Analysis of Lane Level Dynamics for Emergency Vehicles

Transcription

1 Analysis of Lane Level Dynamics for Emergency Vehicles Thesis submitted in partial fulfillment of the requirements for the degree of MASTERS OF SCIENCE BY RESEARCH in COMPUTER SCIENCE by Akash Agarwal International Institute of Information Technology Hyderabad , INDIA July 2017

2 Copyright c AKASH AGARWAL, 2017 All Rights Reserved

3 Scanned by CamScanner

4 To my parents

5 Acknowledgments I would like to sincerely thank my adviser Dr. Praveen Paruchuri, who has been a great mentor to me. It would not have been possible to produce this work without his guidance. His advice on research as well as academics in general have been very helpful for me and will remain with me throughout my life. Dr. Praveen s passion and enthusiasm for excellence has always motivated me to go a step further. I would liked to thank all my lab mates, especially Rashi and Gaurav, for encouraging me throughout my research career at IIIT and helping me with their valuable suggestions. I would also like to thank Prof. Kamal Karlapalem for his valuable contributions and support during the early stages of this work. Finally, I would like to express my gratitude to all my friends and family members. v

6 Abstract Slow moving traffic in heavily populated cities, can many times result in loss of lives due to emergency vehicles not being able to reach their destination hospitals on time. Recent advances in the field of Intelligent Transportation Systems (ITS) makes it increasingly likely that vehicles in the near future will be equipped with advanced systems that allow inter vehicular communication. In this thesis, we assume the usage of such a system to optimize the lane level dynamics for an emergency vehicle (EV), traversing a multi lane stretch of road under a variety of traffic settings. In particular, we present the Fixed Lane Strategy (FLS) and the Best Lane Strategy (BLS) for EV traversal and perform an extensive agent based analysis to study their strengths and weaknesses. Through a series of experiments performed using the well-known traffic simulator SUMO, we could show that: (a) BLS performs better than SUMO strategy on all traffic settings we tested. (b) BLS performs better than FLS in settings that capture real-world traffic conditions involving congestion and uncertainties while FLS performs better in well-behaved conditions and (c) BLS was found to be the best strategy for the setting calibrated using real world data (obtained from NYCDOT). vi

7 Contents Chapter Page 1 Introduction Urban Traffic Emergency Vehicles Vehicle to vehicle (V2V) Communication and VANETS (Vehicular Ad-hoc Networks) Traffic Simulators System Overview Thesis contribution Thesis organization Related Work Improving the efficiency of EV operations Efficient EV-Base station assignment/dispatch Optimal routing problem Lane-level dynamics Control of traffic lights Improving lane level dynamics Lane change model types Strategic lane changing Tactical lane changing Rule/Discrete choice based models Incentive/Utility based models The Strategies The SUMO strategy Assumptions The Fixed Lane Strategy (FLS) The Best Lane Strategy (BLS) BLS Utility Computation Algorithm for BLS FLS as a subset of BLS Experimental Setup Modeling Communication Vehicular parameters Modeling Human behavior vii

8 viii CONTENTS Preferred maximum speed of drivers Speed deviation Other parameters Simulation traffic distribution Environmental Settings Setting Setting Setting Setting Setting Deciding on weights in BLS Mapping NYCDOT data to Setting Modeling groups Modeling buckets Obtaining weights for buckets The SUMO simulation Summary of SUMO parameters used EV run time per bucket Computing mean run time: Other Details Experimental Analysis Comparison between strategies Standard Deviation across different runs Varying the communication distance Analysis of utility computation Effect of varying t, re-computation interval Effect of vehicles not following lane change requests Effect of Communication Delays Effect of deviating from lane speed limits Effect of number of lanes Effect of EV on other vehicles Conclusions and Future Work Conclusion Future Work Related Publications Bibliography

9 List of Figures Figure Page 1.1 System Overview Comparison of strategies Standard Deviation across different runs Variation of Communication Distance, Setting Variation of Communication Distance, Setting Variation of Communication Distance, Setting Variation of Communication Distance, Setting Variation of Communication Distance, Setting Average utility v/s EV time taken, Setting Average utility v/s EV time taken, Setting Average utility v/s EV time taken, Setting Varying recomputation interval, t in BLS - Setting Varying recomputation interval, t in BLS - Setting Varying recomputation interval, t in BLS - Setting Effect of vehicles not following lane change requests, Setting Effect of vehicles not following lane change requests, Setting Effect of Communication Delays, Setting Effect of Communication Delays, Setting Effect of deviating from lane speed limits, Setting Effect of deviating from lane speed limits, Setting Effect of number of lanes, Setting Effect of number of lanes, Setting ix

10 List of Tables Table Page 4.1 Experimental parameters x

11 Chapter 1 Introduction Slow moving traffic in heavily populated cities, can very often result in loss of lives due to emergency vehicles not being able to reach their destination hospitals on time. We assume the usage of inter vehicular communication to optimize the lane level dynamics for an Emergency Vehicle (EV), traversing a multi lane stretch of road. In particular, we present the Fixed Lane Strategy (FLS) and the Best Lane Strategy (BLS) for EV traversal. FLS is a simple strategy that acts as a good baseline while BLS is a sophisticated strategy that can adapt to varying traffic patterns. 1.1 Urban Traffic Roads in heavily populated cities like Hong Kong, Manila, Mumbai, Dhaka and Seoul suffer from extremely slow and dense traffic. The number of cars in such cities are expanding at a much faster pace than the development of road infrastructure. This will only make the situation worsen in the future. [40] presents details statistics on the traffic situation at Hong Kong. According to the report, the growth rate of the total length of public roads in Hong Kong is expected to slow down by a factor of 0.4% p.a. up to Such growth rate clearly cannot keep up with the current growth of vehicle fleet (about 3.4% p.a.). The report adds that building more road transport infrastructure alone cannot resolve traffic congestion, it may actually induce more demand for vehicle usage and fuel vehicular growth. The number of total licensed vehicles grew by about 30% from about 524, 000 in 2003 to about 681, 000 in 2013, with an annual growth rate of 3.4% in recent years. The larger the vehicle fleet size, the slower the car journey speed in the urban areas. Furthermore, even in cities with relatively faster traffic like New York, London and Singapore, there is a wide variation in traffic speeds, which causes the traffic to be quite slow sometimes. Such slow moving traffic can result in loss of lives due to EVs not being able to reach their destination hospitals on time. In this work, we specifically focus on scenarios where the roads are crowded with slow moving traffic but our strategies are general enough to be applied to any type of road settings including free-ways and highways. 1

