A survey on modeling and simulation of vehicular networks: Communications, mobility, and tools

Computer Communications 43 (2014) 1–15

Contents lists available at ScienceDirect

Computer Communications journal homepage: www.elsevier.com/locate/comcom

A survey on modeling and simulation of vehicular networks: Communications, mobility, and tools Francisco J. Ros ⇑, Juan A. Martinez, Pedro M. Ruiz Department of Information and Communications Engineering, University of Murcia, Murcia, Spain

a r t i c l e

i n f o

Article history: Available online 15 February 2014 Keywords: Vehicular networks Mobility model Communication model Simulation

a b s t r a c t Simulation is a key tool for the design and evaluation of Intelligent Transport Systems (ITS) that take advantage of communication-capable vehicles in order to provide valuable safety, traffic management, and infotainment services. It is widely recognized that simulation results are only significant when realistic models are considered within the simulation toolchain. However, quite often research works on the subject are based on simplistic models unable to capture the unique characteristics of vehicular communication networks. If the implications of the assumptions made by the chosen models are not well understood, incorrect interpretations of simulation results will follow. In this paper, we survey the most significant simulation models for wireless signal propagation, dedicated short-range communication technologies, and vehicular mobility. The support that different simulation tools offer for such models is discussed, as well as the steps that must be undertaken to fine-tune the model parameters in order to gather realistic results. Moreover, we provide handy hints and references to help determine the most appropriate tools and models. We hope this article to help prospective collaborative ITS researchers and promote best simulation practices in order to obtain accurate results. Ó 2014 Elsevier B.V. All rights reserved.

1. Introduction and motivation In the last years, the development of collaborative Intelligent Transport Systems (ITS) has been a focus of deep study. Communication-capable vehicles enable a plethora of valuable services targeted at improving road safety, alleviating traffic congestion, and enhancing the overall driving experience [1]. To support these services, different standardization bodies have defined the networking architecture of ITS stations [2,3], including vehicles’ on-board units (OBU) and infrastructure’s road-side units (RSU). Multiple network interface cards of different communication technologies coexist within a same OBU or RSU to support different use cases. Thus, cellular or broadband wireless interfaces provide the vehicle with connectivity to the infrastructure network (V2I), while dedicated short-range communications (DSRC) in the 5.9 GHz frequency band allow for vehicle-to-vehicle (V2V) and vehicleto-roadside (V2R) data transfers. In these cases, vehicles form a vehicular ad hoc network (VANET) in which collaborative services can be deployed. The design and evaluation of ITS services and communication protocols is cumbersome, given the scale of vehicular networks ⇑ Corresponding author. Tel.: +34 868884644; fax: +34 868884151. E-mail address: [email protected] (F.J. Ros). http://dx.doi.org/10.1016/j.comcom.2014.01.010 0140-3664/Ó 2014 Elsevier B.V. All rights reserved.

and their unique characteristics. Some small-scale testbeds have been deployed [4,5] as a proof of concept, but results from small experiments cannot be extrapolated to real networks. In very few cases, field operational tests (FOT) have been implemented to evaluate an ITS platform under real traffic conditions [6]. However, given the high amount of required resources to deploy a FOT, it is only an option for a limited number of researchers and practitioners on the field. As an alternative, simulation models feature a good trade-off between the realism of results and the flexibility of target networks under study. Not surprisingly, most research on VANET and collaborative ITS rely on simulation as the main tool for design and evaluation. Different fields stitch together in the development of collaborative ITS, including wireless communications and civil traffic engineering. In order to gather significant simulation results, a good understanding of the different models involved is required. It is widely recognized that simplistic wireless communication models lead to unreasonable results that do not match reality [7]. In addition, vehicular mobility patterns greatly differ from other networking scenarios and they need specific models to capture the characteristics of vehicles’ movements. Different mobility patterns have a distinct impact onto simulation results [8,9]. Therefore, an ITS researcher must be aware of the different models that can be

2

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

employed for each aspect of the simulation environment and how model parameters should be tuned to obtain realistic results. While, traditionally, mobility and network simulators have been developed by different communities for different end users, both tools merge in collaborative ITS. It is of paramount importance that simulations account for realistic vehicles’ movements as generated by mobility simulators. Then, network simulators must provide realistic simulation of communication models as the vehicles move according to a given traffic pattern. In some cases, an ITS application influences the mobility of vehicles. For instance, a traffic management service might indicate some vehicles to follow an alternative route to avoid a congested road. To support such scenarios, integrated mobility and network simulations are necessary and can be provided by interfacing existing tools or developing new ones. In this paper, we survey the most significant approaches for wireless modeling and mobility modeling in vehicular networks. Specifically, we describe the models that have been often employed to characterize wireless signal propagation (path loss and fading) in the absence or presence of obstacles, including the support provided in available simulation tools. Common configurations of models parameters for both highway and urban environments are provided when applicable. In addition, we cover the simulation models which are available for the simulation of IEEE 802.11p DSRC and related standards. Differences among them are outlined. Regarding vehicular mobility, we briefly review some of the many models that have been proposed for decades and provide references for fine-tune calibration of model parameters when high mobility accuracy is required. Furthermore, we summarize the main features of common mobility simulation tools and describe the process to obtain a realistic vehicular scenario. Finally, we discuss the available options for performing integrated mobility and network simulations. We hope this work to help prospective collaborative ITS researchers and promote best simulation practices in order to obtain accurate results. The reminder of this paper is organized as follows. Section 2 is focused on wireless signals modeling and simulation tools in the context of vehicular networks. Simulation models for DSRC technologies are reviewed in Section 3. In Section 4, we survey some significant vehicular mobility models and different traffic simulation packages. The steps to set up a realistic vehicular scenario are also discussed. Section 5 deals with coupled network and mobility simulations to account for ITS services that influence the behavior of traffic flows. Finally, Section 6 concludes this article.

2. Vehicular wireless communication Understanding the implications of our chosen communication models is key to design appropriate simulation experiments and get insight from their results. In this section we review the most relevant models for vehicular communication systems, as well as the different tools that support them. We highlight the main take-aways that a prospective ITS researcher must keep in mind when conducting a simulation-based study. In addition, valuable information and handy references to help design experiments with different degrees of realism are provided.

2.1. Modeling of wireless links When modeling a wireless ad hoc network, one of the first questions we have to answer is when we can state that a node1 1

Throughout this paper, ‘node’ can be exchanged with ‘vehicle’ or ‘RSU’.

u is able to communicate with another node v. In such case, we say that a link exists from u to v.2 The simplest approach to model a wireless link is derived from a Uniform Disk Graph (UDG). All nodes are assumed to feature a communication range of radius r. In this way, a bidirectional link between u and v exists if and only if ju; v j 6 r, where j  j denotes the Euclidean distance. Note that a UDG represents an ideal network in the sense that perfect communication occurs up to r distance units from the source. This model does not have into account reception errors which might be provoked by radio interferences. It has been often employed in the literature since it provides a rough estimation of network connectivity in a simple way. However, it is well known that real wireless links do not follow this ideal model at all [11]. In order to capture the characteristics of realistic wireless links, signal propagation must be accurately defined (Section 2.2). This determines how signal power dissipates as a function of the distance. In the absence of interferences, the receiver will be able to decode the wireless signal, and therefore reconstruct the original message, whenever the signal to noise ratio (SNR) satisfies the following condition:

SNR ¼

S P b; N

where S is the received signal power, N is the noise power, and b is a threshold dependent on the sensitivity of the wireless decoder. Noise represents the undesired random disturbance of a useful information signal. Since wireless medium is shared by the nodes in an ad hoc network, transmissions from a node interfere with concurrent communications between different nodes. This may cause great disturbance in the resulting signal, so that receivers would not be able to decode the message. In such case, we say that a collision has occurred. Thus, in the most general case, correct reception of a message by a node must satisfy that the signal to interference-noise ratio (SINR) holds the following requirement:

SINR ¼

S P b; IþN

where I is the cumulative power of interfering signals. Next we describe some of the commonly employed propagation models for wireless signals, so that the SINR for a given receiver can be computed. 2.2. Modeling of wireless signal propagation As we have seen in the previous subsection, the signal strength at the receiver is lower than when it leaves the transmitter. Several factors contribute to this phenomenon, such as the natural power dissipation as the signal expands, the presence of obstacles which reflect, diffract and scatter the original signal, and the existence of multiple paths which may lead to signal cancellation at the receiver. The mean signal strength at the receiver as a function of the distance from the transmitter can be estimated by large-scale propagation models, while rapid fluctuations of the signal at the wavelength scale is better represented by small-scale fading models. The following subsections briefly review the most representative models that are relevant to vehicular wireless communications. They can be classified according to different criteria [12]. In Fig. 1, we distinguish among (i) deterministic vs stochastic models, (ii) large-scale path loss vs small-scale fading models, and (iii) whether obstacles (surrounding buildings, the vehicles themselves) are 2 We assume that nodes employ omni-directional antennas. This is the most common scenario, although works on vehicular ad hoc networks with directional antennas have also been undertaken [10].

