggest (positive) difference between the average delay cost at the level of TOC
and the delay cost of each TOC.
P P
f 2F u ðbf ddstn
f Þ f 2F ðbf ddstn
f Þ
wu ¼ 8u 2 U ð19Þ
jF u j jFj
P
f 2F ðbf ddstn
f Þ
/f ¼ bf ddstn 8f 2 F ð20Þ
f
jFj
W P wu 8u 2 U ð21Þ
U P /f 8f 2 F ð22Þ
Constraints (19) and (20) refer to the delay cost deviations of TOCs and trains respectively. The deviation per competitor
(i.e., TOC and train) is calculated by the difference between the delay cost of each competitor and the average delay cost of all
competitors. The maximum delay cost deviations of TOCs and trains, denoted as W and U, are respectively obtained by con-
straints (21) and (22) and then minimized in the objective function. It should be noted that the case of a train arriving ten
minutes early and another case arriving ten minutes late are quite different, i.e., the former has more equity and the latter
has less equity. Thus, the expected delay time is not considered in the objective function when measuring equity, and so do
other models for keeping consistency.
In order to provide a benchmark, we further consider a model (P3) in which delay equity is neglected. The objective func-
tion of model (P3) is formulated in objective function (1), subject to constraints (2)–(15).
Moreover, the above models consider both primary delays and consecutive delays along the routes, i.e., the total delays.
As the consecutive delay is the only factor that can actually be reduced by optimized dispatching, it would be interesting to
explore the impact of minimizing the consecutive delay only on the results. Therefore, Eq. (23) is presented to formulate the
consecutive delay ddstnCsc
f . Additionally, Eq. (15) should be replaced by Eq. (23), and ddstn
f in the objective function and con-
Table 5
Additional parameters and variables for model (P2).
Symbol Description
ka ; kb ; kc 2 R, weights used in the objective function, for average delay cost of trains, equity of TOCs, and equity of trains respectively
wu 2 R, delay cost deviation of TOC u, the maximum value of wu is denoted as W
/f 2 R, delay cost deviation of train f, the maximum value of /f is denoted as U
straints should be changed to ddstnCsc

f . This results in a model (P4) which includes the objective function (18) and constraints
(2)-(14) and (19)–(23), by considering consecutive delays only while generating an equitable train dispatching solution.
ddstnCsc
f ¼ df ði; sf Þ rf dprm
f 8f 2 F; ði; sf Þ 2 Ef ð23Þ
4. Experimental settings
This section provides the description of the experimental settings based on the Dutch railway network, the generation of
train primary delays at their origins, and a scheme to specify the configuration of each experiment corresponding to each
subsection in Section 5.1.
4.1. Description of the realistic dataset based on the Dutch railway network
The realistic dataset used in this paper refers to a line of the Dutch railway network, connecting Utrecht (Ut) to Den Bosch
(Ht), about fifty kilometers long. For simplicity, the rail network is sketched as shown in Fig. 2. The network is composed of
40 nodes and 42 links, with two main tracks, divided into one long corridor for each traffic direction and nine stations. The
two tracks in different directions are independent, so only one direction is considered, i.e., from Utrecht (Ut) to Den Bosch
(Ht). Free flow running and clearance time are computed microscopically based on the typical speed profiles of trains as in
Corman et al. (2011b), and rounded to seconds. Three categories of trains are considered: intercity, local, and freight train,
and each of them is associated to a competing TOC: TOC_InterCity, TOC_Local, and TOC_Freight respectively.
Four global routes (i.e., from Ut to Ht, from Ut to Gdm, from Gdm to Ht, and from Hto to Ht) are determined, which merge
just before Den Bosch (Ht). The routes of intercity, local, and freight trains are graphically presented in the lower part of
Fig. 2, in terms of origin, intermediate stop, destination, and number of trains per hour. Trains can be overtaken when mul-
tiple tracks are present. We consider one hour of traffic based on a regular-interval timetable, with fifteen trains. Local trains
stop at all stations; intercity and freight trains stop only at the origin and destination stations. The dotted line boxes in Fig. 2
show two rerouting zones I and II, where trains can change their usual/planned local routes (tracks).
4.2. Primary delays
Recall that each train is given a randomly generated primary delay time dprm
f at its origin. We consider fifty delay cases of
the primary delays following a 3-parameter Weibull distribution. If trains have different delay distributions, the parameters
below are used:
for intercity trains, scale = 394, shape = 2.27, shift = 315;

for local trains, scale = 235, shape = 3.00, shift = 186;
for freight trains, scale = 1099, shape = 2.62, shift = 885.
If all trains have the same distribution, they follow the one given to intercity trains. The values come from fitting to real
life data, as explained in Corman et al. (2011b).
Rerouting zone I
Rerouting zone II
Ut
Zbm Gdm Cl Htnc Htn Utl

Ht
origin
Hto destination
intermediate stop
2 intercity train
2 local train
1 freight train
Fig. 2. A realistic experimental network adapted from the Dutch railway network.
4.3. Experimental schemes
In this paper, each experiment is denoted with a five-field code hA,B,C,D,Ei, as described in Table 6.
The detailed experimental scheme is illustrated in Table 7 and the property/parameters investigated are highlighted. Each
experiment corresponds to each subsection of Section 5.1.
The proposed models are all solved by CPLEX optimization studio 12.3, on a computer with Intel(R) Core(TM) i7 @
2.00 GHz processor and 16 GB RAM. Recall that fifty delay cases of randomly generated primary delays are considered.
The total number of experimental cases equals to 6250.
Table 6
Description of the five-field code.
Code Code meaning Code Symbol description

