Considering driver behaviour in vehicular delay tolerant networks. Kun-chan Lan, Chien-Ming Chou* and Ching-Fang Yang

152 Int. J. Satellite Communications Policy and Management, Vol. 1, Nos. 2/3, 212 Considering driver behaviour in vehicular delay tolerant networks Kun-chan Lan, Chien-Ming Chou* and Ching-Fang Yang Department of Computer Science and Information Engineering, National Cheng Kung University, No. 1, University Road, Tainan City 71, Taiwan E-mail: klan@csie.ncku.edu.tw E-mail: jensen915@gmail.com E-mail: ycf123@gmail.com *Corresponding author Abstract: A key component for vehicular ad-hoc networks (VANET) simulations is a realistic vehicular mobility model, as this ensures that the conclusions drawn from simulation experiments will carry through to the real deployments. Node mobility in a vehicular network is strongly affected by the driving behaviour such as route choices. While route choice models have been extensively studied in the transportation community, as far as we know, the effects of driver behaviour on vehicular network simulations have not been discussed much in the networking literature. In this work, we set out to understand the effects of destination selection vehicular network simulation. We also discuss how different destination selection models affect event broadcasting in ITS application. We conclude that selecting a sufficient level of detail in the simulations, such as modelling of destination selection, is critical for evaluating VANET protocol design. Keywords: vehicular ad-hoc networks; VANET; driver behaviour; simulation; destination selection; event dissemination. Reference to this paper should be made as follows: Lan, K-c., Chou, C-M. and Yang, C-F. (212) Considering driver behaviour in vehicular delay tolerant networks, Int. J. Satellite Communications Policy and Management, Vol. 1, Nos. 2/3, pp.152 173. Biographical notes: Kun-chan Lan received his PhD from Computer Science at University of Southern California in 24, advised by Professor John Heidemann. From 24 to 27, he joined the Network and Pervasive Computing Programme at NICTA (National ICT Australia) in Sydney as a Researcher. He is currently an Assistant Professor at the Department of Computer Science and Information Engineering of National Cheng Kung University. Chine-Ming Chou received his BS from Computer Science at Vanung University of Technology in 25 and MS from Information and Communication Engineering at Chaoyang University of Technology in 27. He is currently working for his PhD in Department of Computer Science and Information Engineering, National Cheng Kung University. His research interests include network simulation, network measurement, opportunistic routing, WiMAX and DSRC protocol, and mobility model. Copyright 212 Inderscience Enterprises Ltd.

Considering driver behaviour in vehicular delay tolerant networks 153 Ching-Fang Yang received his BS and MS in Electrical Engineering from National Cheng Kung University, Taiwan, in 1999 and 22, respectively. From 23 to 28, he has been with the Industrial Technology Research Institute as a Research Engineer. He obtained his PhD in the Department of Electrical Engineering at National Cheng Kung University in 21. Currently, he is a Postdoctoral Fellow in National Cheng Kung University. His current research interests are in the areas of optical networks, and vehicular network management. 1 Introduction Vehicular ad-hoc network (VANET) communication has recently become an increasingly popular research topic in the area of wireless networking, as well as in the automotive industry. The goal of VANET research is to develop a vehicular communication system to enable the quick and cost-efficient distribution of data for the benefit of passengers safety and comfort. While it is crucial to test and evaluate protocol implementations in a real world environment, simulations are still commonly used as a first step in the protocol development for VANET research. Several communication networking simulation tools already exist to provide a platform to test and evaluate network protocols, such as ns-2 (Breslau et al., 2), OPNET (Chang, 1999) and Qualnet (28 211). However, these tools are designed to provide generic simulation scenarios, without being particularly tailored for applications in the transportation environment. In addition, simulations also play an important role in the field of transportation. A variety of simulation tools, such as PARAMICS (Cameron and Duncan, 1996), CORSIM (Halati et al., 1997) and VISSIM (Fellendorf, 1994), have been developed to analyse transportation scenarios at the micro- and macro-scale levels. However, to date there have been only few attempts (Saha and Johnson, 24; Mahajan et al., 26; Baumann et al., 27; Dressler and Sommer, 21) to create communication scenarios in a realistic transportation simulation environment. One of the most important parameters in simulating vehicular networks is the node mobility. It is important to use a realistic mobility model so that results from the simulation correctly reflect the real world performance of a VANET, as shown in some prior studies (Saha and Johnson, 24; Heidemann et al., 21). Node mobility in a vehicular network is strongly affected by the drivers behaviour, which can change road traffic at different levels. Drivers preferences in path and destination selection can further affect the overall network topology. It has been shown that drivers tend to use certain regular routes for their daily routines (Abdel-aty et al., 1994), and only 15.5% of commuters reported that they did not always choose the same exact route to work. Once a commuter has settled on a habitual route, the route choice strategies they deploy might possibly descend to a subconscious level, unless there are external factors (e.g., accidents or traffic jams) that bring the choice of route back to the conscious level (Tawfik et al., 21). Furthermore, some commuters might select their routes based on the suggestions of some travel guidance system, such as variable message signs. Once a commuter has had a good experience with using a travel guidance system, they might increase their reliance on such advice the next time they travel (Zhao and Shao, 21). While most

