Resilient Composition of Drone Services for Delivery
Babar Shahzaad, Athman Bouguettaya, Sajib Mistry, Azadeh Ghari Neiat
RResilient Composition of Drone Services for Delivery
Babar Shahzaad a, ∗ , Athman Bouguettaya a , Sajib Mistry b , Azadeh GhariNeiat c a The University of Sydney, Sydney NSW 2000, Australia b Curtin University, Perth WA 6102, Australia c Deakin University, Geelong VIC 3220, Australia
Abstract
We propose a novel resilient drone service composition framework for deliveryin dynamic weather conditions. We use a skyline approach to select an opti-mal set of candidate drone services at the source node in a skyway network.Drone services are initially composed using a novel constraint-aware determin-istic lookahead algorithm using the multi-armed bandit tree exploration. Wepropose a heuristic-based resilient service composition approach that adapts toruntime changes and periodically updates the composition to meet delivery ex-pectations. Experimental results prove the efficiency of the proposed approach.
Keywords:
DaaS, Service selection, Service composition, Adaptive lookahead,Service recomposition, Resilient composition
1. Introduction
Drones have gained great attention for civil applications from both aca-demic and industrial domains [1]. The wide range of applications and servicesoffered by drones show the extensive utilization of drones in various sectors in-cluding search and rescue, real-time monitoring, aerial surveillance, structuralinspection, and delivery of goods [2] [3]. Several large corporations such asAmazon, DHL, and Google have shown a growing interest in using drones for ∗ Corresponding author
Email addresses: [email protected] (Babar Shahzaad), [email protected] (Athman Bouguettaya), [email protected] (Sajib Mistry), [email protected] (Azadeh Ghari Neiat)
Preprint submitted to Elsevier September 22, 2020 a r X i v : . [ c s . N I] S e p ackage delivery [4]. The attractive features of commercial drone delivery are higher efficiency, cost-effectiveness, and higher flexibility compared to terrestrialtransportation [5].The service paradigm [6] provides powerful mechanisms to abstract the func-tional and non-functional or Quality of Service (QoS) properties of a drone as
Drone-as-a-Service (DaaS) [7]. The functional property of a DaaS describes thedelivery of a package from a given source to a destination following a skywaynetwork. The non-functional properties of a DaaS are battery capacity, flightrange, payload, and speed. Drone delivery services usually operate in a skywaynetwork to avoid no-fly zones and restricted areas. A skyway network is com-posed of skyway segments between any two particular nodes following the droneflying regulations such as visual line-of-sight [8]. The nodes are assumed to bethe delivery targets or recharging stations.The practicality of drone delivery services is limited by a diverse range of intrinsic and extrinsic factors [9]. The intrinsic factors are the inherited drone’slimitations such as limited battery capacity, limited flight range, and constrainedpayload. The extrinsic factors are related to the drone service environment suchas highly dynamic operating environment and constraints on recharging padsat the stations. The maximum flight range of a delivery drone with full payloadweight varies from 3 to 33 km [10]. The battery capacity, speed, payload weight, and weather conditions influence the flight range of a drone [11].To the best of our knowledge, existing research mainly focuses on the schedul-ing and routing of drones by formulating the problem as Travelling SalesmanProblem (TSP) [12] and Vehicle Routing Problem (VRP) [11]. A single dronerouting problem with fuel constraints is studied to minimize the total fuel con-sumption in [13]. The proposed approach is limited to generating routes foronly a single drone with a finite number of stations. Detailed analysis on max-imizing the profitability and minimizing the drone delivery time is presented in[9]. This approach mainly focuses on battery management of a drone deliveryservice. However, existing approaches do not consider recharging constraints and the stochastic nature of drone delivery services. A drone may need multiple2imes of recharging its battery at intermediate stations for persistent deliveryservices in long-distance areas. The arrival of the drone services at a rechargingstation is usually stochastic in nature [14]. Each station has usually a finitenumber of recharging pads. Therefore, the availability of recharging pads maynot be guaranteed.