symbol
A Model P1 Model (P1) representing equity in constraints
P2 Model (P2) representing equity in objective function
P3 Model (P3) without consideration of equity
P4 Model (P4) considering consecutive delays, instead of total delays
B Primary delays’ Diff Following Weibull distributions with different parameters
distribution Sam Following a Weibull distribution with same parameter
C Weights used in the – Only the average delay cost of trains is considered in the objective function, corresponding to the
objective function model (P1) and the model (P3)
(1,1,1) ka ¼ 1; kb ¼ 1; kc ¼ 1, all weights are set to be one, others have the similar meaning, i.e., (1,2,2),
(10,1,1), and (1,5,5)
ðka ; 0; kc Þ Equity of TOCs is not considered
ðka ; kb ; 0Þ Equity of trains is not considered
ðka ; kb ; kc Þ Combined weights, where ka 2 f1; 2; 3; 5; 10g; kb and kc 2 f0; 1; 2; 3; 5; 10g
D Rerouting option false Rerouting is not considered
true Rerouting is allowed
E Delay cost of trains per (1,1,1) Uniform delay cost per unit time, indicates that intercity, local, and freight trains are treated with
unit time same importance
(2,1,0.5) Variable delay cost per unit time, implies that intercity trains have higher delay cost (importance)
and freight trains are with lower delay cost (importance)
Table 7
Scheme of the experiments presented in Section 5.1.
5. Experimental results and analysis
In this section, we analyze the experimental results with different model and parameter settings, in order to identify the
key factors that influence equity, followed by a summary for providing a global view of the models’ performance and related
findings. Possible applications of the proposed models in practice are further discussed.
5.1. Results analysis for the experiments with the realistic dataset
We report the detailed analyses of the realistic experiments as average result of fifty delay cases with randomly gener-
ated primary delays, in each row of Tables 8–19, as well as in each point value of Figs. 6–9. Figs. 3–5 and 10 give all results of
those fifty delay cases, and then show the distribution of delay cost deviation for competitors (TOCs or trains), i.e., giving the
number of trains with a certain delay cost. TOCs/trains are identified by different colors in some figures, i.e., blue for intercity,
green for local and red for freight. We report the most relevant results in this section, and more detailed experimental results
are presented in Appendix B.
We describe here the contents of the table, which has the same structure with the most subsequent tables. Weights vec-
tor ðka ; kb ; kc Þ represents the weights used in the objective function of the model (P2) and (P4), e.g., weights (1,2,2) means
ka ¼ 1; kb ¼ 2, and kc ¼ 2. Each vector of weights is described by a single Weighting ratio j calculated as the equitable
weight (the largest one of kb and kc ) divided by delay cost weight (ka ). For example, weights (10,1,1) results in a
j ¼ 1=10 ¼ 0:1. An increasing weighting ratio reflects an increasing importance of equity. Computational time is the
Table 8
Results of different scheduling algorithms and optimization methods.
Weights Computation Average delay Max deviation of Max deviation of Average number
vector time (unit: cost (unit: TOCs’ delay cost trains’ delay cost of delayed trains
ðka ; kb ; kc Þ second) second) (unit: second) (unit: second)
FIFO – – 118.68 134.56 482.64 11.42
FSFS – – 209.90 53.80 257.68 12.58
AMCC – – 151.33 79.14 279.89 12.60
Punctuality – 3.28 186.21 263.91 1115.63 6.05
Min_MaxDelay – 2.27 99.43 144.92 327.20 12.16
Without equity, i.e., – 0.39 95.59 112.00 365.10 9.82
solutions of model (P3)
With equity, i.e., (2,3,3) 0.30 436.04 0.00 0.00 15.00
solutions of model (P2)
Table 9
Comparison of solutions considering equity as a hard/soft constraint.
Settinga Percentage of Computation time Average delay cost Max deviation of Max deviation of
feasible solutions (unit: second) (unit: second) TOCs’ delay cost trains’ delay cost
(unit: second) (unit: second)
Solution of model (P1) hf = 300 32% 0.11 95.35 13.94 159.63
hf = 600 76% 0.11 130.06 17.53 307.18
hf = 900 100% 0.12 181.84 18.86 411.98
hf = 1500 100% 0.19 181.15 19.10 547.52
hf = 2100 100% 0.20 181.15 18.78 592.74
Solution of model (P2) Weights (5,3,3) 100% 0.33 199.46 0.00 237.14
Weights (2,3,3) 100% 0.30 436.04 0.00 0.00
a
Note that column 2 lists weights vector ðka ; kb ; kc Þ for the model (P2), or the maximum tolerable delay time hf for the model (P1).
Table 10
Results with primary delays following same and different Weibull distributions.
Experimental configurations Weights vector Computation time Average delay cost Max deviation of Max deviation of
ðka ; kb ; kc Þ (unit: second) (unit: second) TOCs’ delay cost trains’ delay cost
Different distributions Without equity – 0.39 95.59 112.00 365.10
of primary delays With equity (1,1,1) 0.26 222.81 0.00 213.23
Same distribution Without equity – 0.25 89.18 40.76 259.40
of primary delays With equity (1,1,1) 0.19 148.41 0.00 179.05
Table 11
Comparison of solutions with and without rerouting, weights (3,2,2).
Weights vector Computation time Average delay cost Max deviation of TOCs’ Max deviation of trains’
ðka ; kb ; kc Þ (unit: second) (unit: second) delay cost (unit: second) delay cost (unit: second)
Without rerouting (3,2,2) 0.36 200.41 0.00 235.63
With rerouting (3,2,2) 4.11 187.79 0.00 229.39
Table 12
Results with different delay cost per unit time, weights (2,3,3).
*Note that delay cost (2,1,0.5) is the train delay cost per unit time for intercity, local and freight trains respectively.
Table 13
Results considering only consecutive delays, without rerouting options.
Weights vector Weighting Computation time Average delay cost Max deviation of Max deviation of
ðka ; kb ; kc Þ ratio j (unit: second) (unit: second) TOCs’ delay cost trains’ delay cost
FIFO – – – 33.87 60.51 309.75
FSFS – – – 85.10 47.61 293.76
Without equity – 0.00 0.21 10.79 15.30 90.30
With equity (5,1,1) 0.20 0.23 12.61 13.84 83.77
(5,2,2) 0.40 0.21 11.72 14.79 81.74
(1,1,1) 1.00 0.24 27.47 0.00 62.45
(2,3,3) 1.50 0.28 89.25 0.00 1.21
(1,3,3) 3.00 0.28 90.90 0.29 0.78
(1,10,10) 10.00 0.29 91.19 0.49 0.49
required time to find and prove an optimal solution. Average delay cost presents the objective value of the model (P1) and
(P3), and also gives the value of the first objective portion of the model (P2) and (P4). The delay cost deviation is disaggre-
gated per train and per TOC. This is the difference between delay cost for each competitor (TOC or train) and the average for
all competitors (TOCs or trains). The maximum value of those deviations is considered as equitable objective in the model
(P2) and (P4), and shown as Max deviation of TOCs’ delay cost and Max deviation of trains’ delay cost in following tables.
A large deviation means less equity. The deviation of a competitor can be reduced by decreasing its own delay cost or
increasing the delay cost of others. Average number of delayed trains is the number of delayed trains in average for the fifty
delay cases. Each of the fifty instances considered has 4300 variables and more than 14,000 constraints.
5.1.1. Solutions of scheduling algorithms and optimization methods: the model (P2) vs. the model (P3)
This subsection quantifies the trade-off between equity and inequity. On top of the optimization models, we also consider
three other scheduling algorithms, i.e., the well-known FIFO (First-In-First-Out) dispatching rule, common in railway prac-
tice, which simply gives priority to the train arriving first at the current block section; the FSFS (First-Scheduled-First-
Served) rule that follows the train orders of the original timetable; and the AMCC (Avoid Maximum Current Cmax) rule that
consists of forbidding the train orders causing the largest delay at a time, described in D’Ariano et al. (2007). We consider
weights of the model (P2) with equity to be (2,3,3) in this subsection. Moreover, while optimizing, two other objectives
are also used, i.e., ‘‘Punctuality” that focuses on train punctuality to maximize the amount of punctual trains and ‘‘Min_-
MaxDelay” that considers severity of train delay to minimize the maximum train delays.
The results for different scheduling algorithms are reported in Table 8. In column 3 of Table 8, computation time is quite
small, less than four seconds for all solutions. Column 4 clearly shows that the average delay cost increases going from
inequitable solutions to ‘‘Min_MaxDelay” solutions, FIFO solutions, AMCC solutions, ‘‘Punctuality” solutions, FSFS solutions,
and equitable solutions. Equity also results in increasing delays, which can be viewed as a ‘‘price of equity” and quantified as
ð436:04 95:59Þ 15 ¼ 5106:75 seconds for all traffic, in the case with weights (2,3,3).
Fig. 3 shows the distribution of delay cost per competitor. The X-axis represents the delay cost of trains and the Y-axis
shows the number of trains. The vertical lines indicate the average delay cost for different scheduling algorithms, viewed
as the reference for computing deviation (equity). When considering equity, the distribution of deviation results in a much
sharper peak, i.e., all trains have similar delay; while the other scheduling algorithms might result in smaller average values
(i.e., the vertical reference line) but much larger deviation, which reaches as far as 522 s in the FSFS solutions at lease and up
to 3601 s in the ‘‘Punctuality” solutions at most, marked as orange symbols in Fig. 3. Focus on train punctuality results in
large spread of delays (less equity), as shown in Fig. 3(d).
Fig. 4 reports in a similar manner the distribution of delay cost per TOC (with intercity, local, and freight trains shown in
blue, green, and red respectively), for six solutions, as labelled. It is evident that the equitable model results in much less
deviation for all TOCs, while FIFO and FSFS rules result in large spread, for all TOCs. In fact, FSFS rule results in slightly larger
equity than the other rules, while a much better (more equitable) solution is found by the equitable model. Moreover, the
model (P2) always provides the highest quality of equity, but results also in the largest delay cost.
5.1.2. Comparison of considering equity as a hard/soft constraint: the model (P1) vs. the model (P2)
In this subsection, we study the impact of having equity enforced by hard constraints in the model (P1), rather than as an
objective in the model (P2). A strict requirement of TOCs’ equity is used as a hard constraint with the following extra param-
eters: the acceptable delay cost lu ¼ 60, the maximum allowed deviation cu ¼ 1% for all TOCs. Several options of the max-
imum tolerable delay time hf for each train are considered, as listed in column 2 of Table 9. Unlike other tables, column 3 in
Table 9 shows the number of feasible solutions among those fifty delay cases.
Table 9 reports the main results. From the viewpoint of equity, the solution quality of the model (P1) is not as good as that
of the model (P2), especially for trains’ equity in last column. This mostly depends on the loose setting of the parameters.
Setting a more strict value for the maximum tolerable delay time hf reduces this problem, but the feasibility cannot be guar-
anteed for all instances, as shown in the third column. As shown in column 3, by setting the maximum tolerable delay time hf
to 600, thirty-eight of those fifty cases are feasible; by setting hf ¼ 300, only sixteen feasible cases can be obtained.
Fig. 3. Distribution of train delay cost.