12 1.2 Emergency Vehicles Improving the travel time of EVs can potentially save many lives [39, 8, 27]. For example, [39] presents statistics related to thousands of people being affected by EV delays in UK along with detailed location wise statistics. Medical guidance, according to this article, says that immediately lifethreatening (Category A) calls should receive a response within 8 minutes for atleast three quarters of the cases. This target was set because the chance of surviving a heart attack reduces by 10% with every minute that passes. [27] estimates that meeting this guideline could possibly save an additional 3000 heart attack victims plus many other Category A patients in UK. The report adds that only three of England s 32 ambulance services reach a large majority of immediately life-threatening call-outs within eight minutes, according to the latest statistics. [8] presents the following EV response times breakdown for Wales (part of UK): 42.6% of Category A calls received an emergency response within 8 minutes, 47.4% within 9 minutes, 52.0% within 10 minutes, 68.9% within 15 minutes, 79.1% within 20 minutes and 89.8% within 30 minutes. Hence, only 48.5% of the urgent Category A calls received a response within the specified eight minutes while the specified target was 65%. The key takeaway point from all this information is that, we are looking for savings in EV traversal time of the order of few minutes or even seconds, that can result in a life saving difference. 1.3 Vehicle to vehicle (V2V) Communication and VANETS (Vehicular Ad-hoc Networks) Recent advances in the field of Intelligent Transportation Systems (ITS) make it increasingly likely that vehicles in the near future will be equipped with advanced systems that allow inter vehicular communication. According to [30], vehicles will be carrying computing and communication platforms and will have enhanced sensing capabilities. This will enable new versatile systems that enhance transportation safety and efficiency and will provide infotainment. [30] surveys the state-of-the art approaches, solutions, and technologies across a broad range of projects for vehicular communication systems. Vehicles, in the near future, will be equipped with novel computing, communication, and sensing capabilities, and user interfaces. Vehicles are already equipped with multiple processors and microcontrollers dedicated to tasks such as fuel injection, braking, transmission, and battery charging. The vehicular communication and related computing platform(s) will be functionally independent and responsible for running the vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication protocols and the supported applications. The development of such V2V systems and related technologies has been the subject of numerous projects around the globe, as well as for standardization working groups and industrial consortia (e.g., [21], [10]). Vehicle to vehicle (V2V) Communication is being made possible using VANETS (Vehicular Ad-hoc Networks) [46], a key component of ITS. Recent advances in hardware, software, and communication 2

13 technologies are enabling the design and implementation of a whole range of different types of networks that are being deployed in various environments. VANETS are one such networks which have become an active area of research as it has tremendous potential to improve vehicle and road safety, traffic efficiency, and convenience as well as comfort to both drivers and passengers. Recent research efforts have placed a strong emphasis on novel VANET design architectures and implementations. VANETS are envisioned to be based on a variant of the widely known Wi-Fi technology, the IEEE p protocol ([14], [18]). The trial standards are called Wireless Access in Vehicular Environment, WAVE. [14] studies the performance of the IEEE 1609 WAVE and IEEE p trial standards for vehicular communications. The authors implement key components of these standards in a simulation environment also supporting realistic vehicular mobility simulation. The overall capacity of vehicular networks utilizing the said standards, as well as delay performance, which is an extremely important performance metric especially for safety applications, are studied. The results show that even in the presence of multi-channel operation implemented by the IEEE the delay of control messages of highest priority remains on the order of tens of milliseconds. [18] provides an overview of the draft proposed for IEEE p and provides an insight into the reasoning and approaches behind it. Experiments have been performed demonstrating the effectiveness of the WAVE protocol for upto 1000 meters. The IEEE p protocol specifies physical and medium access control protocols designed with the highly volatile vehicular environment in mind. The operation across multiple channels is specified in IEEE There are some concerns around the latency of communication and association with Wi-Fi access points. Customized implementations of based systems allow for very fast association with a Wi-Fi access point, while fast handover techniques allow cellular systems to service nodes moving fast from one base station to another. For transportation safety applications, low-latency communication with neighboring devices is critical. Vehicle to vehicle (V2V) communication will allow for several innovative methods of traffic management. In this thesis, we will study one application of this technology that can improve the traversal time of Emergency Vehicles (EV) [28]. Note that we only assume usage of inter-vehicular communication system (V2V) but not the usage of any other road side infrastructure [38]. The following is a summary of the salient features of using V2V communication: V2V communication can operate beyond the line-of-sight constraints. Vehicles have novel computing, communication, and sensing capabilities, and user interfaces. Accurate information like the speeds and positions of vehicles can be obtained. The p WAVE protocol allows low latency communication up to 1000 meters. 3

14 1.4 Traffic Simulators Simulators have long been useful aids where the understanding of phenomena, that can be simulated, are quite difficult. Simulators help to view the same phenomena at different levels of abstraction and hence aid in easy understanding for various users who have differing knowledge of the phenomena under consideration. Traffic is one such domain where simulations have been found to be very helpful and have a rich history [7, 2]. Traffic simulations facilitate the evaluation of infrastructure and policy changes before implementing them on roads. For example, the effectiveness of dynamic traffic management systems [44] or autonomous intersection and traffic light control mechanisms [9, 4], can be tested and optimized in a simulation before being deployed in the real world. At a broader level, traffic simulators can be classified into macro (involves modeling the general aspects of system like the average speed of vehicles on the road, vehicle density - [16], [32]) and micro (involves modeling each vehicle at an individual level) level simulators ([34], [31], [3]). In this thesis, we perform our analysis using a free and open microscopic traffic simulation suite named Simulation of Urban MObility (SUMO) ([25], [5], [26], [24]), to identify the best possible traversal strategy for an EV in a variety of settings. SUMO allows modeling of traffic systems and has a wealth of supporting tools which can handle tasks such as route finding, visualization and network import. We will use its rich feature set to simulate a variety of realistic traffic scenarios, then introduce the EV and allow its behavior to play out in the simulation. As stated earlier, we assume the presence of a V2V communication infrastructure which interacts with the SUMO simulation to provide V2V related inputs. Our assumption regarding availability of V2V communication infrastructure and its interaction with SUMO ensures that the EV can interact with vehicles upto a particular communication distance. The EV can also find about the speeds and positions of other vehicles, and send them lane change requests. To summarize, this interaction process allows the following: Enables EV to sense the presence of other vehicles. Allows information requests to be sent for obtaining the position and speed details of individual vehicles. Send lane change requests to individual vehicles. Enable other vehicles to provide the requested information or initiate action on lane change requests. This interaction process is enabled by running a communication simulation. The above stated agent based interaction process is decoupled from SUMO in our simulations. We run a communication simulation based on the above stated interaction process which runs in parallel to SUMO and interacts with SUMO using its TraCI interface [42]. For example, upon receiving a lane change request, the vehicle requested only initiates the lane change action while the dynamics of lane change are handled by SUMO. 4