154 K-c. Lan et al. current navigation systems use the shortest path to the destination for selecting routes, some commuters use faster paths instead of shorter ones to avoid congestion and reduce travel time. Some studies also show that path selection could possibly change on a temporal basis (Li et al., 25; Chen and Mahmassani, 1993). For example, when driving in the evening commuters usually have more flexibility in selecting alternate routes than when they drive to work in the morning. In this paper, we set out to understand the effect of path selection (for a particular destination) and destination selection on vehicular network simulations. We also consider an application scenario in which cars want to disseminate their sensor information over the vehicular network via vehicle-to-vehicle communication only. The contributions of this paper are twofold: first, we model destination selection and analyse cluster size in VANET simulation second, we demonstrate how the destination selection could potentially affect the performance of practical ITS application. The remainder of this paper is organised as follows: in Section 2 we review the related literature. In Section 3, we model and analyse the effect of different destination selection models on a ITS application scenario. We use simulations to demonstrate how the destination selection could potentially affect the performance of practical ITS application in Section 4. Finally, in Section 5 we show how destination selection models interact with other parameters in the simulations, and conclude this paper in Section 6. 2 Related work Details in the mobility model may have a critical effect on the fidelity, and thus the usefulness, of the resulting network simulations. Zhang et al. (27) used traces taken from the UMass DieselNet project (Burgess et al., 26) to study the effect of mobility models on the performance of DTN. They showed that a finer-grained route-level model of inter-contact times is able to predict performance much more accurately than a coarser-grained all-bus-pairs aggregated model, which suggests caution should be taken in choosing the right level of detail when modelling vehicle mobility. Random waypoint (RWP) (David and David, 1996) is an earlier mobility model that has been widely used in mobile ad-hoc network (MANET) simulations, such as those in (Das et al., 2; Holland and Vaidya, 1999; Perkins and Royer, 1999). RWP assumes that nodes can move freely in a simulation area, without considering any obstacles. RWP model fails to provide a steady state in that the average nodal speed consistently decreases over time (Yoon et al., 23), and its is inapplicable in VANET simulation since in a VANET environment vehicles are typically restricted by streets, traffic lights and obstacles. Hong et al. (1999) proposed a reference point group mobility (RPGM) model to characterise the relationship between mobile hosts. Bettstetter (21) presented a random direction model, which introduces a stop-turn-and-go behaviour that can mimic vehicle behaviour at intersections. Saha and Johnson (24) proposed a macro mobility model based on the TIGER map database. They considered the use of the Dijkstra shortest path algorithm to select the path from source to destination. Treiber et al. (2) discussed a model that supports car turning at the intersections. Bai et al. (23) undertook a similar study, and further

Considering driver behaviour in vehicular delay tolerant networks 155 introduced Freeway and Manhattan mobility models in which car following and turning behaviours are included. Sven Jaap and Wolf (25) presented a city mobility model that is based the intelligent-driver model (IDM). Stepanov et al. (25) described a spatial model that considered path selection and user movement dynamics (such as road congestion and car following behaviour). However, unlike our study, none of these works considers the effect of traffic lights in their simulations. The street random waypoint (STRAW) (Choffnes and Bustamante, 25) model considers traffic light control and car following and uses a shortest path algorithm to calculate the movement paths. Mahajan et al. (26) discussed the stop sign model (SSM), probabilistic traffic sign model (PTSM), and traffic light model (TLM) in the context of a traffic control system. In SSM, every vehicle stops at the stop sign for a fixed duration time. In contrast, PTSM uses a probability p to decide if a vehicle needs to stop at an intersection. Meanwhile, in TLM, traffic from different directions is considered in order to adjust the traffic light cycle to minimise road congestion. Mahajan et al. (26) showed that a simpler SSM model could have significantly different results when compared to a more sophisticated PTSM model. Finally, Baumann et al. (27) proposed two mobility models, one using vectorised street information from the Swiss geographic information system (GIS), and the other is based on a microscopic, multi-agent traffic simulator (MMTS) (Raney et al., 22) to generate vehicle movement traces. In the GIS-based mobility model, the actual node movement is generated according to the random trip model (Boudec and Vojnovic, 25) on the vectorised street map. In contrast, MMTS models the behaviour of people living in the area, and the travel plans of each individual as well as road congestion are considered in their simulations. However, such a trace-based model, while closely imitating reality, requires a high level of computing power to generate the traces. Huang et al. (21) designed three model parameters: turn probability, road section speed and travel pattern to capture the regularity of the taxi mobility from GPS traces. Travel pattern depicts the regularity of long run trips based on travel grids from origination to destination for probability matrixes. However, to decide the size of the travel grid is challenging since a large size of the travel grid will merge different travel patterns and a small size of the travel gird cannot capture the characteristic of the travel pattern. Most of these prior work selected routes in their simulations based on random decisions and use the same routes throughout the simulation. In reality, drivers might have habitual routes from the source to the destination and change their routes from time to time to avoid congestion. Complementary to previous studies, in this paper we look at the effect of preferred routes on VANET simulations. Many different route choice models have been proposed. For example, Dia and Panwai (27) used fuzzy logic to model the impact of traveller information system on route choices. Liu and Huang (27) investigated day-to-day route paths and modelled them with the logit-based stochastic user equilibrium state. In addition, Shenpei and Xinping (28) discussed how route choices are affected by signal split, while Zhao and Li (21) proposed a route choice model that considers human memory and traffic information factors. Guo et al. (21) proposed a path choice model based on game theory, and assumed that drivers obey traffic information to select their alternate routes. Dingus et al. (1996) discussed how route choices are affected by human factors such as efficiency (e.g., fastest route vs. shortest route), problem avoidance (e.g., safer routes) and road condition (e.g., number of traffic lights). They showed a shortest path is not necessarily the driver s first choice when selecting routes, and that very often commuters