Fig. 4. Distribution of delay cost deviations of trains, shown separately for different TOCs.
Fig. 5. Distribution of train delay cost, solutions with primary delays following same and different distribution.
5.1.3. Impact of primary delays’ distribution: same distribution vs. different distribution
This subsection explores the impact of primary delays’ distribution. As shown in Table 10 and Fig. 5, when primary delays
follow the same distribution for all competitors, a greater equity can be achieved by optimization, even if equity is neglected.
Primary delays following the same distribution decrease the deviation to 36% (for TOC’s equity of the model (P3)) or 84% (for
train’s equity of the model (P2)). This shows the significant effect of the primary delays on equitable operations.
Fig. 5 reports graphically similar findings: the delay cases following different distribution result in a larger maximum
delay cost deviation (i.e., less equity), as shown in the red circles. This justifies the intuition that traffic with very heteroge-
neous delay dynamic leads to more complex situations.
5.1.4. Sensitivity of the solutions to the relative weights

Recall that we use three weights ðka ; kb ; kc Þ to balance the importance of train delay cost, delay equity of competing TOCs,
and delay equity of trains. We here analyze sensitivity of the solutions to those weights. Several combinations of weights
used in this subsection are reported in Tables 6 and 7. We basically consider three cases:
(1) considering both TOC’s and single train’s performance with the same value, denoted as weights (,,);
(2) considering only TOC’s performance and neglecting train equity, denoted as weights (,,0);
(3) considering equity of single train and neglecting equity of TOC, denoted as weights (,0,).
Each vector of weights is described by a single weighting ratio j calculated as the equitable weight (the largest one of kb
and kc ) divided by delay cost weight (ka ). An increasing weighting ratio reflects an increasing importance of equity. The
impact of weighting ratio j on computation time, average delay cost, maximum deviation of TOCs’ and trains’ delay cost
are shown in Figs. 6 and 7.
As shown in Fig. 6, computation time is quite small (less than half second for all cases). In comparison, the cases with
weights (,,) generally have longer computation time. Looking at the plots of Fig. 7(a), delay cost of trains increases with
weighting ratio j, and so does equity. When weighting ratio j is greater than 1.5, the requirements of equity can be fully
satisfied (i.e., the maximum delay cost deviation is zero). Specifically, when weighting ratio j changes from 0.50 to 0.60,
a great reduction of the maximum deviation of competitors’ delay cost, i.e., 27% and 100% for trains’ and TOCs’ equity respec-
tively, is achieved. Moreover, as shown in the red box of Fig. 7(a), ‘‘macroscopic” equity (TOC’s equity) is much easier and
faster to obtain than single train’s equity. When weighting ratio 0:60 6 j < 1:50, the maximum delay cost deviation of TOCs
is zero (i.e., equity is fully achieved), and that of trains remains anyway large. Considering only TOC’s equity allows to aggre-
gate over increasingly large set of trains, and gives more freedom in rescheduling trains.
Next, we discuss the equity for single train, or considering only TOC’s equity. To do so, either weight kb or kc is set to be
zero. As shown in Fig. 7(b), if only TOC’s performance is considered, the maximum delay cost deviation of trains remains any-
way large. In other words, there is still inequity among trains, but for TOC point of view solutions are equitable. When single
train’s equity is considered, as reported in Fig. 7(c), both the delay cost deviations of TOCs and trains approach to zero, i.e.,
the equity of TOCs and trains is both kept well. This means that equity at microscopic level (i.e., the level of train) implies
equity at macroscopic level (i.e., the aggregated level of TOC). The detailed experimental results related to this subsection are
reported in Tables 15–17 of Appendix B.
Fig. 6. Computation time with different weights.
Fig. 7. Average delay cost, max delay cost deviation of TOCs and trains with different weights.
Fig. 8. Computation time with and without rerouting.
5.1.5. Comparison of solutions with and without rerouting

On the Dutch railway network we used, the infrastructure offers a few possibilities of train rerouting. For each train, a set
of local rerouting options is considered and can be exploited by the optimizer. Fig. 8 shows that longer computation time
(still compatible with real time operations, less than five seconds) is needed to find optimized rerouting measures.
In general, train delays and delay equity are conflicting objectives. But if rerouting is considered, the delay cost and the
maximum delay cost deviations of competitors can be both reduced at the same time, by a small factor of 7% and 3% for the
case with weights (3,2,2), as shown in Table 11. This limited improvement probably depends on the simple topology of the
network (only two possible rerouting zones, see Fig. 2). The detailed experimental results related to this subsection are
reported in Table 15 (without rerouting) and Table 18 (with rerouting) of Appendix B respectively.
5.1.6. Impact of train delay cost per unit time