15 By design, our algorithms are abstracted from all the low level operations like vehicle acceleration and deceleration, to focus specifically on high level planning part, i.e., strategies to intelligently pick lanes for the EV to move to during the journey. 1.5 System Overview Figure 1.1 System Overview Figure 1.1 represents the high level system overview. This work consists of following three major components which are designed to be independent of each other: 1. The Emergency Vehicle (EV) module: The EV module is responsible for the computation of the lane change strategies based on the details about the speed and position of other vehicles. Its Physical Module is responsible for initiating the lane changes once a lane change has been decided. The EV module is also responsible for deciding on the lane change requests to be send to other vehicles. The strategies and the related experiments are described in Chapter 3 and Chapter 5 respectively. 2. The V2V communication module: It simulates the V2V communication and the underlying VANET. The EV gets the information about the speed and position of other vehicles (from SUMO), using this intermediate module. The details of the V2V communication module are a part of Chapter 3, The Strategies and Chapter 4, Experimental Setup. 5

16 3. The microscopic traffic simulation, SUMO: SUMO takes care of all low level operations like the acceleration and deceleration of individual vehicles. The actual lane changes of the EV are also performed by SUMO. This decoupled architecture allows the modules to be used in a real environment or in other simulators. Figure 1.1 describes the interactions that take place: Communication to get the speed and position of other vehicles: In a simulation the communication module of the EV gets the details about the speed and position of other vehicle using the TraCI interface which in turn communicates with SUMO. In a real environment the communication module would get it directly from the VANET. Communication of lane change requests to other vehicles: In a simulation the communication module of the EV sends a lane change request to other vehicles using the TraCI interface which in turn communicates with SUMO. In a real environment the communication module would send a lane change request directly to a vehicle using the VANET. Lane change of the EV: In a simulation the Physical Module of the EV communicates a lane change to the TraCI interface which passes it on to SUMO where the actual lane change happens. In a real environment it would directly initiate a lane change. 1.6 Thesis contribution The contributions of this thesis are summarized below: This thesis introduces Lane Level Dynamics for an EV (Emergency Vehicle) and presents two strategies, BLS (Best Lane Strategy) and FLS (Fixed Lane Strategy), for deciding on the lane to travel on while traversing a stretch of road and for clearance of the lane. A realistic simulation of urban roads is presented. We adopt a novel approach to model congestion and uncertainty in traffic for the purpose of testing our strategies. The environmental setting presented in this work can be potentially used for other applications as well. Usage of real time speed data to calibrate traffic in a simulation is demonstrated by capturing live data from New York City traffic. A V2V communication model is simulated. Various nitty-gritty details of communication such as the communication distance and communication delays are analyzed. We present a detailed experimental analysis to demonstrate various realistic behaviors corresponding to lane level dynamics. A variety of human and environmental factors like deviating from speed limits and non adherence to EV requests are presented. This can also be potentially used for other work. 6

17 1.7 Thesis organization The rest of the thesis is organized as follows: Related Work: Chapter 2 presents the related work in this domain. It includes the several approaches used to improve the efficiency of EV operations as well other approaches used to improve lane level dynamics. The Strategies: Chapter 3 presents the description of the SUMO Strategy, the Fixed Lane Strategy (FLS) and the Best Lane Strategy (BLS) algorithms. Experimental Setup: Chapter 4 describes the various aspects related to running our experiments such as Modeling Communication, Vehicular Parameters, Modeling Human behavior, Environmental Settings, etc. Experimental Analysis: Chapter 5 illustrates the results of the experiments conducted in detail. Conclusions and Future Work: Chapter 6 presents the conclusion and some insights on future work. 7

18 Chapter 2 Related Work 2.1 Improving the efficiency of EV operations Our work deals with the subject of improving the efficiency of EV operations. We identified the following different threads of work related to it: Efficient EV-Base station assignment/dispatch The EV-Base station assignment/dispatch problem deals with finding efficient ways for dynamic allocation of EVs to appropriate base stations and redeployment of the fleet so that the EV response time gets minimized. [36] attempts to reduce the response time,i.e., time taken to arrive at the incident location after receiving the emergency call of EVs for as many requests as possible. The authors try to minimize the largest value of response time such that the risk of finding requests that have a higher value is bounded (e.g., Only 10% of requests should have a response time greater than 8 minutes). This is based on a mixed integer linear optimization formulation to learn and compute an allocation from a set of input requests while considering the risk criterion. [45] present an efficient approach to ambulance fleet allocation and dynamic redeployment, where the goal is to position an entire fleet of ambulances to base locations to maximize the service level (or utility) of the Emergency Medical Services (EMS) system. A simulation-based approach is used, where the utility of an allocation is measured by directly simulating emergency requests Optimal routing problem The optimal routing problem deals with finding the appropriate route from base station to the location of the EV, and to other destinations (e.g., hospitals) thereafter. Here the best path to be taken by the EV is identified, but the actual traversal details, once a path is identified, are not described (like selection/changing of lanes). For example, [41] describes a learning based routing system designed to ease the movement of emergency vehicles through a network of congested streets. Information from GPS equipment installed aboard of every emergency vehicle is used. The actual routing algorithm is part of 8

19 the A class and decisions are made with the help of a neural network that estimates the expected time of arrival of every feasible route the emergency vehicles might follow. Real-time traffic data is used to train the neural network and to help the routing algorithm work faster. [12] presents a dynamic path planning algorithm to increase the efficiency of routing an emergency vehicle. The authors implement a graph version of the D Lite informed search algorithm to efficiently and dynamically plan optimal paths for the emergency vehicle while taking into consideration the real-time updates of congestion levels and other delays for computation of travel time. [43] proposes a modified Dijkstra algorithm for optimal routing of vehicles while considering the effective lane-changing time that is a parametric function of the prevailing travel speed and traffic density. However only strategic lane changes (Section ) are considered, while we deal with tactical lane changes (Section ) Lane-level dynamics Lane-level dynamics deals with determining the actual traversal details once an EV enters a stretch of road on its route to destination, including computing the best lane to travel on and the consideration of lane changes. There is sparse work on this field, especially with respect to emergency vehicles. Section 2.2 presents more details Control of traffic lights During emergencies, traffic lights may be controlled to reduce wait times at intersections for a faster journey and enable the EV reach its destination faster. [19] aims to use signal preemption to allow the EV to proceed through signalized intersections as quickly as possible. This is achieved by preempting traffic lights using vehicle to infrastructure communication. The proposed method uses kinematic wave theory (i.e., shock wave theory) to determine when each intersection should be preempted, and the appropriate preemption message is send using vehicle to infrastructure communication. [6] presents a VANETS (Vehicular Ad-hoc Networks) based warning system using radio communication to warn other vehicles and notify traffic lights. 2.2 Improving lane level dynamics We focus specifically on point in this thesis i.e., computing the lane level dynamics for an EV. We envision that the low level traversal strategies demonstrated in this thesis will work in conjunction with planning of EV-Base station assignment/dispatch, high level planning of routes and generation of a green wave for the EV by preempting traffic lights. Most of the previous work on this topic has focused on 2.1.1,2.1.2 and These works do not elaborate on the actual traversal strategies to use, once a particular stretch of road is identified. 9

