Article

Transportation Research Record

1–12
Ó National Academy of Sciences:
Road Network Resilience: How to Transportation Research Board 2018
Reprints and permissions:
Identify Critical Links Subject to sagepub.com/journalsPermissions.nav
DOI: 10.1177/0361198118792115

Day-to-Day Disruptions journals.sagepub.com/home/trr

Pauline Gauthier1, Angelo Furno1, and Nour-Eddin El Faouzi1,2

Abstract
Disruptive events occur on road networks on a daily basis and affect traffic flow. Resilience analysis aims to assess the conse-
quences of such disruptions by quantifying the ability of a network to absorb and react to adverse events. In this paper, we
advance a methodological approach based on resilience stress testing and a dynamic mesoscopic simulator. We aim to identify
and rank the links most critical to the overall performance of the road network, taking into account the dynamic properties
of road traffic and focusing on day-to-day disruptions. As a metric to quantify road network performance in the presence of
such disruptions, we use the increase in overall travel cost. We then compare our approach with purely topological
approaches. We discuss the advantages and drawbacks of the different analyzed metrics. We prove that link ranking varies
greatly depending on the metric. Specifically, the proposed stress testing methodology can produce very accurate results by
taking into account demand and congestion, but requires computationally intensive simulations, being therefore prohibitive
even on medium-sized networks. Conversely, purely static topological metrics can be inaccurate if they do not take into
account traffic demand and network dynamics. Our study highlights the need for joining traditional traffic-agnostic topological
resilience analysis with demand-aware dynamic stress testing techniques.

Roads are vulnerable to being disrupted by adverse theory and network connectivity analysis to study resili-
weather conditions, human errors, or technological ence. Graph theory can make resilience analysis very effi-
breakdowns. Such day-to-day disruptions affect traffic cient in terms of computation costs, especially with the
flow and may profoundly impair the ability of the trans- widespread adoption of big data technologies and cloud-
portation network to guarantee basic mobility services as computing. However, the phenomenon of congestion—
well as management of emergency situations, thus caus- dynamic, spatio-temporal, demand dependent—is tradi-
ing fundamental economic and social strains. tionally not addressed in topological studies on road traf-
Resilience analysis aims to evaluate and predict the fic resilience (3, 4).
consequences of such disruptions, and has become a cru- In this context, our work aims to answer the following
cial research concern in recent years (1, 2). In the field of research questions:
road networks, one possible approach consists in identi-
fying the links that most strongly affect the overall per-  How is it possible to identify the links that are
formance. Operators and planners should be aware of most critical to the operation of the whole road
the consequences deriving from reduced capacities on network, especially with respect to day-to-day
links. They should focus their efforts on improving and disruptions?
maintaining such critical links, since they may cause the  Are topological metrics adequate to measure resili-
most severe consequences for traffic operations when dis- ence for road networks?
rupted. Thus, a methodological approach and a set of
resilience metrics are fundamental to identify critical
links. 1
Similar to other complex systems, a city road trans- University of Lyon, Lyon, France
2
Queensland University of Technology, Brisbane, Queensland, Australia
portation network can be modeled as a graph G = (N , L)
where road intersections are represented as nodes (N ) Corresponding Author:
and roads as links (L). This allows for the use of graph Address correspondence to Angelo Furno: angelo.furno@ifsttar.fr
2 Transportation Research Record 00(0)