156 K-c. Lan et al. use faster paths to avoid congestion and reduce travel time (Jan et al., 2). Finally, Tawfik et al. (21) showed that drivers perceptions are significantly different from their actual experiences, and drivers choices can be better explained by the former rather than the latter. While there are a huge amount of works that model driving behaviour in the transportation literature, there are only a few studies in the networking literature that described the impact of driving behaviour on vehicular network simulations. For example, Mahajan et al. (26) compared SSM, PTSM, and TLM models in simulations. Fiore and Härri (28) simulated different mobility models and observed their effect on cluster size and link duration. Viriyasitavat et al. (29) provided an analysis of how traffic lights affect network topology and connectivity. Dressler and Sommer (21) evaluated route choice strategies via different route choices to show their impact on average speed. They did not consider the effect of route choice on the performance of network communication though. In this work, we set out to understand the effect of destination selection. We also discuss the effects of different destination selection models on ITS applications. 3 Modelling and analysis of destination selection In this section, we show that a destination selection model could have different effects on ITS application, data dissemination in a vehicular network (Spyropoulos et al., 25; Costa et al., 26). We consider a scenario in which a car wants to broadcast an event (e.g., a car accident) over the whole network via vehicle-to-vehicle communication in a flooding-like fashion [e.g., using epidemic routing (Vahdat and Becker, 2)]. For data dissemination applications, the aim is to quickly disseminate data (e.g., accident information) over the whole network. Therefore, the performance of such applications might depend on the concentration of the nodes in a certain area. In other words, the distribution of cluster sizes could play an important role for the effectiveness of such applications. In this section, we first briefly discuss how different destination selection models affect network connectivity and cluster size. 3.1 Modelling of destination selection As described previously, drivers tend to exhibit a bias in their destination selection (Abdel-aty et al., 1994), and thus some locations could potentially be visited more often than others. Different destination selection patterns will result in different network topologies and levels of connectivity. To simplify the discussion, let us assume that the selection of destinations follows a certain probability distribution. Here we consider three different probability distributions:, exponential, and uniform. We model the destination selection process as follows: Given a two-lane, bi-directional roads graph, G = (I, R), which consists of a set of intersections, I(i 1, i 2,,i K ), and a set of roads, R(r 1, r 2,,r L ), where K is the number of intersections and L is the number of roads. The vehicle chooses the shortest path from the source intersection i s to the destination intersection i d, for 1 s K, 1 d K. Each car is initially placed at a random intersection i s within the simulation area. Next, each car chooses a new intersection i k by a probability ζ d (i.e., uniform, exponential, or distribution) to move with a

Considering driver behaviour in vehicular delay tolerant networks 157 velocity v chosen from the interval [, v max ], where v max is the maximum speed allowed on the road. v changes over time depending on several factors, including the acceleration/deceleration ability of the car, car-following and lane-changing behaviours, and encounters of traffic lights. Each car continues this process for the duration of the simulation. When the selection of destinations follows a uniform distribution, it suggests that the probability of a car visiting any location on the map is uniformly distributed. On the other hand, when the selection of destination follows a distribution, it implies that some locations are visited much more often than others. To understand the effects of destination selection, we setup a simulation using a 4 4 grid map with 1 cars. The length of the road segment is 4 m. As shown in Figures 1 and 2, when cars pick their destination following a destination, the network will have a larger cluster coefficient (Fiore and Härri, 28) over time and every car will have more neighbouring nodes, as compared to the cases when cars choose their destinations following an exponential or uniform distribution. Figure 1 Comparison of the number of neighbouring nodes for the three node distributions CDF 1..9.8.7.6.5.4.3.2.1. Exponential 5 1 15 2 25 3 35 4 45 5 55 6 The number of neighboring nodes Figure 2 Comparison of the average cluster coefficient for the three node distributions Average clustering coefficient 1..9.8.7.6.5.4 Exponential.3 5 49 48 47 46 45 44 43 42 41 4 39 38 37 36 35 34 33 32 31 3 29 28 27 26 25 24 23 22 21 2 Time (sec)

158 K-c. Lan et al. However, this is not unexpected, as when more cars select the same destination (e.g., when students select their school as the destination), the chance that some road segments in their selected paths to the destination overlap will become higher. More overlapping road segments suggest a higher node density, shorter inter-car distance and a lower moving speed, which typically leads to a better network connectivity. 3.2 Analysis of destination selection We consider the effect of preferred destination selection on the distribution of network cluster sizes. We define a cluster as a group of cars that can find at least one multipath route to each other and the cluster size as the number of the cars in the same cluster. Thus, cars travelling on two adjacent connected roads will be in the same cluster. If we model the road network as a graph in which that roads are the edges and intersections are the vertexes in the graph, we can then use a greedy algorithm as the following to search for all the adjacent connected roads to computer clusters and their sizes. T i in the following algorithm represents a cluster i. Input: a connected graph G = (V, E), T Φ, V q Φ Output: T ={T1, T2,, Tn}, a set of n spanning trees GT, = ( V, E ), i {1, 2,, n} i T T Initially, i 1, V tmp V While (V tmp Φ) do { v }, v is arbitrarily chosen from V \ V q as the root of the tree, VT i i i ET i Φ for k 1 to V V q 1, do if (finding an segment e* =(u*, v*) with connectivity is 1 among all the edges e(u, v) such that u V Ti and v V \ ( Vq VT i )) { VT V {*} i T v i ET E {( *, *)} i T u v i } end for Vq Vq VT i V tmpq V V q i++; end while n i for i 1 to n do end for V T i N( T) = ϑ( v ), v V i k k Ti k = 1