20 There are several lane changing models that are used by vehicles in microscopic traffic simulation suites like SUMO. [33] presents a detailed review of such models. These models perform very well for vehicles in general, but for an EV we are able to obtain significant traversal time improvements by using BLS in comparison to previous models. This is because EVs can use specialized routing strategies to take advantage of the fact that other vehicles make way for EVs and also to account for the accurate information available about other vehicles such as speed, vehicle position etc., obtainable due to ITS assumptions Lane change model types Most lane changing models that we studied incorporate the following types of lane changes: Strategic lane changing Lane changes due to route requirements of the vehicle under consideration [33, 11]. For example, if a vehicle needs to turn left at an intersection, it needs to be on the leftmost lane when it reached the intersection. Thus, it will require to change its lane to the leftmost lane, if it is not already on it, before it reaches the intersection Tactical lane changing Lane changes to obtain faster speeds. This is mainly based on speed of the leader(s) (the vehicle(s) immediately in front of the vehicle under consideration). For example, models described in [17, 13, 15, 1], which consider lanes with faster leader(s) as better. In this work, we focus only on (2), i.e., Tactical Lane changing. There is a rich history of work related to tactical lane changes. In the models described in prior work, lane change is considered whenever there is a faster leader(s) and appropriate gaps for lane change are available. However, they do not consider the overall advantage of being on a lane where the leader is not the fastest but maybe a better lane on average than a lane with a fast leader and much slower vehicles further ahead. For example, a particular lane may have a faster leader followed by many slow moving vehicles which would soon lead to slower speeds in that lane. Lane change models described in prior work can be broadly classified into the following categories: Rule/Discrete choice based models Decision to change lanes is based on a decision tree with a series for fixed conditions. For example, in [13] the lane changing process is represented as a decision tree with has some fixed conditions typically encountered on urban traffic, and the final output is a lane change/not change decision. [15] incorporates discretionary lane changes (DLC), i.e., lane changes for speed advantages, e.g., a driver may want to change from a lane which has slow moving vehicle by deciding to switch to the right or left lanes. The 10

21 lane change maneuvers are based on the availability of acceptable lead and lag gaps in the target lane. Other models like [35] use a cellular automata where the rules are based on number of empty sites (cells) ahead in the same lane, the forward gap on the other lane and the backward gap on the other lane. [23] describes a game theory based model Incentive/Utility based models [22, 37] introduce incentive/utility based models for lane-changing behavior which considers the incentive/desire to change lanes and the gaps/risks involved. But, the incentive/utility itself is based mainly on attributes like velocity of the leader(s), gap between the vehicle under consideration and each of the leader(s) and the possible acceleration/deceleration if lane change happens. [20] describes a lane changing model which uses information from surrounding vehicles and infrastructure to get the speed and position of vehicles. The model tries to anticipate beforehand, the possibility of a slow leader. But like in other models, it has no provisions to anticipate better lanes based on vehicles clearing out due to EV. In this thesis, we introduce the BLS algorithm which instead of deciding on lanes based on a faster leader(s) uses a sophisticated strategy that can adapt to varying traffic patterns. This is achieved by using a utility function that includes average speeds, slowest speeds and normalized free space, and considers both the possibility of an immediate faster lane and the clearing time of other vehicles. We also introduce the FLS algorithm which acts as a good baseline. This thesis includes extensive experimentation on a number of settings developed to model real world traffic environments including one setting calibrated using real-world data obtained from NYCDOT, which hasn t been focused in prior works. We compare our strategies with the SUMO strategy, which includes a state of the art Tactical lane-changing model (lane-change maneuvers where a vehicle attempts to avoid following a slow leader) [11] and show that our strategies perform much better as compared to the SUMO model. 11

22 Chapter 3 The Strategies We present here two strategies for the EV to handle lane level dynamics while traversing on a multilane stretch of road namely FLS (Fixed Lane Strategy) and BLS (Best Lane Strategy). Our strategies were developed, specifically for the EV, to take advantage of the higher priority an EV has and the accurate information available about other vehicles due to the presence of inter-vehicular communication. Hence, these strategies only specify the lane that the EV should travel in while the low-level dynamics of traversal namely speed, acceleration/deceleration and the dynamics involved in changing the lane for the EV are handled by SUMO. Another task these strategies perform is to identify the appropriate vehicles to send lane change requests. In a similar vein, the traffic behavior for all the other vehicles is entirely generated by SUMO depending on the initial parameters which are picked using five different settings described in Section 4.6 (one of these is calibrated on data from NYCDOT). Once the initial parameters are picked, we do not make any modifications to the general traffic behavior except for the two specific interventions presented in Section 4.9 to incorporate lane change requests and lane modeling related parameters which SUMO does not provide otherwise. The key idea here is to let traffic patterns evolve using the underlying models SUMO has with minimal intervention and focus on lane picking strategies. This is also the same reason for why we let SUMO control the low level dynamics of EV. We then study the strengths and weaknesses of the lane picking behavior generated by these two strategies under a variety of experimental settings and benchmark it against SUMO strategy with suitable modifications described below. 3.1 The SUMO strategy SUMO provides an emergency vehicle class which, combined with the default simulator strategy to control the dynamics of vehicular traversal, allows the modeling of EV traversal behavior. SUMO is a well proven simulator that simulates realistic traffic patterns. Thus, the EV run times obtained from the default simulator strategy are reasonably close to the actual time taken by an EV in real traffic. To bring in further realism into our simulation we also added communication to the default simulator 12

23 strategy, i.e., vehicles on the current lane of the EV (using the default simulator strategy) are sent lane change requests. This is to model the effect of sirens and lights with the assumption that other vehicles clear from the path of the EV upon hearing siren (or seeing lights). Similar communication facility is assumed for FLS and BLS strategies (not to model the effect of sirens/lights but) due to availability of V2V communication. Note that we assume, (except in the experiment 5.5) that all vehicles within the communication distance (default value of 100 meters), always get the lane change request and always clear the lane upon obtaining it. In present day real traffic scenarios, sirens may not always be heard or drivers may not always know the lane to clear until the EV is in sight. Hence, the SUMO strategy we use for bench-marking is expected to be favorable to the EV in terms of travel time. 3.2 Assumptions For our strategies(fls and BLS), we make the following assumptions: The EV can get the information about the position and speed of vehicles upto the communication distance, c d (100 meters is the default in our experiments). By design, FLS and BLS decide only on whether to make a lane change or not. The actual lane change maneuver involving acceleration/deceleration of the EV, creation of appropriate gaps,etc, is handled by the simulator. 3.3 The Fixed Lane Strategy (FLS) The idea behind FLS is that the EV should identify the lane, that is fastest on an average, based on prior information and pick that as the fixed lane for its entire journey. This would on an average ensure the minimum possible traversal time unless exceptional situations arise. Assuming a right handed driving system, in most cases leftmost lane is the fastest, as faster vehicles tend to move on the left lanes while the slower ones move on the right lanes (the vice-verse typically holds true for left handed driving systems). When using FLS the EV therefore moves to the leftmost lane from its current position and then tries to clear out the vehicles from that lane. As described earlier, we assume V2V communication is feasible in the ITS framework. The EV can therefore clear a lane, by sending lane change requests to other vehicles traveling on its lane within its communication distance, c d. It sends a lane change request after every fixed re-computation interval, t. We also analyze situations when clearance of other vehicles is not feasible. 3.4 The Best Lane Strategy (BLS) The idea behind BLS is that the EV should identify the best lane at the current time and move to that lane from its current position. In order to do this the EV must calculate the utilities of the current lane 13