 Which metrics are the most effective in assessing Sullivan et al. define road network robustness as the
road network resilience, taking into account their degree to which the network can function in the presence
dynamic, spatio-temporal, and demand-dependent of capacity disruptions on links. A robust road network
properties? can face disruptions on links with only slight increases in
overall network-wide travel costs. Conversely, a non-
The main contributions of this paper are the follow- robust road network is subject to substantial increases in
ing: (a) a methodology based on link-based stress testing costs (7).
and a dynamic mesoscopic simulator is proposed for In the rest of the paper, we tackle the problem of resi-
identifying critical links and quantifying road network lience assessment according to the perspective provided
robustness; (b) the methodology is compared with multi- by Woods’s second concept, that is, robustness (5), and
ple topology-based metrics to clearly identify their lim- the definition proposed by Sullivan et al (7). The paper
itations on a simplified road network; (c) the approach is proposes an approach based on stress testing for asses-
evaluated on a real-world network to identify and rank sing the impact of day-to-day disruptions on network
the most critical links in a realistic scenario; (d) guide- links by measuring overall travel costs, aiming to quan-
lines are provided to decide when one approach can be tify the ability of the whole road network to absorb such
preferred to another. an impact. We remark, as an additional contribution of
The paper is outlined as follows. The next section this paper, that our approach also allows for assessment
deals with the related literature. In the third section, we of resilience under the occurrence of unexpected and pos-
review different resilience metrics and describe our meth- sibly very rare events, which corresponds to Woods’s
odological approach, by presenting the stress testing third concept that is, graceful extensibility when surprise
approach and our dynamic mesoscopic simulator. The challenges boundaries (5). Our stress testing methodol-
fourth section describes our two case studies. The fifth ogy is also applied for evaluating the impact of sudden
section reports on the evaluation of our approach in the variations in travel demand on the overall travel costs.
considered case studies. A final discussion is then pre-
sented, while some directions for further work are pro-
posed in the final section. Approaches to Assessment of Resilience
The majority of the approaches aimed at quantifying
resilience are based on topology models and static net-
Literature Review work connectivity analysis (3, 4, 8). Among the plethora
of metrics proposed to perform connectivity analysis,
In this section, we firstly introduce the very general con-
betweenness centrality (BC) is traditionally the best
cept of resilience based on previous works, by adapting it
choice for traffic network analysis purposes, as it
to the context of road network analysis. We then present
expresses the frequency with which a point falls between
relevant road network resilience approaches for identify-
pairs of other points on the shortest paths connecting
ing critical links.
them (9). BC has therefore been widely adopted to assess
road resilience, by identifying topologically vulnerable
links and intersections (10, 11). Very recently, approaches
Resilience Definition
based on machine learning and big data solutions (12)
A wide diversity of definitions has been introduced in the have been proposed to significantly reduce computation
literature to characterize resilience. Woods identified four time of BC on several kinds of very large complex net-
major concepts (5): (i) resilience as rebound from trauma works (13).
and return to equilibrium; (ii) resilience as a synonym for Even though a large majority of studies focus on static
robustness; (iii) resilience as the opposite of brittleness, topological features of the network to identify its vulner-
that is, graceful extensibility when surprise challenges abilities, some authors have tried to join network topol-
boundaries; and (iv) resilience as network architectures ogy features to more dynamic traffic-related information.
that can sustain the ability to adapt to future surprises as Augmented definitions of BC take into account time-
conditions evolve. varying origin–destination (OD) travel demand (14–17),
Bruneau et al. (6) define resilience as the ability to: (i) as well as travel times (16, 18, 19) for shortest paths com-
mitigate hazards (robustness or pre-perturbation resili- putation. Such augmented metrics have been adopted for
ence); (ii) contain the effects of disasters during their traffic flow analysis and prediction (14, 15, 18, 19) and
occurrence (reactivity); and (iii) identify disruptions rap- traffic assignment (16) as well as network performance
idly as they occur and mobilize resources to recover an monitoring in the presence of extreme events (17). To the
acceptable traffic flow quickly (recovery or post-perturba- best of our knowledge, none of the previous approaches
tion resilience). has been evaluated with respect to day-to-day disruptions
Gauthier et al 3

(i.e., reduced link capacity), which represents one of the In the literature, two general approaches are com-
core aspects of our paper. Also, their applicability to resi- monly used to quantify resilience (26). The first is purely
lience assessment is still neither widespread nor fully topological and usually demand-insensitive. The second
understood, these metrics often being inaccurate in takes into account traffic network performance and
highly dynamic environments (20) and prohibitive to includes demand variability via simulation.
compute for large-scale networks (12). It is worth noting the lack of studies combining the
Another contribution of our paper is a methodology stress testing approach with the topological one. We ana-
based on stress testing, that is, pushing a system beyond lyze and compare the two different approaches to assess
its normal operational capacity and observing how it the resilience of road networks, and to determine their
responds to the applied stress. Stress tests have been advantages and drawbacks in assessing road network
widely used in banking (21, 22), medical (23, 24), and resilience. For the first approach, that is, the topological
hydro-geology domains (25). Also referred to as one, we consider and evaluate several metrics from graph
network-disruption analysis, this approach has been theory. As for the second approach, by means of a simu-
occasionally leveraged in the field of transportation to lator, we inject perturbations in the network and quan-
identify critical links in a road network, but it is still at a tify to what degree it can adjust to them.
very early stage (17, 26, 27). Sullivan et al. (7) set differ-
ent link-based capacity-disruption values for identifying
and ranking the most critical links and quantifying road
Methodology
network resilience. Jenelius and Mattsson (28) developed In the following, we report on the topological metrics
the notion of the importance of a link, which is a func- considered in our analysis as well as the stress testing
tion of the increase in travel time when the link is dis- technique proposed to perform demand-aware, dynamic
rupted. Their method is demand-aware as they weight disruption analysis.
travel time by demand. Stress testing allows intra-
network comparison; links are ranked based on their
contribution to the overall network resilience (29, 30). Graph Theory Metrics
Other authors have focused on identifying only the most The topological metrics used in this study are based on
critical nodes to be improved (30). Some studies use BC, originally proposed by Freeman (8). BC measures
probability-based models to calculate the likelihood that the importance of the generic link l of a graph by consid-
a network continues functioning after a given stress (31, ering the number of shortest paths that traverse it, and is
32). To the best of our knowledge, the large majority of defined as follows:
research works based on stress testing for assessing the
resilience of road networks mainly deal with disasters X sij (l)
BC(l) = , ð1Þ
and extreme events (17, 33, 34), instead of day-to-day sij
i6¼l6¼j
disruptions as targeted in this paper.
where:
Assessment of the Literature
 sij (l) is the number of shortest paths from node i
The definition and quantification of resilience varies
to node j that traverse link l;
greatly depending on the context and application
 sij is the total number of shortest paths from node
domain. No universal and totally agreed definition or
i to node j.
metric of resilience exists. Despite the plethora of work
on resilience in various domains, relatively small num-
In shortest path computation, links can be unweighted
bers of studies have targeted resilience of road networks
or weighted (e.g., in terms of the associated estimated
explicitly, and very few applications on real-world roads
travel time). In this study, we test and compare both
have been proposed. Our paper deals instead with a real
cases. We also consider multiple variants of the BC,
network in France, in the Paris agglomeration area.
reported in the following, to model different aspects of a
The concept of resilience can be divided into two
road traffic network.
parts: pre-perturbation resilience (robustness), and post-
perturbation resilience. In this study, focused on traffic
modeling and analysis, we evaluate pre-perturbation resi- BC for Entry and Exit Nodes Only. We propose to use an
lience only. We do not account for post-perturbation alternative definition of BC consisting in calculating the
socio-technical actions. As the aim of this paper is to shortest paths from entry to exit nodes only. (The com-
rank and identify the most critical links in a given net- putation of BC on a subset of nodes [entries and exits] is
work with respect to day-to-day disruptions, we focus on based on the function edge_betweenness_centrality_subset
intra-network comparisons. from the NetworkX Python library.) This definition
4 Transportation Research Record 00(0)