Considering driver behaviour in vehicular delay tolerant networks 159 Figure 3 The statistic of clusters for different node densities when different destination selection models are used, (a) the number of nodes is 25 (b) the number of nodes is 5 (c) the number of nodes is 1 Normalized cluster size Normalized cluster size Normalized cluster size 1..9.8.7.6.5.4.3.2.1. 1..9.8.7.6.5.4.3.2.1. 1..9.8.7.6.5.4.3.2.1. Exponential 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 Cluster index (a) Exponential 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 Cluster index (b) Exponential 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 Cluster index (c)

16 K-c. Lan et al. Accordingly, we can get the size of each cluster (i.e., the summation of the number of cars in the connected tree), N(T i ) i {1, 2,,n}. We run the simulation again using the same grid topology as discussed previously, and orderly plot the normalised cluster size for different destination selection models. As shown in Figure 3, the axis represents the index of the cluster sorted by the cluster size and y-axis is the normalised cluster size (i.e., cluster size/total number of cars). Our results show that, when the distribution is used to model the destination selection process, most of the cars will concentrate in the same cluster. Hence, we can expect that a data dissemination application will have smaller forwarding delays when the drivers choose their destinations following a distribution, as compared to when uniform or exponential distribution is used. In addition, we observe that, when the distribution is employed, the cluster size distribution is less affected by the node density. Next, we further use simulations to demonstrate how the destination selection could potentially affect the performance of practical ITS application. 4 Performance evaluation To understand the effect of destination selection on vehicular network simulation, we use MOVE (28 211) to simulate various driving behaviours. MOVE runs on top of an open-source micro-traffic simulator called SUMO (Krajzewicz et al., 22), and allows users to simulate road dynamics on the fly by utilising the traffic control interface (TraCI) (Wegener et al., 28). More details about MOVE can be found in our prior work (Karnadi et al., 27). MOVE also supports modelling obstacles on roads by allowing users to specify the shape of the obstacle and their penetration loss. We used Cramer s Rule (Leon, 26) to check if there is an obstacle between the sender and the receiver and adjust the radio signal attenuation accordingly based on the obstacle s penetration loss. The roads in our simulations have two lanes and are bi-directional. Each simulation runs for 2, seconds and the maximum radio transmission range is 25 m. We enabled CSMA/CA in our simulations and used the two ray ground model to simulate the radio propagation. The road topology generated in our simulation is based on the TIGER database (Marx, 1986) using real-world traces. All nodes employ 82.11 MAC operating at 2 Mbps. Here we consider that the store-and-forward mechanism is employed so that a car can carry a packet around when it cannot immediately find the next forwarder. The performance metric we are interested in here is how long the message takes to reach every car in the network. We use a 4 4 grid map. The length of each grid is 4 m. Each car periodically (i.e., every one second) broadcasts hello message to discover neighbouring node. We employ simple MAC in ns-2 as the underlying MAC protocol. Simple MAC supports CSMA (without the backoff mechanism when data needs to be retransmitted). The sender is randomly selected. For simplicity, here we only consider three different node densities: 2, 8 and 14. Generally speaking, the delay in data dissemination is a function of the inter-contact time of vehicles. Here, we define data dissemination delay as the duration from when the originating car sends out the event until when the other cars receives the event, and inter-contact time as the time interval between two contacts (i.e., the duration from when one car encounter finishes and the next one begins). Karagiannis et al. (27) showed that inter-contact time tends to follow a power law and exponential decay distribution.

Considering driver behaviour in vehicular delay tolerant networks 161 As shown in Figure 4, when cars select their destinations following a uniform distribution, most of the MAC-layer collisions occur at the intersections and the central part of the map. In contrast, when cars select destinations following an exponential or distribution, more than 9% of collisions happen in the bottom part of the map. Figure 4 Distributions of MAC-layer collisions when different destination selection models are used, (a) uniform distribution (b) exponential distribution (c) distribution (see online version for colours) (a) (b) (c)

162 K-c. Lan et al. Figure 5 Effect of different destination selection models on data dissemination delay, (a) uniform distribution (b) exponential distribution (c) distribution (see online version for colours) Average dissemination delay (sec) 21 2 19 18 17 16 15 14 13 12 11 1 9 8 7 6 5 4 3 2 1 Nodes = 2 Nodes = 8 Nodes = 14 3 6 9 12 15 18 21 Average inter-contact time (sec) 24 27 3 (a) Average dissemination delay (sec) 21 2 19 18 17 16 15 14 13 12 11 1 9 8 7 6 5 4 3 2 1 Nodes = 2 Nodes = 8 Nodes = 14 3 6 9 12 15 18 21 Average inter-contact time (sec) 24 27 3 (b) Average dissemination delay (sec) 21 2 19 18 17 16 15 14 13 12 11 1 9 8 7 6 5 4 3 2 1 Nodes = 2 Nodes = 8 Nodes = 14 3 6 9 12 15 18 21 Average inter-contact time (sec) 24 27 3 (c)