We now study the relative importance of mixed traffic by considering different delay costs per unit time for intercity,
local, and freight trains, for example, equal to 2, 1 and 0.5 respectively. The computation results are shown in Table 12.
Unlike other tables, columns 3–5 in Table 12 show the average delay and for each TOC, and columns 6–7 give
the maximum deviation of competitors’ delay and (for TOCs and trains respectively).
Fig. 9 reports the distribution of deviations of competitors’ delay and respectively (on the X-axis). The Y-axis
shows the number of trains. The vertical lines (zero lines) indicate the average delay / , viewed as the reference of
deviation (equity). The intercity, local, and freight are represented in blue, green, and red respectively. Fig. 9(a) and (c) show
the total distribution, while Fig. 9(b) and (d) show the distribution per TOC (in different colors). Fig. 9(a) and (b) report the
fx1 deviation, while Fig. 9(c) and (d) report the fx2 deviation.
Fig. 9. Distribution of delay / deviations of trains and TOCs with different delay cost per unit time.
TOC_Freight faces more delay times, as its delay cost per unit is the lowest. This is evident in Fig. 9(a) and (b). When con-
sidering delay cost (and not delay time), in Fig. 9(c) and (d), the deviations of trains’ delay cost are more concentrated to the
Y-axis, and the delay costs of three TOCs are quite similar.
5.1.7. Solutions considering only consecutive delays

