DASC: Towards A Road Damage-Aware Social-Media-Driven Car Sensing Framework for Disaster Response Applications
DDASC: Towards A Road Damage-Aware Social-Media-Driven Car SensingFramework for Disaster Response Applications
Md Tahmid Rashid, Daniel (Yue) Zhang, Dong Wang
Department of Computer Science and EngineeringUniversity of Notre DameNotre Dame, IN 46556
Abstract
While vehicular sensor networks (VSNs) have earned the stature of a mobile sensing paradigmutilizing sensors built into cars, they have limited sensing scopes since car drivers only opportunis-tically discover new events. Conversely, social sensing is emerging as a new sensing paradigm wheremeasurements about the physical world are collected from humans. In contrast to VSNs, socialsensing is more pervasive, but one of its key limitations lies in its inconsistent reliability stemmingfrom the data contributed by unreliable human sensors. In this paper, we present DASC, a road D amage- A ware S ocial-media-driven C ar sensing framework that exploits the collective power ofsocial sensing and VSNs for reliable disaster response applications. However, integrating VSNswith social sensing introduces a new set of challenges: i) How to leverage noisy and unreliablesocial signals to route the vehicles to accurate regions of interest? ii) How to tackle the inconsis-tent availability (e.g., churns) caused by car drivers being rational actors? iii) How to efficientlyguide the cars to the event locations with little prior knowledge of the road damage caused by thedisaster, while also handling the dynamics of the physical world and social media? The DASCframework addresses the above challenges by establishing a novel hybrid social-car sensing systemthat employs techniques from game theory, feedback control, and Markov Decision Process (MDP).In particular, DASC distills signals emitted from social media and discovers the road damages toeffectively drive cars to target areas for verifying emergency events. We implement and evaluateDASC in a reputed vehicle simulator that can emulate real-world disaster response scenarios. Theresults of a real-world application demonstrate the superiority of DASC over current VSNs-basedsolutions in detection accuracy and efficiency. Keywords: vehicle sensor networks, social sensing, bottom-up game theory, incentive control,Markov Decision Process.
Preprint submitted to Elsevier June 5, 2020 a r X i v : . [ c s . S I] J un . Introduction Vehicular sensor networks (VSNs) have evolved into a robust networked sensing paradigmfor obtaining situational awareness in disaster response applications [1]. VSNs incorporate carsequipped with arrays of on-board sensors (e.g. dashboard cameras) to opportunistically identifyevent occurrences like gas unavailability at nearby gas stations or accidents on the roads [2]. Socialsensing, on the other hand, is a permeating sensing paradigm for collecting real-time measurementsabout the physical world from observations reported by social media users [3]. Examples of socialsensing applications include monitoring air quality in smart cities [4], studying human mobility inurban areas [5], and obtaining situation awareness in the aftermath of disasters [6] using onlinesocial media (e.g., Twitter, Instagram).While VSNs render greater reliability in the discovery of the ground truth of events usingphysical sensors, one limitation is that the information collected by the vehicles is restricted to onlythose regions traversed by car drivers. Such a limitation vastly limits the scope of sensing for VSNsand their adaptability in unraveling new events. Moreover, during a disaster situation, roads couldbecome inaccessible due to damages caused by the disaster, rendering VSNs ineffective in certainscenarios. In contrast to VSNs, the scale of social sensing is broader and any individual possessing asmart device with Internet connectivity can potentially report an event on social media. However,an inherent limitation of social sensing is the inconsistent reliability of the sensing data that areoften contributed by unreliable human sensors [7]. In this paper, we exploit the complementarynature and collective strengths of VSNs and social sensing to develop DASC, a Damage-AwareSocial-media-driven Car sensing framework.Consider Hurricane Harvey that occurred in Southern Texas in August 2017 as an example.In the aftermath of the hurricane, access to critical resources (e.g., gas stations, pharmacies, food)became crucial for the victims affected by the disaster [8, 9]. Figure 1 shows tweets posted duringHurricane Harvey. If these tweets could be utilized to direct car drivers to desired locations, thedisaster recovery process could be facilitated by obtaining facts regarding the reported events usingvehicular sensors. However, a few key technical challenges need to be addressed for developing areliable social-media-driven car sensing system. 2 igure 1: Tweets Posted During Hurricane Harvey
The first challenge is leverag-ing the sparse and unreliable so-cial media data to guide cars todesired locations. A key challeng-ing task in social sensing appli-cations is the accurate identifica-tion of reliable sources and truth-ful claims from the sparse and un-certain social sensing data, otherwise known as truth discovery [10]. To discover truthful informa-tion from unvetted social media users, existing truth discovery solutions primarily rely on the postspresented on social media. These solutions may yield unreliable truth discovery results, makingit difficult to decide where to dispatch the cars [11]. In addition to that, existing truth discoveryalgorithms that output probabilistic distributions for the classification cannot definitively confirmthe truthfulness of an event. Therefore, it is intrinsically difficult to extract reliable social signalsto guide cars to accurate locations of interests. We deem this challenge as the cyber challenge ofthe problem.The second challenge is the inconsistent availability, otherwise known as churn , caused by therational car drivers who may drop sensing tasks midway during explorations. We assume thatat times, car drivers may focus on personal goals instead of the given assignment and abandonthe sensing tasks abruptly in the middle of a trip. This inconsistent availability issue of churn has been explored in the field of distributed systems [12, 13, 14]. Existing literature has proposedmethods to reduce churns by increasing the total number of devices in the system [15, 16] or byreallocating tasks to more reliable devices [17]. While those solutions may work for networkeddevices, they are application-specific and are hard to be applied to cars. For example, it maynot be possible to rigorously control the number of cars in the system since the cars are privatelyowned by individuals. Moreover, while allocating tasks to more reputable drivers (i.e., drivers whoare less likely to drop a task) may improve the performance of the system, it could be impracticalto switch the tasks to a car that is in the middle of an ongoing exploration on the road. Hence,it remains an open challenge to decisively allocate sensing tasks to cars (drivers) in our DASCframework. We deem this challenge as the human challenge of the problem.3he third challenge is allocating the cars to the event locations with little or no prior knowl-edge of road damages caused by the disaster. After a major natural disaster, it is likely that acertain proportion of roadways are unreachable [18], as exemplified in Figure 1 (marked by redstars). Road damages inflicted by the disaster would greatly limit the maneuverability of the cars.Furthermore, the extent of road damages is unpredictable and cannot be known beforehand. A fewrouting strategies have been proposed to model the road damages and route cars safely [19, 20].However, existing solutions assume that the system has global knowledge of all the road damages,making feasible routing decisions. In contrast, with damaged infrastructures after a disaster, theinformation about the road conditions cannot be readily determined and disseminated. For exam-ple, at any point in time, new road damages may appear and also existing road may get repaired,introducing a level of dynamism into the system. Current solutions in the discipline of routingalgorithms consider the issue of unexplored and incomplete information. However, in the contextof social-media-driven car sensing systems, the dynamics of social media combined with that of thephysical world make the problem more challenging. We deem this challenge as the physical chal-lenge of the problem. Figure 2 illustrates the three challenges in developing a social-media-drivencar sensing system.
Figure 2: The three challenges in social-media-driven car sensing system
In this paper, we develop DASC, a damage-aware so-cial media-driven car sensing system to address the abovechallenges. To address the first challenge, we develop abottom-up game-theoretic task allocation model to judi-ciously dispatch the cars to reported locations and verifythe event information extracted from unreliable social me-dia data. To address the second challenge, we design atop-down incentive control mechanism to dynamically ad-just the incentives for exploration of the event locationsbased on the aggregated reputations of the cars (i.e., thehistorical behavior of the cars in attempting to successfullycomplete the tasks). To address the third challenge, we develop a Markov Decision Process (MDP)-based damage discovery scheme to locate the roads affected by damage and leverage the obtainedknowledge to make optimal route planning decisions. To the best of our knowledge, DASC is the4rst solution that melds the realms of social sensing and VSNs for a robust car sensing (SCS)application with explicit consideration of the road damages in the aftermath of a disaster. Weimplemented the DASC framework and evaluated it with CARLA, a reputed car emulation systemthat is capable of closely replicating road networks and vehicles in real-world scenarios. We com-pared our framework against representative vehicular based sensing systems on a real-world datasetcollected from Twitter during a natural disaster: Hurricane Harvey in August 2017. The resultsshow that DASC significantly outperforms the compared baselines in both detection effectivenessand deadline hit rate during the aftermath of a disaster.A preliminary version of this work has been published in [21]. We refer to the scheme developedin the prior work as the SocialCar scheme. This paper is a significant extension of the previous workin the following aspects. First, we define a new problem of guiding cars to the event locations usingsocial media signals while explicitly considering the damaged roads left by a disaster. In contrastto the SocialCar scheme, the new problem is much more challenging because immediately after thedisaster, the knowledge of the road damage is not available, making it much more difficult to decidethe route for sending the cars to the event locations. Second, we allocate a portion of participatingcars for “scouting” the routes to explore the damaged roads. Third, we introduce a Markov DecisionProcess (MDP)-based technique for utilizing the knowledge of the road damage obtained in theprior step for feasible route planning. Fourth, we employ an exploration-exploitation strategy tolearn the optimal routing decision over time. Fifth, we carry out a new set of experiments toexplicitly evaluate the performance of all schemes in terms of detection effectiveness and deadlinehit rates in the new problem setting. Sixth, we include additional baselines to further exhibitthe performance gains achieved by DASC. Finally, we extend the related work by adding a newdiscussion on the damage-aware routing schemes and highlight the difference between DASC andthose schemes (Section 2).
2. Related Work
The emergence of modern cars equipped with advanced sensors has opened new domains invehicular networked sensing [2]. For example, Nekovee presented a comprehensive study of severalroadside monitoring systems that integrate the data of nearby vehicular sensors [22]. Lee et al. pro-5osed Mobeyes, an urban surveillance system that manages a dedicated number of sensor-equippedcars to opportunistically explore events [23]. In combination with built-in sensors, crowdsourcedVSNs using smartphones have also gained popularity. Commercial solutions like Waze [24] andGasBuddy [25] provide users with valuable information about the availability of critical resourcesas reported by car drivers. However, the above approaches primarily depend on the opportunisticnature of the car drivers since cars only “sense” incidents when they come across them. In contrast,this paper presents the DASC framework that leverages the social media signals to guide cars todesirable locations for a greater sensing scope.