Considering driver behaviour in vehicular delay tolerant networks 163 As shown in Figure 5 (each point in the graph is the result from one simulation), we find that the average inter-contact time and dissemination delay are larger when cars choose their destinations following a uniform distribution, as compared to when they follow an exponential or distribution. This is because many cars might cluster together as their selected destinations and the paths to them are similar when they choose destinations following either of these distributions. As a result, the inter-contact distance between cars will be shorter and, consequently, the inter-contact time and dissemination delay are smaller. In addition, we observe that the dissemination delay is not significantly affected by the changing node density when the node density is high (e.g., when nodes = 8 and 14), if the selected destinations follow a distribution (i.e., most of the cars are in the same cluster). Note that some large delays in Figure 5 (e.g., > 13 seconds) are because we happened to select a sender that was far from all the other nodes in that simulation. Next, we consider the effect of destination selection on the network protocol. Many different forwarding techniques have been previously proposed for data dissemination of vehicular network. In this work, we consider the following three different forwarding strategies including epidemic (Vahdat and Becker, 2), probability-based (Lindgren et al., 23) and distance-based (Karp and Kung, 2) forwarding protocols. For our application scenario, we assume that some cars want to send data to a road-side unit (RSU) which is located at the centre of the map. All the protocols we consider use the concept of store-carry-forward to manage the intermittent connectivity between nodes. The number of nodes in our simulations is 5 from which we randomly select a number of cars as the sources (Note that we also try other node densities and the results are similar). As shown in Figure 6, due to its aggressive flooding, epidemic forwarding performs best regardless which destination selection scenario is used. On the other hand, the results of the probability-based protocol are the worst for all three destination selection scenarios because from time to time it might select wrong forwarders which lead to an unnecessary longer delay. Figure 6 Comparison of forwarding delays when cars select the destinations following different probability distributions and use different forwarding strategies (see online version for colours) Average forwarding delay (sec) 35 3 25 2 15 1 5 Epidemic Probability Distance Exponential Different probability distributions for destination selection

164 K-c. Lan et al. Figure 7 Neighbouring nodes of the RSU (when the RSU is located at the centre of the map), (a) uniform distribution (b) exponential distribution (c) distribution The number of RSU neighbors 14 12 1 8 6 4 2 1 15 2 25 3 35 4 45 5 55 Time (sec) 6 65 7 75 8 85 9 The number of RSU neighbors 14 12 1 8 6 4 2 (a) 1 15 2 25 3 35 4 45 5 55 Time (sec) 6 65 7 75 8 85 9 The number of RSU neighbors 14 12 1 8 6 4 2 (b) 1 15 2 25 3 35 4 45 5 55 Time (sec) 6 65 7 75 8 85 9 (c)

Considering driver behaviour in vehicular delay tolerant networks 165 Figure 8 CDF of forwarding delay on different forwarding strategies for three distributions of destination selection, when the RSU is lied at the centre of the map, (a) epidemic (b) probability-based (c) distance-based (see online version for colours) CDF CDF CDF 1..9.8.7.6.5.4.3.2.1. 1..9.8.7.6.5.4.3.2.1. 1..9.8.7.6.5.4.3.2.1. Exp Exp 5 1 15 2 25 3 35 4 45 Forwarding delay (sec) Exp (a) Exp 5 1 15 2 25 3 35 4 45 Forwarding delay (sec) Exp (b) Exp 5 1 15 2 25 3 35 4 45 Forwarding delay (sec) (c)

166 K-c. Lan et al. We find that, when the source vehicles select their destinations following the uniform distribution, nodes tend to cross the centre of the map with a relatively high frequency, a phenomenon previously observed in the RWP mobility model. As shown in Figure 7, when nodes selects their destinations following the uniform distribution, the RSU (which is located at the centre of the map) have more neighbouring nodes, as compared to the cases when the other two distributions are used. As a result, as shown in Figure 8, regardless which forwarding strategy is employed, when nodes selecting the destination following the uniform distribution, the forwarding delay is lower as compared to the other two scenarios because it is easier to find a forwarder to carry the data toward the RSU. On the other hand, if the RSU is moved from the centre to the bottom of the map, we then obtain an opposite result as shown in Figure 9, in which when the distribution is used to model the destination selection has the shortest forwarding delay. This is due to that, in our simulations, when nodes choose their destinations following a distribution, the bottom part of the map will have a higher node density as described previously. Figure 9 Forwarding delay when the epidemic forwarding is employed (RSU is located in the bottom of the map) (see online version for colours) CDF 1..9.8.7.6.5.4.3.2.1. Exp Exp 5 1 15 2 25 3 35 4 45 Forwarding delay (sec) Replication strategy is an important parameter when designing a vehicular network protocol. Multiple-copy forwarding is commonly used to cope with intermittent network connectivity in a vehicular network. For example, epidemic routing floods the network to exploit the best possible delivery delay brought by mobility. This scheme achieves optimal delay assuming unlimited bandwidth and relay buffers. In general, a multiple-copy scheme that allows multiple replicates of the same message to be distributed on multiple relay nodes can improve delivery delay by providing more diverse delivery paths. However, it also incurs significant overhead on the storage space and communication bandwidth requirements of the relay nodes. When multiple messages have to be delivered simultaneously across a vehicular network, these common network resources are under contention. In this work, we show that the distribution of node density can be a function of the destination selection model. For example, as previously in our simple grid network example, more nodes concentrate at the centre of map when the uniform distribution is employed, but cluster in the bottom part of the map when the distribution is used. One might be able to utilise such insights to design a protocol