In this subsection, we study the model (P4) which considers only consecutive delays (i.e., the sum of the delays encoun-
tered by solving train conflicts). Table 13 shows the results considering only consecutive delays with different scheduling
algorithms and different weights. Fig. 10 reports the distribution of delay cost deviations of trains, considering all delay cases
and scheduling algorithms, in an analogous way to Fig. 3.
Graphically, the solution of the model (P4) has a smaller spread compared to the other scheduling algorithms. The aver-
age delay cost approaches the value of FSFS solution, and is larger than the other scheduling algorithms, but much smaller
deviation is found, i.e., the equity is improved greatly. What is also interesting is that this model results in better perfor-
mances compared to the model (P2), see Fig. 3 and Table 8. As marked in orange symbols of Fig. 10, the maximum delay cost
deviations in this case for the FIFO, and FSFS are respectively 670 and 625 s; for the model (P3) is 528 s, while for the model
(P4) considering equity is as low as 16 s of consecutive delay. This phenomenon can be explained by the fact that consecutive
delays are the only factor that can actually be reduced by optimized dispatching. The model (P2) achieves uniform equitable
delays for all trains, but this might mean that if a train is delayed a lot, all other traffic will be delayed by a similar amount.
Focusing on the consecutive delay only, trains will face different delays (that can be seen as a form of inequity), but the delay
incurred in the control process will be spread more uniformly, and with overall smaller delays.
5.2. Summary of results analysis
We here derive the main conclusions, sketched qualitatively in Fig. 11 from the point of view of delay cost (Y-axis) and
inequity among competitors (X-axis). The original timetable (red dot) is assumed to have no delays, and a reference value for
what concerns equity. When delays occur, the FSFS solution (purple dot) results in high delay costs and large inequity, so
does the AMCC solution (orange dot). The FIFO solution (associated to equity in air traffic control, pink dot) is relatively
bad for both equity and delays.
When optimizing, the objectives of train delays and delay equity are in conflict. Traditional ‘‘inequitable” optimized dis-
patching model (P3, blue dot) allows to greatly decrease delay costs at the expense of a large inequity (blue dot). Similarly,
the ‘‘Min_MaxDelay” solution (light blue dot) attains a little higher delay cost and a little improved equity. The ‘‘Punctuality”
solution (dark green dot) results in the largest inequity, which gives priority to the punctual trains and make the delayed
trains to face more delays. The newly proposed equitable model (P2) determines and explores the trade-off between train
delays and delay equity. This results in the black line spanning the entire plot, depending on the weights used, defining
the trade-off between the two performance indicators. Equity of running traffic is improved at the only expense of larger
delay costs, which could be identified a ‘‘price of equity”. In our cases, a weighting ratio j larger than 1.50 leads to the full
satisfaction of equity, represented as the black dot. The model (P2) performs better than FSFS (purple dot) and AMCC (orange
dot) for both delays and equity: given an equity target, less delays are found (the vertical dotted line); given a delay target,
higher equity of operations is achieved (the horizontal dotted line). Rerouting (green dot) can improve both performance
indicators at once, at the only costs of increased computational complexity.
Equity is also the result of primary delays faced by trains; similar distributions of delays (for instance, TOCs competing in
the same market, trains of the same category) result in higher equity. Considering equity at the level of single train performs
better than considering equity aggregated at the level of TOC. The former implies the latter in our experimental results. The
choice for one or the other setting is an interesting policy issue.
Results are consistently better if equity is considered as an objective in model (P2) and not as a constraint in model (P1).
This latter setting depends on extra parameters (i.e., lu ; cu , and hf ) to keep solutions feasibility, whose value has to be fur-
ther carefully chosen and justified to the stakeholders.
Fig. 10. Distribution of delay cost deviations of trains, considering only consecutive delays.
Fig. 11. Overview of solutions obtained by different models and scheduling algorithms.
If different delay costs of train categories are considered, equity is associated to limited deviation (and inequity) in costs,
while delay times can still vary. The optimization model exploits trains with lower delay costs and delays them more than
other trains. The delay costs can be further adjusted to match the economic value of train punctuality of real operations for
different train categories (i.e., the different punctuality targets of passenger and freight trains).
If only consecutive delays are considered, see the model (P4), the performance concerning equity is similar to that con-
sidering total delays. Anyway, the total delay cost is smaller compared to the model (P2). This is due to the fact that primary
delays (over which limited to no control is possible) are not counted in the objective function. When considering consecutive
delays only, trains will face different total delays (that can be seen as a form of inequity), but the delay incurred in the control
process will be spread more uniformly, and with overall smaller delays.
5.3. Possible application of the proposed models in practice
In spite of the wishes for dispatching trains in a non-discriminatory (equitable) way, only limited steps forward have been
made to ensure this behavior for every single delay case and instance. According to the conclusions summarized in Sec-
tion 5.2, we provide some ideas and suggestions for managing railway traffic in a non-discriminatory fashion, from a variety
of point of views.
Equitable railway traffic management may suffer some expense of extra delay cost (compared to traditional dispatching
which neglects equity), which can be viewed as the ‘‘price of equity”. We proved and quantified that the ‘‘price of equity”
depends on the network complexity, the intensity of perturbations, the starting/original timetable and the number of trains
in operation. Exploring and determining an acceptable price of equity is a political/regulatory choice.
Implications on system performance and usability of equitable models. An example of how to practically deal with the gap
between system performance and equity can be seen in Fig. 12. First, the model (P2) and (P3) are simultaneously solved with
the same dataset, in order to obtain two extreme solutions of equitable case and inequitable case. If the extra delays asso-
ciated with equitable dispatching are acceptable, the equitable solution would be chosen right away. If instead the extra
delays are unacceptable, the equity can be incorporated to a certain extent only, by using weighing ratio j in the optimiza-
tion; or it can be considered at the level of TOC only, and not at the level of individual trains. If only the inequitable solution
with the minimum delays is to be adopted in practice, a last resort would be to incorporate actions that offset the inequitable
traffic. As stated in Directive 2001/14/EC (Commission of the European Communities, 2001), penalty and compensation may
be included for actions which disrupt the operation of the network and for TOCs that suffer from disruption respectively. For
instance, the TOC with less delays may provide compensation to others that have more delays. The value of compensation
can be measured by the difference of delay cost between the inequitable solution and the absolute equitable solution. This
leads to a typical cooperative game-theory setting of redistributing welfare.
Implications from a policy point of view. The trade-off between system optimum (inequitable solution) and the equitable
solution might be large, and sometimes even prescribe traffic to be delayed only for achieving equity. This behavior might be
acceptable only under the strictest application of a non-discriminatory policy, but degrades performance of the system to a
large extent. Defining a threshold for acceptable equity in terms of weighing factor can find a balance between the two objec-
tives; or typical Multi Criteria Decision Making techniques can be applied. Finally, the resulting equity depends on factors
that can be regulated by policy, such as capacity allocation and the original timetable. An open challenge for policy makers
is how to define timetables that naturally lead to equitable solutions in practice.
Fig. 12. Illustration of the steps for applying the proposed models in practice.
6. Conclusions and future research
This paper addresses the problem of determining real-time non-discriminatory (enforcing equity of train traffic against all
possible traffic conditions and delays) train dispatching solutions, providing a set of mathematical formulations and compre-
hensive experiments. The non-discriminatory train dispatching problem is respectively formulated by a set of constraints in
the model (P1) and by objectives in the model (P2). The performance of non-discriminatory models is assessed in compar-
ison with the traditional model (P3) which neglects equity, the models considering objective of train punctuality and sever-
ity of train delay respectively, and also with three dispatching rules (i.e., FIFO, FSFS, and AMCC rules), on a real-world dataset
adapted from the Dutch railway network. Conclusions made from the experimental results describe how to manage railway
traffic in a non-discriminatory fashion, from the policy and practice point of view. We studied and quantified the trade-off
between equity and system performance, based on an extended sensitivity analysis. According to the experimental results,
the optimization model (P2) yields better performance than the FIFO, FSFS, and AMCC scheduling rules for both delays and
equity. The minimization of train delays and delay inequity are two conflicting objectives; generally, equity of running traffic
is improved at the expense of larger delays. Similar distributions of the primary delays result in higher equity. Moreover,
considering only TOC’s equity allows to aggregate over increasingly large set of trains and gives more freedom in reschedul-
ing trains.
Our future research focuses on the following main extensions. First, rerouting or consecutive delay allows to reach improved
equity and reduced delays in this current study, so a natural extension is to further promote them in the model. Second, com-
plex interlocking systems can be incorporated further in the model, by refining the concept of cells. This would allow to enlarge
the set of routes in station areas, as well as including more processes at stations, like turn-around or shunting. A third direction
is to study how to best structure the original timetable, with the objective of ensuring equitable traffic control in operations.
This would describe the impact of the timetable beyond robustness and resiliency against small delays in operations (see
Bešinović et al., 2016). The trade-off between equity and heterogeneity of the timetable should also be explored. Finally, a com-
prehensive framework can be defined where equitable planning, capacity allocation and equitable control can be considered at
once, to reach non-discriminatory operations at system level, with impact on decision making and policy regulations.
Acknowledgments
The work of the first and third authors is jointly supported by National Natural Science Foundation of China (71571012)
and the State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University (RCS2016ZT022). The work of the
first author is also supported by China Scholarship Council under Grant 201507090058.
Appendix A. Summary of the studies addressing equity in air traffic control
Table 14 reports on the studies addressing equity in air traffic control (discussed in Section 2.4), including some allocation
rules (principles), formalizing methods of representing equity as a hard/soft constraint (i.e., represented in constraints and in
the objective function respectively) in optimization models.
Table 14
Summary of the studies addressing equity in air traffic control.
Description Publication(s)
Allocation rules/ FSFS, or Ration-by- A planning oriented rule that follows the orders of flights Hoffman and Davidson (2003),
principles schedule (RBS) in the planned timetable Manley and Sherry (2008, 2010),
and Glover and Ball (2013)
FIFO, or FCFS An operation oriented rule that gives high priority to the Milan (1997), Sherry et al. (2008),
flight arriving first and Rebollo and Cruz (2009)
Ration-by-Aircraft Size Flight in the heavy category is assigned slot first, followed Manley and Sherry (2008, 2010)
(RBAc) by flight in the large, then the small one
Ration-by-Passengers A service oriented rule that the flight carrying more Manley and Sherry (2008, 2010)
(RBPax) passenger is first served
Ration-by-distance A distance oriented rule the flight with longest flying Hoffman et al. (2007), Ball et al.
(RBD) distance is given precedence over the others (2010), and Manley and Sherry
(2010)
Ration-by-Fuel Flow A cost oriented rule that first serves the aircraft with Manley and Sherry (2010)
High/Low Precedence higher/lower fuel burn rate
(RBFF-High/RBFF-Low)
Congestion charge, A market-based approach using congestion charge or Brueckner (2005), Ball et al.
auctioning allocation auction to allocate resource, e.g., high priority is given to (2006), and Harsha (2009)
the competitor who pays more money.
P v 2
Equity formulation as a min a ga vg or Minimizing the variance/absolute value of the difference Zhong (2012)
P vaa v
soft constraint (in min a g g between the proportion of delayed flights for airline a and
a
objective function) h i the proportion of all delayed flights
v
min maxa ga vg Minimizing the maximum difference between the Zhong (2012)
a
proportion of delayed flights for airline a and the
proportion of all delayed flights
P 2
min a ð-a -Þ Minimizing the variance value of the difference between Chen (2009)
the total delay time of passengers on flights of airline a
and the total delay time of all passengers
va ga
Equity formulation as a v 6 g þ K or Any airline cannot incur too large deviations in the Zhong (2012)
va ga
hard constraint (in v 6 g ð1 þ KÞ number of delayed flights, compared with other airlines
qa
constraints) na
n 6 q þK Any airline cannot incur too large deviations in delay Rios and Ross (2007)
times, compared with other airlines. The benchmark of
deviation is determined by flight travel time, i.e., flight
with longer travel time is likely to face more delay times
P
Symbol descriptions for Table 14: Let ga be the number of flights for airline a; g ¼ ga is the total amount of flights; va is the number of delayed flights for
P P
airline a; v ¼ va indicates the total amount of delayed flights; nP a denotes the total delay time of flights for airline a; n ¼ na is the total delay time of all
flights; qa denotes the total travel time of flights for airline a; q ¼ qa is the total travel time of all flights; -a represents the total delay time of passengers
P
on the flights of airline a; - ¼ -a denotes the total delay time of all passengers; K is the overall deviation upper bound.
Appendix B. The network layout and the detailed experimental results
The layout of the Dutch railway network, which is composed of 40 nodes and 42 links, is illustrated in Fig. 13.
The detailed experimental results related to Section 5.1 are reported here. Tables 15–17 provide the average experimen-
tal results of fifty delay cases with different weights vector ðka ; kb ; kc Þ, considering both TOC’s and single train’s equity, only
single trains’ equity and only TOC’s equity respectively, denoted as weights (,,), weights (,0,) and weights (,,0). Rerouting is
not considered in these experimental results.
Table 18 reports the average experimental results of fifty delay cases with different weights vector ðka ; kb ; kc Þ, considering
both TOC’s and single train’s equity. Rerouting is adopted in these experiments.
Table 19 lists the average experimental results of fifty delay cases obtained by model (P4), considering only consecutive
delays and no rerouting options.
Fig. 13. Layout of the Dutch railway network.