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

3

Fig. 1. Taxonomy of wireless signal propagation models for V2V/V2R communications.

explicitly accounted for or not. At the expense of higher computation cost, increased realism is achieved by considering more characteristics of the wireless channel. In order to achieve this, some models build upon simpler ones to provide a more realistic framework for vehicular communications. For a deeper coverage of large-scale and small-scale fading models in general, please refer to [13].

ground reflection model [13] provides more accurate predictions by explicitly accounting for both the direct path between sender and receiver and the ground-reflected path. For large distances, the following expression estimates the received signal strength:

2.2.1. Large-scale path loss Given the transmission power P t , large-scale models predict the received signal power P r as a function of the distance d between transmitter and receiver. The attenuation of the signal strength at the receiver with respect to the transmitter is called the path loss (PL). Regardless the propagation model we employ to obtain P r , the path loss can be computed (in dB) as in the following expression:

where ht and hr are the heights of the transmit and receive antennas, respectively, and P r is measured in the same units as Pt (usually W or mW). Note that according to Eq. (2), when transmitter and receiver pffiffiffiffiffiffiffiffiffi are far away (d  ht hr ) the power decays with the distance raised to the fourth power (much faster fall than in free space). In addition, the path loss becomes independent of the signal frequency. The former conclusions do not hold for a short distance d, in which case a different expression must be used to compute Pr . A common model employed in such cases is the Friis’ free space propagation. For instance, such approach can be found in both the ns-2 and ns-3 network simulators. Thus, a cross-over distance dc is calculated: when d < dc , Eq. (1) is employed; when d > dc , Eq. (2) is used; at the cross-over distance dc , Friis and two-ray ground models provide the same result. Therefore, dc can be computed as follows:

PLðdÞ ¼ 10log10

Pt Pr

The Friis’ free space propagation model [14] assumes the ideal condition that there is just one clear line-of-sight (LOS) path between sender and receiver. If we consider isotropic antennas and nodes located in a plane, this model represents the communication range as a circle around the transmitter. Friis proposed the following expression to compute Pr :

Pr ðdÞ ¼

P t Gt Gr k2 2

ð4pÞ2 d L

;

ð1Þ

where Gt and Gr are the antenna gains of the transmitter and receiver respectively, L P 1 is the system loss (because of electronic circuitry), k is the signal wavelength, and P r is measured in the same units as Pt (usually W or mW). As you can see, signal strength decays as a quadratic power law of the distance. In typical communications within a vehicular network, ideal conditions to apply the Friis model are rarely achieved. For instance, signal reflection is not being considered. The two-ray

2 2

Pr ðdÞ ¼

dc ¼

Pt Gt Gr ht hr 4

d L

;

ð2Þ

4pht hr k

Note that for the 5.9 GHz DSRC band for vehicular communications, and assuming that antennas are mounted on cars’ roofs at 1.5 meters high, dc  556 m. This means that if you are employing the two-ray ground model in an ns-2 or ns-3 simulation of a IEEE WAVE/802.11p vehicular ad hoc network, communications between vehicles far away less than 556 m will experience a quadratic power decay typical of free space communications. This could lead to non-accurate simulation results.

4

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

In any case, the former models are not able to capture the different subtleties of wireless communications in general, and vehicular environments in particular. Therefore, most models employed in practice follow an empirical approach in which analytical expressions for the path loss are fitted according to a set of measurements performed in the target scenario. Such expressions approximate the signal strength at an arbitrary distance d by taking as input the signal strength at a reference close distance d0 . Pr ðd0 Þ can be either empirically determined or computed according to the free space model (Eq. (1)). For DSRC vehicular communications, commonly employed values for d0 are 10 m [15] and 1 m [16]. The log-distance path loss model [13] includes a path loss exponent c that indicates the rate at which the path loss increases with the distance:

Pr ðdÞ ¼ Pr ðd0 Þ  10clog10

  d ; d0

Table 1 Configuration of path loss models for vehicular networks. Legend: P  reference to model proposal. S  reference to model usage in simulation. Scenario

Model

Parameters

P

Highway

dual-slope log-distance

c1 ¼ 1:9 c2 ¼ 3:8

[17]

Urban Urban Urban

log-normal log-normal dual-slope log-normal

Urban

dual-slope log-normal

S [16] [15]

dc ¼ 200 m dc ¼ 80 m c ¼ 2:75 r ¼ 5:5 c ¼ 2:32 r ¼ 7:1 c1 ¼ 2:1 c2 ¼ 3:8 r1 ¼ 2:6 r2 ¼ 4:4 dc ¼ 100 m c1 ¼ 2 c2 ¼ 4 r1 ¼ 5:6 r2 ¼ 8:4 dc ¼ 100 m

[18] [18] [18]

[18]

[19]

ð3Þ

where P r is measured in dB. Given a set of measurements, the path loss exponent c can be fitted to let Eq. (3) approximate the real data set. In practice, dual-slope piecewise-linear models like the one in Eq. (4) provide a better fit. Such model has been employed to represent large-scale path loss in highways [15] by adjusting its parameters according to a set of experiments carried out at highway 101 in the Bay Area [17]. Afterwards, this dual-slope log-distance model was implemented within ns-2 [16] and later on integrated within the official simulator codebase (starting from ns-2.34).

8   > < Pr ðd0 Þ  10c1 log10 dd d0 6 d 6 dc 0     P r ðdÞ ¼ > : Pr ðd0 Þ  10c1 log10 ddc  10c2 log10 dd d > dc c 0

(a) Power loss ð4Þ

Pr ðdÞ ¼ Pr ðd0 Þ  10clog10

  d þ Xr; d0

Free space Two−ray ground Log−normal 2

1.5

1

ð5Þ

where X r is a zero-mean normally distributed random variable with standard deviation r. By means of sets of experiments, parameters c and r can be obtained via linear regression, adjusting the model to produce realistic random values for a given scenario. Also in this case, more accurate results have been found by using dual-slope piecewise-linear models such as the one in Eq. (6). This model has been employed for urban scenarios [18], where the authors conducted two sets of experiments in Pittsburgh. The obtained configuration has been employed afterwards [19] to perform vehicular simulations in ns-2.3

  8 > Pr ðd0 Þ  10c1 log10 dd0 þ X r1 > > > <   P r ðdÞ ¼ Pr ðd0 Þ  10c1 log10 dd0c  > >   > > : 10c log d 10 dc þ X r2 2

−3

Density

So far, all reviewed models ignore the fact that two receivers at the same distance d from the transmitter could sense very different signal strengths depending on the environment the signal encounters on its path. Different measurements have shown that the signal strength (in dB) is random and log-normally distributed about the mean distance-dependent value. Therefore, log-normal shadowing models better capture this fact:

x 10

d0 6 d 6 dc d > dc

ð6Þ

In order to provide the reader with a handy reference, Table 1 summarizes the configuration of the main path loss models that have been employed for characterizing vehicular networks, both in highway and urban scenarios. Fig. 2 shows the impact of free 3 Code available at http://masimum.inf.um.es/fjrm/development/ lognormalnakagami6.

0.5

0

200

400

600

800

1000

1200

1400

1600

1800

2000

Distance (m)

(b) Reception PDF Fig. 2. Path loss models for vehicular DSRC.

space, two-ray ground and single-slope log-normal shadowing (urban environment, third row of Table 1) on vehicular communications at the 5.9 GHz DSRC band. The results have been obtained in an interference-free unobstructed scenario with IEEE 802.11p communications at 3 Mbps. The transmission power is set to 20 dB, and antennas are mounted at 1.5 m height. Destination vehicles are from 10 m to 2000 m away from the source, which issues 500 broadcast frames. For each one, we record the power level each receiver senses the frame, including whether it can be successfully decoded. Note that free space and two-ray ground models provide the same power loss up to the cross-over distance dc (Fig. 2(a)) in common implementations of these simulation models. In addition, both approaches feature a non-realistic deterministic radio range,

5

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

while the log-normal model provides a non-deterministic radio range. This leads to very distinct frame reception probabilities, as can be observed in Fig. 2(b). Such figure shows the probability density function (PDF) of successfully decoding a frame with respect to the distance between sender and receiver. In other words, this ‘‘Reception PDF’’ characterizes the likelihood that the resulting SINR at the receiver is enough to recover the original frame given the sensitivity of the decoder. While deterministic models offer a binary probability distribution in the absence of interference (frames are always decoded if the distance is below the effective communication range, and not decoded otherwise), stochastic approaches introduce randomness typical of wireless communications. The main take-aways from this subsection can be summarized as:  Conditions to convey significant simulation results from free space and two-ray ground simulation models are rarely (if ever) achieved in practice for vehicular networks.  Log-normal path loss models capture the randomness that logdistance models lack.  Usually, multi-slope models provide a better fit than singleslope ones.  Whatever the path loss model employed, it must be fine-tuned to the particular vehicular scenario (highway, urban) under evaluation. 2.2.2. Small-scale fading Path loss models discussed so far do not capture the effect of the multiple waves of a signal that arrive at the receiver through different paths. Since the traveled distance is different for each wave, as well as the environment it traverses, each version of the original signal reaches the receiver with different amplitude at a different time instant. In addition, if the transmitter/receiver is in motion as it is the case in vehicular networks, the Doppler effect also causes frequency dispersion with respect to the original signal. These aspects make the receiver face a heavily distorted version of the original signal. There are abrupt changes in amplitude and phase that can be modeled by means of small-scale fading models. Such kind of fading occurs at the scale of a wavelength, and it might be the dominant component in a severe multi-path environment like the one encountered in vehicular networks. In order to account for small-scale fading, we usually rely on statistical distributions that model the envelope of the signal over time [13]. The Ricean distribution models the amplitude of a multi-path envelope when there exists a stronger wave with LOS between transmitter and receiver. As the distance between sender and destination increases in the 5.9 GHz band, the probability density function of the signal amplitude is better captured by the Rayleigh distribution. In case there is no line-of-sight (NLOS), higher-than-Rayleigh fading is observed (the Weibull distribution can be a good fit in such case [20]). In general, the Nakagami distribution [21] is able to capture different severities of fading depending on the chosen parameters. In fact, Ricean fading can be approximated by Nakagami, Rayleigh fading can be seen as a special case of the Nakagami model, and higher-than-Rayleigh fading can be modeled with Nakagami. This distribution approximates the amplitude of the wireless signal according to the following probability density function f ðx; l; xÞ.

f ðx; l; xÞ ¼

2ll x2l1 lxx2 e ; xl CðlÞ

ð7Þ

where l is a shape parameter, x ¼ E½x2  is an estimate of the average power in the fading envelope, and C is the Gamma function. Therefore, x can be computed by employing any of the path loss

Table 2 Configuration of the Nakagami fading model for vehicular networks. Legend: P  reference to model proposal. S  reference to model usage in simulation. Scenario

x

l

P

S

Highway

Dual-slope Log-distance Dual-slope Log-distance

l1 ¼ 1:5 for ð0; 80 m l2 ¼ 0:75 otherwise l1 ¼ 3 for ð0; 50 m l2 ¼ 1:5 for ð50; 150 m l3 ¼ 1 otherwise l1 ¼ 4:07 for ð0; 5:5 m l2 ¼ 2:44 for ð5:5; 13:9 m l3 ¼ 3:08 for ð13:9; 35:5 m l4 ¼ 1:52 for ð33:5; 90:5 m l5 ¼ 0:74 for ð90:5; 230:7 m l6 ¼ 0:84 for ð230:7; 588 m l1 ¼ 3:01 for ð0; 4:7 m l2 ¼ 1:18 for ð4:7; 11:7 m l3 ¼ 1:94 for ð11:7; 28:9 m l4 ¼ 1:86 for ð28:9; 71:6 m l5 ¼ 0:44 for ð71:6; 177:3 m l6 ¼ 0:32 for ð177:3; 439 m

[17]

[19]

[15]

[15]