Considering driver behaviour in vehicular delay tolerant networks 167 that can dynamically adjust the replication policy that adapts to the node distribution, assuming that the distribution of drivers preferred destinations can be known in advance. 5 Discussion Selecting an appropriate level of details for the mobility model used in VANET simulation is a challenging task. Unrealistic mobility models can produce misleading or incorrect results. On the other hand, adding details requires more time to implement and debug the system. In this section, we discuss the interaction between the destination selection and other parameters in VANET simulations. Specifically, we consider different simulation topology and MAC protocol and observe the effect of destination selection in VANET simulations. Finally, we show the distributions of destination selection obtained from real-world vehicle traces. In Section 4, we used a simple grid topology to evaluate the performance of ITS application. In order to understand if our results will be affected by different road topologies, we re-run our simulations for the event broadcasting application using real-world traces obtained from TIGER map database. The results are shown in Figure 1, which has similar delay distributions as those in Figure 5. That is, generally, when drivers select their destinations following a distribution, the application has a better performance as compared to when uniform or exponential distribution is employed. Furthermore, we compare fixed backoff (simple MAC) and dynamic backoff (82.11 MAC) in NS2 to observe the effects of destination selection on different MAC protocols. In both MAC protocols, the contention window (CW) is randomly selected from a specific to avoid collision and reduce collision probability. The difference between these two protocols is that the former uses a fixed backoff time while in the latter the CW is exponentially increased (in addition, RTS/CTS is disabled in this case). As shown in Figure 11, when network density is sparse (nodes number = 2), the dissemination delay are similar in uniform and exponential distributions due to that there are only few collisions happened. A dynamic backoff MAC has a better performance in the case of distribution because a fixed backoff MAC will result in more collisions when nodes are clustered together. When we increase network density, the dynamic backoff MAC results in a shorter delay in uniform and exponential distributions because it is able to adjust CW effectively and reduce collision probability. On the other hand, when drivers select their destination following a distribution, a dynamic backoff MAC has a poor performance when nodes are highly concentrated in a small area (i.e., in the bottom part of the map). This is because the dynamic backoff cannot effectively reduce collision and only lead to a larger and larger waiting time when the collision repeatedly occurs. In other words, when the collision probability is small, a dynamic backoff MAC can achieve a shorter dissemination delay than a fixed backoff MAC. On the other hand, when collisions repeatedly happen, a fixed MAC protocol might be a better choice. We also find that most of the collisions occur at intersections due to traffic light introduce clustering nodes. One might want to consider this phenomenon when designing a MAC protocol for vehicular network, for example, by employ big CW at intersection to reduces collision. Previously, Jerbi et al. (28) proposed a virtual infrastructure to disseminate information at intersections.

168 K-c. Lan et al. Figure 1 Effect of different destination selection models on data dissemination delay using real-world map, (a) uniform distribution (b) exponential distribution (c) distribution (see online version for colours) Average dissemination delay (sec) 44 42 4 38 36 34 32 3 28 26 24 22 2 18 16 14 12 1 8 6 4 2 3 Nodes = 2 Nodes = 8 Nodes = 14 6 9 12 15 18 21 24 27 3 33 36 Average inter-contact time (sec) 39 42 45 48 51 (a) Average dissemination delay (sec) 54 52 5 48 46 44 42 4 38 36 34 32 3 28 26 24 22 2 18 16 14 12 1 8 6 4 2 Nodes = 2 Nodes = 8 Nodes = 14 3 6 9 12 15 18 21 24 27 3 Average inter-contact time (sec) 33 36 39 42 (b) Average dissemination delay (sec) 3 28 26 24 22 2 18 16 14 12 1 8 6 4 2 Nodes = 2 Nodes = 8 Nodes = 14 3 6 9 12 15 18 Average inter-contact time (sec) 21 24 27 (c)

Considering driver behaviour in vehicular delay tolerant networks 169 Figure 11 CDF of data dissemination delay using different destination selection models and MAC protocols, (a) uniform distribution (b) exponential distribution (c) distribution (see online version for colours) CDF 1..9.8.7.6.5.4.3.2.1. Nodes 2 w/ backoff Nodes 8 w/ backoff Nodes 14 w/ backoff Nodes 2 w/o backoff Nodes 8 w/o backoff Nodes 14 w/o backoff 1 1 1 1 Dissemination delay (sec) (a) CDF 1..9.8.7.6.5.4.3.2.1. Nodes 2 w/ backoff Nodes 8 w/ backoff Nodes 14 w/ backoff Nodes 2 w/o backoff Nodes 8 w/o backoff Nodes 14 w/o backoff 1 1 1 1 Dissemination delay (sec) (b) CDF 1..9.8.7.6.5.4.3.2.1. Nodes 2 w/ backoff Nodes 8 w/ backoff Nodes 14 w/ backoff Nodes 2 w/o backoff Nodes 8 w/o backoff Nodes 14 w/o backoff 1 1 1 1 Dissemination delay (sec) (c)

17 K-c. Lan et al. In order to understand the distributions of destination selection in the real world, we observe nodes visited locations in a large taxi trace (Lai et al., 29). For simplicity, the map is divided into many blocks and each block has a size of 1 m 1 m. The number of visited times is counted when a node enters a new block. As shown in Figure 12, the results show that most of the blocks are visited only for a few times. On the other hand, a small percentages of blocks (most are in the downtown area) are visited by most of the taxies. We find that we can fit the distribution of block visited times into a distribution. Figure 12 The distribution of the block visited times observes in Shanghai taxi trace data, (a) PDF of block visited times (b) QQ plot for fitting into a distribution (see online version for colours).4.35.3.25 PDF.2.15.1.5. 1 1 1 1 1 The number of block visited times (a) (b) 6 Conclusions and future works Performing a realistic VANET simulation is challenging, since many factors could affect the node mobility in real life situations. In this paper, we discuss the effect of destination selection on the network topology and application performance. We model destination