24 and the other lanes using the utility computation function described below and then take a decision to switch if it is beneficial to do so. Computing this utility function would need the EV to make use of V2V communication and obtain the needed information. Similar to FLS the EV tries to clear out vehicles from the lane it currently is in (and like for FLS we analyze when such clearance is not feasible). Traffic being a dynamic domain once the EV moves onto the best lane, the traffic patterns on different lanes may change. BLS should therefore be able to switch to better lanes. In case the lane on which the EV is currently moving is no longer the fastest, it may be beneficial for the EV to change lanes depending on the trade-off between the advantage gained and the overhead of changing lanes BLS Utility Computation A key step in BLS is to compute the utility u l, of a lane l. We envision u l to be a function of the following factors: (a) Normalized speed of the slowest vehicle (since traffic on a lane eventually moves at speed of the slowest vehicle) (b) Normalized average speed (many times a vehicle (or vehicles) might be temporarily slow since it is just about to change a lane or near an intersection i.e., give some weight to average rather than decide entirely on temporary phenomenon) and (c) Normalized free space (since not all vehicles may be able to switch lanes immediately after a clear lane message is received). It is a function of number of vehicles present upto distance c d in front of the vehicle and the maximum number possible when the lane upto c d is full. Here, normalized lowest speed is calculated as (speed/maximum possible speed) and is denoted by a m. Normalized average speed, b m, is calculated as (average speed/maximum possible speed). Here, maximum possible speed, is the speed limit of the road (different lanes can have additional speed restrictions). Normalized free space is an approximation computed as n c, where c is the number of n vehicles present on the lane l upto distance c d, and n is the maximum number of vehicles that can be on the lane upto c d. (n-c) therefore represents the number of vehicles that can be added in the free space available upto c d. To compute n we assume an average length for vehicles (l v ) which makes the computation an approximation. Combining the terms: u l = w a a m + w b b m + w c n c (3.1) n where w a, w b, w c are the weights of each of the terms. At the beginning of the simulation an EV starts on the lane with maximum value of u l. Utilities are recomputed every t seconds and lane changes happen when the utility of the best lane u b, exceeds the utility of the current lane u c, by at least compensate for lane switching overheads: Summarizing the notation used so far: u b u c > ( Condition for lane change ) to 14

25 a - Lowest speed of a vehicle on lane l within the communication distance, c d b - Average speed of vehicles on lane l (within communication distance c d ) c - Number of vehicles on lane l (within distance c d ) w a, w b, w c are the weight-age of the factors a,b and c respectively where 0 <= w a,w b,w c <=1 and w a + w b + w c =1 m - Maximum speed limit of the road (Different lanes may have different speed limits) n - Maximum number of vehicles possible within distance c d, n = gap between vehicles and l v is length of a vehicle c d m v+l v, where m v is minimum If there are no vehicles on the lane up to c d, i.e., c =0, then, a = b = s l (the speed limit of the lane). u b, u c - Utility of moving on the best lane and the current lane respectively - Minimum utility difference for lane switch to happen Algorithm for BLS Compute initial utility u l for each lane Pick the lane c with maximum u l Set Current Lane = c for every t seconds do if EV is not on the road then BLS Ends; end Perform u l computation for each lane Set d = Lane with maximum u l if u d u c > then c = d; end if EV is not on lane c then Change Lane of EV to c; end if Lane clearing is allowed then for Vehicles in lane c upto c d in front do Send lane change requests; end end end Algorithm 1: Best Lane Strategy Algorithm 1 summarizes the steps of the BLS strategy. 15

26 Initially the EV computes utility of all the lanes using equation 3.1 and assigns current lane to the one with highest utility. Using appropriate values for w a, w b and w c (Section 4.7), we weigh the different factors suitably to identify the lane with the highest utility (and hence the fastest traffic). Every t steps, the EV re-computes the utilities. If the difference between the maximum utility, u b and the current lane utility, u c exceeds, the EV changes its lane to the one with utility u b. When the utility u b is just slightly better, it may not actually be beneficial to change lanes. A value of =0, means that a lane change will take place whenever there is a lane with better utility than the current lane. However, changing a lane has some cost in terms of changing the traffic pattern that was used to compute the utilities of lanes e.g., deceleration of EV/other vehicles when making a lane change and other disturbances to traffic. This cost incurred due to the changed traffic pattern need not always be worth the advantage gained from increased utility due to being on a faster lane, hence we use an overhead cost,. Additionally, if lane clearing is allowed, it sends lane change requests (every t seconds) to vehicles on the current lane, that are within its communication distance, c d. The best case scenario for the EV is to be in a lane that is empty till distance c d, i.e., c =0. In a this case, a and b both equal s l. If s l = m, m = b m = n c n =1, hence, u l = w a + w b + w c =1, the maximum utility possible for a lane. If s l 6= m, i.e., when different lanes have different speed limits, then a m = b m = s l m. Thus, different lanes will have different utilities depending on the values of s l. When a lane has some vehicles running on it, lane speed limits are already incorporated in the speed of the slowest moving vehicle and average speeds. 3.5 FLS as a subset of BLS While FLS may appear to be a strict subset of BLS it should be noted that, BLS uses only locally available information from V2V communications to compute the best lanes while FLS uses historical information to pick the fastest lane for the entire journey. Also, in some of the environmental settings described later, the simulation setup results in much higher average speeds in the leftmost lane, giving FLS a significant advantage by design over BLS. 16

27 Chapter 4 Experimental Setup We present here our experimental set-up for testing the EV strategies in a variety of environmental settings and parameters using the SUMO simulator. As mentioned earlier SUMO is a well-acclaimed microscopic traffic simulator with a rich parameter set to try out different settings. SUMO has been extensively used by the research community to answer a large variety of research questions such as evaluating performance of traffic lights, evaluating vehicle route choices, providing traffic forecasts to authorities during soccer world cup 2006 and many other applications [25]. Our simulations are run on an environment that is common on many roads of heavily populated cities like Hong Kong, Manila, Mumbai, Dhaka and Seoul. Road segments in such cities, have multiple lanes that are quite crowded, and are in general 1 km to 10 km in length, at a stretch, before hitting an intersection. The speeds of vehicles on such roads are relatively low, ranging from 30 km/h to 90 km/h. Thus, unless stated otherwise, for all the experiments shown in this thesis, we use a 2 km one way stretch of road with 4 lanes (except in Experiment 5.8, where we conclude varying the number of lanes between 2-6 does not cause significant difference in the results). We tested with several traffic patterns on the road and our results are reproducible in all reasonably long road segments (> 500 meters). 4.1 Modeling Communication We simulate a V2V single hop communication model for our experiments. In particular, the EV can get the position and speed of a vehicle in any lane up to a fixed distance, c d via V2V communication. It can also send lane change requests upto c d. We model communication delay i.e., the time taken by the EV s request to reach a vehicle, using the variable c j for the j th request. We fix c j s to 1 second for all the requests unless stated otherwise (except in Experiment 5.6, where we study the effect of communication delays), as most realistic delays are less than 1 second. 17