Social sensing is transcending as a recent sensing paradigm that uses humans as sensors to reportabout observations in the physical world [26, 27]. Examples of social sensing applications includeidentifying traffic risks [28], tracking social unrest [29] and disasters [30, 31], sensing points ofinterest in large cities [32, 33], and detecting car plates of suspects [34, 14]. Several challenges havebeen studied in social sensing. Examples include data reliability [35], human uncertainty [36], datasparsity [37], privacy preservation [38], and real-time requirements [39]. A comprehensive survey ofsocial sensing schemes is provided in [7]. Xu et al. developed a framework for semantic and spatialanalysis of urban emergency events using social sensing [40]. Chen et al. proposed a road trafficcongestion monitoring system using social media data [41]. Imran et al. developed a machinelearning-based disaster identification system capable of classifying and analyzing information fromcrisis-related tweets in real-time [42]. A key limitation of current social sensing systems is that theyonly rely on the social media data which could be unreliable [43, 44, 45]. More recently, there is aninception of social media driven UAV-based sensing approaches that address the data reliabilityissue of social sensing by using physical drones [46, 47]. However, these solutions require dedicateddrones for the sensing purpose that are typically known to be expensive and limited in numbers.In contrast, the SocialCar framework integrates social media with existing vehicular-based sensingsystems to provide data reliability assurance in scalable social sensing.
Recent literature has presented methods to diminish the issue of high churn in participatorysensing and distributed systems. For example, Gao et al. [48] presented a study of different6ncentive mechanisms for participatory sensing to lower the possibility of churn. Godfrey et al. [15] proposed a technique to reduce churn by intelligently selecting only the most reliable devicesfrom all participating devices in a system. Haeberlen et al. proposed a solution to reduce churnby increasing the total number of existing devices [16]. One major drawback of these approachesis that they are all either bottom-up or top-down and do not holistically consider the objectives ofthe individual devices and the server. In contrast, our DASC framework uses both a bottom-uptask preference and a top-down dynamic incentive control to collectively consider the objectiveson both ends (i.e. the cars and the SVS application) to better control the churn issue.
Road damage-aware vehicle routing is a well-studied topic in vehicular networks. For example,Hsueh et al. [49] presented a comprehensive study of road damage-aware dynamic vehicle routing forrelief logistics during natural disasters. Korkmaz et al. [50] proposed a road damage-aware satellite-imagery based path planning framework for rescue vehicles during disasters. Mahmoudabadi etal. [51] developed a damage severity-aware route planning for transporting hazardous substancesduring emergencies. Kuntze et al. [52] explored the possibility of collision-free path planning ofunmanned ground vehicles (UGVs) in road-damage prone areas. While the above approachesintend to solve the critical challenge of guiding vehicles through road damage, our problem ofbuilding a damage-aware social-media-driven car sensing system is even more challenging due toboth the dynamics of the social media and the physical world. In this paper, we develop the DASCframework to address this challenge by designing an MDP-based damage discovery technique tolocate damaged roads and use the knowledge for optimal routing.
3. Problem Formulation
In this section, we present the fundamental definitions and assumptions of our model and definethe objective of our problem. In a damage-aware social-media-driven car sensing (SCS) application,we inspect a physical region of interest (ROI) for a specific duration of sensing. The sensing timelineis discretized into T periodic intervals, namely response cycles . In particular, t ∈ [1 , T ] indicatesthe t th response cycle. 7 EFINITION 1. Sensing Cells:
We divide the sensing region reachable by cars into H disjoint sensing cells . Each cell represents a real-world location, connected by roads and accessible to cars At any given time, a sensing cell can be occupied by multiple cars, increasing the chance of a taskbeing completed. In particular, we define SC t,h to be the h th sensing cell at the t th response cycle. DEFINITION 2. Road Damage D t,h for sensing cell SC t,h : We assume that cars cannottraverse a sensing cell if it contains road damage. We identify a sensing cell’s damage state by abinary variable D t,h , where a value of 0 indicates no damage and a value of 1 indicates damage.We consider that a set of social media users reports a collection of independent events atdifferent sensing cells as defined below: DEFINITION 3. Event E t,n : An event is assumed to represent a physical variable of interestwithin a sensing cell in the SCS application. Examples of reported events include gas availabilityat a gas station or a person trapped under a vehicle. We let a binary variable E t,n denote the n th event in the t th response cycle with a total of N t events. For each reported event E t,n , it either“exists” (i.e., E t,n = 1) or “does not exist” (i.e., E t,n = 0).We use (cid:100) E t,n to denote the truth of event E t,n estimated by our DASC system. An essentialattribute of each event is the sensing deadline as defined below: DEFINITION 4. Sensing Deadline δ t,n for event E t,n : Each event is assigned a deadlinerepresenting the urgency by analyzing the content of the social media report [53]. We assume thatthe deadlines of events are shorter than the duration of the sensing cycle [34]. For an event with alonger deadline than the sensing cycle, we split the event into multiple events, each resulting eventhaving a deadline shorter than the sensing cycle. The split events also inherit the priority fromthe original event to avoid the potential problem of “priority inversion” [47].We define the data from social media (e.g., tweets from Twitter) as follows:
DEFINITION 5. Social Media Data S : A set of social media posts that reports events in thephysical world in the aftermath of a disaster. An example of such a report is shown in Figure 1. Please note that we only focus on the events in the sensing cells defined above and ignore the events thathappened in locations which are not accessible to cars given the scope of this work.
8e consider a set of tasks that are broadcast for the car drivers to explore the reported events.We formally define a task as:
DEFINITION 6. Task V t,n : A task for a car at each response cycle refers to the location of anevent (i.e., a sensing cell) where the car should be dispatched to investigate the event.Figure 3 presents an illustrative example of the concepts defined above.
Figure 3: Snapshot of the sensing grid across response cycles. Theyellow boxes signify sensing cells , the blue boxes represent unreachableregions, the red boxes signify the cells with road damage, the blackicons indicate the events, and the green car icons represent the cars.
We make two important as-sumptions about the unique com-pliance and churn issues in theDASC system.
Voluntary Compliance : Weassume that a car driver may ormay not be willing to accept anytask offered.
Dynamic Churn : We alsoassume that even if car drivers arewilling to pick up events for investigation, they may randomly abort the tasks (e.g., the driverdecides to go home after taking a task), causing the churn in the system.Using the above definitions, we, therefore, define the goal of our DASC framework. Given thesocial media data input S , a set of cars C , the corresponding deadlines for the events δ t,n , theroad damage D , as well as sensing cells SC t,h , the objective of the DASC framework is to dispatchthe cars to a set of sensing cells to maximally recover the true states of the events reported bysocial media users while considering the road damage caused by the disaster. We formally solve aconstrained optimization problem as follows:arg min (cid:100) E t,n M t (cid:88) n =1 ( abs ( (cid:100) E t,n − E t,n ) |D , S , C , δ t,n , SC t,h ) , ∀ ≤ t ≤ T, ∀ ≤ h ≤ H (1)The definitions of the notations are summarized in Table 1.9 able 1: Summary of Notations t The t th response cycle, t ∈ { , , ..., T } E t,n The n th event in response cycle t (cid:100) E t,n The estimated truth of event E t,n δ t,n The sensing deadline for event E t,n SC t,h The h th sensing cell in response cycle tD t,h The road damage for sensing cell SC t,h V t,n The n th task in response cycle t S The social media data C b The c th car, c ∈ { , , ..., G }
4. The DASC Framework
In this section, we present the DASC framework that integrates the social media and thevehicular sensing system for a reliable road damage-aware SCS application. An overview of DASCis shown in Figure 4. The DASC consists of four major components: i) a Social Signal Distillation(SSD) module; ii) a Road Damage Discovery (RDD) module; iii) a Vehicle Dispatch (VD) module; and iv) a Dynamic Incentive Control (DIC) module.
Figure 4: Overview of the DASC Framework. Theblue car icons represent the cars sent out for locatingthe road damage, the black car icons represent thecars dispatched to event locations, and the red starsindicate the discovered road damages along the routes.
The SSD module collects and extracts reli-able event reports from unreliable social mediausers. Concurrently, the RDD module assignsa portion of the participating cars as scout carsto explore the routes to the event locations forroad damage. The VD module allows the cardrivers to pick their preferred tasks based ontheir individual payoffs and leverages the knowl-edge from road damage data to guide the carsto the event locations. The DIC module assignsand adjusts the incentives for the tasks to max-imize the chance of the cars completing all thetasks (i.e. locating road damage and exploring all the reported events). To further maximize the10hance of the tasks being completed, once each task is selected by at least one car driver, the VDmodule allows the tasks to be selected by multiple car drivers. A detailed discussion of each moduleis presented in the following subsections.