Highway

Urban

Log-normal

Urban

Log-normal

[18]

[18]

[19]

models that we discussed above. If l ¼ 1, Rayleigh fading is obtained (higher-than-Rayleigh fading for l < 1, and less severe fading for l > 1). Given that signal amplitude is Nakagamidistributed with parameters ðl; xÞ, signal power obeys a Gamma distribution with parameters ðl; x l Þ. A set of experiments conducted at highway 101 in the Bay Area [17] were employed to fit this model. Estimate x is computed by means of the dual-slope log-distance model discussed in the previous subsection. For a better fit of the data, distances between sender and receiver are grouped in a set of bins and different fits for parameter l are estimated on a per-bin basis. Table 2 shows the parameters employed in two different highway setups. The Nakagami fading model has also been used for urban scenarios. In the Pittsburgh experiments [18], x is calculated by means of log-normal path loss models (both single-slope and dual-slope). The value of parameters l on a per-bin basis on two different data sets are provided in Table 2. This model is available for simulation as a separate patch for the ns-2 network simulator.3 Fig. 3 compares the power loss and reception probability of highway (first row of Table 2) and urban (fourth row of Table 2) models in DSRC (same setup as in the other figures of this section). Note the higher dispersion that fading models provoke with respect to large-scale path loss (Fig. 3(a) and (c)), as well as the higher challenge that urban environments impose on vehicular communications (Fig. 3(b) and (d)). Fig. 4 shows the cumulative density function (CDF) of the reception probability in IEEE 802.11p when different path loss and fading models are employed in simulation. Such figure summarizes the results we have seen so far. Specifically, deterministic path loss models are shown to give a fixed communication range, which is lower for two-ray ground than free space given that the former models a higher power decay beyond the cross-over distance. On its hand, the log-normal path loss model provides a probabilistic communication range with higher reception likelihood at short distances than at long distances. When small-scale fading is also taken into account, the effective communication range decreases (the reduction is higher in urban scenarios than in highways). The main take-aways from this subsection are:  Small-scale fading is the dominant component in dynamic multi-path scenarios like those comprised of communicating vehicles. Therefore, it must be taken into account to convey realistic simulation results.  The Nakagami model is general enough to capture different levels of fading.

6

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

(a) Power loss - Urban 3.5

(b) Power loss - Highway vs Urban 3.5

Urban path loss

Highway Urban

3

3

2.5

2.5 Density

Density

Urban path loss + fading

2 1.5

2 1.5

1

1

0.5

0.5

0

0 0

500

1000 1500 Distance (m)

2000

(c) Reception PDF - Urban

0

200

400

600

800 1000 1200 1400

Distance (m)

(d) Reception PDF - Highway vs Urban

Fig. 3. Fading models for vehicular DSRC.

1 0.9

Cumulative probability

0.8 0.7 0.6 0.5 0.4 0.3 0.2

Free space Two−ray ground Log−normal

0.1

Log−distance + Nakagami (highway) Log−distance + Nakagami (urban)

0 0

200 400 600 800 1000 1200 1400 1600 1800 2000 Distance (m)

Fig. 4. CDF of reception probability in IEEE 802.11p.

 A better fit is obtained if different distance bins between sender and receiver are considered.  Whatever the fading model employed, it must be fine-tuned to the particular vehicular scenario (highway, urban) under evaluation. 2.2.3. Propagation through obstacles The former models do not consider obstacles in an explicit way. Given that many of them consist of stochastic processes fine-tuned according to real experiments, the effect of obstacles that reflect, diffract and scatter the original signal is incorporated in an indirect way. However, they do not accurately cover the shadowing of the signal by a given obstacle between two vehicles, since the environment map is not incorporated into the model. Therefore, different works have focused on explicitly accounting for the impact of obstacles into vehicular communications.

Obstacles reduce channel congestion at the cost of a greater (more realistic) number of NLOS situations. This also fosters the appearance of hidden terminals, challenging the performance of medium access control schemes. Given the existence of numerous obstacles in real vehicular communication scenarios, they should be considered in simulations since they have a great impact onto the accuracy of simulation results. Ray tracing techniques have been employed to accurately account for the effect of obstacles in wireless communications [22]. However, traditional ray tracing is not appropriate for the simulation of vehicular networks due to the high processing requirements it imposes. The approach is not able to scale to large networks. Hence, different works have proposed simplified ray tracing techniques that employ pre-processing steps to reduce the simulation time without heavily impacting the accuracy of the results. Along these lines, a general methodology for generating urban channel models for a given 3D map was proposed recently [23]. Nevertheless, simpler solutions are often employed in practice. For instance, the ns-2 simulator incorporates a so-called shadowing visibility model which actually consists of two log-normal models that can be independently configured. One of them is used when there is LOS between the communicating entities, and the other is employed for NLOS cases. In order to determine what model shall be used for a given transmission, the user must provide a bitmap file that represents the obstacles in the simulation scenario. The former approach is hard to generalize because it does not account for the number of traversed obstacles, their size, and their shape. In general, it is better to rely on a path loss model and compute the extra attenuation which is due to the obstacles which are in the path of the signal. Such scheme is adopted by the inexpensive empirical model [24] for urban simulation, in which the extra loss Lo is related to the number of times n the border of a obstacle is

7

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

intersected by the straight wave from transmitter to receiver, and the total length p of penetration within obstacles:

2.3. Simulation support

Lo ¼ an þ bp;

When choosing a simulation tool for a vehicular network, the list of supported propagation models is key to decide which simulator is more appropriate depending on our particular requirements. We focus on the three most popular simulators for vehicular ad hoc networks, namely ns-2 [32], ns-3 [33] and OMNet++ [34]. Table 3 summarizes the support they provide (as of this writing) for different wireless propagation models which are relevant to vehicular ad hoc networks. The network simulator ns-2 has been many popular within the research community for years, and it therefore features many simulation models in its official distribution, while still more are available as separate patches. For the official release, it provides decent support for wireless propagation in vehicular environments, although obstacle modeling is quite limited. The simulator employs two different programming languages: C++ and OTcl. The core is mostly written in C++, while simulation scripts use the OTcl interpreter to avoid recompiling the simulator when changing the experiment setup. This design decision allows for an optimized core that runs faster while providing flexibility to the user. On the down side, two different (but linked) object hierarchies must be maintained in memory: user-created OTcl objects and their corresponding shadow objects in the C++ domain. On the other hand, ns-3 has been designed from scratch to provide a flexible and efficient modern network simulator. It is completely written in C++, although Python bindings are also available for end users. By means of a convenient attribute system, C++ simulation scripts are easily configurable through command line arguments without requiring recompiling the simulator. In addition, ns-3 provides powerful tracing and data collection facilities, emulation support, and the ability to run existing user-space and kernel-space network protocols or applications without changing the source code (DCE – Direct Code Execution). Compared to ns-2, ns-3 models the network stack with more accuracy and allows for increased functionality while improving the scalability of the simulator in terms of memory footprint and processing efficiency. It has experienced a great level of activity and supports many different models. This can be assessed against the wide range of propagation models that are incorporated within its official release. Many of them were designed for different target networks (e.g. cellular), although support for V2V/V2R environments is also provided. In particular, an interesting Buildings model facilitates the evaluation of vehicular communications in urban scenarios. Additional models related to vehicular ad hoc communications are included, as shown in Table 3. Regarding OMNeT++, this tool is actually a simulation framework to build network simulators. It basically provides an eventbased network simulation kernel where different models can be built on top of it. In this way, different projects focus on a given set of related features to serve a particular group of users. For instance, the INET project develops a communications framework

ð8Þ

where Lo is measured in dB, a (dB per wall) is the attenuation a transmission experiences due to the exterior wall of a building, and b (dB/m) is the attenuation due to the internal structure of the building. This model has been employed along with the free space path loss model and a set of experiments were conducted to fit parameters a and b. For most of the considered scenarios, the parameters can be approximated by a  9:2 dB per wall and b  0:4 dB/m [24]. However, there are other cases in which different combinations of parameters must be used to capture the characteristics of propagation through the obstacles. In addition, this inexpensive approach does not model near-LOS scenarios in which the wireless signal hits the corner of the building. This leads to optimistic results that do not match reality. As an alternative, models that consider the extra attenuation in both NLOS and near-LOS cases have been proposed [25]. They have been used along with log-distance path loss and Nakagami fading models [15] to provide a more comprehensive characterization of signal propagation in urban environments. In this model, Lo is computed as follows:

Lo ¼

8 > < 0; > :

LOS

c1 ; NLOS c2 ; near-LOS

ð9Þ

Extra attenuation exponents are set as c1 ¼ 30 dB and c2 ¼ 13 dB in [15]. This configuration has been shown to match real data and theoretical analyses based on ray tracing. Similar approaches have been employed in many other works, some of them also considering the vehicles themselves as obstacles. Initial experiments showed that vehicles heavily obstruct the wireless signal at short distances, with drops of over 20 dB for two static cars communicating at a distance of 10 m [26]. Similar results are observed when vehicles are in motion. For longer distances, the relative impact of vehicles as obstacles is lower, but still very high: the distance at which two vehicles can communicate with 90% chance of success is halved if other vehicles obstruct the line of sight. Therefore, the implications for simulations become clear. In order to provide a high degree of realism, the simulation model must consider the vehicles themselves as obstacles that shadow the wireless signal. Note that this applies to all vehicular scenarios (highways, suburban, urban canyons), while the discussion so far has been focused on urban environments. The former observation has encouraged works that also consider vehicles as obstacles in simulation [27,28], where the latter reference implements the model proposed by Boban et al. [29] in the Veins vehicular simulator [30]. This model first computes the probability that any two communicating vehicles are in LOS. For those vehicles obstructing the LOS of the communication, the multiple knife-edge diffraction model [31] is employed. The interested reader is referred to the original paper [29] and the ITU-R recommendation [31] of the multiple knife-edge model for the details of this approach. To conclude with this subsection, let us enumerate its main take-aways:  In urban scenarios it is of paramount importance to consider the extra attenuation of wireless signals due to the presence of obstacles.  In all cases, high fidelity simulation models should consider that the vehicles themselves shadow wireless communications and provoke a significant drop of the signal strength. This might have a great impact on the performance of upper layers.