Table 15
Results with different weights vector ðka ; kb ; kc Þ and no rerouting options.
Weights Weighting Computation time Average delay cost Max deviation of Max deviation of
vector ratio j (unit: second) (unit: second) TOCs’ delay cost trains’ delay cost
ðka ; kb ; kc Þ (unit: second) (unit: second)
Without equity – 0.00 0.39 95.59 112.00 365.10
With equity (10,1,1) 0.10 0.38 95.93 107.43 350.41
(5,1,1) 0.20 0.43 98.41 102.62 338.67
(10,3,3) 0.30 0.42 98.75 101.76 337.85
(3,1,1) 0.33 0.40 99.02 101.10 337.62
(5,2,2) 0.40 0.38 99.46 100.22 337.14
(2,1,1) 0.50 0.31 114.44 83.80 322.94
(5,3,3) 0.60 0.33 199.46 0.00 237.14
(3,2,2) 0.67 0.36 200.41 0.00 235.63
(1,1,1) 1.00 0.26 222.81 0.00 213.23
(2,3,3) 1.50 0.30 436.04 0.00 0.00
(3,5,5) 1.67 0.27 436.04 0.00 0.00
(1,2,2) 2.00 0.28 436.04 0.00 0.00
(2,5,5) 2.50 0.26 436.04 0.00 0.00
(1,3,3) 3.00 0.27 436.04 0.00 0.00
(3,10,10) 3.33 0.27 436.04 0.00 0.00
(1,5,5) 5.00 0.31 436.04 0.00 0.00
(1,10,10) 10.00 0.25 436.04 0.00 0.00
Table 16
Results considering only single trains’ equity and no rerouting options.
Considering only equity for (10,0,1) 0.10 0.22 96.09 348.77 110.82
single train, i.e., weights (,0,) (5,0,1) 0.20 0.22 97.10 343.44 109.24
(10,0,3) 0.30 0.20 98.36 338.78 107.58
(3,0,1) 0.33 0.20 98.40 338.72 107.44
(5,0,2) 0.40 0.21 98.51 338.09 105.94
(2,0,1) 0.50 0.21 98.43 338.21 106.76
(5,0,3) 0.60 0.21 99.02 337.02 107.10
(3,0,2) 0.67 0.20 99.02 337.02 107.10
(1,0,1) 1.00 0.18 143.87 292.17 97.20
(2,0,3) 1.50 0.16 436.04 0.00 0.86
(1,0,2) 1.67 0.17 436.04 0.00 0.86
(1,0,2) 2.00 0.16 436.04 0.00 0.86
(2,0,5) 2.50 0.15 436.04 0.00 0.86
(1,0,3) 3.00 0.18 436.04 0.00 0.86
(3,0,10) 3.33 0.17 436.04 0.00 0.86
(2,0,10) 5.00 0.15 436.04 0.00 0.86
(1,0,10) 10.00 0.17 436.04 0.00 0.86
Table 17
Results considering only TOC’s equity and no rerouting options.
Considering only TOC’s equity, (10,1,0) 0.10 0.19 95.68 108.74 361.90
i.e., weights (,,0) (5,1,0) 0.20 0.21 96.19 105.39 356.38
(10,3,0) 0.30 0.23 96.50 104.14 358.92
(3,1,0) 0.33 0.22 96.59 103.88 358.90
(5,2,0) 0.40 0.23 96.59 103.62 358.92
(2,1,0) 0.50 0.23 98.65 98.96 359.28
(5,3,0) 0.60 0.21 98.61 98.85 358.70
(3,2,0) 0.67 0.22 98.74 98.64 358.58
(1,1,0) 1.00 0.20 114.15 81.79 383.76
(2,3,0) 1.50 0.32 195.94 0.00 485.10
(3,5,0) 1.67 0.30 195.94 0.00 486.50
(1,2,0) 2.00 0.28 195.94 0.00 482.84
(continued on next page)