introduces two advantages: computation time is reduced,  c0ij is the mean travel cost from nodes i to j in the
and the definition seems more realistic from a demand- base case, i.e., without disruption.
aware perspective, since individuals tend to start and fin-
ish their trips over a subset of intersections. This corre-
sponds to the OD representation of the traffic demand. Traffic Model and Algorithms
The formula is the same as Equation 1 with some To model traffic dynamics we use a dynamic mesoscopic
exceptions: simulator based on the Lighthill-Whitham-Richards
model (35, 36) and implemented in Matlab by our
 i is selected from the entry-nodes subset, i.e., inter- research group (37, 38). The Lighthill-Whitham-
sections used by vehicles to enter the network; Richards model is formulated in Lagrangian-space coor-
 j is selected from the exit-nodes subset, i.e., inter- dinates and uses both Lagrangian and Eulerian observa-
sections used by vehicles to leave the network. tions. It represents individual vehicles but only records
their transit times at network nodes. A dynamic traffic
In conclusion, we consider four different formulations of assignment procedure distributes vehicles along all the
the BC: possible alternative paths in the network, according to
the traffic conditions at the moment the vehicle is gener-
 Unweighted BC (BC) ated. More precisely, travel times on all paths are calcu-
 Travel-time weighted BC (TTWBC) lated based on traffic flow, and the vehicle chooses the
 Unweighted BC on entry/exit nodes only (BC path that requires the smallest travel time. The following
entries–exits) parameters have to be specified before running simula-
 Travel-time weighted BC from entry to exit nodes tions: simulation duration, OD demands, and link capa-
only (TTWBC entries–exits) cities. This simulator is adequate for our stress testing
approach: as opposed to static topological indicators, it
Spatio-temporal traffic properties and phenomena, includes traffic dynamic properties such as demand, con-
like demand, congestion and dynamic re-routing, are gestion, traffic-based route assignment, dynamic shortest
typically not addressed in graph-based models. Thus, path computation, and queues. Moreover, travel costs
graph-based metrics are usually incapable of capturing are calculated for each vehicle and can be extracted eas-
these aspects in turn. ily to compute network performance metrics, such as
importance from Equation 2.

Demand-Sensitive Metric Stress Testing

Jenelius et al. (2) introduced the demand-aware metric of The aim of our stress testing methodology is to identify the
importance (I) to characterize transportation vulnerabil- most critical links in the road network and to assess their
ity. This metric allows measurement of network perfor- resilience by considering the dynamic, spatio-temporal, and
mance loss by using travel costs weighted by the traffic demand-dependent properties of network traffic.
demand. Such metric is adequate for our methodological Stress testing can be leveraged to quantify the adverse
approach as it includes demand and the dynamic phe- impacts associated with a reduction of capacity on specific
nomenon of congestion (i.e., travel costs increase when links. Disruptive road events such as flooding, obstacles
traffic is congested). This metric uses a generic notion of on the road, and traffic accidents are likely to reduce the
travel cost, which can be specified depending on the capacity of a given link and negatively affect network per-
study context and aim. In this paper, we define travel formance. Measuring network performance loss when
cost as travel time divided by travel distance (in seconds/ reducing the capacity of a given link provides the critical-
kilometer). The importance of a link l is the following: ity of this link to the operation of the whole network.
P P Therefore, stress testing is an adequate methodological
d 0
i j6¼i xij (cij (l)  cij ) approach to identify and rank the most critical links. It
I(l) = P P , ð2Þ
i j6¼i xij
captures the relative importance of the disrupted link to
the other links and assess the overall resilience of the
where: whole road network from an intra-network comparison.
Our methodology for road network stress testing is
 xij is the demand from origin node i to destination composed of the following steps:
node j (number of vehicles);
1. Simulating disruptive road events:
 cdij (l) is the mean travel cost from origin node i to
destination node j when link l is disrupted at We propose two strategies to perform this step. In the
level d; first one, we simulate day-to-day road disruption as link
Gauthier et al 5