Our previous work [15] is the first to focus on the recharging constraints ofdrone services using the service paradigm . In our previous work, we proposeda novel DaaS composition framework considering the recharging constraints ofdrone services . In this context, recharging at intermediate stations leads to the composition of DaaS services. The composition provides a means to aggregatethe skyway segment services from source to destination [16]. We formulatedthe problem of constraint-aware DaaS composition as a multi-armed bandit treeexploration problem . We assumed that both the intrinsic and extrinsic factorsare deterministic , i.e., we know a priori about the available drone services, theirQoS properties, and the service environment. Multiple DaaS services instan-tiated by different drones, operating in the same skyway network at the sametime, may cause congestion in the network. We defined congestion as the totalwaiting time require a drone for the availability of recharging pad at a certainstation [10]. To avoid congestion within the network, we proposed a looka-head heuristic-based multi-armed bandit approach to compose drone servicesminimizing the delivery time and cost.However, our previous work does not consider the failures in drone servicesin dynamic weather conditions.
In real-world settings, the drone service envi-ronment is highly dynamic in nature [17]. The QoS properties of drone servicesmay fluctuate due to the changes in the airflow pattern [18]. For example, adrone service may arrive late due to strong headwind or may not find a recharg-ing pad available on a certain recharging station due to recharging constraintsand stochastic arrival of other drone services. As a result, the drone servicemay no longer provide the required QoS and fail. Therefore, the initial deter-ministic composition plan may become non-optimal and need to be replanned to deal with changing weather conditions and recharging constraints.
Failure
3s a natural phenomenon in service composition.
To the best of our knowledge,no prior work has addressed the failure of drone service composition during thedelivery operation.
In this paper, we extend our previous DaaS compositionframework [7] [15] by adapting the failures in DaaS composition. Our objectiveis to propose a resilient DaaS composition framework .We compose the DaaS services and build an initial composition plan usingour deterministic approach. The drone services are required to reach certainintermediate stations at a specific time during the delivery operation. Theposition and time of a drone service are of paramount importance for the smoothexecution of the delivery operation.
Failure in DaaS composition means to failin executing the initial deterministic composition plan.
For example, a droneservice
DaaS needs to reach a recharging station S at 02:30 pm. The smoothexecution of subsequent drone services depends on the current drone serviceand the movement of other drone services. The early or late arrival of DaaS may affect the other drone services and require to change the initial plan. Theearly arrival of a drone service at an intermediate recharging station does notnecessarily mean to support the initial composition plan. This early arrival mayresult in long waiting time for the availability of recharging pad. Failure to meetconstraints of a composite plan may result in the failure of partial or completecomposite drone service.We propose a resilient composition of drone services for delivery consideringthe recharging constraints and uncertain weather conditions . In this context,resilient means that DaaS composition eventually delivers the package to thedestination by adapting failures in the initial deterministic composition plan.The recharging time, weather conditions, and arrival (or departure) of one droneinfluence the execution plan of other drones at each station. We assume thatthe available drone services are initially deterministic, i.e., there is a knowl-edge about the availability of drone services and their QoS values a priori. Thereal-time delivery operation transforms the deterministic drone services to dy-namic and stochastic drone services. The service environment is dynamic andthe availability of recharging pads may not be guaranteed. We analyze the local4mpact of a failed drone service. We then locally recompose the initial com-position plan using a novel adaptive lookahead heuristic-based approach. Ourproposed approach finds the best composition plan from the current position tothe next intermediate station where no change to the initial plan has occurred.This process continues until the delivery of the package to the destination.The main contributions of this paper are as follows: • A formal model to represent constraint-aware DaaS services. • A Skyline approach for DaaS selection in delivery. • A resilient drone service composition approach considering recharging con-straints and uncertain weather conditions. • A new heuristic-based local service recomposition algorithm using adap-tive lookahead approach. • A custom drone simulation model for simulating the experiments. • An evaluation using a real-world dataset to show the efficiency and effec-tiveness of the proposed model.