The SSD module is designed to collect, pre-process, and analyze noisy social media posts toestimate the possibility of critical events in the physical world. The SSD module uses a real-timedata crawler engine to obtain social media data (e.g. Tweets) with geo-location tags indicatingdisaster-related events. The collected data is filtered by running keyword searches on it (e.g., gas,fuel, oil, medicine, healthcare, and pharmacy), and afterward clustered and labeled using a mi-croblog data clustering tool [11]. A key issue with the above generated social media data lies in thetrustworthiness of the reported events since these events are often reported by unvetted grass-rootusers, of whom the credibility is unknown a priori. Without carefully excluding the misinformationand rumors provided by unreliable users, the performance of the DASC system can be significantlydegraded. Another frequent issue observed in social media data is data sparsity, whereby a ma-jority of the users contribute only a small number of event reports, providing insufficient evidenceto accomplish the truth estimation task. In light of such challenges, the SSD module incorporatesa truth discovery (TD) solution to estimate the truthfulness of the reported events along withobtaining the estimation confidence/uncertainty [54].While there is an abundant number of TD solutions, we select a particular approach called theRobust Truth Discovery (RTD) [11] algorithm for our SSD module to filter useful signals from socialsensing data. Our main motivation for selecting this algorithm lies in its design philosophy to berobust against misinformation spread and data sparsity in social media applications. In particular,the RTD scheme handles widespread misinformation by explicitly quantifying different degreesof attitude that a source may express on a claim and incorporating the historical contributionsof a source using a principled approach. The fine-grained source attitude facilitates an effectivedetection of misinformation, which is based on the observation that the misinformation is morelikely to attract opposite opinions and intensive debates. Moreover, the RTD scheme addresses thedata sparsity issue by computing the claim truthfulness based on a function of the source attitude,the sources historical contributions, and the source reliability. The estimation is more robustsince it does not solely rely on the source reliability estimation, which is challenging to estimate11ccurately in a sparse dataset [32]. The RTD scheme measures the historical claims of each sourceby computing a metric called the contribution score that determines a source’s contribution to anevent report based on several factors [11]. The algorithm also utilizes a metric called the sourceattitude score to fully capture the reporting behavior of sources. We define the output of the RTDalgorithm as event veracity and estimation confidence which are defined below:
DEFINITION 7. Event Veracity Λ t,n for event E t,n : A score in the range (0,1] that indicatesthe chance of an event being true. Intuitively, the greater the value of Λ t,n , the more likely event E t,n is true (i.e., E t,n exists). To obtain Λ t,n for event E t,n , the RTD algorithm iteratively sumsup all the contribution scores from the set of social media users who contribute to E t,n . DEFINITION 8. Estimation Confidence Score EC t,n for event E t,n : A score in the range(0,1] that signifies the estimation confidence for an event. Intuitively, the greater the value of EC t,n ,the more confident the RTD algorithm is its estimation. Formally, it is defined as the absolutedifference between the event veracity score and the midpoint of event veracity score’s range (i.e.,the neutral point for determining the truthfulness).Leveraging social sensing and truth discovery, the SSD module of the DASC framework notonly provides the signals of the critical events for the vehicles to verify but also helps to quantify thepriorities of these events based on their confidence. Both the event truthfulness and the estimationconfidence are critical inputs to the DASC framework that guide the dispatching strategies of thevehicles in the VD module. The decision to dispatch the cars rely on the values of Λ t,n and EC t,n .For a given response cycle, if the value of EC t,n is above an adjustable threshold, DASC truststhe RTD algorithm’s decision without dispatching the vehicles and concludes upon event E t,n ’struthfulness based on the value of Λ t,n . For cases otherwise, where the value of EC t,n is belowthe threshold deeming the veracity doubtful, DASC incorporates the value of EC t,n into the VDmodule (the process of which is detailed in Section 4.3) and allocate tasks for the car drivers toexplore the event. Once the cars travel to the event destination and collect the actual truth withgreater reliability using the onboard sensors, DASC finally determines event E t,n ’s truthfulness. The RDD module is designed to incentivize and assign a fraction of cars (from the pool ofavailable cars that are willing to participate) to explore the available routes for road damages. In12articular, we assign Q % of all the available cars, that are willing to participate in the sensingprocess, as scout cars , where Q is adjusted according to the application scenario. We model eachroad intersection as the node of a graph and each road branching out of an intersection as the edge of a graph. Each exploration is modeled as a task that can be picked up by one or multiple carsand assigned a reward r t,n to incentivize the drivers to pick it up. The reward r t,n is determinedby the Dynamic Incentive Control (DIC) module discussed later in this section. The rationale isthat if we can traverse the maximum number of roads using the scout cars, we may have a betterchance of locating road damages.Once a scout car is adjacent to a cell with road damage, the damage information (i.e., D t,h )is recorded by the scout car and then it proceeds to explore a different route. The edges of thegraph are basically a series of contiguous sensing cells accessible by cars. While the road damagevariable D t,h indicates whether a cell has damage or not, we acknowledge that all the road damageinformation cannot be readily obtained or updated in a given sensing cycle. Therefore, it is areasonable assumption to determine the possibility of road damage across a sensing cell based onthe historical damage condition of the cell. In order to accomplish this, all the sensing cells thatmake up the edges of the graph are assigned a score called accessibility index , which is definedbelow. DEFINITION 9. Accessibility Index X t,h : A score in the range of [0 ,
1] to indicate the possi-bility of road damage across a sensing cell (i.e., how likely damage may occur again in a cell in thefuture). Intuitively, a lower accessibility index indicates that a route is less likely to be traversabledue to the possibility of containing damaged roads. Initially, all the cells are considered to have aninitial accessibility index X ,h , the value of X ,h is discussed in Section 5. Over response cycles,only when a sensing cell is visited by a scout car, the accessibility index is calculated as: X t,h = X t − ,h − κ t , D t,h = 1 X t − ,h + κ t , otherwise , ≤ X t,h ≤ κ t is an adjustable parameter called accessibility penalty, which is determined by a slidingwindow correlation [55] between the total number of detected road damages in t th sensing cycle, D totalt and the total number of reported events in t th sensing cycle, N t . The value of κ t is computed13y the following equation: κ t = (cid:80) ti = t − j +1 { ( D totalt − (cid:92) D totalt ) × ( N t − (cid:99) N t ) } (cid:113)(cid:80) ti = t − j +1 ( D totalt − (cid:92) D totalt ) × (cid:80) ti = t − j +1 ( N t − (cid:99) N t ) (3)where j is the sliding window of the number of cells to look back, (cid:92) D totalt is the average number ofroad damages during the sliding window, and (cid:99) N t is the average number of events during the slidingwindow. Intuitively, if there is a strong correlation between the number of events and the roaddamages in the current sensing cycle, the value of κ t will increase, thereby making the accessibilityindex more sensitive.In the beginning, the road damage D t,h = 0 for all the cells, assuming that all the roads aretraversable. If a cell is not visited by a car in a sensing cycle, the accessibility index for that cell isretained (i.e. X t,h = X t − ,h ). If a cell is found to have damage at a sensing cycle (i.e. D t,h = 1),the accessibility index is decremented by κ . Conversely, if a cell is detected to have no damagein a sensing cycle (i.e. D t,h = 0), the accessibility index is increased by κ . Intuitively, a lower X t,h means that a cell has a higher chance of being damaged based on prior history and should bedisregarded from routing decisions. On the other hand, a higher X t,h indicates that a cell has lesschance of damage based on historical damage information.We employ an established graph traversal algorithm, the A* search algorithm [56], to direct thescout cars in traversing the path covering the highest number of non-repeating edges between pairsof farthest nodes. The accessibility index is used by the Vehicle Dispatch (VD) module discussedin Section 4.3.2 to decide the route selection strategy for cars involved in event exploration. The VD module is designed to take the filtered social signals from the SSD module and theroad damage information in the form of accessibility index from the RDD module for appropriatelydispatching a group of interested vehicles to probable event locations. In particular, we use aBottom-Up Game-Theoretic (BGT) policy to prioritize and allocate tasks to the cars based on theassigned task rewards, the distance between the cars and the event locations, the remaining timeof the tasks, and the event uncertainty. Once the allocation of the tasks to the cars is completedby the BGT module, a Markov Decision Process (MDP)-based approach is employed to select thebest available routing strategy for the cars while incorporating the road damage information.14 .3.1. Bottom-Up Game-Theoretic Task Allocation
The bottom-up game-theoretic (BGT) task allocation approach is designed to allow the cardrivers to make choices of event locations to travel to. The key motivation behind this designprinciple of the VD module is to let the car drivers express their individual task preferences in theallocation process. This allows the cars to determine the strategy that maximizes their individualpayoffs [57].In game theory, congestion games are typically used to mitigate resource conflicts (e.g., eventlocations) among a set of players (e.g., cars). We adopt singleton weighted congestion games [58], avariant of congestion games where the expected utility of each task uniformly decreases as the sumof players (cars) that picked the task increases. Moreover, each car only picks one task at a timeaccording to the singleton property. The Pure Strategy Nash Equilibrium is guaranteed to existunder the above singleton weighted congestion game protocol [21]. This property enables the carsto make conclusive task allocation decisions. In particular, where are four core components in oursingleton weighted congestion game protocol: the reputation , the reward , the weighted congestionrate , and the utility function . We elaborate on them below.Similar to the damage discovery module, we assume that a task can be picked up by multiplecars and assign a reward r t,n for each task to incentivize the drivers to pick it up. We maintain a reputation score π t,p for each car C p based on the historical performance till the response cycle t .We use ν t,p to count the number of tasks successfully completed by car C p , and τ t,p to count thetasks marked as unsuccessful (i.e. a car not being able to perform a task by the sensing deadline)up to response cycle t , respectively. Intuitively, if a car picks up a task and successfully completesit, the reputation score will increase. If the car fails to reach the destination on time or drops thetask, the score will decrease accordingly. The reputation score π t,p is based on an initial reputation π ,p at t = 0 and is subsequently computed as: π t,p = π t − ,p + η × ( (cid:88) ν t,p − (cid:88) τ t,p ) , t > η is an adjustable parameter called reputation coefficient . If η is set high, the reputationscore will be more sensitive to the success and failure in the completion of tasks.We define a key component of our congestion game called weighted congestion rate as follows: DEFINITION 10. Weighted Congestion Rate γ mt,n for task V t,n for car C m : A score in therange of (0, ∞ ) that indicates the level of contention on a task. It serves as a discounting factor15f the utility function to dissuade cars to pick the same task already selected by several cars. Theweighted congestion rate is computed by: γ mt,n = G (cid:88) p =1 p (cid:54) = m S × ( π t,m − π t,p ) k S = sgn ( π t,m − π t,p ) , k is even1 , otherwise (5)where k is an exponential scaling factor to adjust the intensity of the congestion property. If k isset to be high, the congestion rate will be more sensitive to the difference in the reputation scores.The intuition here is that if several cars with reputation scores greater than car C m ’s reputationscore have already picked up a particular task, the congestion is higher. On the contrary, if a fewcars have already picked the event and have lower reputation scores, the congestion will be lower.We anticipate that once all the cars select all the tasks, a churn situation can occur. Forexample, a car may drop a task at