capacity drops. The capacity-disruption level is defined as network. To this purpose, we use the notion of road net-
the reduction in link capacity, expressed as a fraction of work performance, measured via the importance metric
the original one. In many studies, the capacity-disruption of Equation 2. Specifically, we consider travel time
level is total, that is, 100% of the original value, which increase divided by travel distance as a measure of cost.
means that the capacity of the link is reduced to 0 vehi- Below, we provide the formulation of the overall perfor-
cles per hour (10). The link is then completely removed mance loss (PL), based on the notion of importance,
from the road network. However, a 100% capacity- when a link l is disrupted:
disruption level does not accurately reflect the actual link
capacity resulting from frequent day-to-day disruptions X
n
cd (l)  c0
v v
PL(l, d) = ð4Þ
or minor events (e.g., number of lane reductions, adverse n
v=1
weather, etc.) that can affect the network. That is why
we gradually reduce the capacity to analyze the evolution where:
of the performance depending on the capacity-disruption
level. We consider five possible capacity-disruption lev-  PL(l, d) is the overall performance loss when link l
els, denoted as d, on each examined link, that is, is disrupted at level d (seconds/kilometer);
d 2 f0%, 20%, 40%, 60%, 80%g. Therefore, we can con-  cdv (l) is the travel cost of vehicle v when link l is dis-
sider the following equation to compute the maximum rupted at level d (seconds/kilometer);
capacity of the generic link l in the presence of a  c0v is the travel cost of vehicle v in the base case
capacity-disruption level d: (seconds/kilometer);
 n is the number of vehicles in the network.
qd (l)
d = 100  (1  max ), ð3Þ
q0max (l)
where: 3. Analyzing the results:

 d is the capacity-disruption level applied to link l For each link we know the performance loss correspond-
(percentage) with d 2 f0, 20, 40, 60, 80g; ing to the considered capacity-disruption level (i.e., 0%,
 qdmax (l) is the capacity of link l when it is disrupted 20%, 40%, 60%, and 80% of the original link capacity).
at level d (in vehicles/hour); As an example, PL(2, 40%) represents the overall perfor-
 q0max (l) is the capacity of link l in the base case (in mance loss on link 2 when it is disrupted at 40% of its
vehicles/hour); initial capacity. In the evaluation section, we present and
discuss the overall performance loss depending on the
As a second strategy to simulate disruptive road events, capacity-disruption level of the link according to two dif-
we consider increases in the traffic demand on ferent strategies.
specific entry/exit nodes of the network. By this
approach, it is possible to simulate exceptional situations
like city evacuations following extreme events (e.g., Traffic Robustness Index
flooding, attacks, etc.) which typically put significant The stress testing methodology reported in the previous
strain on the road infrastructure and result in total con- section allows us to compute the overall performance loss
gestion. This strategy consists in changing the OD for each link of the network with respect to five different
matrix, that is, increasing the traffic flow from given capacity-disruption levels. To compare and identify criti-
entries, and comparing the stress testing results with cal links, we need a unique value of criticality associated
another demand level. to each link. To this purpose, a global metric is required
Based on the selected strategy for disruptive road to aggregate the performance loss values in the five dif-
events, we set the parameters of our mesoscopic simula- ferent capacity-disruption levels. We propose the stress
tor (e.g., link capacity, traffic demand) and we simulate test criticality (STC) metric, defined as follows, for the
the network in the specific setting. For both strategies, generic link l:
travel costs are collected for all vehicles in order to com-
ð
pute the performance metrics described in the following
point. STC(l) = PL(l, d), ð5Þ
d

2. Computing overall performance loss: where:

This step is about quantifying the consequences of the  STC(l) is STC when link l is stress tested (seconds/
simulated disruptive event on the operation of the whole kilometer)
6 Transportation Research Record 00(0)

Figure 1. Simple virtual network. Network structure (a) and OD travel demand (b, c) for the two considered strategies.

 d is the capacity-disruption level of link l flow directions. The network has been tested with two
(percentage) different demand levels reported as OD matrices in
 PL(l, d) is the overall performance loss (seconds/ Figure 1b and c.
kilometer)

We use the trapezoid rule to approximate the integral DIRIF: A Real-World Road Network
in Equation 5 from the (five) overall performance values The DIRIF network is situated in the south of Paris,
computed on link l. France, and includes 868 links and 827 nodes (657 inter-
sections, 86 entries, and 84 exits). Its roads are mostly
highways. Each simulation duration is fixed to 15 min-
Case Studies utes. Since the network is much bigger than the virtual
The methodology and metrics described in the previous one described above, and traffic demand is extremely
sections have been evaluated in two different case studies. low on some links, we specify a higher simulation dura-
The first one is related to a simple virtual network, used tion to ensure that enough vehicles travel through the
as a basic testbed for our approach. The second one is a whole network and to collect a significant number of
real road network in France, which we use to confirm the travel cost observations. Simulation is performed with
validity of our results in a realistic scenario and to sup- real demand data from 9:00 a.m. to 9:15 a.m., that is, the
port the discussion on both simulation-based stress test- morning peak-time, in order to increase the probability
ing and topological metrics. of observing some performance loss in our stress tests.
The network is graphically presented in Figure 3a and b.