Considering driver behaviour in vehicular delay tolerant networks 171 selection with uniform, exponential, and distributions and analyse the effect of destination selection on the cluster size. Furthermore, we find that our simulation results are not significantly affected by different node density settings when cars pick their destination following a distribution. Finally, we show that a fixed backoff MAC might perform better than a dynamic backoff MAC when a large number of nodes are clustered in the same area (e.g., traffic jam). In summary, selecting an appropriate level of details in the mobility model for a VANET simulation is important. As for our future work, we plan to look at how other driving behaviours, such as lane changing, car following and intersection behaviour, affect the results of vehicular simulations. In addition, we plan to use real-world vehicle traces to derive a trace-based destination selection model. Acknowledgements This work was supported by the National Science Council, Taiwan under Grant NSC 1-2219-E-6-5 and by the Ministry of Education, Taiwan under the NCKU Project of Promoting Academic Excellence and Developing World Class Research Centres. The data of Shanghai trace was obtained from Wireless and Sensor Networks Lab (WnSN), Shanghai Jiao Tong University. References Abdel-aty, M.A., Vaughn, K.M., Kitamura, R., Jovanis, P.P. and Mannering, F.L. (1994) Models of commuters information use and route choice: initial results based on southern California commuter route choice survey, Transportation Research Record, Vol. 1453, pp.46 55. Bai, F., Sadagopan, N. and Helmy, A. (23) The important framework for analyzing the impact of mobility on performance of routing for ad hoc networks, AdHoc Networks Journal, Vol. 1, No. 4, pp.383 43, Elsevier Science. Baumann, R., Heimlicher, S. and May, M. (27) Towards realistic mobility models for vehicular ad-hoc networks, Proc. Mobile Networking for Vehicular Environments, pp.73 78. Bettstetter, C. (21) Smooth is better than sharp: a random mobility model for simulation of wireless networks, Proc. ACM MSWiM. Boudec, J-Y.L. and Vojnovic, M. (25) Perfect simulation and stationarity of a class of mobility models, Proc. IEEE INFOCOM. Breslau, L., Estrin, D., Fall, K., Floyd, S., Heidemann, J., Helmy, A., Huang, P., McCanne, S., Varadhan, K., Xu, Y. and Yu, H. (2) Advances in network simulation, IEEE Computer, Vol. 33, No. 5, pp.59 69. Burgess, J., Gallagher, B., Jensen, D. and Levine, B.N. (26) Maxprop: routing for vehicle-based disruption-tolerant networks, Proc. IEEE INFOCOM. Cameron, G.D.B. and Duncan, G.I.D. (1996) Paramics: parallel microscopic simulation of road traffic, Journal of Supercomputing, Vol. 1, No. 1, pp.25 53. Chang, X. (1999) Network simulations with OPNET, Proc. WSC 1999, Vol. 1, pp.37 314. Chen, P-T. and Mahmassani, H.S. (1993) Dynamic interactive simulator for studying commuter behavior under real-time traffic information supply strategies, Transportation Research Record, Vol. 1413, pp.12 21. Choffnes, D.R. and Bustamante, F.E. (25) An integrated mobility and traffic model for vehicular wireless networks, Proc. ACM VANET, Cologne, Germany.

172 K-c. Lan et al. Costa, P., Frey, D., Migliavacca, M. and Mottola, L. (26) Towards lightweight information dissemination in inter-vehicular networks, Proc. ACM VANET, pp.2 29, New York, USA. Das, S.R., Perkins, C.E. and Royer, E.M. (2) Performance comparison of two on-demand routing protocols for ad hoc networks, Proc. IEEE INFOCOM. David, B. and David, A. (1996) Dynamic source routing in ad hoc wireless networks, Proc. Mobile Computing, Vol. 35, pp.153 181, Kluwer Academic. Dia, H. and Panwai, S. (27) Modelling drivers compliance and route choice behaviour in response to travel information, Nonlinear Dynamics, Vol. 49, No. 4, pp.493 59. Dingus, T., Hulse, M., Jahns, S., Alves-Foss, J., Confer, S., Rice, A., Roberts, I., Hanowski, R. and Sorenson, D. (1996) Development of human factors guidelines for advanced traveler information systems and commercial vehicle operations: literature review, Technical Report, FHWA-RD-95-153, Federal Highway Administration. Dressler, F. and Sommer, C. (21) On the impact of human driver behavior on intelligent transportation systems, Proc. IEEE VTC-Spring. Fellendorf, M. (1994) VISSIM: a microscopic simulation tool to evaluate actuated signal control including bus priority, Proc. ITE Annual Meeting. Fiore, M. and Härri, J. (28) The networking shape of vehicular mobility, Proc. ACM MobiHoc, pp.261 272. Guo, D., Li, X., Liu, M. and Zhang, L. (21) Model of traffic path choice based on game theory and induction mechanism, Proc. DCABES, pp.495 498. Halati, A., Lieu, H. and Walker, S. (1997) Corsim-corridor traffic simulation model, Proc. the Traffic Congestion and Traffic Safety in the 21st Century Conference, pp.57 576. Heidemann, J., Bulusu, N., Elson, J., Intanagonwiwat, C., Lan, K.C., Xu, Y., Ye, W., Estrin, D. and Govindan, R. (21) Effects of detail in wireless network simulation, Proc. CNDS. Holland, G. and Vaidya, N.H. (1999) Analysis of TCP performance over mobile ad hoc networks, Proc. ACM MobiCom. Hong, X., Gerla, M., Pei, G. and Chiang, C-C. (1999) A group mobility model for ad hoc wireless networks, Proc. ACM MSWiM. Huang, H., Zhu, Y., Li, X., Li, M. and Wu, M.Y. (21) Meta: a mobility model of metropolitan taxis extracted from GPS traces, IEEE WCNC. Jan, O., Horowitz, A.J. and Peng, Z-R. (2) Using global positioning system data to understand variations in path choice, Transportation Research Record, Vol. 1725, pp.37 44. Jerbi, M., Senouci, S.M., Doudane, Y.G. and Beylot, A-L. (28) Geo-localized virtual infrastructure for urban vehicular networks, Proc. IEEE ITST, pp.35 31. Karagiannis, T., Boudec, J-Y.L. and Vojnović, M. (27) Power law and exponential decay of inter contact times between mobile devices, Proc. ACM MobiCom, pp.183 194. Karnadi, F.K., Mo, Z.H. and Lan, K.C. (27) Rapid generation of realistic mobility models for VANET, Proc. IEEE WCNC, pp.256 2511. Karp, B. and Kung, H.T. (2) GPSR: greedy perimeter stateless routing for wireless networks, Proc. ACM MobiCom, New York, NY, USA. Krajzewicz, D., Hertkorn, G., Rőssel, C. and Wanger, P. (22) Sumo (simulation of urban mobility); an open-source traffic simulation, Proc. MESM, pp.183 187. Lai, P., Wang, X., Lu, N. and Liu, F. (29) A reliable broadcast routing scheme based on mobility prediction for VANET, Proc. IEEE IV, pp.183 187. Leon, S.J. (26) Linear Algebra with Applications, Pearson, New Jersey. Li, H., Guensler, R. and Ogle, J. (25) An analysis of morning commute route choice patterns using GPS based vehicle activity data, Proc. the 84th Annual Meeting of the Transportation Research Board.