Table 3 Wireless propagation models supported by different network simulators. Legend: I  INETMANET. M  MiXiM. V  Veins. Models

ns-2

ns-3

OMNeT++

Free space Two-ray ground Log-distance

U U Dual slope (within Nakagami) Single slope

U U Single slope Triple slope Single slope (within Buildings) 3 bins Buildings

I I

Log-normal Nakagami Obstacles

3 bins Shadowing Visibility

Single slope (I) Single bin (I) Inexpensive Empirical (V,M)

8

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

for wired and wireless networks. The INETMANET project extends the former to support the simulation of mobile ad hoc networks. And MiXiM combines different frameworks for the simulation of wireless mobile networks. The most relevant project for vehicular communications is Veins, which features the inexpensive empirical model [24] for obstacle simulation (also available for INET and MiXiM). An interesting characteristic of this implementation is that it can directly import 2.5D maps as provided by OpenStreetMap [35]. This saves a lot of effort when trying to set up a realistic urban scenario for simulation. In addition, Veins and MiXiM incorporate a more realistic two-ray model that does not rely on the free space model for communications below the cross-over distance [36]. To summarize, signal propagation is relatively well supported within the official releases of widely employed network simulators. The decision of choosing one tool or another might be influenced by the range of propagation models that are supported and we are interested to experiment with. In some cases, other models are supported by means of external patches (e.g. dual-slope log-normal along with Nakagami 6-bins in ns-23). And for other cases, our target degree of realism might lead us to a model which is not directly provided by any of them (e.g. if we want to consider vehicles as obstacles). Thus, we would be required to implement the solution ourselves. Fortunately, all reviewed simulators are open source, freely available, and well documented, so that it should be reasonably easy to incorporate new propagation models as required (the complexity lies on the particular model being implemented, not on simulator issues).

3. Vehicular wireless technologies 3.1. DSRC models Once we understand the different choices we have to model signal propagation in vehicular environments, let us take a look to the actual communication technology that we want to employ. No doubt, the most prominent standard for V2V and V2R communications is the IEEE 802.11p amendment [37] for Wifi networks, which specifies the lower layers (PHY and MAC) of the protocol stack. Most vehicular networking research assumes 802.11p as communication technology, unless the objective is to develop an alternative PHY and/or MAC solution. There are several factors that led to modifications in the baseline 802.11 specification. First, longer ranges of operation (approximately 1000 m) were desired in order to support some ITS applications. Also, high relative speeds of vehicles and high influence of multi-path fading due to the environment challenge wireless signal decoding. In addition, the 5.9 GHz frequency band was allocated for dedicated short-range communications (DSRC) in different regions of the world. The spectrum was partitioned into a number of 10 MHz channels, so that the standard must accommodate to this band and channel bandwidth. To fulfill the aforementioned requirements, the IEEE 802.11a physical layer [38] was taken as reference. It employs Orthogonal Frequency Division Multiplexing (OFDM), which is well-known to prevent inter-symbol interference and inter-carrier interference in multi-path environments. In 802.11p, channel bandwidth (10 MHz) is half the 802.11a channel bandwidth (20 MHz). This directly affects the maximum data rate that can be achieved (27 Mbps instead of 54 Mbps) and some protocol parameters (symbol guard period is 1.6 l s instead of 0.8 l s, symbol duration is 8 l s instead of 4 l s). Fig. 5 shows the impact that different data rates (equivalently, modulation schemes and coding rates) have onto reception probability.

IEEE 802.11p MAC layer is based on 802.11 CSMA/CA Distributed Control Function (DCF) and 802.11e Enhanced Distributed Channel Access (EDCA) traffic prioritization [38]. For the latter, 802.11p defines different contention windows (CW) and arbitration inter-frame spaces (AIFS) for voice, video, best effort, and background traffic. Given the opportunistic nature of vehicular communications, in which short-lived transmission opportunities might arise intermittently, the MAC layer is designed to work outside the context of a BSS (OCB). In this way, vehicles can communicate without the added overhead associated to network scanning, authentication, and association. Taking 802.11p as the main technology for V2V and V2R communications, several standardization bodies provide a whole protocol stack on top of it. Such layers deal with a variety of functions like securing communications, defining multi-channel operation, providing a management plane, and supporting ITS applications through appropriate network and transport layers. Multi-channel operation is particularly relevant to vehicular networks because they are comprised of a great number of nodes. Hence, such a mechanism is required to overlap different communications in time and space without degrading the perceived quality of service at the face of network congestion. Along these lines, the IEEE provides the 1609 family of standards [39–42] that, along with 802.11p, are collectively known as WAVE (Wireless Access for Vehicular Environments). According to IEEE 1609.4 [42], each second is split into ten sync intervals. Every sync interval contains alternating control channel (CCH) and service channel (SCH) intervals. During the CCH interval, all WAVE stations monitor the CCH. If any station needs to transmit or receive at a given SCH, it switches to such channel during the SCH interval. Traditional IPv6 communications are borne in SCHs. On its hand, data from the specific transport and networking protocol (WSMP – WAVE Short Message Protocol) defined in IEEE 1609.3 [41] can employ either CCHs or SCHs. WAVE transmits high priority frames and WAVE Announcement frames (beacons) during the CCH interval. Low priority messages can also be transmitted on the CCH during the SCH interval for those stations that do not switch to any SCH. IEEE 802.11p is also employed as the communication technology for the ITS-G5 European profile defined within the ETSI ITS communications architecture [3]. Such architecture takes into account the set of ISO standards for vehicular communications [2], known as CALM (Communications Access for Land Mobiles). Support for these simulation models varies across simulators. Therefore, the availability of communications technology models or the ease to implement them ourselves, is an interesting point to take into consideration when choosing a simulation tool. Furthermore, the researcher must decide what technology is more relevant for the study under consideration. In some cases, just focusing on the lower layers (802.11p) is enough, while for other evaluations we might consider including the multi-channel operation as provided by WAVE or any other protocol stack. Finally, it is always recommended to check what parts of the standard are actually implemented, and what parameters can be configured. In the following we discuss the kind of support which is provided by different well-known network simulators to conduct experiments which involve vehicular communications. 3.2. Simulation support IEEE 802.11p PHY and MAC layers are the ones with broadest support across simulators. Since the release of ns-2.34, this simulator includes a highly improved IEEE 802.11 model called 802_11Ext [16]. The physical layer is accurately modeled, incorporating different modulation schemes and coding rates that fit the needs of different Wifi standards, including the amendment for

9

F.J. Ros et al. / Computer Communications 43 (2014) 1–15 1 3 Mbps

5

0.9

6 Mbps 12 Mbps

0.8

27 Mbps

Cumulative probability

Density

4

3

2

0.7 0.6 0.5 0.4 0.3 3 Mbps

0.2

1

6 Mbps 12 Mbps

0.1 0

27 Mbps

0 0

200

400

600

800

1000

1200

1400

Distance (m)

(a) PDF

0

200

400

600

800

1000

1200

1400

Distance (m)

(b) CDF

Fig. 5. Reception probability of IEEE 802.11p with different data rates in a highway scenario.

vehicular communications. In addition, the user can simulate preamble and frame capture features, as provided by several network interface cards. When enabled, an on-going frame reception can be discarded whenever another frame is detected with significant higher power. In such case, the wireless card switches to decode the newly arrived frame. This simulation model also accounts for the cumulative SINR in order to determine whether a frame can be successfully decoded depending on a threshold value. A strong point of this Wifi implementation is that it is easily configurable to support a variety of IEEE standards. Hence, 802.11p can be simulated by just providing the required parameters, such as the header duration, symbol duration, or modulation scheme. Unfortunately, there is no support for traffic prioritization through EDCA. Since upper layers of vehicular communication architectures are not provided, it does not feature multi-channel operation either. Compared to ns-2, ns-3 provides a more sophisticated method to model the impact of interference. After determining the cumulative SINR, the physical model computes the corresponding bit error rate (BER) and packet error rate (PER) for the modulation and coding schemes which are being used. Once the probability that the frame is received with any error is known, the packet is randomly discarded according to such probability [43]. The ns-3 network simulator does not have support for 802.11p in its official release yet. However, it has been implemented recently thanks to a Google Summer of Code project that finalized in September 2013 [44]. In this work, an 802.11p MAC entity that operates in OCB mode has been developed. In addition, support for WAVE devices has been added by implementing different features specified in IEEE 1609 standards. Specifically, this simulation model provides multi-channel operation through one CCH and several SCH (IEEE 1609.4 [42]), while leveraging existing EDCA support for traffic prioritization. This model also implements WSMP (IEEE 1609.3 [41]) at the upper layer. Both the data and management planes of 1609.3 and 1609.4 are provided. On the other hand, an independent project called iTETRIS [45] implements the European ITS communications architecture within ns-3, including the ITS-G5 profile for DSRC communications. In this way, iTETRIS provides a simulation model for 802.11p with a realistic physical layer and 10 MHz-width CCH and SCH channels. Channel switching and EDCA are both supported. In addition, upper layers are allowed to control transmissions parameters such as transmission power and data rate on a per-packet basis [46]. Finally, OMNeT++ implements IEEE 802.11p/WAVE through the Veins vehicular network simulator. This model accounts for an accurate physical layer, CCH and SCHs of 10 MHz each (1609.4), and traffic prioritization through the EDCA function [47]. Regarding interference modeling, Veins takes advantage of the cumulative

SINR computation performed by the MiXiM module [48] to obtain the BER of the frame header and payload separately [47]. Putting all together, each simulation model has its own advantages and disadvantages. In general, interference modeling is better captured in ns-3 and OMNeT++ than in ns-2. In any case, prospective ITS researchers that perform its work above the data link layer must analyze what model is more convenient depending on the target of their study. 4. Vehicular mobility So far in this paper, we have dealt with modeling and simulation of vehicular communications. This is a key topic to perform significant studies on ITS. However, it alone is not enough. The missing part in the equation is the simulation of realistic vehicular mobility patterns, since they influence the design of ITS services and protocols and their performance. Because of this, most research in the field relies on mobility simulation tools that provide realistic mobility traces. Such traces are fed afterwards into a network simulator (e.g. any of the reviewed simulators in the previous section) so that accurate communications are simulated as vehicles perform realistic movements. In this section, we survey the most representative vehicular mobility models that have been proposed in the literature. In addition, we review the most prominent mobility tools from the viewpoint of an ITS researcher focused on developing communication-based services, and describe the challenges that must be faced to obtain realistic vehicular mobility datasets. Finally, the key concepts that must be retained are highlighted to conclude the section. 4.1. Modeling of vehicular mobility Vehicles do not move freely. They are subject to mobility rules which are determined by the road topology, traffic signs, other vehicles’ movements, and physics laws that limit the motion of the vehicle according to sensible values of acceleration and deceleration. These rules and constraints have made researchers make a great effort at developing accurate mobility models that capture their essence with the objective of generating the most realistic vehicles movements in simulations. Depending on their application we can find three kinds of mobility models: macroscopic, microscopic and mesoscopic models. Macroscopic models are aimed at dealing with traffic density, traffic flows and initial vehicle distribution modeling. On the other hand, microscopic models are in charge of modeling the location, velocity, and acceleration of each vehicle that participates in the