any instant abruptly, new cars may join or existing cars mayleave the system. This may necessitate a reallocation of the tasks. We keep track of the remainingtime ρ t,n for each task at any time instant and define it as: ρ t,n = δ t,n − τ t (6)where τ t is the elapsed time from the beginning of the response cycle t .Given the definitions above, we can now derive the utility function based on which the carsdecide their best strategies and define it as: DEFINITION 11. Utility Function u mt,n for task V t,n : the utility function represents thebenefit for picking a specific task (i.e., event location) for car C m .In our model, we devised a customized utility function for car C m , referred to as event priorityscore as follows: u mt,n = r t,n × ( λ × ω mt,n + λ × ρ t,n + λ × h (Λ t,n )) γ mt,n , ρ t,n > , ρ t,n = 0 (7)The above utility function prioritizes the tasks for car task allocation based on four factors: i)the reward for the task, r t,n ; ii) the distance from the car to the event location, denoted as ω mt,n ; iii)the remaining time of the task, ρ t,n ; and iv) the uncertainty of an event, as captured by a function16f the estimation confidence score (i.e. f ( EC t,n )) from Definition 8. Given the rewards for thetasks, each car tries to prioritize the tasks with higher rewards. In particular, the remaining timefactor prioritizes the tasks with tighter remaining deadlines while the distance factor priorities taskswith shorter distances from the cars in order to reach nearby tasks first. λ , λ , and λ representthe weights of each factor. Their values are computed using proportional control , a widely usedcontrol technique [59]. In Section 5, we discuss how the three parameters are determined. Finally,the congestion rate, γ mt,n on the denominator of the utility function is designed to avoid contentionof cars for a task. We highlight that multiple cars have the freedom to select the same task afterall the tasks are allocated to at least one car. This is to increase the value of the congestion rate,thereby reducing the utility for each car. However, this approach also increases the chance of atask being completed. Additionally, if the remaining time ρ t,n is 0, the utility is 0.Afterward, each car decides on its best strategy towards maximizing its utility in the congestiongame, until a Nash Equilibrium is reached. The Nash Equilibrium (NE) exists in the proposedgame where each car is assumed to have determined its optimal decision (i.e., picking the taskhas the highest utility) and no car has anything to gain by only changing its preferred tasks. Weexploit the best-response dynamics algorithm to find the NE [60].We use an array U t,n for each task V t,n to record all the cars that pick the task in the t th response cycle after the NE is reached. Once the BGT sub-component determines the destinations for the cars, the MDP scheme incorporates the road damage information from the RDD moduleto assign the best routes to destinations for the cars. As the accessibility index (indicating the possibility of road damage) is obtained from the RDDmodule and the destinations for the cars are derived from the BGT sub-component, the knowledge isutilized to perform routing decisions. We found that our problem of allocating traversable routes forcars nicely fits into the principle of Online Markov Decision Process (MDP) [61]. For our problem,we consider the starting location of each car at every response cycle as a source and the eventlocation assigned to the car as a destination . We assume that most pairs of source-destination areconnected by multiple routes. Based on this assumption, we consider the source-destination pairsas the states for our MDP model. In our model, we map the actions to the list of available routes foreach state and the penalties (i.e., or equivalently negative reward) to the sum of the road damage17or each action (i.e., route). Afterward, we develop a custom action selection scheme to determinethe best actions and solve our MDP problem. Our choice for developing our own approach forsolving the MDP problem is driven by the rationale that our environment is highly uncertaindue to the dynamics of the social media and the physical world, which makes the determinationof the best actions challenging. While our states and actions do not change, the values of thepenalties (i.e., road damages) for corresponding actions often exhibit a dynamic behavior acrossresponse cycle due to the constantly changing number of event reports in the social media, theirlocations in the real world, and the road damage situation along the routes. This dynamism makesthe determination of the best actions challenging. Moreover, while the scout cars provide theframework with limited information related to the road damages in a prior response cycle, we maynot have the complete information of all the damaged roads across all the routes with a limitednumber of scout cars.Our scheme for determining the best actions is built on a feedback control mechanism that usesthe penalties as a feedback signal. At the end of every response cycle, the penalties from the currentsensing cycle are updated based on Equation 9 which is a function of the predicted and actual roaddamage for the current sensing cycle. The actions to take in the next sensing cycle are determinedby Equation 8 which is a function of the prior road damage and the normalized difference betweenthe sums of the penalties across successive response cycles. The details of the states, actions, andrewards, as well as the mechanism of our algorithm for selecting the best actions, are discussedbelow.
DEFINITION 12. States W t : A set of tuples W t = { W t , W t , ..., W tj t } denoting source-destinationpairs at the t th response cycle. An example of a source-destination pair in response cycle t = 3 is W = ( SC , , SC , ), where SC , is the source sensing cell (i.e., the position of the car) and SC , is the destination sensing cell (i.e., the event location). DEFINITION 13. Actions A tk for state W tk : A set of ordered lists A tk = { A tk, , A tk, , ..., A tk,l t } representing all the available routes for each state k at sensing cycle t . Each action set A tk , consistingof a set of contiguous sensing cells, maps to each state W tk . For example, actions set A for state W at t = 3 can have an action A , with sensing cells [ SC , , SC , , SC , ].All the available routes for each state are obtained by considering the task deadlines and sensingcell constraints for the cars while reaching the assigned sensing cells. We leverage a route planning18lgorithm based on a Contraction Hierarchies technique from graph theory [62] to generate all theavailable actions for each state.The probability function for the v th action, A tk,v in an action set is given by: P ( A tk,v ) = σ t − (cid:89) h ∈ a tk,v X t − ,h (8)where σ t − is a parameter called penalty differential that is determined by the penalty from theprior response cycle in Equation 10 discussed later. The term a tk,v represents the set of sensingcells that are part of the route for action A tk,v .At the beginning of every sensing cycle t , the accessibility index indicating the possibility ofroad damage from the last sensing cycle (i.e. t −
1) is used in Equation 8 to generate the probabilityof all the actions that can be taken in the current sensing cycle (i.e. t ). Intuitively, the greater theaccessibility indices for the associated cells, the higher the probability for that action consisting ofthe particular cells to be taken.We anticipate that it might not possible to locate all the road damage by the scout cars in theRDD module. Moreover, the road damage could be encountered later by the other cars assignedfor event exploration. DEFINITION 14. Penalties R t : A set of penalty scores R t = { R t , R t , ..., R tj t } obtained bysumming the road damages discovered by the car drivers tasked with event exploration for eachaction that is selected for each state. Each element of the set represents a particular action selectedfor the corresponding state and is thereby computed by: R tu = (cid:88) h ∈ a tu ( D t,h ∨ D t,h ) (9)where D t,h represents the road damage discovered by the cars exploring events at sensing cell SC th and a tu represents the set of all the sensing cells in the selected action. Initially, D t,h = D t − ,h for allthe cells at the beginning of every sensing cycle. Once an event exploration car discovers damage ina sensing cycle, D t,h = 1. In a future sensing cycle if the damage appears to be repaired, D t,h = 0.Intuitively, if R l = 0, the route is assumed to be fully accessible. Otherwise if R l >
0, the routemay contain damage and should be avoided. We adjust the penalty differential σ t discussed earlierbased on the normalized difference between the sums of the penalties across successive response19ycle: σ t = σ t − − (cid:80) R t − (cid:80) R t − (cid:80) R t + (cid:80) R t − (10)The rationale is that if the total magnitude of discovered damage increases in the current responsecycle, we lower the penalty differential to “penalize” the action taken earlier.We formally define our MDP model below. We consider an MDP with a state set W t mappedto the source-destination pairs, an action set A tk mapped to the choice of routes, and a penalty set R t mapped to the aggregate road damages discovered by all the car drivers for the selected actions.The goal of the MDP is to derive an optimal collection of actions that minimizes the encounter ofroad damages by the cars assigned for event exploration. Formally the objective is:argmin A tk T (cid:88) t =1 R t , ≤ t ≤ T (11)The optimal action set for a corresponding state can be obtained by incorporating the contextualepsilon-greedy strategy [63]. The first step is to initiate a learning phase (i.e., exploration ), for acertain duration of response cycles to determine the optimal actions A tk for each state. Duringthe learning phase, the state is determined based on the source-destination pairs of each car andthe action sets are generated by enumerating through all the possible routes. The scout carsare dispatched by the RDD module which continuously feeds the damage information to the VDmodule. An action A tk ∈ A tk is selected for each corresponding state W tk that has not been previouslyexplored. Cars are then dispatched for exploration using the selected action based on the samplingprobability and the penalty R t is observed. If the observed penalty changes, the penalty differential σ t is adjusted. Once the learning phase is complete, the probability of the actions in the action set A tk is updated. Subsequent response cycles use the present information to make route selections(i.e., exploitation ). The Dynamic Incentive Control (DIC) module is incorporated to complement the VD moduleand mitigate the churn issue that may prevail in the system. We utilize a top-down optimal controlto adjust the rewards for the events based on the attribute of the cars U t,n for selecting the tasksat every response cycle. Intuitively, if a lot of tasks are being dropped, we may want to increasethe rewards for the particular tasks to encourage more drivers to pick up those tasks.20 .4.1. Top-Down Optimal PID Controller The reward for each task is assigned based on an initial reward r and a reward adjustmentfunction q t,n as expressed below: r t,n = r + q t,n (12)A na¨ıve solution to decide the value of q t,n would be to set it proportional to the numberof dropped tasks after each response cycle t . However, this approach may not be optimal asit would infrequently set rewards and make the system unstable (i.e. the rewards may fluctu-ate uncontrollably if too many tasks get dropped). To address this problem, we incorporate aproportional-integral-derivative (PID) controller, a robust control loop feedback mechanism usedin industrial control systems as well as applications requiring continuously modulated control. ThePID controller nicely maps to our problem of determining the value of q t,n . The process variablein the PID controller is the aggregate reputation of the cars that pick a particular task which isformally defined as: DEFINITION 15. Aggregate Reputation e t,n for task V t,n : The sum of the reputations ofall the cars that selected task V t,n . e t,n = (cid:88) p ∈ U t,n π t,p (13)Intuitively, the higher the value of e t,n , the greater the chance for the cars to make successfulattempts to complete the task.We consider that the framework has a settable parameter called base reputation score e (cid:48) , whichdefines the worst-case aggregate reputation for all the tasks acceptable by the system at anyresponse cycle. If the aggregate reputation falls below this threshold for any task, the system aimsto recover the performance by increasing the rewards assigned for the specific task. On the otherhand, if the score is above the threshold, the system makes a decision to lower the reward for theparticular task, and the surplus could be allocated elsewhere with other tasks. We map the basereputation score e (cid:48) as the set point for the PID controller and the aggregate reputation score e t,n as the measured process variable. Thus, the error for the PID controller is given by: e t,n = e (cid:48) − e t,n (14)The system constantly monitors the number of tasks that are dropped by the cars (i.e. churn).Every time the system observes that the number of dropped tasks exceeds a certain threshold21 , it reruns the algorithm for computing the rewards for the tasks and the utilities for the cars.Otherwise, the algorithm is run periodically at every response cycle for allocating the tasks and tocater to updated event reports.