Considering driver behaviour in vehicular delay tolerant networks 173 Lindgren, A., Doria, A. and Schelén, O. (23) Probabilistic routing in intermittently connected networks, ACM SIGMOBILE Mobile Computing and Communications Review, Vol. 7, No. 3, pp.19 2. Liu, T-L. and Huang, H-J. (27) Multi-agent simulation on day-to-day route choice behavior, Proc. ICNC, Vol. 5, pp.492 498. Mahajan, A., Potnis, N., Gopalan, K. and Wang, A-I.A. (26) Urban mobility models for VANETs, Proc. WoNGeN. Marx, R.W. (1986) The tiger system: automating the geographic structure of the United States Census, Government Publications Review, Vol. 13, No. 2, pp.181 21. Mobility Model Generator for Vehicular Networks (MOVE) (28 211) Available at http://www. lens.csie.ncku.edu.tw/move/ (accessed on December 211). Perkins, C.E. and Royer, E.M. (1999) Ad-hoc on-demand distance vector routing, Proc. IEEE Workshop on Mobile Computing Systems and Applications. QualNet network simulator (Qualnet) (28 211) Available at http://www.scalable-networks. com/ (accessed on January 211). Raney, B., Cetin, N., Vollmy, A., Vrtic, M., Axhausen, K. and Nagel, K. (22) Towards a microscopic traffic simulation of all of Switzerland, Proc. ICCS, pp.371 38. Saha, A.K. and Johnson, D.B. (24) Modeling mobility for vehicular ad hoc networks, Proc. ACM VANET. Shenpei, Z. and Xinping, Y. (28) Drivers route choice model based on traffic signal control, Proc. IEEE ICIEA, pp.2331 2334. Spyropoulos, T., Psounis, K. and Raghavendra, C.S. (25) Spray and wait: efficient routing in intermittently connected mobile networks, Proc. ACMSIGCOMM Workshop on Delay Tolerant Networking (WDTN). Stepanov, I., Marron, P-J. and Rothermel, K. (25) Mobility modeling of outdoor scenarios for MANETS, Proc. the 38th Annual Symposium on Simulation, pp.312 322. Sven Jaap, M.B. and Wolf, L. (25) Evaluation of routing protocols for vehicular ad hoc networks in city traffic scenarios, Proc. ITST. Tawfik, A.M., Rakha, H.A. and Miller, S.D. (21) An experimental exploration of route choice: identifying drivers choices and choice patterns, and capturing network evolution, Proc. IEEE ICST, pp.15 112. Treiber, M., Hennecke, A. and Helbing, D. (2) Congested traffic states in empirical observations and microscopic simulations, Physical Review, Vol. 62, No. 2, pp.185 1824. Vahdat, A. and Becker, D. (2) Epidemic routing for partially-connected ad hoc networks, Technical Report, Duke University. Viriyasitavat, W., Tonguz, O. and Bai, F. (29) Network connectivity of VANETs in urban areas, Proc. IEEE SECON, pp.1 9. Wegener, A., Piórkowski, M., Raya, M., Hellbrűck, H., Fischer, S. and Hubaux, J-P. (28) TraCI: an interface for coupling road traffic and network simulators, Proc. CNS, pp.155 163. Yoon, J., Liu, M. and Noble, B. (23) Random waypoint considered harmful, Proc. IEEE INFOCOM. Zhang, X., Kurose, J., Levine, B., Towsley, D. and Zhang, H. (27) Study of a bus-based disruption-tolerant network: mobility modeling and impact of routing, Proc. ACM MobiCom, pp.195 26. Zhao, D. and Shao, C. (21) Empirical study of drivers learning behavior and reliance on VMS, Proc. IEEE ITSC, pp.1614 1619. Zhao, J. and Li, Q. (21) A method for modeling drivers behavior rules in agent-based traffic simulation, Proc. the 18th International Conference on Geoinformatics.