10

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

simulated scenario. Finally, as an intermediate approach, mesoscopic models aggregate the movements of different nodes. In this paper, we focus on the behavior of each vehicle as an independent unit. Therefore, we constrain our review to microscopic models, namely car following models responsible for the motion of a vehicle being aware of the vehicle in front of it and going in the same lane and direction, and lane-changing models that determine how vehicles switch from one lane to another. Such models have been shown to mimic real world vehicular dynamics [9], contrary to other simplistic approaches like stochastic and traffic steam models. Using realistic models in simulation is of paramount importance because it has a great impact onto the underlying network topology. Specifically, car-to-car interactions at intersections, as captured by car following models, have a great impact on networking metrics. 4.1.1. Car following One of the most studied tasks involved in driving, is that of a vehicle following the vehicle ahead along a lane of the roadway. Car following is simpler than other facets of driving, and it has a great impact onto the macroscopic characteristics of traffic flow. Therefore, this topic has been focus of deep study for several decades. This relatively simple and common driving task can be categorized in three specific sub-tasks:  Perception: the driver gathers relevant information usually through the visual channel. The information is obtained from the motion of the leading and follower vehicles, such as vehicle speeds, accelerations, inter-vehicle spacing, or relative speeds.  Decision making: the driver interprets the information and integrates it with the knowledge of the vehicle characteristics and its own driving experience.  Control: according to the decision that has been made, the driver takes some actions to control the vehicle by maintaining a ‘good’ distance with the leading vehicle. The more experienced the driver, the more smoothness and coordination will be achieved while following the leader. It is not clear how exactly this process is carried out by the driver in detail, so different car following models were defined so as to disclose this behavior. Car following models can be classified into six different groups [49,50]: stimulus–response models; safe distance models; psychophysical models; cell-based models; models based on the optimum velocity and models based on the trajectory of the leading vehicle. The first approaches are encompassed in stimulus–response models. They were developed in 1958 and 1961 by Chandler et al. [51] and General Motors [52] respectively. Assuming that the reaction of a driver is proportional to the stimulus he perceives, Chandler et al. proposed a simple model in which the relative speed with respect to the leading vehicle is the only stimulus that the driver receives. The corresponding response takes place after a given response time T. Since not every driver reacts at the same time given the same stimulus, this model also introduces a sensitivity factor k. Later on, General Motors conducted additional research on the subject, introducing new parameters into the Chandler model such as the speed of the vehicles and the distance between them. This extended model is commonly known as the GM model. Treiber et al. proposed an evolution of this approach in the Intelligent Driver Model (IDM) [53]. In this case, driver’s behavior is characterized by the instantaneous acceleration of vehicles. On his hand, Krauß [54] models the vehicle’s speed at each time instant and introduces a stochastic term that accounts for imperfections in driving. Alternatively, a driver’s behavior might be determined by his perception of the safe distance. That is, each driver will maintain

a safe distance with the leading vehicle that will depend on the speed of his car, its acceleration, and deceleration, among others. This idea was first introduced into a car following model by Kometani and Sasaki [55]. In 1981, Gipps extended the former model by making some common sense assumptions about the acceleration, deceleration and allowed maximum speed [56]. Leutzbach et al. proposed the psychophysical model [57] in 1986. It considers the acceleration of the vehicle ahead as a stimulus for the following vehicle. In addition, it also accounts for the difference between the current inter-vehicle spacing and the desired following distance. The cell based model was introduced in 1992 by Nagel and Schereckenberg [58] (also referred to as the N-S model). It considers two parameters to be optimized, namely the acceleration and the desired maximum speed. The particularity of this model is that it divides the traffic scenario into a set of cells of equal size. The size of the cells normally do not exceed the size of a vehicle, therefore only one vehicle will be in each cell at a time. This model can be seen as a set of rules that control the movement of a vehicle from a cell to the next one. In 1995, Bando et al. [59] proposed an optimum velocity-based model. The optimum velocity consists of the required speed to maintain a given distance with the vehicle ahead. In this model, at any time, the response of the following driver is proportional to the difference between his optimum speed and his current speed. More recently, Newell introduced a new trajectory-based model [60] in 2002. In this approach, the trajectory of the leading vehicle as well as the following one are considered the same, except for a translation in space and time. Thus, the following vehicle drives along a shifted trajectory of the vehicle ahead. In general, car following models are able to reproduce realworld traffic phenomena when properly calibrated (Section 4.2). It has been shown [9] that IDM, Krauß and N-S models all match with the fundamental diagram of traffic flow, this is, with the relationship between vehicular outflow, speed and density observed in reality. They also capture the effect of traffic perturbations such as the reaction when a slow vehicle is encountered ahead. Finally, slow-speed waves typical of traffic congestion are also faithfully reproduced by these models. Given that the N-S model is spacediscrete, the wave profile is sensibly different than the other models, although it also captures the same macroscopic effect. 4.1.2. Lane-changing models In addition to following the leading car, common driving behavior includes lane changes when the appropriate conditions are met. Therefore, lane changing must be considered in both highways and urban streets where multiple lanes per road exist. In this subsection we survey some models that account for this behavior, so that they can be incorporated into microscopic mobility simulators to obtain realistic movement patterns. In general, changing from one lane to another can be modeled as a sequence of three steps: decision to consider a lane change, choice of the target lane, and determination of gap acceptance (whether the state of the system during and/or after the maneuver is consistent with basic safety constraints). The two most popular model categories that prescribe these steps are rule-based (RB) models and discrete choice-based (DCB) models. RB models provide deterministic lane change behavior. Reasons for changing lane are explicitly enumerated, and afterwards an algorithm evaluates whether the maneuver must proceed. Gap acceptance criteria usually depend on threshold values regarding space gaps or speed variations. For instance, in the Gipps model [61] the maneuver is considered feasible if the deceleration required for a lane change is within an acceptance range. After the Gipps model, other RB approaches have been proposed [62–64],

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

although the Gipps algorithm is still widespread among vehicular mobility simulators. On their hand, DCB solutions [65–67] model the driver behavior by a probabilistic function (probit or logit) to estimate the effects of specific significant attributes. Among such attributes, we can find neighborhood variables (presence of neighboring vehicles and their current state, target lanes, or target gaps, among others), path plan variables (distance to a freeway off-ramp or allowed turns at intersections, among others), network knowledge, driver experience and style, and so on. When the decision of changing to a target lane has been made, the DCB model evaluates gap acceptance criteria to finally determine if the maneuver must proceed. For this, gap acceptance is usually formulated as a multinomial probit model affected by the spatial relations between the subject vehicle and the lead/lag vehicles in the target lane. When an acceptable gap exists in terms of the positions and relative speeds of the lead and lag vehicles, the maneuver takes place. Among the different DCB approaches, let us focus on the MOBIL4 proposal [68], given that it is widely implemented in mobility simulators. The model provides an incentive criterion that determines if changing to a different lane improves the situation of the vehicle in its current lane. It also imposes some safety restrictions that must be accomplished in order to make a safe lane change. For the gap acceptance criterion, MOBIL takes advantage of car following models to get the difference of speed between vehicles. Larger gaps between the new following vehicle in the target lane and the own position are required if the lead vehicle is faster than the own one. On the other hand, lower gap values are allowed if the speed of the following car is lower. In addition, MOBIL also formulates an asymmetric lane-changing criterion to implement the ‘‘keep right’’ directive:  Overtaking rule. Overtaking is forbidden using the right lane, unless traffic flow is congested. In this case, a symmetric lanechanging rule is applied. If a vehicle is driving at a speed lower than some specified velocity v crit , congested traffic will be assumed.  Lane usage rule. The default lane is the right lane. The left lane should be only used in case of overtaking. Overall, MOBIL intends to reflect the driver behavior of considering whether an overtaking is dangerous by taking a look at the rearview mirror. The decision is made according to the perceived speed of the approaching vehicles.

4.2. Calibrating microscopic traffic simulation models The former car following and lane-changing models, as well as their variations, feature a different but rich set of model parameters (for example, the driver’s reaction time, vehicle’s maximum speed, normal deceleration, and the like) that influence the resulting traffic patterns. The main objective of any given model is to provide realistic movements that match real traffic characteristics. In order to achieve this, a previous calibration phase is required to fine-tune the different model parameters. Such phase is of paramount importance for traffic engineering. In the context of an ITS researcher focused on communication-based services, the level of mobility accuracy varies depending on the objective of the study. Calibration of microscopic traffic models can be formulated as an optimization problem. The objective is to find the parameter values that minimize the discrepancy between the model output and a given ground truth with respect to a certain measure of performance that depicts traffic behavior (e.g. follower’s speed or 4

Minimizing Overall Braking Induced by Lane change.