A Simple Virtual Road Network

Evaluation
This network is composed of eight nodes (4 road inter-
sections, 2 entries, and 2 exits) and nine links. The dura- In this section, we present the results of our stress testing
tion of each simulation is fixed to 10 minutes. Figure 1a methodology and discuss the link ranking derived from
depicts the network structure with numbered links and the different selected metrics. We show that the ranking
Gauthier et al 7

Figure 2. Performance evaluation on the simple virtual network.

of critical links on the same network can vary signifi- measures of STC that result from the two strategies
cantly when different indicators are used, thus proving above are distinguished as STC A and STC B, respec-
that simple modifications of one centrality indicator can tively. STC A and B are calculated with the same for-
have a relevant impact on the capacity of the metric to mula, but different parameters are set before stress
capture different facets of resilience. Moreover, we dis- testing.
cuss the advantages and drawbacks of each different First we discuss the results of strategy A. In our simu-
approach in assessing road network resilience, and pro- lations, we applied sequentially five capacity-disruption
vide guidelines that can be helpful towards the definition levels (i.e., 0%, 20%, 40%, 60%, 80%) to each link. We
of a new enhanced centrality metric. then measured the network-wide performance loss (i.e.,
Equation 4) consequent to the disruption applied to the
link. The overall performance loss from our stress tests is
Application to a Simple Virtual Network reported on the y-axis of Figure 2a, while the correspond-
In the scenario of the simple network described in the ing capacity-disruption levels (d) are reported on the x-
previous section, we measured STC and all of the pro- axis. Results for different links are depicted with different
posed topological metrics on all the links. To perform colors and markers in the figure, using a linear interpola-
stress testing, we used both strategies described above, tion. For readability, the figure only reports the five most
that is, link capacity drop (referred to as A in the follow- critical links (i.e., those with the highest overall perfor-
ing) and traffic demand increase (referred to as B). The mance loss).
8 Transportation Research Record 00(0)

Intuitively, a link capacity drop translates into an the longest one, thus demanding more time to be tra-
increase of network-wide travel cost. velled than the other links. Metrics like BC and BC from
According to the definition provided in Equation 5, entries to exits are not weighted, that is, all links are val-
the STC of a given link corresponds to the area below its ued equally, and consequently they are unable to grasp
curve. Thus, in strategy A, link 9 is the most critical to this important aspect. In contrast, links with high travel
the operation of the whole road network, followed by times are not considered critical by the analyzed weighted
links 5, 1, 2, and 8. By using link ranking from STC as a approaches, because they are not often part of shortest
baseline, we compare in the following the other link rank- paths. The same consideration applies to link 7, which is
ings as derived from the different selected topological the second-longest link of the network.
metrics. Figure 2c reports on such link rankings for both Finally, it is worth noting that BC values are often
STC and the whole set of topological metrics. clustered. In particular, the BC from entries to exits has
As a preliminary consideration, it can be observed many equal values and only two link ranks. Traditional
that the rankings of critical links on the same network topological metrics appear to have very limited capability
may dramatically change depending on the metric, due to discriminate link criticality at a fine level. In this case,
to the different properties of the network captured by STC does not differentiate all links either, but this is due
each of them. As an example, link 5 is on top of all the to capacity-disruption levels. For links 8, 6, 4, 7, and 3
topological rankings whereas, with STC, it is ranked sec- which are all ranked in fifth place with the same value,
ond, below link 9. The top-rank of link 5 by all the topo- both capacity-disruption and demand levels are not high
logical metrics can be explained by the large number of enough to observe a significant performance loss com-
shortest paths traversing this link: e.g., paths (4, 5), (7, pared with the base case. For example link 8 capacity dis-
5), (7, 5, 8), (7, 5, 9), (6, 4, 5), (7, 5, 8) are all shortest ruption does not affect the overall network performance
paths. (see Figure 2a). The overall travel cost remains at its base
The different ranking issued by STC A can be easily case value. That is why some links have the same critical-
explained. If link 5 is disrupted, alternative paths through ity value. STC differentiation between links then depends
link 2 exist for all individuals heading to exit 1 or exit 2. on capacity-disruption and demand levels.
Conversely, when link 9 is disrupted, no alternative path To investigate this aspect further, we use our second
exists for users willing to travel to exit 2 from both entry stress testing strategy B, that is, with different traffic
1 and entry 2, thus resulting in serious congestion and demand as reported in Figure 1c. Results are shown in
increased travel time for all individuals heading to exit 2. Figure 2b and c (i.e., STC B). Link ranking changes sig-
Additionally, traffic demand for exit 2 is very high (see nificantly when different demand levels are used. As an
OD matrix in Figure 1b). This explains why link 5 is example, Link 8 becomes the most critical link, whereas
more critical than link 9 in terms of topology, but less in the previous case it involved no performance loss.
critical than link 9 when considering demand data, as This is due to the large increase in demand level associ-
made possible by our stress testing methodology (based ated with exit 1, which is directly connected to link 8 (see
on dynamic simulations) and captured by the related cri- Figure 1a).
ticality metric. This simple test clarifies how traditional Takeaways: Critical link ranking is highly variable as
demand-agnostic approaches may fail in properly rank- different approaches are used. Resilience analysis via topo-
ing edge criticality. logical metrics is limited in the sense that such metrics do
Our simple test also shows that alternative paths may not usually take into account traffic demand and network
become shortest paths of the network as links are dis- re-configurations following disruptive events. Conversely,
rupted by adverse event, thus attracting traffic flow pre- the simulation-based stress testing approach is able to cap-
viously directed through the disrupted links. This ture these aspects thus providing more realistic rankings
represents another fundamental aspect that is impossible via the proposed performance loss metric. Stress testing
to capture with a static graph-based approach. However, can also be used to compare different road networks and
this does not necessarily mean that topological metrics sub-networks by analyzing their response to similar stres-
are not good resilience indicators, but rather that road ses. Travel-time weighted BC produces better estimations
graph modeling should include a dynamic component of link criticality compared with unweighted BC, which
(e.g., edge weights), and that BC metrics should be rap- treats all links equally.
idly re-computed after relevant network disruptions.
Another striking difference worth analyzing concerns
link 2: considered as one of the most critical ones accord- Application to a Real Road Network
ing to the BC metric from entries to exits, it is the least To confirm the results of our previous analysis in a rea-
critical one for the TTWBC and the TTWBC from listic scenario, we considered the Paris DIRIF road net-
entries to exits. The peculiarity of link 2 is its length; it is work, described above. Given the large size of this
Gauthier et al 9