5. Evaluation
In this section, we evaluate the performance of DASC through a real-world post-disaster casestudy involving road damage scenarios. The evaluation results exhibit significant performancegains of DASC over the compared baselines in terms of both detection effectiveness and deadlinehit rate in verifying the disaster events while considering the road damage.
We acknowledge the fact that the deployment of vehicles in an actual disaster scenario is eitherimpossible or immensely difficult because a real-world disaster is hard to predict and cannot bereproduced. As such, we carry out a real-world data-driven emulation to evaluate our system. Theevaluation platform consists of three key components: 1) the CARLA simulator; 2) a real-worldmapping interface; and 3) the DASC system. CARLA is a widely used car simulator that canclosely imitate physical models of cars traveling in the physical world along with congestion andtraffic signals at intersections [64]. Figure 5 shows a snapshot of our emulation environment.The real-world mapping interface integrates the CARLA simulator with OpenStreetMaps [65]to replicate the real-world map. This enables CARLA to simulate dispatching cars to real-worldlocations (i.e. addresses reported in the social media). The DASC scheme generates the tasksand rewards for dispatching the cars along with their corresponding routes. The DASC schemeconnects with the CARLA simulator using a Python API [64] to send commands for simulatingthe actual car route in the real-world. Figure 5 illustrates a snapshot of the CARLA simulatorinterfaced with the DASC framework.
In order to obtain the optimized values of the parameters λ , λ , λ , K p , K i , K d , ψ , and κ t ,we carry out a parameter tuning process in the first 1 / th of the response cycles. We select theF1 score as the optimization objective as it can give a better measure of the incorrectly classifiedcases with imbalanced distributions [66] such as in our case with social media data. Since the input22arameters cannot be directly modeled on the F1 score (i.e., using a mathematical equation), weincorporate a non-linear optimization [67] approach for obtaining the values of the parameters.We first set the initial values of all the parameters to a maximum value of 1. At each responsecycle during the training, cars are dispatched using the BGT task allocation scheme. We locatethe parameter values that yield the maximum F1 score by using the Nelder-Mead method [68].Specifically, we split the training phase into 3 equal segments and after a series of car observationsare collected for every segment, the values of the parameters that increase the F1 score are retained.We then apply a non-linear optimization on the values of the parameters and assign the final valuesof the parameters. We determined the values of the parameters as: λ = 0 . λ = 0 . λ = 0 . K p = 0 . K i = 0 . K d = 0 . ψ = 0 .
62, and κ t = 0 .
65. For the initial accessibility index X ,h ,our objective is to determine a value that minimizes the difference between the accessibility index X t,h and the actual road damage D t,h . Therefore, instead of the complex Nelder-Mead method, weuse a linear time optimization [69] approach to determine X ,h , which is found to operate optimallyapproximately at the mid-range of X t,h (i.e. X ,h = 0 . Figure 5: CARLA interfaced with DASC. Thetop pane displays the third-person view of asingle vehicle while the bottom pane shows theview of multiple vehicles.
We collected a real-world dataset using Twitterdata feeds posted immediately after the 2017 Hurri-cane Harvey, a hurricane marked as the costliest trop-ical cyclone, causing $125 billion in damage. The hur-ricane originated from a destructive rainfall-triggeredflood in the Houston metropolitan area and SoutheastTexas in August 2017 . To obtain the road damageinformation of the disaster, we analyzed the 2017 Hur-ricane Harvey damage report map published by FEMAand deduced the roads affected by damage during thedisaster [70]. We then replicated the road damage in the CARLA simulator.We obtained the Twitter data using the Apollo data collection tool . For the evaluation http://apollo.cse.nd.edu/index.html Table 2: Data Statistics
Start Date August 27, 2017Time Duration 3 daysLocation Houston, Texas, USANo. of tweets 1,691No. of tweet users 1,446No. of event locations 106
The social sensing component for our framework aswell as the baselines use the Robust Truth Discovery(RTD) [11] algorithm for deciding whether an eventoccurs or not. We replay the obtained data trace toemulate the disaster event. We sort all the reportedevents based on their timestamps and distribute themacross different response cycles. For our particular experiment, we selected the duration of eachresponse cycle to be 100 minutes based on the frequency of the events observed in our dataset.There are a total of 36 response cycles. Within each response cycle, a set of data preprocessing stepsare performed. In particular, we extract the relevant tweets by first running keyword searches (e.g.,gas, fuel, oil, medicine, healthcare, and pharmacy) and discard the irrelevant ones. We then clustersimilar tweets into the same groups using the state-of-the-art online tweet clustering tool [11] andobtain claims that report events at particular locations. Also, we only keep the tweets that havevalid geo-location tags for our experiments.
We compare the performance of DASC with a few representative baselines. We first acknowledgethe fact that we have not come across any solution that guides vehicles for sensing using social mediasignals and simultaneously incorporates the dynamics of the social media (i.e., evolving numberof events and social media users), the physical world (i.e., the road damage and the deadline ofthe events), and the rationale behavior of the car drivers (i.e., churn). Therefore, we includedfive established VSN-based event discovery schemes from current literature. Since the schemesdo not incorporate any social sensing component, we included our previous SocialCar frameworkas a baseline as well a simplified version of the DASC framework called “DASC w/o MDP” todemonstrate the impact of the social signals. • Random Allocation : tasks are allotted randomly to cars on the roads. Once cars comeacross an event, they record and report it. Commercial crowdsourcing platforms like Waze2424] use this technique. • Fixed Route : a fixed number of dedicated cars traverse along designated patrol routes. Apatrol route is designed by covering the maximum number of sensing cells using a HamiltonianCycle-based approach [72]. • Shortest Distance Based : cars that are in closer proximity to event locations are prior-itized first with the assumption that they have higher chances of reaching the destinationsfaster [73]. • Reputation Based : tasks with the shortest deadlines are assigned to cars with the highestreputation first [74]. • Incentive Based : tasks with the shortest deadline get the highest rewards in the taskallocation process [48]. • SocialCar : a simplified version of the DASC scheme that assigns tasks to cars solely basedon the social media reports and the reputation of cars, and adjusts the incentives based onour previous work [21]. The SocialCar scheme does not consider the road damage along anyroute. • DASC w/o MDP : a simplified version of the DASC scheme without the MDP component.Once road damage is discovered, the cars are na¨ıvely assigned the routes with the highestaccessibility indices.
We conduct four different sets of experiments to extensively assess the performance of all theschemes using the real-world dataset. We considered three types of drivers in our evaluation: i)drivers who accept tasks and attempt to successfully complete the tasks; ii) drivers who accepttasks and randomly abort midway; iii) drivers who are unwilling to participate in the sensingapplication. We maintain an equal proportion of cars across all three categories in the first threesets of experiments to observe the impact of other variables. However, in the fourth experiment,we analyze the effect of varying the ratios of the three types of drivers on the performance of allschemes. 25 .5.1. Detection Effectiveness
In the first set of experiments, we assess the performance of all schemes across the entiredataset. The detection effectiveness is evaluated using common metrics for binary classification:
Accuracy , Precision , Recall , and
F1-Score . We utilized a set of 90 cars in our system. The resultsare presented in Table 3. We discover that DASC outperforms the other schemes in identifyingthe truthful events (i.e., gas station and pharmacy availability) in the aftermath of HurricaneHarvey. In terms of classification accuracy, precision, recall, and F1 score, the performance gainsachieved by DASC compared to the best-performing baseline (i.e., the
SocialCar scheme) are4.4%, 13.4%, 1%, and 9.5%, respectively. Such increased performance highlights the importanceof incorporating the top-down incentive control in the task allocation process. Since the rewardsare adjusted dynamically based on the reputation of the cars that select the events, the systemmaximizes the possibility of completing the tasks. In addition, we observe that DASC outperformsother baselines by a fairly large margin. We accredit this performance gain to the design of DASCthat explicitly considers the road damages along the cars’ routes and seamlessly integrates thesocial sensing and vehicular sensing system.