11

inter-vehicle spacing in car following models). Different goodness of fit criteria can be employed to evaluate the appropriateness of the resulting model instance, such as the root mean square error, the mean absolute error, or the Theil’s inequality coefficient, among others. The ground truth can be obtained from real traffic data [69–71] or synthetic data from the model itself [72–74]. Different algorithms can be employed to solve the aforementioned problem [75]. However, they might get stuck in a local minimum and provide an undesired result. Determining what algorithm should be employed is not an easy task, since calibration belongs to the framework of ‘‘no free lunch’’ theorems [76,77]: there is no algorithm which outperforms the others over the entire problem domain. This is, the choice of the most appropriate algorithm depends on the specific problem under investigation. In particular, many variables influence the calibration process in microscopic traffic modeling [78]: the measure of performance, location of traffic detectors, subset of parameters that are being optimized, specific micro-simulation model under consideration, and traffic conditions under investigation. Given this fact, sound methodologies are required to provide a systematic framework for model calibration. Detailing any given methodology is out of the scope of this paper, although the interested reader can refer to [75]. In the following we focus on some simulation tools that allow us to obtain mobility traffic patterns, according to a set of microscopic models, that can be employed to simulate vehicular networks. 4.3. Simulation support Some of the former models have been implemented within fullfeatured vehicular mobility simulators. They are key to evaluate the design of traffic infrastructure for modern intra-city and inter-city transportation, which is of paramount importance nowadays. Luckily, current mobility simulators are able to generate vehicular traffic for very complex scenarios. In this section, we review some of the most relevant simulation products and their main features. There are several commercial solutions that provide a high degree of accuracy to model a vehicular setup. They are very powerful and can model nearly everything related to road building, such as the number of lanes per road, the shape of each road, on-ramps, off-ramps, traffic lights, and so on. In addition, simulations include different types of cars and motorbikes, as well as support for pedestrians in some cases. They also provide graphical user interfaces (2D or 3D) that facilitate the scenario configuration and allow to visualize animations of the simulated traffic model. In terms of vehicles mobility, they usually implement a car following model, as well as a multi-lane changing model to simulate overtaking among vehicles. Among the most relevant tools, we find CORSIM [79], PARAMICS [80], VISSIM [81], and AIMSUN [82] simulators. A comparative evaluation of the first three simulators showed that setting up a realistic scenario is time consuming (about three-four days for a highway) in all cases [83]. Moreover, it is not easy to calibrate the scenario in order to convey realistic simulation results. However, as outlined in the previous subsection, the simulation model should be calibrated according to the particular application that is to be evaluated. For instance, Hadi et al. perform a calibration process in CORSIM, VISSIM and AIMSUN in order to analyze the impact of incidents on traffic flow operations [84]. After calibration is performed, these tools provide accurate mobility patterns that mimic the macroscopic behavior of traffic flows, at the expense of complex setups. While such simulators are key for traffic engineers and transportation researchers, they were not designed with a focus on vehicular communications. In the following we concentrate on the most widely employed

12

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

mobility simulator for vehicular networking research, SUMO (Simulation of Urban Mobility) [85], which is freely available and released under an open source license (contrary to previous tools). The reader is referred to [86,87] for surveys that cover additional tools. SUMO is a microscopic road traffic simulation package designed to handle large road networks. It is mainly developed by employees of the Institute of Transportation Systems at the German Aerospace Center. This simulator accounts for space-continuous, time-discrete movements of individual vehicles of different types. It supports multi-lane roads, on-ramps and off-ramps, roundabouts, traffic lights, and stop, yield and variable speed signs. In addition, SUMO includes a visualization tool that shows the simulated road topology and the movement of vehicles as the simulation goes along. In general, SUMO scenarios are defined in XML files. A third party tool called MOVE [88] provides a graphical interface to design the road map and help define the simulation setup. However, the process of creating the simulation scenario by hand is tedious and only applicable for small road topologies. To increase the simulation realism, most researchers characterize a slice of a real road topology, such as part of a city. Fortunately, SUMO facilitates this task by importing maps in different formats, including the OSM format of the freely available OpenStreetMap database [35]. However, it must be noted that this automatic process is not completely error-free and human inspection must be undertaken to solve some common issues that affect simulation accuracy [89]. In particular, OSM maps are very accurate from a topology viewpoint but they enforce some movement restrictions that do not match reality. In addition, some OSM attributes are just not understood by SUMO. Once we have a realistic simulation scenario in SUMO, we need to model representative traffic flows [89]. For this, first we need to characterize the traffic demand by means of a Origin/Destination (O/D) matrix. Sociologic and demographic information about the target scenario can be employed to determine an appropriate O/D, and some traffic authorities directly provide the O/D matrix associated with a given area or traffic information as collected by induction loops in highways. We can provide SUMO with the O/D matrix or data gathered from traffic detectors and then run a routing assignment algorithm that computes the path that each vehicle follows from origin to destination. The simplest approach consists of using shortest path routing (Dijkstra’s algorithm) over a weighted graph of the road network. Road segments are assigned a fixed cost that depends on the road’s length and the maximum allowed speed. However, this causes the unrealistic effect that most drivers with similar origin/destination choose the same route. Luckily, SUMO also provides so-called dynamic user assignment, in which an iterative algorithm (based on a logit criterion or the Gawron’s method) reroutes vehicles by considering the actual road usage from the previous iteration. This is, as the algorithm iterates, vehicles tend to employ different available routes and the overall congestion is reduced. Such approach has been shown to better align with the macroscopic behavior of real traffic [89]. At this stage, SUMO is ready to employ the car-following and lane-changing models it implements to perform the vehicular mobility simulation. The default and most employed car-following model in SUMO is the Krauß model, that can be fine-tuned by setting the parameters shown in Table 4. A lane-changing model developed by Krajzewicz [90] is also provided. Depending on the accuracy level required by the researcher, a calibration phase as discussed in Section 4.2 might be required. As the simulation output, SUMO can generate different files that provide information about vehicles’ movements and traffic statistics. These files can be exported to different trace formats that

Table 4 Krauß car-following model parameters in SUMO. Parameter

Description

Units

Driver imperfection Driver’s reaction time Minimum distance to the leading vehicle in a jam

m/s2 m/s2 [0, 1] s m

Acceleration Deceleration

r s Gap

are afterwards employed as input for a network simulator. In particular, our three simulators of interest, namely ns-2, ns-3 and OMNet++, are supported. 4.4. Mobility datasets Depending on their needs, researchers might avoid the burden of defining accurate mobility scenarios (road topology, movement restrictions, traffic demand, traffic assignment, micro-mobility calibration) by reusing an existing mobility dataset for network simulation. Both real traffic tracking and simulation-generated datasets have been created during the last years. Some of them are publicly available, and others can be obtained on request. Of course, not every dataset provides the same level of accuracy, so a previous study must be undergone in order to determine the most suitable option. Uppoor et al. provide a comprehensive analysis of existing datasets and propose a new one that models the urban area of Köln [89]. Although the microscopic model has not been calibrated, the resulting mobility traces match the macroscopic behavior of real traffic. 4.5. Summary In order to conclude this section, we summary the main takeaway messages that a researcher in vehicular communications should get. To obtain significant simulation results, we must count on realistic mobility patterns that retain the network topology properties of real vehicular networks. There are some datasets that could be reused if they fit the requirements of the particular study to be performed. Otherwise, an additional effort must be made to design a convenient simulation scenario. Such effort includes (at least) defining the road topology and movement constraints, getting a realistic traffic demand (O/D matrix), and running an appropriate traffic assignment algorithm. SUMO is a great tool that helps us in this process and that incorporates simulation models for carfollowing and lane-changing. As we have pointed out, such microscopic models reproduce realistic macroscopic behaviors of traffic flows when they are properly calibrated. On the other hand, calibration can be a complex procedure that requires a good understanding of the particular model being employed. Systematic methodologies have been proposed in the literature to achieve good fits between simulated movements and realistic ones. 5. Integrated simulation of mobility and communications The simulation workflow described so far is depicted in Fig. 6(a). This is, first a mobility simulation is conducted, it generates a trace file with the movement of vehicles, and such trace is fed into a network simulator. In this way, mobility and network simulations are completely decoupled through the trace file. While this approach is appropriate for certain ITS applications, other vehicular services have a direct impact onto the vehicles’ movements. In particular, this is the case of many safety and traffic management applications. For instance, a crash warning service might make on-coming vehicles slow down or change lane. Similarly, a

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

13

Fig. 6. Relationships between mobility and network simulators.

traffic congestion-aware navigator should reroute vehicles along less congested roads. In these cases, independent mobility and communication simulations are not an option to evaluate the system. Two different types of solutions have been adopted to solve this issue: (i) bind existing network simulator and mobility simulator by means of a common two-way interface, as in Fig. 6(b); or (ii) develop a single integrated simulator that handles both aspects simultaneously (Fig. 6(c)). The former is much more flexible and allows for the independent evolution of both simulators, while the latter is usually more efficient from a simulation runtime viewpoint. Let us first discuss solutions of the first type that leverage the SUMO road traffic simulation package. The Traffic and Network Simulation Environment5 (TraNS) [91] provides two-way online communications between SUMO and ns-2 through the Traffic Control Interface (TraCI) [92]. SUMO issues mobility updates to ns-2 through TraCI, and the application logic implemented in ns-2 sends back series of atomic mobility commands (stop, change lane, change speed, change route and change target, all on a per-vehicle basis) to SUMO. The rationale is that the ITS application can implement high level operations like ‘avoid crash’ by issuing series of these atomic commands. SUMO and ns-2 communicate over a TCP connection, in which the mobility simulator acts as the TraCI server and the network simulator as the TraCI client. In this way, both simulators can be collocated on the same host or distributed among two different machines. As a final note, please be aware that TraNS development has been discontinued and it only works with a specific snapshot of SUMO. ns-2 has been also employed to perform joint simulations with commercial mobility tools like VISSIM [93]. Similar to TraNS, Veins [30] provides SUMO and OMNeT++ twoway coupling by means of TraCI. In this way, the ITS service logic is allowed to stop vehicles, change their speed, or change their route. SUMO is expected to run with TraCI server support. iTETRIS introduces a 3-blocks architecture [94] in which a mediator called iTETRIS Control System (iCS) is in charge of configuring, synchronizing and controlling the simulations performed in SUMO and ns-3. This

5

http://lca.epfl.ch/projects/trans/.

architecture allows for a straightforward mapping of the ETSI communications standards for ITS stations to the simulation framework. Specifically, SUMO deals with the mobility simulation, ns-3 hosts the different communications-related layers defined by the ETSI (management, access, transport & network, and communication facilities), and the iCS implements the application-related facility layer that supports ITS services. Hence, the iCS constitutes the user interface to application developers. Each block can be independently evolved by different development communities and end users just need to interact with the iCS to deploy a certain simulation scenario. A different approach is undertaken by the remaining simulators that we review next. In these cases, network and mobility simulation are both integrated within a single tool. This is just what Application-aware SWANS with Highway mobility (ASH) [95] provides. It extends an existing network simulation tool, SWANS6, with a two-way communication facility between the network and mobility models. ASH only supports highway scenarios, in which vehicles move according to the IDM car-following model and the MOBIL lane-changing model. The simulator does not provide IEEE 802.11p support, and the range of wireless propagation models is quite limited. The VGSim [96] vehicular simulator is also based on SWANS. In this case, an enhanced N-S car-following model that features finer spatial and temporal resolution as well as lane-changing capability is implemented. As in the case of ASH, VGSim is limited to the network simulation models provided by SWANS. Finally, the NCTUns network simulator and emulator [97] provides fully integrated mobility and communications simulations. It supports WAVE/802.11p communications [98] and a variety of wireless propagation models. Hence, NCTUns users can choose among several theoretical channel models or empirical channel models. Among the latter, there are specific propagation models for wireless vehicular networks [99]. Regarding mobility support, this tool implements the IDM car-following model and MOBIL lane-change. A summary of this integrated simulator and the previously reviewed ones ca be found in Table 5.