Figure 3. The DIRIF road network in Paris agglomeration (a) and (b). Evaluation results (c).

network and the high computation time associated with simple network case. Therefore, we performed stress tests
each network simulation (stress testing one link from on a limited set of representative links: the three links
such a large network with five capacity drops takes more with the highest demand, the three with the highest BC,
than one hour on an Intel Xeon E3 CPU equipped with and three randomly selected edges with BC in three
8 GB of RAM running a Matlab implementation of the classes of values (high, medium, and low). We discuss in
mesoscopic simulator, properly configured to handle the the following only our simulations related to strategy A.
DIRIF network in Scenarios A and B), it was prohibitive (Simulations for strategy B were in line with the results
to perform an exhaustive stress test analysis as in the reported in the previous section and are not discussed
10 Transportation Research Record 00(0)

due to space limitation.) Figure 3c reports the actual val- be chosen, since it produces reliable results by taking into
ues of the considered metrics for the analyzed links. account traffic demand and congestion phenomena. The
Unlike Figure 1, we do not report metric rankings but drawback is that it requires many computationally inten-
actual values of the metrics for each analyzed link. This sive simulations, thus being recommended only in appli-
is motivated by the impossibility of obtaining the full cation scenarios that allow for larger computation time,
ranking for performance loss (i.e., STC). or that address small-sized (sub-)networks. Conversely,
Consistently with our previous analysis on the simple in domains with very stringent requirements on response
network, Figure 3c shows that rankings of critical links time (e.g., on-line vulnerability monitoring), topological
vary significantly. As an example, links 95, 93, 94 have a indicators could be the only valuable option. However, it
very high value of STC A, whereas the topological is worth remarking that efficient solutions are still
metrics rate them much less critical than links 802, 803, required to compute these metrics on very large networks
607. As pointed out in the previous section, the STC A within reasonable computation time. For future work,
ranking appears to be more realistic since it captures the we are currently working on the switch from a Matlab
higher criticality of links 95, 93, 94 due to the associated implementation of our simulation-based tools to a new
higher demand (not reported due to space limitations). implementation based on faster programming languages
On the small link subset considered in our analysis, and approaches explicitly designed for big data process-
taking into account travel times (TTWBC and TTWBC ing and real-time computation (e.g., Python and Scala
entries–exits) does not significantly change rankings, on top of the Spark processing framework). In particu-
since link lengths (and therefore the average travel times) lar, we believe that implementing real-time advanced
happen to be very similar on all considered links. Finally, solutions for data-driven, on-line monitoring of road
it is worth noting that in the DIRIF network BC values traffic resilience will constitute a fundamental research
(especially in the entries/exits variations) tend to cluster problem to investigate.
significantly themselves (i.e., many edges have very simi- We advance that, in order to improve road network
lar values of BC), thus exhibiting a lower discriminant resilience analysis, future research work is needed that
power than in the case of the simple virtual network. should consider joining graph-based approaches with
Takeaways: In a real-world scenario, stress testing demand-aware dynamic stress testing techniques. In this
proved to be a realistic and reliable approach to evaluate context, we believe that a further improvement with
network resilience. Our evaluation confirms the importance topological metrics could be achieved by modeling the
of traffic demand and network dynamics for fine-grained road network as a dynamic graph, whose link weights
ranking of the most vulnerable road network links. Stress may change over time depending on actual traffic condi-
testing has however the drawback of requiring high execu- tions and both structural and performance-related net-
tion times due to computationally intensive network work properties (e.g., road capacity, real-time traffic
simulations. information, etc.).
Finally, we argue that future work should also con-
sider area-wide disruptions in addition to single link-
Discussion and Perspectives based disruptions, especially in the light of measuring the
From the previous results, we summarize in the following impact of extreme events.
a few guidelines for properly characterizing critical links
with respect to day-to-day disruptions, by means of an
Conclusion
intra-network approach in different application contexts.
Firstly, if resilience has to be evaluated in a relatively Identifying links critical to the overall network perfor-
static context (e.g., network maintenance or planning), mance is part of road network resilience and intra-
BC and TTWBC appear to be adequate. In particular, if network analysis. To this purpose, we have analyzed in
data about demand and travel times are not available, this paper several topological metrics based on BC and
we recommend BC, BC from entries and exits, and BC proposed a stress testing approach exploiting a dynamic
on all paths from entries and exits. These indicators do simulator. Stress testing appears to be a very promising
not require special knowledge of network performance solution for resilience analysis, allowing for measurement
and demand data, but only the basic network topology of resilience in terms of the overall performance loss of
(links and intersections). If traffic demand is the only the whole network consequent to simulated link
missing information, STC, TTWBC, and TTWBC from disruptions.
entries to exits should be preferred, since they also take Our analysis shows that link ranking varies greatly
into account travel time information. when different metrics are used. As opposed to purely
If the goal is instead to achieve a more accurate char- topological metrics, the proposed stress testing approach
acterization of network resilience, stress testing should takes into account demand levels and dynamic
Gauthier et al 11