Table 3: Overall Performance with Hurricane Harvey Dataset
Algorithm Accuracy Precision Recall F1-ScoreRandom Allocation 0.111 0.131 0.323 0.186Fixed Route 0.239 0.267 0.517 0.352Shortest Distance 0.375 0.483 0.581 0.528Reputation Based 0.343 0.423 0.582 0.490Incentive Based 0.338 0.445 0.568 0.498SocialCar 0.613 0.527 0.818 0.641DASC w/o MDP 0.430 0.502 0.668 0.573
DASC 0.657 0.661 0.828 0.736
We further split the dataset across two different categories: i) gas station availability and ii)pharmacy availability in the Houston region for a more fine-grained assessment of all comparedschemes. Table 4 shows the results. We observe that DASC continues to outperform all baselinesacross the split dataset. 26 able 4: Performance with Hurricane Harvey Dataset Across Different Categories
Gas Station Availability Pharmacy AvailabilityAlgorithm Accuracy Precision Recall F1-Score Accuracy Precision Recall F1-ScoreRandom Allocation 0.126 0.148 0.334 0.205 0.093 0.110 0.303 0.162Fixed Route 0.244 0.272 0.519 0.357 0.234 0.261 0.513 0.346Shortest Distance 0.402 0.507 0.598 0.549 0.352 0.463 0.569 0.511Reputation Based 0.362 0.437 0.592 0.503 0.327 0.413 0.576 0.481Incentive Based 0.355 0.476 0.566 0.517 0.315 0.407 0.561 0.472SocialCar 0.625 0.540 0.839 0.657 0.606 0.519 0.803 0.630DASC w/o MDP 0.435 0.499 0.685 0.577 0.415 0.494 0.635 0.555
DASC 0.678 0.689 0.839 0.757 0.628 0.626 0.808 0.706
In the second set of experiments, we investigate the effect of the number of cars on the per-formance of all the tested schemes. For this assessment, we varied the number of cars acrossall the schemes within the City of Pasadena in the Houston Metropolitan area from our dataset.Figures 6, 7, 8, and 9 show the results for accuracy, precision, recall, and F1 scores, respectivelyfor all the compared schemes. We start with 10 cars and scale up gradually in increments of 10cars for each round. We observe that the benefit obtained by increasing the number of cars startsto slowly plateau when the total number of cars reaches 100. We investigated this phenomenonand found two possible causes for this. Firstly, our dataset encompasses the relatively small cityof Pasadena in Houston, Texas which is about 44.52 sq. mi. in size [75]. Secondly, due to thehurricane, most road networks were rendered unusable leaving only a limited number of availableroutes for the cars. Given the small region and constrained road networks, increasing the numberof cars for the sensing would not necessarily increase the performance. We observe that despitethis, DASC manages to outperform all the baselines when changing the number of cars. This isbecause when churn issue occurs (i.e., cars drop tasks), the proposed reputation-based incentiveadjustment, jointly with the deadline-aware bottom-up task allocation, allows the DASC to ensurethe maximum number of events been covered in time. This eventually leads to better performanceof DASC compared to other baseline schemes when the total number of cars are varied across the27egion. In addition to this, the damage-aware route allocation in DASC avoids routes that mayhave higher chances of road damage, thereby maintaining consistent performance.
Figure 6: Accuracy vs. Number of CarsFigure 7: Precision vs. Number of CarsFigure 8: Recall vs. Number of Cars
In the third set of experiments, we assess the deadline hit rate of all the compared schemeswhile varying the number of cars. Figure 10 shows the results. We observe that DASC achievesthe highest deadline hit rate when the number of cars changes. Compared to the best-performingbaseline, the SocialCar scheme, DASC achieves a 5.87% better deadline hit rate with 90 cars. This28 igure 9: F1 Score vs. Number of Cars is accredited to the combined effort of the bottom-up game-theoretic task allocation, the top-downincentive control, and the damage-aware route selection of the DASC system. At every responsecycle, tasks having tighter deadlines are prioritized.
Figure 10: Deadline Hit Rate vs. Number of Cars
In the fourth set of experiments, we study the churn issue of the cars in the system andevaluate the robustness of the schemes against different behavior of the drivers. In our study,we found that it is very difficult to apply a concrete mathematical model to explicitly model thebehavior of rational car drivers. The actions of car drivers are often unpredictable in real-worldapplications [76]. In addition to that, during our findings, we did not come across any existingpublicly available dataset that summarizes the behavior of the drivers who participate in roadsidesensing. As such, it is impossible to predetermine what action each driver would take [77]. Thus,we design a set of simulation experiments to study the effect of driver behavior on the performanceof DASC where we separately vary the proportion of car driver: i) who attempt to successfullycomplete the given tasks, ii) who abort tasks midway during exploration, and iii) who are unwillingto participate in the first place. 29or each experiment, at any given time we vary the proportion of cars across one categoryand equally distribute the rest of the cars across the other two categories. Figures 11-13 show thedeadline hit rate while varying the proportions of car drivers that i) successfully complete tasks,ii) randomly drop tasks, and ii) are unwilling to participate, respectively. Likewise, Figures 14-16 illustrate the corresponding accuracy, Figures 17-19 illustrates the corresponding precision,Figures 20-22 illustrate the corresponding recall, and Figures 23-25 illustrates the correspondingF1 scores while varying the three types of drivers. We note that DASC achieves the highestdeadline hit rate, accuracy, precision, recall, and F1 scores when the combination of cars acrossdifferent categories changes. This improvement is primarily attributed to the dynamic incentivecontrol (DIC) module that helps to adjust the rewards considering the aggregate reputation of thecars. If the aggregate reputation for a task falls, the module increases the reward which encouragesthe cars to select the task. This implicitly ensures the selection of the maximum number of taskswithin their given deadlines. The performance gain is also imputed to the road damage discovery(RDD) module that sends out the scout cars for locating the road damage, which in turn helps tomake better routing decisions by the vehicle dispatch (VD) module.
Figure 11: Deadline Hit Rate vs.Proportion of car drivers that suc-cessfully complete tasks Figure 12: Deadline Hit Rate vs.Proportion of car drivers that ran-domly drop tasks Figure 13: Deadline Hit Rate vs.Proportion of car drivers that areunwilling to participate
6. Discussion
DASC is designed to only operate in post-disaster scenarios but not during the disaster itselfand in safe environments where at least some road networks are operational allowing cars to beable to traverse safely along them. The priority of DASC is to ensure the safety of the participatingcar drivers and thus DASC is not intended to operate in life-threatening environments where thelives of car drivers are at risk (e.g., during a nuclear explosion).30 igure 14: Accuracy vs. Propor-tion of car drivers that successfullycomplete tasks Figure 15: Accuracy vs. Propor-tion of car drivers that randomlydrop tasks Figure 16: Accuracy vs. Propor-tion of car drivers that are unwill-ing to participateFigure 17: Precision vs. Propor-tion of car drivers that successfullycomplete tasks Figure 18: Precision vs. Propor-tion of car drivers that randomlydrop tasks Figure 19: Precision vs. Propor-tion of car drivers that are unwill-ing to participateFigure 20: Recall vs. Proportion ofcar drivers that successfully com-plete tasks Figure 21: Recall vs. Proportionof car drivers that randomly droptasks Figure 22: Recall vs. Proportionof car drivers that are unwilling toparticipate
The DASC framework is also designed to be a general-purpose sensing response frameworkthat can not only operate in the aftermath of disaster scenarios but could be seamlessly extendedto other sensing applications. For example, the DASC can be applied to smart city applicationssuch as urban noise mapping, detecting air pollution, free parking spot locating, traffic conges-tion detection. Depending on the application scenario, the search criteria for the Tweets can be31 igure 23: F1 Score vs. Propor-tion of car drivers that successfullycomplete tasks Figure 24: F1 Score vs. Proportionof car drivers that randomly droptasks Figure 25: F1 Score vs. Proportionof car drivers that are unwilling toparticipate modified. For example, in an urban noise mapping application, the SSD module can be adaptedto look for events indicating “noisy streets” instead of “pharmacy availability”. Furthermore, theDIC module’s sensitivity to the dropping of tasks can be adjusted by tuning the PID constants.For example, in a free parking spot locating application, the rewards adjustment can be madeless sensitive. Likewise, the Utility Function on the VD module can include additional factors forassigning the tasks to the cars. For example, the speed of the cars can be incorporated in theUtility Function in a traffic congestion detection application.The flexibility of DASC allows it to be extended to incorporate information from multiplesources beyond cars. Based on the application scenario, the cars can be complemented with otherinformation sources such as weather databases, official news agencies, field agents (i.e., rescuers,firefighters), crowdsourcing, etc. The DASC framework could be enhanced by incorporating theadditional sensing information into the Vehicle Dispatch (VD) Module. Specifically, the UtilityFunction in the VD module can be modified accordingly to use the additional sensing data as atask allocation factor to lower dependencies on the car. For example, if there is an additionalinput signal from official news agencies, Equation 7 can be updated with a fourth factor indicatingthe assertion of the news agency along with its weight (e.g., λ ). The news agency can be givengreater credibility than the event veracity score from social media posts. Moreover, depending onthe scenario, the DASC framework could be able to conclude the actual truth of certain eventswithout dispatching cars, thereby saving resources that can be utilized elsewhere to cover otherevents.We acknowledge that it is likely that the network connectivity could be lost during a disaster.The basis for DASC’s successful operation is to have network connectivity between the car drivers32nd the framework’s backend so that they may establish communication. Therefore, DASC isintended to be only operated post-disaster scenarios and safe environments where some form ofnetwork connectivity is present. It is reasonable to assume that after a disaster at least some degreeof cellular or Wi-Fi connectivity could be available depending on the type of the disaster [78].The DASC framework can also be enhanced to handle the temporal loss of connectivity (i.e.,intermittent) and sparse connectivity. For example, in applications where an intermittent networkloss may occur, a layered vehicular delay-tolerant network (DTN) can be implemented throughDTN gateways on proximal vehicles (i.e., nearby cars) to provide persistent storage for storingsensing data [79]. Moreover, in applications involving spotty connectivity, a multi-hop wirelessmesh network could be established between the vehicles to strengthen areas with weaker signalstrength [80].
7. Conclusion
In this paper, we develop the DASC scheme for a road damage-aware social-media-driven carsensing framework in reliable SCS applications. DASC addresses three intrinsic challenges inintegrating the social media with cars: i) utilizing unreliable social signals to drive cars to reportedevent locations; ii) mitigating the adverse effect of the churn introduced by the rational car drivers;and iii) handling the road-damages caused by disasters to optimally guide cars to destinations.The results from a rigorous evaluation with a real-world disaster recovery case study reveal thatthe DASC achieves remarkable performance gains over the state-of-the-art VSNs-based sensingsystems. We envision the outcomes of this paper to pave the road for a novel road damage-awaresocial-media-driven car sensing system to expedite the recovery phases of unexpected calamities.