28 4.2 Vehicular parameters For each vehicle i, we have the parameters v i, a i and d i - The maximum speed, acceleration and deceleration of the vehicle respectively in addition to other parameters we introduce later. 4.3 Modeling Human behavior To make the strategies work with autonomous vehicles and human drivers, we model some parameters in SUMO. The values of these parameters are set depending on whether a vehicle is autonomous or human driven. These can be set accordingly for scenarios where there are only autonomous vehicles on the road or for a combination of both automated and manual cars: Preferred maximum speed of drivers The preferred maximum speed of drivers, for each vehicle i, is set as v i p <= v i (A driver can t go faster than the vehicle maximum speed, hence the upper bound). For an autonomous vehicle we set v i p = v i Speed deviation Speed deviation (s dev, speeddev in SUMO) models the deviation of vehicles from lane speed limits. The maximum allowed speed of vehicles, v i mas is generated using a normal distribution with a mean of s l (the speed limit of the lane) and a standard deviation of s l s dev with 0 <= s dev <=1. If s dev =0, v i mas = s l and is a distribution otherwise. Typically, s dev =0for autonomous vehicles i.e., they follow the lane speed limits while > 0 for humans (or a mix). Humans are also restricted by their own preferred maximum speed, v i p. Thus, the maximum speed of i th vehicle is the minimum of v i mas and v i p. 4.4 Other parameters Our experiments also use a parameter, re-computation time, t, which is the time interval after which lane change requests can be sent again. In BLS, t also represents the time interval after which the utility of each lane, u l, is recomputed for taking a lane change decision. We also have the parameters w a, w b, w c and as described in the BLS section. All the parameters and their default values are summarized in Table 4.1, which we arrived at via domain knowledge and experimentation. 18

29 Parameter Description Defaults v i Vehicle maximum speed 120 km/h v i p Preferred maximum speed variable a i Vehicle acceleration 0.8 m/s 2 d i Vehicle deceleration 4.5 m/s 2 c j Request delay 1 second s dev Speed deviation 0, 0.2 Driver imperfection 0, 0.5 w a Lowest speed weight-age 0.4 w b Average speed weight-age 0.4 w c Free space weight-age 0.2 Min utility difference between lanes 0.10 t Recomputation Interval 10 seconds c d Communication Distance 100 meters Table 4.1 Experimental parameters 4.5 Simulation traffic distribution New vehicles get introduced in our simulation any time before 250 seconds (i.e., between 1 and 250 seconds) and the start times for these new vehicles are generated randomly using a uniform distribution. The distribution is such that a new vehicle enters the simulation every second with a 60% probability. This corresponds to 60 vehicles entering the simulation, every hundred seconds, on average. The EV enters the simulation at 200 seconds. The simulation run ends when all the vehicles that had entered during the 250 second period reach the end of the road segment. 4.6 Environmental Settings For experimentation purposes, we modeled the following different environmental settings Setting 1 In this setting, we try to model an idealistic traffic pattern by setting all the vehicles to have the same preferred maximum speed, v i p of 90 km/h. The speed limit of all lanes is 60 km/h. The speed deviation factor, s dev, is set to 0 and the driver imperfection factor is also set to 0. Hence, we expect a behavior where vehicles enter the road and travel at the lane speed limits throughout the journey. This is a possible scenario, when all the vehicles on the road are autonomous vehicles of the same kind, traveling without any obstructions or road damages. 19

30 4.6.2 Setting 2 In this setting, we try to add realism into a traffic pattern in the following ways: (a) Typically all the vehicles on a dense road do not travel at same speed. Hence, using a uniform distribution, we generate v i p of vehicles to vary from 20 to 60 km/h. The lane speed limits are 60 km/h here. (b) While speed deviation, s dev is set to 0 which is an idealistic behavior we do model driver imperfections by setting to 0.5. To sum up, we expect this to be more realistic than Setting 1, due to the distribution over maximum preferred speeds and parameter Setting 3 We try to make this setting closest to most roads in densely populated cities, by adding the following behaviors: (a) As in Setting 2, v i p of vehicles varies from 20 to 60 km/h. (b) We model deviation from lane speed limits by setting s dev to 0.2. (c) Driver imperfections are modeled by setting to 0.5. (d) To model congestion, we put limits on traffic speeds in different lanes by using a Gaussian distribution for lane speed limits with mean = 60 km/h and standard deviation = 30 km/h. We expect this to result in traffic speeds that vary between very low sometimes (as happens during congestion) to slow moving traffic range most times. (e) Modeling different speed limits for different lanes as in (d), also introduces significant planning uncertainties since vehicles would not know beforehand the traffic patterns that may arise. As described earlier, the maximum speed of a vehicle is derived as min(v i mas,v i p ) where v i mas is generated using lane speed from distribution in (d) and s dev Setting 4 In this setting, we model the traffic patterns corresponding to cities with relatively faster moving traffic like New York City. We test here on a wider range of lane speeds than setting 3, modeled using actual traffic speeds of a few roads. In particular, we use the data available from the City of New York Department of Transportation(NYCDOT) [29] to perform the modeling. More details in Section

31 4.6.5 Setting 5 In this setting, we do not allow communication of lane change requests. Other than this all other parameters are same as Setting 2. This setting represents the situation where typically the system does not have the infrastructure to facilitate V2V communication for sending lane change requests but can receive the information about speeds and positions of vehicles. Even if communication is allowed, this represents the situation for (lower priority) vehicles such as cars, buses etc., not having the permission to clear a road. 4.7 Deciding on weights in BLS As described in Section 3.4, weights in BLS are very important to determine the utility of different lanes. We therefore performed experiments using several values of w a, w b and w c which combined with domain knowledge led to default weights of 0.4, 0.4 and 0.2, as they resulted in best traversal time for EV in most environmental settings. Although these weights are suitable for most scenarios, a different set of weights maybe used by the EV in some radically different road settings (e.g., w c could be set to 0 for a road which is always very sparse, causing the effect of free space to be insignificant). We verified the robustness of solution for the weights we picked, by perturbing the weights by (-5%,+5%) from defaults. The maximum difference in EV traversal times, on average, was 1.64 seconds or 0.704%. 4.8 Mapping NYCDOT data to Setting 4 The NYCDOT data feed that we use to model Setting 4, contains real-time traffic information from sensor feeds, mostly from major arterials and highways of New York City (NYC). This data feed is updated every minute for each road over a total of 153 roads. We developed our model using the following: (a) We collected data for about a week (9670 minutes) from the real-time data feed. (b) For any road we simulate, we use a segment of 2 km, irrespective of the actual length of the road segment. (c) For (a few) roads with varying number of lanes, we use the maximum number of lanes for the entire length we simulate. (d) We included 151 roads having >= 2lanes each (2 single lane roads were excluded). (e) The road speed data we have is converted into individual lane speed using the procedure described below that would capture the salient features of NYC traffic. We use the following procedure to calibrate Setting 4 for New York City road speeds: 21