Table 17 (continued)
(2,5,0) 2.50 0.27 195.94 0.00 494.98
(1,3,0) 3.00 0.28 196.03 0.00 492.00
(3,10,0) 3.33 0.26 195.94 0.00 490.78
(2,10,0) 5.00 0.25 195.94 0.00 491.32
(1,10,0) 10.00 0.26 195.96 0.00 498.86
Table 18
Results with different weights vector ðka ; kb ; kc Þ and considering rerouting options.
Without equity – 0.00 2.20 26.47 224.88 460.68
With equity (10,1,1) 0.10 2.21 25.57 216.01 444.19
(5,1,1) 0.20 2.27 25.25 213.76 442.43
(10,3,3) 0.30 2.56 24.95 212.96 442.15
(3,1,1) 0.33 2.18 24.49 212.59 442.09
(5,2,2) 0.40 2.45 24.23 212.43 441.41
(2,1,1) 0.50 3.78 143.41 44.39 273.77
(5,3,3) 0.60 3.70 187.79 0.01 229.39
(3,2,2) 0.67 4.11 187.79 0.00 229.39
(1,1,1) 1.00 3.13 235.91 0.00 181.27
(2,3,3) 1.50 4.05 417.18 0.00 0.00
(3,5,5) 1.67 4.24 417.18 0.00 0.00
(1,2,2) 2.00 4.63 417.18 0.00 0.00
(2,5,5) 2.50 4.30 417.18 0.00 0.00
(1,3,3) 3.00 4.15 417.18 0.00 0.00
(3,10,10) 3.33 3.95 417.18 0.00 0.00
(1,5,5) 5.00 4.22 417.18 0.00 0.00
(1,10,10) 10.00 4.12 417.18 0.00 0.00
Table 19
Results of the model (P2) and (P3) considering only consecutive delays, without rerouting options.
Without equity – 0.00 0.21 10.79 15.30 90.30
With equity (10,1,1) 0.10 0.21 12.37 14.39 83.55
(5,1,1) 0.20 0.23 12.61 13.84 83.77
(10,3,3) 0.30 0.21 12.33 15.06 82.35
(3,1,1) 0.33 0.23 12.33 15.06 82.35
(5,2,2) 0.40 0.21 11.72 14.79 81.74
(2,1,1) 0.50 0.21 11.55 10.15 80.63
(5,3,3) 0.60 0.30 22.60 0.01 68.56
(3,2,2) 0.67 0.27 22.72 0.01 68.64
(1,1,1) 1.00 0.24 27.47 0.00 62.45
(2,3,3) 1.50 0.28 89.25 0.00 1.21
(3,5,5) 1.67 0.24 89.23 0.00 1.21
(1,2,2) 2.00 0.27 89.23 0.00 1.21
(2,5,5) 2.50 0.25 90.35 0.00 1.21
(1,3,3) 3.00 0.28 90.90 0.29 0.78
(3,10,10) 3.33 0.28 90.09 0.49 0.49
(1,5,5) 5.00 0.28 90.09 0.49 0.49
(1,10,10) 10.00 0.29 91.19 0.49 0.49
References
Access Rights Policy, 2016. Network Rail Access Rights Policy <https://www.networkrail.co.uk/wp-content/uploads/2016/ 12/Network-Rail-Access-Rights-
Policy.pdf>.
Ahmed, Q.I., Lu, H., Ye, S., 2008. Urban transportation and equity: a case study of Beijing and Karachi. Transp. Res. Part A: Policy Practice 42 (1), 125–139.
Albrecht, A.R., Panton, D.M., Lee, D.H., 2013. Rescheduling rail networks with maintenance disruptions using problem space search. Comput. Oper. Res. 40
(3), 703–712.
Alexandersson, G., Hultén, S., Nilsson, J.E., Pyddoke, R., 2012. The Liberalization of Railway Passenger Transport in Sweden-Outstanding Regulatory
Challenges. Working paper in transport economics. Central for Transport Studies KTH Sotockholm.
Ball, M., Donohue, G., Hoffman, K., 2006. Auctions for the safe, efficient, and equitable allocation of airspace system resources. Comb. Auctions 1.
Ball, M.O., Hoffman, R., Mukherjee, A., 2010. Ground delay program planning under uncertainty based on the ration-by-distance principle. Transp. Sci. 44
(1), 1–14.
Bešinović, N., Goverde, R.M., Quaglietta, E., Roberti, R., 2016. An integrated micro–macro approach to robust railway timetabling. Transp. Res. Part B:
Methodol. 87, 14–32.
Boccia, M., Mannino, C., Vasilyev, I., 2013. The dispatching problem on multitrack territories: heuristic approaches based on mixed integer linear
programming. Networks 62 (4), 315–326.
Borndorfer, R., Grotschel, M., Lukac, S., Mitusch, K., 2006. An auctioning approach to railway slot allocation. Compet. Reg. Network Indus. 7, 163.
Brueckner, J.K., 2005. Internalization of airport congestion: a network analysis. Int. J. Ind. Organ. 23 (7), 599–614.
Cacchiani, V., Huisman, D., Kidd, M., Kroon, L., Toth, P., Veelenturf, L., Wagenaar, J., 2014. An overview of recovery models and algorithms for real-time
railway rescheduling. Transp. Res. Part B 63, 15–37.
Canadian rail operating rules, 2015. <http://www.tc.gc.ca/eng/railsafety/rules-tco167.htm/>.
Chen, S., 2009. The key technique and algorithms in air traffic flow management (ATFM) based on collaborative decision making (CDM) theory PhD Thesis.
Nanjing University of Aeronautics and Astronautics, China.
Commission of the European Communities, 1991. Council Directive 91/440/EEC on the development of the Community’s railways.
Commission of the European Communities, 1995. Council Directive 95/18/EC on the licensing of railway undertakings.
Commission of the European Communities, 2001. Council Directive 2001/14/EC on the allocation of railway infrastructure capacity and the levying of
charges for the use of railway infrastructure and safety certification.
Commission of the European Communities, 2013. The Fourth Railway Package - Completing The Single European Railway Area to Foster European
Competitiveness and Growth, COM 25.
Corman, F., D’Ariano, A., Pacciarelli, D., Pranzo, M., 2009. Evaluation of green wave policy in real-time railway traffic management. Transp. Res. Part C 17 (6),
607–616.
Corman, F., D’Ariano, A., Hansen, I.A., Pacciarelli, D., 2011a. Optimal multi-class rescheduling of railway traffic. J. Rail Transp. Plan. Manage. 1 (1), 14–24.
Corman, F., D’Ariano, A., Pranzo, M., Hansen, I.A., 2011b. Effectiveness of dynamic reordering and rerouting of trains in a complicated and densely occupied
station area. Transp. Plan. Technol. 34 (4), 341–362.
Corman, F., D’Ariano, A., Pacciarelli, D., Pranzo, M., 2012. Bi-objective conflict detection and resolution in railway traffic management. Transp. Res. Part C 20
(1), 79–94.
Corman, F., Meng, L., 2014. A review of online dynamic models and algorithms for railway traffic management. IEEE Trans. Intell. Transp. Syst. 9, 1–11.
Del Pozo y de Poza, I., Vilaplana Ruiz, M., Goodchild, C., 2009. Assessing fairness and equity in trajectory based operations. In: 9th AIAA Aviation Technology,
Integration, and Operations Conference (ATIO), South Carolina, USA.
D’ariano, A., Pacciarelli, D., Pranzo, M., 2007. A branch and bound algorithm for scheduling trains in a railway network. Eur. J. Oper. Res. 183 (2), 643–657.
Fang, W., Yang, S., Yao, X., 2015. A survey on problem models and solution approaches to rescheduling in railway networks. IEEE Trans. Intell. Transp. Syst.
16 (6), 2997–3016.
Fuchsberger, M., 2012. Algorithms for real traffic management in complex central station areas Dissertation 20398. ETH Zurich, Switzerland.
Ghoseiri, K., Szidarovszky, F., Asgharpour, M.J., 2004. A multi-objective train scheduling model and solution. Transp. Res. Part B 38 (10), 927–952.
Glover, C.N., Ball, M.O., 2013. Stochastic optimization models for ground delay program planning with equity–efficiency tradeoffs. Transp. Res. Part C:
Emerging Technol. 33, 196–202.
Harrod, S., 2013. Auction pricing of network access for North American railways. Transp. Res. Part E: Logistics Transp. Rev. 49 (1), 176–189.
Harsha, P., 2009. Mitigating airport congestion: market mechanisms and airline response models PhD Thesis. Massachusetts Institute of Technology (MIT),
USA.
Hoffman, R., Davidson, G., 2003. Equitable Allocation of Limited Resources (EALR) – Defining, Measuring, and Implementing Equity. FAA Free Flight Program,
Metron aviation, Virginia, USA.
Hoffman, R., Ball, M.O., Mukherjee, A., 2007. Ration-by-distance with equity guarantees: a new approach to ground delay program planning and control. In:
The 7th ATM R&D Seminar, Barcelona, Spain.
IBM Global Business Services, 2011. Rail Liberalisation Index 2011-Market Opening: Rail Markets of the Member States of the European Union. Switzerland
and Norway in comparison, Brussel.
Karsu, Ö., Morton, A., 2015. Inequity averse optimization in operational research. Eur. J. Oper. Res. 245 (2), 343–359.
Kim, A., Hansen, M., 2013. A framework for the assessment of collaborative en route resource allocation strategies. Transp. Res. Part C: Emerging Technol.
33, 324–339.
Klabes, S.G., 2010. Algorithmic railway capacity allocation in a competitive European railway market. Fakultat fur Bauingenieurwesen der Rheinisch-
Westfalischen Technischen Hochschule Aachen zur Erlangung des akademischen Grades eines Doktors der Ingenieurwissenschaften genehmigte
Dissertation. Manuscript. 2010. 209 p.
Kuhn, K.D., 2013. Ground delay program planning: delay, equity, and computational complexity. Transp. Res. Part C 35, 193–203.
Kurosaki, F., 2008. An analysis of vertical separation of railways Doctoral dissertation. University of Leeds, NY.
Lalive, R., Schmutzler, A., 2008. Exploring the effects of competition for railway markets. Int. J. Ind. Organ. 26 (2), 443–458.
Litman, T., 2002. Evaluating transportation equity. World Transp. Policy Practice 8 (2), 50–65.
Manley, B., Sherry, L., 2008. Impact of ground delay program rationing rules on passenger and airline equity. In: Integrated Communications, Navigation and
Surveillance Conference, ICNS. IEEE, pp. 1–11.
Manley, B., Sherry, L., 2010. Analysis of performance and equity in ground delay programs. Transp. Res. Part C 18 (6), 910–920.
Mazzarello, M., Ottaviani, E., 2007. A traffic management system for real-time traffic optimization in railways. Transp. Res. Part B: Methodol. 41 (2), 246–
274.
Meng, L., Zhou, X., 2014. Simultaneous train rerouting and rescheduling on an N-track network: a model reformulation with network-based cumulative flow
variables. Transp. Res. Part B 67, 208–234.
Milan, J., 1997. The flow management problem in air traffic control: a model of assigning priorities for landings at a congested airport. Transp. Plan. Technol.
20 (2), 131–162.
Nash, C., Rivera-Trujillo, C., 2004. Rail regulatory reform in Europe–principles and practice. In: STELLA Focus Group 5 Synthesis Meeting, Athens (June).
Pacciarelli, D., 2013. Introduction and classification of existing rescheduling literature. In: Lecture in the 5th International Seminar on Rail Operations
Modeling and Analysis (RailCopenhagen 2013), Copenhagen, Denmark.
Pellegrini, P., Rodriguez, J., 2013. Single European Sky and Single European Railway Area: a system level analysis of air and rail transportation. Transp. Res.
Part A 57, 64–86.
Rebollo, J.J., Cruz, A., 2009. Hybrid demand and capacity balance model for the future air traffic management concept of operations. In: Eighth USA/Europe
Air Traffic Management Research and Development Seminar.
Rios, J., Ross, K., 2007. Delay optimization for airspace capacity management with runtime and equity considerations. In: AIAA Guidance, Navigation and
Control Conf. Exhibit, Hilton Head, SC.
Santos, B., Antunes, A., Miller, E., 2008. Integrating equity objectives in a road network design model. Transp. Res. Rec.: J. Transp. Res. Board 2089, 35–42.
Schachtebeck, M., Schöbel, A., 2010. To wait or not to wait-and who goes first? Delay management with priority decisions. Transp. Sci. 44 (3), 307–321.
Schlechte, T., 2011. Railway Track allocation: Models and Algorithms PhD Thesis. Technischen Universität Berlin, Germay.
Sherry, L., Kumar, V., Mezhepoglu, B., Consiglio, M., 2008. Sensitivity of system performance & equity to user cooperation in the arrival flow: guidelines for
nextgen. In: Transportation Research Board 87th Annual Meeting, 08–0926.
UNESCAP, 2003. The Restructuring of Railways. United Nations Economic and Social Commission for Asia and the Pacific (UNESCAP).
Van de Velde, D., Nash, C., Smith, A., Mizutani, F., Uranishi, S., Lijesen, M., Zschoche, F., 2012. EVES-Rail: Economic Effects of Vertical Separation in the
Railway Sector. CER (Community of European Railways and Infrastructure Companies), Brussels, Belgium.
Vossen, T., Ball, M., Hoffman, R., Wambsganss, M., 2003. A general approach to equity in traffic flow management and its application to mitigating
exemption bias in ground delay programs. Air Traffic Control Quart. 11 (4), 277–292.
Wong, W.G., Han, B.M., Ferreira, L., Zhu, X.N., Sun, Q.X., 2002. Evaluation of management strategies for the operation of high-speed lines in China. Transp.
Res. Part A 36, 277–289.
Wu, D., Yin, Y., Lawphongpanich, S., Yang, H., 2012. Design of more equitable congestion pricing and tradable credit schemes for multimodal transportation
networks. Transp. Res. Part B: Methodol. 46 (9), 1273–1287.
Wu, J., Liu, M., Sun, H., Li, T., Gao, Z., Wang, D.Z., 2015. Equity-based timetable synchronization optimization in urban subway network. Transp. Res. Part C:
Emerging Technol. 51, 1–18.
Wu, J.H., Nash, C., 2000. Railway reform in China. Transp. Rev. 20 (1), 25–48.
Xu, X., Li, K., Yang, L., Ye, J., 2014. Balanced train timetabling on a single-line railway with optimized velocity. Appl. Math. Model. 38 (3), 894–909.
Yang, H., Zhang, X., 2002. Multiclass network toll design problem with social and spatial equity constraints. J. Transp. Eng. 128 (5), 420–428.
Yin, Y., Yang, H., 2004. Optimal tolls with a multiclass, bicriterion traffic network equilibrium. Transp. Res. Rec.: J. Transp. Res. Board 1882, 45–52.
Zhong, M., 2012. Models and solution algorithms for equitable resource allocation in air traffic flow management PhD Thesis. University of Maryland, USA.