6

http://jist.ece.cornell.edu/.

14

F.J. Ros et al. / Computer Communications 43 (2014) 1–15

Table 5 Comparative table of integrated simulators. Simulators

Mobility

Path loss

Fading

Scenarios

TraNS iTETRIS Veins ASH

SUMO SUMO SUMO IDM MOBIL N-S

ns-2 ns-3 OMNeT++ Free space Two-ray ground Free space Two-ray ground Free space Two-ray ground Shadowing Empirical

ns-2 ns-3 OMNeT++ Rayleigh Ricean Rayleigh Ricean Rayleigh Ricean

Unrestricted Unrestricted Unrestricted Highway

VGSim NCTUns

IDM MOBIL

Highway Unrestricted

6. Conclusion We have surveyed the most prominent simulation models for wireless propagation, communication technologies, and mobility within the context of vehicular networks. Sample configurations and/or useful references for most of them have been provided. In addition, we have reviewed widely employed simulation tools and discussed the different models they support. With this information, a researcher should be able to make a decision about the most appropriate simulators, models and configurations that are required for a particular study. In this way, realistic simulation results for an ITS solution can be obtained and interpreted. We hope this paper to encourage good simulation practices and provide a handy reference to set up appropriate scenarios for vehicular networks.

Acknowledgment This work has been partially funded by the ‘‘Technologies for Cyber-Physical Systems’’ Strategic Action for Preferential Research of the University of Murcia.

References [1] M. Sichitiu, M. Kihl, Inter-vehicle communication systems: a survey, IEEE Commun. Surv. Tutor. 10 (2) (2008) 88–105. [2] International Organization for Standardization, ISO 21217, Intelligent Transport Systems – Communications access for land mobiles (CALM) – Architecture, 2010. [3] European Telecommunications Standards Institute, ETSI EN 302 665 v1.1.1, Intelligent Transport Systems (ITS); Communications, Architecture, 2010. [4] M. Tsukada, I. Ben-Jemaa, H. Menouar, W. Zhang, M. Goleva, T. Ernst, Experimental evaluation for ipv6 over vanet geographic routing, in: IWCMC 2010: Proceedings of the 6th International Wireless Communications and Mobile Computing Conference, 2010, pp. 736–741. [5] F.J. Ros, Scalable data delivery in mobile multi-hop ad-hoc networks – application to gateway discovery protocols (Ph.D. thesis), University of Murcia, Spain, 2011. [6] Fot net, . [7] D. Kotz, C. Newport, R.S. Gray, J. Liu, Y. Yuan, C. Elliott, Experimental evaluation of wireless simulation assumptions, in: Proceedings of the 7th ACM International Symposium on Modeling, Analysis and Simulation of Wireless and Mobile Systems, MSWiM ’04, 2004, pp. 78–82. [8] A. Jardosh, E.M. Belding-Royer, K.C. Almeroth, S. Suri, Towards realistic mobility models for mobile ad hoc networks, in: Proceedings of the Ninth Annual International Conference on Mobile Computing and Networking, MobiCom ’03, 2003, pp. 217–229. [9] M. Fiore, Vehicular Mobility Models, Chapman & Hall/CRC, 2009. Ch. 12, pp. 1–57. [10] Y. Li, Z. Wang, D. Jin, L. Zeng, S. Chen, Collaborative vehicular content dissemination with directional antennas, IEEE Trans. Wirel. Commun. 11 (4) (2012) 1301–1306. [11] J. Zhao, R. Govindan, Understanding packet delivery performance in dense wireless sensor networks, in: SenSys03: Proceedings of the First International Conference on Embedded Networked Sensor Systems, ACM Press, 2003, pp. 1–13.

[12] C.-X. Wang, X. Cheng, D.-I. Laurenson, Vehicle-to-vehicle channel modeling and measurements: recent advances and future challenges, IEEE Commun, Mag. 47 (11) (2009) 96–103. [13] T.S. Rappaport, Wireless Communications: Principles and Practice, second ed., Prentice Hall, 2002. [14] H. Friis, A note on a simple transmission formula, in: I.R.E. and Waves and Electrons, 1946, pp. 254–256. [15] Intelligent transport systems (its); stdma recommended parameters and settings for cooperative its; access layer part, Technical Report ETSI TR 102 861 V1.1.1, ETSI, 2012. [16] Q. Chen, F. Schmidt-Eisenlohr, D. Jiang, M. Torrent-Moreno, L. Delgrossi, H. Hartenstein, Overhaul of ieee 802.11 modeling and simulation in ns-2, in: MSWiM’07: Proceedings of the 10th ACM/IEEE International Symposium on Modeling, Analysis, and Simulation of Wireless and Mobile Systems, 2007, pp. 159–168. [17] V. Taliwal, D. Jiang, H. Mangold, C. Chen, R. Sengupta, Empirical determination of channel characteristics for dsrc vehicle-to-vehicle communication, in: VANET’04: Proceedings of the First ACM International Workshop on Vehicular, Ad Hoc Networks, 2004, pp. 88–88. [18] L. Cheng, B. Henty, D. Stancil, F. Bai, P. Mudalige, Mobile vehicle-to-vehicle narrow-band channel measurement and characterization of the 5.9 ghz dedicated short range communication (dsrc) frequency band, IEEE J. Sel. Areas Commun. 25 (8) (2007) 1501–1516. [19] F.J. Ros, P.M. Ruiz, I. Stojmenovic, Acknowledgment-based broadcast protocol for reliable and efficient data dissemination in vehicular ad-hoc networks, IEEE Trans. Mob. Comput. 11 (1) (2012) 33–46. [20] I. Sen, D.-W. Matolak, Vehicle-vehicle channel models for the 5-ghz band, IEEE Trans. Intell. Transp. Syst. 9 (2) (2008) 235–245. [21] M. Nakagami, The m-distribution, a general formula of intensity distribution of the rapid fading, Statistical Methods in Radio Wave Propagation. [22] S. Moser, F. Kargl, A. Keller, Interactive realistic simulation of wireless networks, in: RT’07: Proceedings of the Second IEEE/EG Symposium on Interactive Ray Tracing, 2007, pp. 161–166. [23] L. Pilosu, F. Fileppo, R. Scopigno, Radii: a computationally affordable method to summarize urban ray-tracing data for vanets, in: WiCOM’11: Proceedings of the Seventh Int. Conf. on Wireless Communications, Networking and Mobile Computing, 2011, pp. 1–6. [24] C. Sommer, D. Eckhoff, R. German, F. Dressler, A computationally inexpensive empirical model of ieee 802.11p radio shadowing in urban environments, in: WONS’11: Proceedings of the 8th Int. Conf. on Wireless On-Demand Network Systems and Services, 2011, pp. 84–90. [25] R. Scopigno, H. Cozzetti, Signal shadowing in simulation of urban vehicular communications, in: IWCMC’10: Proceedings of the 6th Int. Wireless Communications and Mobile Computing Conference, 2010, pp. 131–138. [26] R. Meireles, M. Boban, P. Steenkiste, O. Tonguz, J. Barros, Experimental study on the impact of obstructions in vehicular ad hoc networks, in: VNC’10: Proceedings of the IEEE Vehicular Networking Conference, 2010. [27] M. Baguena, C. Calafate, J. Cano, P. Manzoni, Towards realistic vehicular network simulation models, in: Wireless Days, IEEE, 2012, pp. 1–3. [28] C. Sommer, S. Joerer, M. Segata, O. Tonguz, R.L. Cigno, F. Dressler, How shadowing hurts vehicular communications and how dynamic beaconing can help, in: INFOCOM’13: Proceedings of 32nd IEEE Conference on Computer Communications, Mini-Conference, 2013, pp. 110–114. [29] M. Boban, T.T. Vinhoza, M. Ferreira, J. Barros, O.K. Tonguz, Impact of vehicles as obstacles in vehicular ad hoc networks, IEEE J. Sel. Areas Commun. 29 (1) (2011) 15–28. [30] D. o. C.S. Veins project (University of Erlangen), Vehicles in network simulation, . [31] Propagation by diffraction, Tech. Rep. Recommendation P.526, International Telecommunication Union Radiocommunication Sector, 2007. [32] The network simulator ns-2, . [33] The network simulator ns-3, . [34] Omnet++ simulator framework, . [35] Openstreetmap, . [36] C. Sommer, S. Joerer, F. Dressler, On the Applicability of Two-Ray Path Loss Models for Vehicular Network Simulation, in: VNC’12: Proceedings of the 4th IEEE Vehicular Networking Conference, 2012, pp. 64–69. [37] Institute of Electrical and Electronics Engineers, 802.11p-2010, IEEE Standard for Information technology – Telecommunications and information exchange between systems – Local and metropolitan area networks – Specific requirements Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications Amendment 6: Wireless Access in Vehicular, Environments, 2010. [38] Institute of Electrical and Electronics Engineers, 802.11-2007, IEEE Standard for Information technology – Telecommunications and information exchange between systems – Local and metropolitan area networks – Specific requirements – Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, 2007. [39] Institute of Electrical and Electronics Engineers, 1609.1-2006. IEEE Trial-Use Standard for Wireless Access in Vehicular Environments (WAVE) – Resource, Manager, 2006. [40] Institute of Electrical and Electronics Engineers, 1609.2-2006. IEEE Trial-Use Standard for Wireless Access in Vehicular Environments (WAVE) – Security Services for Applications and Management Messages, 2006.