Acknowledgment
This research is supported in part by the National Science Foundation under Grant No. CNS-1845639, CNS-1831669, Army Research Office under Grant W911NF-17-1-0409. The views andconclusions contained in this document are those of the authors and should not be interpreted asrepresenting the official policies, either expressed or implied, of the Army Research Office or theU.S. Government. The U.S. Government is authorized to reproduce and distribute reprints forGovernment purposes notwithstanding any copyright notation here on.33 eferences [1] C. Zhang, R. Lu, X. Lin, P.-H. Ho, X. Shen, An efficient identity-based batch verification scheme for vehicularsensor networks, in: IEEE INFOCOM 2008-The 27th Conference on Computer Communications, IEEE, 2008,pp. 246–250 (2008).[2] S. Park, J. Kim, R. Mizouni, U. Lee, Motives and concerns of dashcam video sharing, in: Proceedings of the2016 CHI Conference on Human Factors in Computing Systems, ACM, 2016, pp. 4758–4769 (2016).[3] D. Wang, B. K. Szymanski, T. Abdelzaher, H. Ji, L. Kaplan, The age of social sensing, Computer 52 (1) (2019)36–45 (2019).[4] D. Zhang, Y. Ma, Y. Zhang, S. Lin, X. S. Hu, D. Wang, A real-time and non-cooperative task allocationframework for social sensing applications in edge computing systems, in: 2018 IEEE Real-Time and EmbeddedTechnology and Applications Symposium (RTAS), IEEE, 2018, pp. 316–326 (2018).[5] A. Noulas, S. Scellato, R. Lambiotte, M. Pontil, C. Mascolo, A tale of many cities: universal patterns in humanurban mobility, PloS one 7 (5) (2012) e37027 (2012).[6] D. Y. Zhang, D. Wang, Y. Zhang, Constraint-aware dynamic truth discovery in big data social media sensing,in: Big Data (Big Data), 2017 IEEE International Conference on, IEEE, 2017, pp. 57–66 (2017).[7] D. Wang, T. Abdelzaher, L. Kaplan, Social sensing: building reliable systems on unreliable data, MorganKaufmann, 2015 (2015).[8] M. Lichtveld, Disasters through the lens of disparities: elevate community resilience as an essential public healthservice (2018).[9] D. Zhang, Y. Zhang, Q. Li, T. Plummer, D. Wang, Crowdlearn: A crowd-ai hybrid system for deep learning-based damage assessment applications, in: 2019 IEEE 39th International Conference on Distributed ComputingSystems (ICDCS), IEEE, 2019, pp. 1221–1232 (2019).[10] D. Wang, D. Zhang, Y. Zhang, M. T. Rashid, L. Shang, N. Wei, Social edge intelligence: Integrating humanand artificial intelligence at the edge, in: 2019 IEEE First International Conference on Cognitive MachineIntelligence (CogMI), IEEE, 2019, pp. 194–201 (2019).[11] D. Y. Zhang, R. Han, D. Wang, C. Huang, On robust truth discovery in sparse social media sensing, in: 2016IEEE International Conference on Big Data (Big Data), IEEE, 2016, pp. 1076–1081 (2016).[12] S. Rhea, D. Geels, T. Roscoe, J. Kubiatowicz, et al., Handling churn in a dht, in: Proceedings of the USENIXAnnual Technical Conference, Vol. 6, Boston, MA, USA, 2004, pp. 127–140 (2004).[13] N. Vance, M. T. Rashid, D. Zhang, D. Wang, Towards reliability in online high-churn edge computing: Adeviceless pipelining approach, in: 2019 IEEE International Conference on Smart Computing (SMARTCOMP),IEEE, 2019, pp. 301–308 (2019).[14] D. Zhang, T. Rashid, X. Li, N. Vance, D. Wang, Heteroedge: taming the heterogeneity of edge computingsystem in social sensing, in: Proceedings of the International Conference on Internet of Things Design andImplementation, 2019, pp. 37–48 (2019).[15] P. Godfrey, S. Shenker, I. Stoica, Minimizing churn in distributed systems, Vol. 36, ACM, 2006 (2006).[16] A. Haeberlen, A. Mislove, A. Post, P. Druschel, Fallacies in evaluating decentralized systems., in: IPTPS, 2006, p. 1–6 (2006).[17] Q. Zhang, L. Cheng, R. Boutaba, Cloud computing: state-of-the-art and research challenges, Journal of internetservices and applications 1 (1) (2010) 7–18 (2010).[18] F. Bono, E. Guti´errez, A network-based analysis of the impact of structural damage on urban accessibilityfollowing a disaster: the case of the seismically damaged port au prince and carrefour urban road networks,Journal of Transport Geography 19 (6) (2011) 1443–1455 (2011).[19] B. M. Baker, M. Ayechew, A genetic algorithm for the vehicle routing problem, Computers & OperationsResearch 30 (5) (2003) 787–800 (2003).[20] C. Prins, A simple and effective evolutionary algorithm for the vehicle routing problem, Computers & OperationsResearch 31 (12) (2004) 1985–2002 (2004).[21] M. T. Rashid, D. Zhang, D. Wang, Socialcar: A task allocation framework for social media driven vehicularnetwork sensing systems, in: The 15th International Conference on Mobile Ad-hoc and Sensor Networks (MSN),IEEE, 2019, pp. 125–130 (2019).[22] M. Nekovee, Sensor networks on the road: the promises and challenges of vehicular ad hoc networks and grids,in: Workshop on ubiquitous computing and e-Research, Vol. 47, 2005, pp. 1–6 (2005).[23] U. Lee, B. Zhou, M. Gerla, E. Magistretti, P. Bellavista, A. Corradi, Mobeyes: smart mobs for urban monitoringwith a vehicular sensor network, IEEE Wireless Communications 13 (5) (2006).[24] M. Galeso, Waze: An Easy Guide to the Best Features, Lulu Press, Inc, 2016 (2016).[25] Y. F. Dong, S. Kanhere, C. T. Chou, N. Bulusu, Automatic collection of fuel prices from a network of mobilecameras, in: DCoSS 2008, Springer, 2008, pp. 140–156 (2008).[26] D. Wang, L. Kaplan, H. Le, T. Abdelzaher, On truth discovery in social sensing: A maximum likelihoodestimation approach, in: Proc. ACM/IEEE 11th Int Information Processing in Sensor Networks (IPSN) Conf,2012, pp. 233–244 (Apr. 2012). doi:10.1109/IPSN.2012.6920960 .[27] D. Wang, M. T. Amin, S. Li, T. Abdelzaher, L. Kaplan, S. Gu, C. Pan, H. Liu, C. C. Aggarwal, R. Ganti,Using humans as sensors: an estimation-theoretic perspective, in: Information Processing in Sensor Networks,IPSN-14 Proceedings of the 13th International Symposium on, IEEE, 2014, pp. 35–46 (2014).[28] Y. Zhang, X. Dong, L. Shang, D. Zhang, D. Wang, A multi-modal graph neural network approach to traffic riskforecasting in smart urban sensing, in: International Conference on Sensing, Communication, and Networking(SECON), IEEE, 2020, p. to appear (2020).[29] M. T. Al Amin, T. Abdelzaher, D. Wang, B. Szymanski, Crowd-sensing with polarized sources, in: 2014 IEEEInternational Conference on Distributed Computing in Sensor Systems, IEEE, 2014, pp. 67–74 (2014).[30] J. Marshall, D. Wang, Mood-sensitive truth discovery for reliable recommendation systems in social sensing, in:Proceedings of International Conference on Recommender Systems (Recsys), ACM, 2016 (2016).[31] D. Wang, T. Abdelzaher, L. Kaplan, C. C. Aggarwal, Recursive fact-finding: A streaming approach to truth es-timation in crowdsourcing applications, in: 2013 IEEE 33rd International Conference on Distributed ComputingSystems, IEEE, 2013, pp. 530–539 (2013).[32] D. Zhang, Y. Zhang, Q. Li, D. Wang, Sparse user check-in venue prediction by exploring latent decision contextsfrom location-based social networks, IEEE Transactions on Big Data (2019).