Transportation Research Part C: Xiaojie Luan, Francesco Corman, Lingyun Meng

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Transportation Research Part C: Xiaojie Luan, Francesco Corman, Lingyun Meng

Uploaded by

Copyright:

Available Formats

Transportation Research Part C 80 (2017) 148–174

Contents lists available at ScienceDirect

Transportation Research Part C

Non-discriminatory train dispatching in a rail transport market

2.1. Network access and competition policies

2.2. Equitable allocation of capacity in timetabling

2.3. Real-time train dispatching without considering delay equity

2.4. Equitable control of air traffic and road traffic

Publications Objective(s) Degree of Model Solution algorithm

X. Luan et al. / Transportation Research Part C 80 (2017) 148–174

2.5. Statement of contribution

3.1. Explanations on relevant terms

double-track siding track

In the optimization models, following inputs are considered:

3.3. Mathematical model

Group I: Route choice constraints

Group II: Time-space network constraints

af ðof ; jÞ P ðef þ dprm

df ði; jÞ P af ði; jÞ 8f 2 F; ði; jÞ 2 Ef ð6Þ

df ði; jÞ ¼ af ðj; kÞ 8f 2 F; j 2 V n fof ; sf g; ði; jÞ 2 Ef ; ðj; kÞ 2 Ef ð7Þ

sf ði; jÞ ¼ df ði; jÞ af ði; jÞ 8f 2 F; ði; jÞ 2 Ef ð8Þ

Group V: Capacity constraints

Group VI: Train delays and delay equity

straints should be changed to ddstnCsc

4.2. Primary delays

for intercity trains, scale = 394, shape = 2.27, shift = 315;

Zbm Gdm Cl Htnc Htn Utl

4.3. Experimental schemes

Code Code meaning Code Symbol description

5. Experimental results and analysis

Fig. 3. Distribution of train delay cost.

5.1.4. Sensitivity of the solutions to the relative weights

Fig. 6. Computation time with different weights.

Fig. 8. Computation time with and without rerouting.

5.1.5. Comparison of solutions with and without rerouting

5.1.6. Impact of train delay cost per unit time

5.1.7. Solutions considering only consecutive delays

5.2. Summary of results analysis

5.3. Possible application of the proposed models in practice

6. Conclusions and future research

Appendix A. Summary of the studies addressing equity in air traffic control

Appendix B. The network layout and the detailed experimental results

Fig. 13. Layout of the Dutch railway network.

(continued on next page)

You might also like