F.J. Ros et al. / Computer Communications 43 (2014) 1–15 [41] Institute of Electrical and Electronics Engineers, 1609.3-2007. IEEE Trial-Use Standard for Wireless Access in Vehicular Environments (WAVE) – Networking Services, 2007. [42] Institute of Electrical and Electronics Engineers, 1609.4-2006. IEEE Trial-Use Standard for Wireless Access in Vehicular Environments (WAVE) – MultiChannel Operation, 2006. [43] M. Lacage, T.R. Henderson, Yet another network simulator, in: Proc. of the 2006 Workshop on Ns-2: The IP Network Simulator, WNS2 ’06, ACM, 2006. [44] J. Bu, Implementation of wave 1609.4/802.11p, 2013, . [45] itetris – an integrated wireless and traffic platform for real-time road traffic management solutions, . [46] R. Bauza, J. Hernandez, J. Gozalvez, S. Vaz, D. Valiente, I. Aguado, A. Gonzalez, itetris heterogeneous wireless communication platform for the large-scale evaluation of cooperative road traffic management policies, online at: . [47] D. Eckhoff, C. Sommer, A multi-channel ieee 1609.4 and 802.11p edca model for the veins framework, in: Fifth ACM/ICST International Conference on Simulation Tools and Techniques for Communications, Networks and Systems: Fifth ACM/ICST International Workshop on OMNeT++ (OMNeT++ 2012), Poster Session, ACM, Desenzano, Italy, 2012. [48] A. Köpke, M. Swigulski, K. Wessel, D. Willkomm, P.K. Haneveld, T. Parker, O. Visser, H.S. Lichte, S. Valentin, Simulating wireless and mobile networks in omnet++ – the mixim vision, in: Proc. of the 1st International Workshop on OMNeT++ (hosted by SIMUTools 2008), ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering), 2008. [49] S.R. Mousa, Report on car following models and their evaluation, cairo University, Faculty of Engineering, Department of Civil Engineering. [50] T.P. Ranjitkar, A. Kawamua, Car-following models: an experiment based benchmarking, J. East. Asia Soc. Trans. Stud. 6 (2005) 1582–1596. [51] R.E. Chandler, R. Herman, E.W. Montroll, Traffic dynamics: studies in car following, Oper. Res. 6 (2) (1958) 165–184. [52] P. Chakroborty, S. Kikuchi, Evaluation of the general motors based carfollowing models and a proposed fuzzy inference model, Trans. Res. Part C Emerg. Technol. 7 (4) (1999) 209–235. [53] M. Treiber, A. Hennecke, D. Helbing, Congested traffic states in empirical observations and microscopic simulations, Rev. E 62 (2000) 2000. [54] S. Krau, Microscopic modeling of traffic flow: Investigation of collision free vehicle dynamics, Deutsches Zentrum fuer Luftund Raumfahrt. Forschungsberichte. [55] E. Kometani, T. Sasaki, On the stability of traffic flow, J. Oper. Res. Soc. Jpn. (1958) 11–26. [56] G.G. Peter, A behavioral car-following model for computer simulation, Trans. Res. B 15 (2) (1981) 105–111. [57] W. Leutzbach, R. Wiedemann, Development and application of traffic simulation models at karlsuhe institut fur verkehrwesen, Traffic Eng. Control (1986). [58] K. Nagel, M. Schreckenberg, A cellular automaton model for freeway traffic, J. de Physique I 2 (12) (1992) 2221–2229. [59] M. Bando, K. Hasebe, A. Nakayama, A. Shibata, Y. Sugiyama, A dynamical model of traffic congestion and numerical simulation, Phys. Rev. Part E 51 (2) (1995) 1035–1042. [60] G. Newell, A simplified car-following theory: a lower order model, Transp. Res. Part B method. 36 (2002) 195–205. [61] P.G. Gipps, A model for the structure of lane-changing decisions, Transp. Res. Part B Method. 20 (5) (1986) 403–414. [62] Y. Zhang, L.E. Owen, J.E. Clark, Multiregime approach for microscopic traffic simulation, Transp. Res. Record J. Transp. Res. Board 1644 (1) (1998) 103–114. [63] P. Hidas, Modelling vehicle interactions in microscopic simulation of merging and weaving, Transp. Res. Part C Emerg. Technol. 13 (1) (2005) 37–62. [64] R. Liu, D. Van Vliet, D. Watling, Microsimulation models incorporating both demand and supply dynamics, Transp. Res. Part A Policy Pract. 40 (2) (2006) 125–150. [65] Q. Yang, H.N. Koutsopoulos, A microscopic traffic simulator for evaluation of dynamic traffic management systems, Transp. Res. Part C Emerg. Technol. 4 (3) (1996) 113–129. [66] K.I. Ahmed, M. Ben-Akiva, H.N. Koutsopoulos, R.G. Mishalani, Models of freeway lane changing and gap acceptance behavior, in: Proceedings of the 13th international symposium on the theory of traffic flow and transportation, Pergamon, New York, NY, 1996, pp. 501–515. [67] T. Toledo, H.N. Koutsopoulos, M. Ben-Akiva, Integrated driving behavior modeling, Transp. Res. Part C Emerg. Technol. 15 (2) (2007) 96–112. [68] A. Kesting, M. Treiber, D. Helbing, General lane-changing model mobil for carfollowing models, Transp. Res. Record J. Transp. Res. Board 2007 (1999) 86–94. [69] J. Hourdakis, P.G. Michalopoulos, J. Kottommannil, Practical procedure for calibrating microscopic traffic simulation models, Transp. Res. Record J. Transp. Res. Board 1852 (1) (2003) 130–139. [70] Y. Hollander, R. Liu, Estimation of the distribution of travel times by repeated simulation, Transp. Res. Part C Emer. Technol. 16 (2) (2008) 212–231.

15

[71] T. Ma, B. Abdulhai, Genetic algorithm-based optimization approach and generic tool for calibrating traffic microscopic simulation parameters, Transp. Res. Record J. Transp. Res. Board 1800 (1) (2002) 6–15. [72] S. Ossen, S.P. Hoogendoorn, Validity of trajectory-based calibration approach of car-following models in presence of measurement errors, Transp. Res. Record J. Transp. Res. Board 2088 (1) (2008) 117–125. [73] V. Vaze, C. Antoniou, Y. Wen, M. Ben-Akiva, Calibration of dynamic traffic assignment models with point-to-point traffic surveillance, Transp. Res. Record J. Transp. Res. Board 2090 (1) (2009) 1–9. [74] B.F. Ciuffo, V. Punzo, Verification of traffic micro-simulation model calibration procedures: Analysis of goodness-of-fit measures, in: Transportation Research Board 89th Annual Meeting, no. 10–3229, 2010. [75] B.F. Ciuffo, V. Punzo, M. Montanino, The calibration of traffic simulation models, report on the assessment of different goodness of fit measures and optimization algorithms, Tech. Rep. JRC68403, European Commission, Joint Research Centre, Institute for Environment and Sustainability. Institute for Energy and Transport, 2012. [76] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, IEEE Trans. Evol. Comput. 1 (1) (1997) 67–82. [77] J.C. Spall, S.D. Hill, D.R. Stark, Theoretical framework for comparing several stochastic optimization approaches, in: Probabilistic and Randomized Methods for Design under Uncertainty, Springer, 2006, pp. 99–117. [78] V. Punzo, B. Ciuffo, How parameters of microscopic traffic flow models relate to traffic conditions and implications on model calibration, Transportation Research Record. [79] M.M. Technology, Traffic software integrated system - corridor simulation, . [80] Q.P. Ltd., Paramics, . [81] P. America, Vissim 5.3, . [82] Tss–transport simulation systems, aimsun, tss, barcelona, spain, 2013. . [83] F. Choa, R.T. Milam, D. Stanek, Corsim, paramics, and vissim: What the manuals never told you, in: Ninth TRB Conference on the Application of Transportation Planning, Methods, 2004, pp. 392–402. [84] M. Hadi, P. Sinha, A. Wang, Modeling reductions in freeway capacity due to incidents in microscopic simulation models, Transp. Res. Record J. Transp. Res. Board 1999 (1) (2007) 62–68. [85] I. of Transportation Systems at the German Aerospace Center, Simulation of urban mobility, . [86] N.T. Ratrout, S.M. Rahman, A comparative analysis of currently used microscopic and macroscopic traffic simulation software, Arab. J. Sci. Eng. 34 (1B) (2008) 121–133. [87] Y. Gongjun, I. Khaled, W. Michele, Vehicular Network Simulators, Chapman & Hall/CRC, 2009. Ch. 13, pp. 1–22. [88] Realistic mobility generator for vehicular networks, . [89] S. Uppoor, O. Trullols-Cruces, M. Fiore, J. Barcelo-Ordinas, Generation and analysis of a large-scale urban vehicular mobility dataset, IEEE Transactions on Mobile Computing (2013). [90] D. Krajzewicz, Kombination von taktischen und strategischen einflüssen in einer mikroskopischen verkehrsflusssimulation, 2009. [91] M. Piórkowski, M. Raya, A.L. Lugo, P. Papadimitratos, M. Grossglauser, J.-P. Hubaux, Trans: realistic joint traffic and network simulator for vanets, SIGMOBILE Mob. Comput. Commun. Rev. 12 (2008) 31–33. [92] A. Wegener, M. Piórkowski, M. Raya, H. Hellbrück, S. Fischer, J.-P. Hubaux, Traci: an interface for coupling road traffic and network simulators, in: Proceedings of the 11th Communications and Networking Simulation Symposium, CNS ’08, ACM, 2008, pp. 155–163. [93] C. Lochert, A. Barthels, A. Cervantes, M. Mauve, M. Caliskan, Multiple simulator interlinking environment for ivc, in: K.P. Laberteaux, H. Hartenstein, D.B. Johnson, R. Sengupta (Eds.), Vehicular Ad Hoc Networks, ACM, 2005, pp. 87– 88. [94] V. Kumar, R. Bauza, F. Filali, J. Gozalvez, L. Lin, M. Rondinone, iTETRIS: a large scale integrated simulation platform for V2X communications: Application to real-time traffic management, in: ITST 2009, Ninth International Conference on ITS Telecommunications, Lille, France, 2009. [95] K. Ibrahim, M.C. Weigle, Ash: application-aware swans with highway mobility, in: INFOCOM Workshops 2008, IEEE, IEEE, 2008, pp. 1–6. [96] B. Liu, B. Khorashadi, H. Du, D. Ghosal, C.-N. Chuah, M. Zhang, Vgsim: an integrated networking and microscopic vehicular mobility simulation platform, Commun. Mag. 47 (2009) 134–141. [97] S.Y. Wang, C.L. Chou, Nctuns tool for wireless vehicular communication network researches, Simul. Modell. Pract. Theory 17 (7) (2009) 1211–1226. [98] S.-Y. Wang, C.-C. Lin, Nctuns 5.0: a network simulator for ieee 802.11(p) and 1609 wireless vehicular network researches, in: VTC Fall, IEEE, 2008, pp. 1–2. [99] S.-Y. Wang, P.-F. Wang, Y.-W. Li, L.-C. Lau, Design and implementation of a more realistic radio propagation model for wireless vehicular networks over the nctuns network simulator, in: WCNC, IEEE, 2011, pp. 1937–1942.