32 4.8.1 Modeling groups We first classify roads into groups based on number of lanes in the road. Hence all roads with 2 lanes are classified into one group g 2, roads with 3 lanes into group g 3 and so on till g 6 (maximum number of lanes for any road in the data). We therefore obtain 5 groups in total g 2 g Modeling buckets Within each group g j, the average road speeds are classified into buckets using intervals of 5 m/s (18 km/h) each. Therefore we have buckets b 1 j for 0-5 m/s, b 2 j for 5-10 m/s till b 10 j for m/s. The average speed for each bucket, bavs i j, is taken as the mean of the bucket. For example, for the bucket 0-5 m/s, the average bucket speed, bavs 1 j is considered 2.5 m/s. In a similar way the average road speed for 5-10 m/s bucket is considered 7.5 m/s Obtaining weights for buckets We then obtain the weights for each bucket in the following fashion: We have information for 151 roads having 9670 minutes of data collected per minute. Therefore, for each road, we have 9670 data points corresponding to average speed of the road at that minute. Each of these 9670 points are then classified into buckets depending on the speed the data point represents. The weight of a bucket is incremented by 1 for each data point that falls under this bucket. For example, if the average speed of a road (with j lanes) is 0-5 m/s for 500 minutes (out of the 9670 minutes) then we increment the weight w 1 j by 500. This procedure is repeated for all the roads to obtain the total weight for each bucket The SUMO simulation Each b i j has an average bucket speed bavs i j and a weight w i j. bavs i j is taken as the average road speed in simulation. To simulate lanes, bavs i j is converted into lane speeds: Each lane is set a maximum speed, picked using a uniform distribution between the average road speed ± 40%. Hence, the mean of maximum speed across lanes is the average road speed on expectation. We then let the EV traverse on a 2 km stretch of road in the SUMO simulation and calculate it s run time. As mentioned in Chapter 5, each setting is run for 100 times, hence the same lane will have different maximum speeds across the 100 runs and we obtain 100 different EV run times Summary of SUMO parameters used (a) The number of lanes in this setting ranges from 2 to 6. (b) The maximum preferred speed of vehicles, v i p, varies from 60 to 120 km/h. 22

33 (c) As in Setting 3, we model deviation of drivers from lane speed limits by setting s dev to 0.2 and model driver imperfections by setting to 0.5. (d) All the other parameters were set to default values mentioned in Table EV run time per bucket We average the 100 different EV run times obtained to get the EV run time EV r i j for a bucket i and group j. It represents the time an EV would take (on average) if the number of lanes is j and the road speed corresponds to bavs i j. This is repeated for all buckets in every group Computing mean run time: For each group j, we compute gr j = EV r 1 j * w 1 j + EV r 2 j * w 2 j EV r 10 j * w 10 j. We repeat this procedure to obtain gr 2, gr 3,..., gr 6. Here (gr j )/(w 1 j w 10 j ) represents the average time an EV needs to travel a 2 km road with j lanes. The mean run time is obtained as (gr gr 6 )/((w w 10 2 )+(w w 10 3 )+... +(w w 10 6 )). This mean run time represents the average time an EV needs to cover a 2 km stretch of road with speeds corresponding to NYC roads and is used as run time for different strategies and experimental parameters for Setting Other Details As described earlier, apart from the EV lane picking strategy we let SUMO control most of the traffic behavior (to generate realistic traffic patterns), except for the following changes/interventions: (a) In Settings 1,2,3 and 4 our strategies involve changing the lanes of some of the vehicles to implement lane change requests. (b) In Setting 3 and 4 we (also) modify lane speed limits to implement the varying speed limits for different lanes. Modification of behavior of vehicles and lane speed limits is done using the TraCI interface of SUMO [42]. Apart from these specific interventions, SUMO handles all the behaviors and we do not interfere in any way. Another detail related to experiments is that, we compare the FLS and BLS strategies against two baselines: One is the SUMO Strategy described earlier. The second is an Empty road baseline (ERB), which is the time taken by the EV when there are 0 vehicles on the road apart from the EV for the entire simulation period. This acts as a lower bound for the EV traversal time. 23

34 Chapter 5 Experimental Analysis In this chapter we present a detailed experimental analysis of the SUMO, BLS and FLS strategies. It includes several realistic factors that would effect an EV on roads like human drivers ignoring lane change requests, communication delays, effect of deviations from lane speed limits,etc. We also analyze the improvement over current practices and effect on other vehicles traversing the road. All the results of our experiments presented in this chapter are an average over 100 runs. Each run has vehicles entering the simulation of a 2 km stretch of road with different starting times, generated using the description in Section 4.5. In Setting 4, we generate the results for a representative stretch of road, as described in Section Comparison between strategies Environmental Setting Figure 5.1 Comparison of strategies 24

35 The first experiment studies how the four different strategies ERB, SUMO, FLS and BLS perform on each of the five settings described earlier. Results of this experiment are summarized in Figure 5.1. The figure shows the five settings on x-axis, and for each setting the four different strategies are shown as bars. The y-axis represents the time taken in seconds by the EV to travel from start to destination using each of the strategies. Our results show that in Setting 1, all the strategies need similar time (as expected) as the empty road baseline (ERB) since vehicles are able to move at the lane speed limits throughout the journey due to the idealistic nature of this setting. Thus, there is no advantage gained or lost due to the use of lane level dynamics. In Setting 2, FLS performs better than BLS on an average (12.39% faster EV travel time). Both of them perform significantly better than the SUMO Strategy (BLS is 6.18% faster while FLS is 17.81% faster). FLS is able to perform better here as the simulation setup leads to leftmost lane (chosen by FLS) being the fastest. The average speed of vehicles in the rightmost lane was km/h while those in the leftmost lane was km/h (31.7% more). As described earlier, FLS selects a fixed lane based on historical data. Hence, FLS is a useful strategy for roads which are very predictable and the fastest lane can be determined using historical behaviors, with a high accuracy. However, we observe that in urban cities traffic is highly unpredictable and dynamic in nature. In Settings 3 and 5, BLS performs better than FLS (with 41.37% and 5.17% improvement) and the SUMO Strategy (with 29.44% and 6.13% improvement). Setting 3 allows deviation from lane speeds limits and has congestion, thus adding a lot of dynamicity to the traffic patterns. BLS is handling it better as it can select the best lane for EV traversal given the current traffic situation. In other words using the information obtained from V2V communication, BLS is able to perform a better optimization of its lane selection, given the current traffic situation, in the face of everyday occurring traffic phenomenon such as congestion, uncertainty and dynamicity of traffic. On the other hand, FLS due to its static nature, is not able to handle this setting well and has a lower performance than even the SUMO strategy (SUMO strategy is 16.90% faster here). The travel time trends for Setting 4 are similar to Setting 3. BLS performs significantly better than FLS and the SUMO strategy (with 18.90% and 16.42% improvement). FLS has a lower performance here too (SUMO strategy is 2.96% faster). Hence, we see that BLS performs better on the setting modeled using the New York city traffic data. In Setting 5 we do not allow clearance of vehicles. Here, BLS performs slightly better than FLS as it is able to find gaps on the road. In particular, BLS performs significantly better than FLS on less dense roads due to its ability to pick up the gaps (gaps have more utility) and about the same in empty roads as picking of gaps is not really needed. In this setting, both FLS and BLS perform a little better than the SUMO strategy (1.01% improvement and 6.13% improvement). 25