33] D. Y. Zhang, D. Wang, H. Zheng, X. Mu, Q. Li, Y. Zhang, Large-scale point-of-interest category predictionusing natural language processing models, in: 2017 IEEE International Conference on Big Data (Big Data),IEEE, 2017, pp. 1027–1032 (2017).[34] D. Zhang, N. Vance, Y. Zhang, M. T. Rashid, D. Wang, Edgebatch: Towards ai-empowered optimal taskbatching in intelligent edge systems, in: 2019 IEEE Real-Time Systems Symposium (RTSS), IEEE, 2019, pp.366–379 (2019).[35] D. Wang, L. Kaplan, T. Abdelzaher, C. C. Aggarwal, On credibility estimation tradeoffs in assured socialsensing, IEEE Journal on Selected Areas in Communications 31 (6) (2013) 1026–1037 (2013).[36] D. Wang, L. Kaplan, T. Abdelzaher, C. C. Aggarwal, On scalability and robustness limitations of real andasymptotic confidence bounds in social sensing, in: 2012 9th Annual IEEE Communications Society Conferenceon Sensor, Mesh and Ad Hoc Communications and Networks (SECON), IEEE, 2012, pp. 506–514 (2012).[37] Y. Zhang, H. Wang, D. Zhang, D. Wang, Deeprisk: A deep transfer learning approach to migratable trafficrisk estimation in intelligent transportation using social sensing, in: 2019 15th International Conference onDistributed Computing in Sensor Systems (DCOSS), IEEE, 2019, pp. 123–130 (2019).[38] N. Vance, D. Y. Zhang, Y. Zhang, D. Wang, Privacy-aware edge computing in social sensing applications usingring signatures, in: 2018 IEEE 24th International Conference on Parallel and Distributed Systems (ICPADS),IEEE, 2018, pp. 755–762 (2018).[39] D. Y. Zhang, D. Wang, An integrated top-down and bottom-up task allocation approach in social sensing basededge computing systems, in: IEEE INFOCOM 2019-IEEE Conference on Computer Communications, IEEE,2019, pp. 766–774 (2019).[40] Z. Xu, H. Zhang, V. Sugumaran, K.-K. R. Choo, L. Mei, Y. Zhu, Participatory sensing-based semantic andspatial analysis of urban emergency events using mobile social media, EURASIP Journal on Wireless Commu-nications and Networking 2016 (1) (2016) 44 (2016).[41] P.-T. Chen, F. Chen, Z. Qian, Road traffic congestion monitoring in social media with hinge-loss markov randomfields, in: 2014 IEEE International Conference on Data Mining, IEEE, 2014, pp. 80–89 (2014).[42] M. Imran, C. Castillo, J. Lucas, P. Meier, S. Vieweg, Aidr: Artificial intelligence for disaster response, in:Proceedings of the 23rd International Conference on World Wide Web, ACM, 2014, pp. 159–162 (2014).[43] D. Wang, M. T. Al Amin, T. Abdelzaher, D. Roth, C. R. Voss, L. M. Kaplan, S. Tratz, J. Laoudi, D. Briesch,Provenance-assisted classification in social networks, IEEE Journal of Selected Topics in Signal Processing 8 (4)(2014) 624–637 (2014).[44] D. Wang, T. Abdelzaher, L. Kaplan, Surrogate mobile sensing, IEEE Communications Magazine 52 (8) (2014)36–41 (2014).[45] D. Zhang, N. Vance, D. Wang, When social sensing meets edge computing: Vision and challenges, in: 2019 28thInternational Conference on Computer Communication and Networks (ICCCN), IEEE, 2019, pp. 1–9 (2019).[46] M. T. Rashid, D. Zhang, Z. Liu, H. Lin, D. Wang, Collabdrone: A collaborative spatiotemporal-aware dronesensing system driven by social sensing signals, in: 2019 28th International Conference on Computer Commu-nication and Networks (ICCCN), IEEE, 2019, pp. 1–9 (2019).[47] M. T. Rashid, D. Y. Zhang, L. Shang, D. Wang, Sead: Towards a social-media-driven energy-aware drone sensing ramework, in: 2019 IEEE 25th International Conference on Parallel and Distributed Systems (ICPADS), IEEE,2019, pp. 647–654 (2019).[48] H. Gao, C. H. Liu, W. Wang, J. Zhao, Z. Song, X. Su, J. Crowcroft, K. K. Leung, A survey of incentivemechanisms for participatory sensing, IEEE Communications Surveys & Tutorials 17 (2) (2015) 918–943 (2015).[49] C.-F. Hsueh, H.-K. Chen, H.-W. Chou, Dynamic vehicle routing for relief logistics in natural disasters, in:Vehicle routing problem, IntechOpen, 2008, pp. 71–84 (2008).[50] S. A. Korkmaz, M. Poyraz, Path planning for rescue vehicles via segmented satellite disaster images and gpsroad map, in: 2016 CISP-BMEI, IEEE, 2016, pp. 145–150 (2016).[51] A. Mahmoudabadi, S. M. Seyedhosseini, Solving hazmat routing problem in chaotic damage severity networkunder emergency environment, Transport policy 36 (2014) 34–45 (2014).[52] H.-B. Kuntze, C. W. Frey, I. Tchouchenkov, B. Staehle, E. Rome, K. Pfeiffer, A. Wenzel, J. W¨ollenstein,Seneka-sensor network with mobile robots for disaster management, in: 2012 IEEE Conference on Technologiesfor Homeland Security (HST), IEEE, 2012, pp. 406–410 (2012).[53] N. Singh, N. Roy, A. Gangopadhyay, Analyzing the sentiment of crowd for improving the emergency responseservices, in: 2018 IEEE International Conference on Smart Computing (SMARTCOMP), IEEE, 2018, pp. 1–8(2018).[54] Y. Zhang, X. Dong, M. T. Rashid, L. Shang, J. Han, D. Zhang, D. Wang, Pqa-cnn: Towards perceptual qualityassured single-image super-resolution in remote sensing, in: The 17th Annual IEEE Communications SocietyConference on Sensor, Mesh and Ad Hoc Communications and Networks (SECON 2020), IEEE, 2020 (2020).[55] F. Mokhtari, M. I. Akhlaghi, S. L. Simpson, G. Wu, P. J. Laurienti, Sliding window correlation analysis:Modulating window shape for dynamic brain connectivity in resting state, NeuroImage 189 (2019) 655–666(2019).[56] F. Duchoˇn, A. Babinec, M. Kajan, P. Beˇno, M. Florek, T. Fico, L. Juriˇsica, Path planning with modified a staralgorithm for a mobile robot, Procedia Engineering 96 (2014) 59–69 (2014).[57] M. T. Rashid, D. Zhang, L. Shang, D. Wang, An integrated social media and drone sensing system for reliabledisaster response, in: IEEE INFOCOM 2020-IEEE Conference on Computer Communications, IEEE, 2020(2020).[58] S. Ieong, R. McGrew, E. Nudelman, Y. Shoham, Q. Sun, Fast and compact: A simple class of congestion games,in: AAAI, Vol. 5, 2005, pp. 489–494 (2005).[59] J. C. Doyle, B. A. Francis, A. R. Tannenbaum, Feedback control theory, Courier Corporation, 2013 (2013).[60] M. T. Rashid, Y. Zhang, D. Y. Zhang, D. Wang, Compdrone: Towards integrated computational model andsocial drone based wildfire monitoring, in: 16th International Conference on Distributed Computing in SensorSystems, (DCOSS20), IEEE, 2020, accepted (2020).[61] E. Even-Dar, S. M. Kakade, Y. Mansour, Online markov decision processes, Mathematics of Operations Research34 (3) (2009) 726–736 (2009).[62] R. Geisberger, P. Sanders, D. Schultes, D. Delling, Contraction hierarchies: Faster and simpler hierarchicalrouting in road networks, in: International Workshop on Experimental and Efficient Algorithms, Springer,2008, pp. 319–333 (2008).
63] V. Raykar, P. Agrawal, Sequential crowdsourced labeling as an epsilon-greedy exploration in a markov decisionprocess, in: Artificial intelligence and statistics, 2014, pp. 832–840 (2014).[64] A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, V. Koltun, Carla: An open urban driving simulator, arXivpreprint arXiv:1711.03938 (2017).[65] M. Haklay, P. Weber, Openstreetmap: User-generated street maps, IEEE Pervasive Computing 7 (4) (2008)12–18 (2008).[66] R. Joshi, Accuracy, precision, recall & f1 score: Interpretation of performance measures, Retrieved April 1(2016) 2018 (2016).[67] C. A. Floudas, Nonlinear and mixed-integer optimization: fundamentals and applications, Oxford UniversityPress, 1995 (1995).[68] M. A. Luersen, R. Le Riche, Globalized nelder–mead method for engineering optimization, Computers & struc-tures 82 (23-26) (2004) 2251–2260 (2004).[69] M. Pilanci, M. J. Wainwright, Newton sketch: A near linear-time optimization algorithm with linear-quadraticconvergence, SIAM Journal on Optimization 27 (1) (2017) 205–245 (2017).[70] R. Murphy, Where harveys effects were felt the most in texas, Online: https://apps.texastribune.org/harvey-fema-damage-analysis/ (2017).[71] Multi-state fleet response working group (fwg) report (2017).URL [72] D. Portugal, R. Rocha, Msp algorithm: multi-robot patrolling based on territory allocation using balanced graphpartitioning, in: Proceedings of the 2010 ACM symposium on applied computing, ACM, 2010, pp. 1271–1276(2010).[73] C. E. Palazzi, F. Pezzoni, P. M. Ruiz, Delay-bounded data gathering in urban vehicular sensor networks,Pervasive and Mobile Computing 8 (2) (2012) 180–193 (2012).[74] H. Gong, L. Yu, X. Zhang, Social contribution-based routing protocol for vehicular network with selfish nodes,International Journal of Distributed Sensor Networks 10 (4) (2014) 753024 (2014).[75] Pasadena texas city guide (2020).URL https://kristinamorales.com/guides/pasadena-tx/ [76] G. J. Wilde, Social interaction patterns in driver behavior: An introductory review, Human factors 18 (5) (1976)477–492 (1976).[77] W. Wang, J. Xi, H. Chen, Modeling and recognizing driver behavior based on driving data: A survey, Mathe-matical Problems in Engineering 2014 (2014).[78] M. Casoni, C. A. Grazia, M. Klapez, N. Patriciello, A. Amditis, E. Sdongos, Integration of satellite and lte fordisaster recovery, IEEE Communications Magazine 53 (3) (2015) 47–53 (2015).[79] V. N. Soares, F. Farahmand, J. J. Rodrigues, A layered architecture for vehicular delay-tolerant networks, in:2009 IEEE Symposium on Computers and Communications, IEEE, 2009, pp. 122–127 (2009).[80] R. Draves, J. Padhye, B. Zill, Routing in multi-radio, multi-hop wireless mesh networks, in: Proceedings of the10th annual international conference on Mobile computing and networking, 2004, pp. 114–128 (2004).[76] G. J. Wilde, Social interaction patterns in driver behavior: An introductory review, Human factors 18 (5) (1976)477–492 (1976).[77] W. Wang, J. Xi, H. Chen, Modeling and recognizing driver behavior based on driving data: A survey, Mathe-matical Problems in Engineering 2014 (2014).[78] M. Casoni, C. A. Grazia, M. Klapez, N. Patriciello, A. Amditis, E. Sdongos, Integration of satellite and lte fordisaster recovery, IEEE Communications Magazine 53 (3) (2015) 47–53 (2015).[79] V. N. Soares, F. Farahmand, J. J. Rodrigues, A layered architecture for vehicular delay-tolerant networks, in:2009 IEEE Symposium on Computers and Communications, IEEE, 2009, pp. 122–127 (2009).[80] R. Draves, J. Padhye, B. Zill, Routing in multi-radio, multi-hop wireless mesh networks, in: Proceedings of the10th annual international conference on Mobile computing and networking, 2004, pp. 114–128 (2004).