characteristics of road traffic. However, it requires more Analysis Zones. Transportation Research Record:
computation time and data than traditional graph-based Journal of the Transportation Research Board, 2011. 2256:
metrics. The choice of a relevant metric for assessing 16–24.
road network resilience should depend on the context 10. King, D., A. Shalaby, and P. Eng. Performance Metrics
and the specific application requirements. and Analysis of Transit Network Resilience in Toronto.
Presented at 95th Annual Meeting of the Transportation
Merging static topological metrics and demand-based
Research Board, Washington, D.C., 2016.
approaches could be of further research interest. It could
11. Berche, B., C. von Ferber, T. Holovatch, and Y. Holo-
be relevant to adopt dynamic graphs modeling, using link vatch. Resilience of Public Transport Networks Against
weights to include dynamic information on the network. Attacks. The European Physical Journal B, Vol. 71, No. 1,
In such an approach, topological metrics should be com- 2009, pp. 125–137.
puted dynamically by means of efficient quasi real-time 12. Furno, A., N. E. El Faouzi, R. Sharma, and E. Zimeo.
solutions. Two-level Clustering Fast Betweenness Centrality Compu-
tation for Requirement-driven Approximation. IEEE Big
Author Contributions Data 2017 Conference, 2017.
The authors confirm contribution to the paper as follows: study 13. Furno, A., N. E. El Faouzi, R. Sharma, and E. Zimeo.
conception and design: Gauthier, Furno, El Faouzi; data col- Reducing Pivots of Approximated Betweenness Computa-
lection: Gauthier, Furno, El Faouzi; analysis and interpretation tion by Hierarchically Clustering Complex Networks. In
of results: Gauthier, Furno, El Faouzi; draft manuscript pre- International Conference on Complex Networks and their
paration: Gauthier, Furno, El Faouzi. All authors reviewed the Applications, Springer, 2017, pp. 65–77.
results and approved the final version of the manuscript. 14. Jayasinghe, A., K. Sano, and H. Nishiuchi. Explaining
Traffic Flow Patterns Using Centrality Measures. Interna-
tional Journal for Traffic & Transport Engineering, Vol. 5,
References No. 2, 2015, pp. 134–149.
1. Berdica, K. An Introduction to Road Vulnerability: What 15. Kazerani, A., and S. Winter. Modified Betweenness Central-
Has Been Done, Is Done and Should Be Done. Transport ity for Predicting Traffic Flow, Vol. 2, 2009.
Policy, Vol. 9, No. 2, 2002, pp. 117–127. 16. Puzis, R., Y. Altshuler, Y. Elovici, S. Bekhor, Y. Shiftan,
2. Jenelius, E., T. Petersen, and L.-G. Mattsson. Importance and A. Pentland. Augmented Betweenness Centrality for
and Exposure in Road Network Vulnerability Analysis. Environmentally Aware Traffic Monitoring in Transporta-
Transportation Research Part A: Policy and Practice, Vol. tion Networks. Journal of Intelligent Transportation Sys-
40, No. 7, 2006, pp. 537–560. tems, Vol. 17, No. 1, 2013, pp. 91–105.
3. Di Gangi, M., and A. S. Luongo. Measures of Network 17. Wisetjindawat, W., A. Kermanshah, S. Derrible, and M.
Vulnerability Indicators for Risk Evaluation and Exposure Fujita. Stochastic Modeling of Road System Performance
Reduction. Environmental Health Risk III, Vol. 9, 2005, During Multihazard Events: Flash Floods and Earth-
p.1151. quakes. Journal of Infrastructure Systems, Vol. 23, No. 4,
4. Tu, Y., C. Yang, and X. Chen. Methodology for Evaluat- 2017, 04017031.
ing and Improving Road Network Topology Vulnerability. 18. Altshuler, Y., R. Puzis, Y. Elovici, S. Bekhor, and A. B.
In 2010 International Conference on Intelligent Computation Pentland. Augmented Betweenness Centrality for Mobility
Technology and Automation (ICICTA), Vol. 2, IEEE, Prediction in Transportation Networks. Finding Patterns of
2010, pp. 664–669. Human Behaviors in Network and Mobility Data (NEMO),
5. Woods, D. D. Four Concepts for Resilience and the Impli- 2011.
cations for the Future of Resilience Engineering. Reliability 19. Cheng, Y.-Y., R. Ka-Wei Lee, E.-P. Lim, and F. Zhu.
Engineering & System Safety, Vol. 141, 2015, pp. 5–9. Measuring Centralities for Transportation Networks
6. Bruneau, M., S. E. Chang, R. T. Eguchi, G. C. Lee, T. D. Beyond Structures. Applications of Social Media and Social
O’Rourke, A. M. Reinhorn, M. Shinozuka, K. Tierney, Network Analysis, Springer, 2015, pp. 23–39.
W. A. Wallace, and D. Von Winterfeldt. A Framework to 20. Duan, Y., and F. Lu. Robustness of City Road Networks
Quantitatively Assess and Enhance the Seismic Resilience at Different Granularities. Physica A: Statistical Mechanics
of Communities. Earthquake Spectra, Vol. 19, No. 4, 2003, and its Applications, Vol. 411, 2014, pp. 21–34.
pp. 733–752. 21. Haldane, A. Why Banks Failed the Stress Test. BIS Review,
7. Sullivan, J. L., D. C. Novak, L. Aultman-Hall, and D. M. Vol. 18, 2009, 2009p.
Scott. Identifying Critical Road Segments and Measuring 22. Schuermann, T. Stress Testing Banks. International Journal
System-wide Robustness in Transportation Networks with of Forecasting, Vol. 30, No. 3, 2014, pp. 717–728.
Isolating Links: A Link-based Capacity-reduction 23. Goldschlager, N., A. Selzer, and K. Cohn. Treadmill Stress
Approach. Transportation Research Part A: Policy and Tests as Indicators of Presence and Severity of Coronary
Practice, Vol. 44, No. 5, 2010, pp. 323–336. Artery Disease. Annals of Internal Medicine, Vol. 85, No.
8. Freeman, L. C. A Set of Measures of Centrality Based on 3, 1976, pp. 277–286.
Betweenness. Sociometry, Vol. 40, No. 1, 1977, pp. 35–41. 24. Baertschi, S. W., K. M. Alsante, and R. A. Reed. Pharma-
9. Zhang, Y., X. Wang, P. Zeng, and X. Chen. Centrality ceutical Stress Testing: Predicting Drug Degradation. CRC
Characteristics of Road Network Patterns of Traffic Press, 2016.
12 Transportation Research Record 00(0)