36 5.1.1 Standard Deviation across different runs The results in Figure 5.1 were presented as an average over 100 runs. We analyze here the variation across the different runs captured using standard deviation (SD). Figure 5.2 shows the five settings on x-axis, and for each setting the four different strategies are shown as bars. The y-axis represents the standard deviation obtained from the individual times taken by the EV (in seconds), to travel from start to destination, using each of the strategies. Environmental Setting Figure 5.2 Standard Deviation across different runs As expected, Setting 1 has a negligible SD while Setting 3 has the highest. Across all settings ERB has no variations and always takes 146 seconds. In Setting 1, BLS and SUMO strategy have a negligible SD of 0.51 seconds and 0.50 seconds respectively. The SD for FLS is 2.17 seconds (the EV sometimes has a small delay to move to leftmost lane from random initial position). In Settings 2 and 5, we observe the SD to be seconds across all the strategies. This is because we generate a new traffic pattern for every run ( Section 4.5 ), hence it is expected that the EV takes slightly varied times. Settings 3 and 4 have a higher standard deviation. It is more for FLS ( seconds in Setting 3 and seconds in Setting 4 ) than BLS ( seconds in Setting 3 and seconds in Setting 4 ). This is because FLS is not able to adjust well to the dynamicity of traffic caused by varying congestion in different runs. The deviation for SUMO strategy is slightly more than BLS. To sum up, across the Settings 2,3,4 and 5, FLS, SUMO and BLS have a standard deviation of seconds, seconds and seconds respectively (this is the SD accross different environmental 26

37 setting, considering the averages of each setting). This shows that BLS is much more robust to a wide variety of traffic settings in comparison to both FLS and SUMO strategies. 5.2 Varying the communication distance This experiment studies the effect of changing the communication distance, c d, in different environmental settings. Each of the figures, represents a setting and shows the different values picked for the parameter c d on the x-axis (in meters) and the time taken by the EV on the y-axis (in seconds). As described earlier, c d represents the distance upto which the EV can send lane change requests to vehicles in front of it (if allowed to). In addition, c d also represents the distance in BLS upto which requests for information about other vehicles are sent to compute the utilities of lanes. Figure 5.3 Variation of Communication Distance, Setting 1 For Setting 1 in Figure 5.3, there is no benefit from additional communication as vehicles are able to travel at the lane speed limits (and cannot go any faster). In this Setting, the only difference in traversal times is due to the small overhead for changing lanes. As a result, there is no effect of communication on the traversal times of the EV. Setting 2 is represented in Figure 5.4. We see that FLS, BLS and the SUMO strategies gain advantage from an increase in c d till a certain threshold value, with maximum benefit gained by BLS and least by SUMO. FLS, BLS and SUMO strategies gain advantage from an increase in c d as the EV can send lane 27

38 Figure 5.4 Variation of Communication Distance, Setting 2 Figure 5.5 Variation of Communication Distance, Setting 3 28

39 Figure 5.6 Variation of Communication Distance, Setting 4 change requests to father vehicles. BLS gains additional advantage as it also uses the information about speed and position of other vehicles present in different lanes upto distance c d, to compute better lanes. For Settings 3 (represented in Figure 5.5) and 4 (represented in Figure 5.6) also, we see that the traversal time for FLS, BLS and the SUMO strategies decrease with an increase in c d, till a certain threshold value. We believe this happens at a lower threshold for Settings 3 and 4 (as compared to Setting 2) due to the following reason: There is lesser relevant information to be gained by communicating with farther vehicles, since due to congestion, there are more vehicles on the road within relatively short distances. This makes it easier to get accurate information about traffic speeds from nearer vehicles, thus sufficing to plan for shorter distances. Here also, BLS gains maximum benefit with increase in c d since apart from sending lane change requests to farther vehicles, the EV also uses the information about speed and position of other vehicles, to compute better lanes. We found that the EV travel time trends for Setting 4 are similar to Setting 3 except that the run time difference between the FLS and SUMO strategies is less and becomes close to zero at higher communication distances. This is because in Setting 4, there is less congestion (compared to Setting 3), thus the EV using a fixed lane can achieve better clearance of the lane since adjacent lanes might have more free space to clear than in Setting 3. Setting 5 is represented in Figure 5.7. In Setting 5, lane clearing is not allowed. Since FLS and SUMO strategies use communication only for lane clearance, c d has no effect here. BLS uses communication for utility computation too. Hence, for small values of c d (less than 30 meters), the EV 29

40 Figure 5.7 Variation of Communication Distance, Setting 5 performance suffers as there is information about too few vehicles to analyze the lane dynamics. However, c d > 100 meters does not lead to much better utility computation since vehicles further away may not add much information. In fact we observed a slight decrease in EV performance at 200 meters as compared to 150 meters due to addition of not so relevant or possibly irrelevant information. In the above experiments, Setting 2 represented idealistic movement on roads, while Setting 3 and 4 represent more realistic scenarios. We found that Setting 3 and 4 follow similar trends. Since, its much easier to run Setting 3 on scale, for the following experiments we only show the results for Setting 3. Setting 1 is very idealistic and its analysis does not provide any significant insights. Hence, the following experiments contain the results for Setting 2, 3 and 5 respectively. 5.3 Analysis of utility computation In this experiment, we study the relationship between utility computation and the EV performance (traversal time). The goal is to verify that our utility function indeed results in lower run times for higher utilities i.e., inversely correlated. The figures show the average reported utilities over an EV run on the x-axis and time taken by the EV in that run on y-axis for Setting 2 (Figure 5.8), Setting 3 (Figure 5.9) and Setting 5 (Figure 5.10) respectively. Note that utility values range from 0 to 1. We plotted the best fit curve for the points plotted and observed that, over the entire curve, on an average, an 30

41 Reported utility Figure 5.8 Average utility v/s EV time taken, Setting 2 Reported utility Figure 5.9 Average utility v/s EV time taken, Setting 3 31