25. Zhang, L. M., L. Gao, S. Y. Zhou, R. W. M. Cheung, and Disaster. In: Murray A.T., Grubesic T.H. (eds) Critical
S. Lacasse. Stress Testing Framework for Managing Land- Infrastructure. Advances in Spatial Science. Springer, Ber-
slide Risks Under Extreme Storms. Workshop on World lin, Heidelberg, 2007, pp. 221–241.
Landslide Forum, Springer, 2017, pp. 17–32. 33. Donovan, B., and D. B. Work. Using Coarse GPS Data to
26. Mattsson, L.-G., and E. Jenelius. Vulnerability and Resili- Quantify City-scale Transportation System Resilience to
ence of Transport Systems – A Discussion of Recent Extreme Events. arXiv preprint arXiv:1507.06011, 2015.
Research. Transportation Research Part A: Policy and 34. Mayada, O., A. Mostashari, and R. Nilchiani. Assessing
Practice, Vol. 81, 2015, pp. 16–34. Resilience in a Regional Road-based Transportation Net-
27. Clarke, J., and E. O’Brien. A Multi-hazard Risk Assess- work. International Journal of Industrial and Systems Engi-
ment Methodology, Stress Test Framework and Decision neering, Vol. 13, No. 4, 2013, pp. 389–408.
Support Tool for Transport Infrastructure Networks. 35. Lighthill, M. J., and G. B. Whitham. On Kinematic Waves.
Transportation Research Procedia, Vol. 14, 2016, pp. II. A Theory of Traffic Flow on Long Crowded Roads. In
1355–1363. Proceedings of the Royal Society of London A: Mathemati-
28. Jenelius, E., and L.-G. Mattsson. Road Network Vulner- cal, Physical and Engineering Sciences, Vol. 229, The Royal
ability Analysis of Area-covering Disruptions: A Grid- Society, 1955, pp. 317–345.
based Approach with Case Study. Transportation Research 36. Richards, P. I. Shock Waves on the Highway. Operations
Part A: Policy and Practice, Vol. 46, No. 5, 2012, pp. Research, Vol. 4, No. 1, 1956, pp. 42–51.
746–760. 37. Duret, A., L. Leclercq, and N.-E. El Faouzi. Data Assimi-
29. Scott, D. M., D. C. Novak, L. Aultman-Hall, and F. Guo. lation Using a Mesoscopic Lighthill–Whitham–Richards
Network Robustness Index: A New Method for Identifying Model and Loop Detector Data: Methodology and Large-
Critical Links and Evaluating the Performance of Trans- scale Network Application. Transportation Research
portation Networks. Journal of Transport Geography, Vol. Record: Journal of the Transportation Research Board,
14, No. 3, 2006, pp. 215–227. 2016. 2560: 26–36.
30. Berdica, K., and M. Lars-Göran. Vulnerability: A Model- 38. Laval, J. A., and L. Leclercq. The Hamilton–Jacobi Partial
based Case Study of the Road Network in Stockholm. In: Differential Equation and the Three Representations of
Murray A.T., Grubesic T.H. (eds) Critical Infrastructure. Traffic Flow. Transportation Research Part B: Methodolo-
Advances in Spatial Science. Springer, Berlin, Heidelberg, gical, Vol. 52, 2013, pp. 17–30.
2007, pp. 81–106.
31. Zhen-Ping, D., and A. Nicholson. Degradable Transporta- The Standing Committee on Critical Transportation
tion Systems: Sensitivity and Reliability Analysis. Trans- Infrastructure Protection (ABR10) peer-reviewed this paper (18-
portation Research Part B: Methodological, Vol. 31, No. 3, 03204).
1997, pp. 225–237.
32. Church, R., and M. P. Scaparra. Analysis of Facility Sys-
tems Reliability When Subject to Attack or a Natural