Motivating Scenario
We use a typical drone delivery scenario as our motivating scenario. Dronesdeliver the packages within Sydney, Australia. Suppose a drone delivery serviceprovider company is planning to deliver a package from
Richmond to Cronulla (89 km). The maximum service distance of a typical delivery drone ranges from3 to 33 km. The payload weight and wind speed also affect the flight range of adrone. The Bureau of Meteorology (BoM) provides the real-time informationof wind speed and direction for the Sydney area which helps in determining theflight range of a drone. Multiple times of recharge may be required to servethe delivery request. Avoiding the strong wind areas and the congestion of igure 1: Skyway network for drone-based package delivery considering failures drones at recharging stations is of paramount importance for time-optimal andcost-effective delivery services.We construct a skyway network following the Civil Aviation Safety Authority(CASA) drone flying regulations such as avoiding no-fly zones and restrictedareas. The nodes of the skyway network are the rooftops of high-rise buildingswithin the Sydney area. Each node can be a recharging station or a deliverytarget . Each rooftop has a finite number of recharging pads where a dronecan land and recharge. To avoid compatibility issues and present a realisticscenario, we assume that there is no handover of packages at the intermediatestations, i.e., the same drone delivers the package from source to destination.The stochastic arrival of drone services may cause dynamic congestion at certainnodes, i.e., all recharging pads are occupied. Avoiding the congested nodeswould result in faster delivery services.Suppose Yasir needs a package to be delivered within the shortest periodof time . If we ignore the uncertainties, e.g., wind effect and the congestionat the stations (i.e., busy recharging pads) and assume that the services aredeterministic, the problem would be reduced to finding the shortest path (the smart resilient approach to deal with the effectsof failures in the initial deterministic composition plan. This smart resilientapproach adapts to the failures automatically, handles the effects of failures, andensures the on-time package delivery. The brute-force approach considers all thepossible compositions to find the best composition plan. However, this approachis highly time-consuming as finding all the possible compositions may produceexponential search space. As a result, we consider the local recomposition approach which updates the initial composition plan when a failure occurs in theinitial plan at an intermediate recharging station. We use an adaptive lookaheadheuristic-based approach which performs the local optimizations instead of the replanning from scratch or global optimization , i.e., finding the impact of failurein next couple of nodes (relative to the direction of the destination).
2. Related Work
To the best of our knowledge, there exists no similar resilient drone servicecomposition approach in the literature considering the dynamic weather con-ditions. The proposed framework combines concepts from two separate areas:(1) routing and scheduling of drones and (2) failure detection and recovery incomposite services. In this section, we overview related work in these two areas.
Several studies address the routing and scheduling problems for drone de-livery services. Most of the existing research work focuses on using drones in7ombination with ground vehicles for last-mile delivery. A hybrid frameworkfor ground vehicle and drone was first studied in [19]. They proposed two newapproaches for drone-assisted parcel delivery problem to minimize the total de-livery time. In the first approach, a drone is launched from the ground vehicleto serve a customer while a ground vehicle is serving another customer. Afterserving the customer, the drone meets with the ground vehicle in a rendezvouslocation. In the second approach, the ground vehicle and the drone are sepa-rately operated, i.e., the ground vehicle and drone perform dedicated deliveries.
It is concluded that the speed of a drone is an important consideration in de-termining its flight range [19]. The proposed approach is tested for small-sizedcustomer instances up to 20.
The proposed hybrid approach requires road accessfor ground vehicles to make deliveries, i.e., not suitable for remote areas wherethere is no road infrastructure .A single drone routing problem is examined considering multiple refuellingdepots in [13] where a drone can refuel at any depot. The objective of thisstudy is to minimize the total fuel consumption for visiting all the customers.
It is assumed that drone will never run out of fuel during the journey to acustomer . The problem is modelled using Mixed Integer Linear Programming(MILP) formulation. An approximation algorithm is proposed for solving theproblem. The proposed approach is tested for 6 depots and 25 targets only.
Theproposed model does not consider the temporal logic constraints . The proposedapproach is restricted to generating delivery routes for only a single drone, i.e.,not scalable to be used for multiple drones.Two multi-trip VRPs problem is proposed considering solely drones to per-form deliveries [11]. The objective is to minimize delivery time and operationalcost. They proposed an energy consumption model based on the relationshipbetween battery capacity and payload weight. Simulated annealing (SA) meta-heuristic and MILP solver are used to find sub-optimal solutions for the dronedelivery problem. The service area for drone deliveries is limited because alldrones are restricted to dispatch from and return to a single depot. The actualflight time, drone speed, and uncertain weather conditions are not taken into8ccount in the proposed model.
The proposed approach does not consider thefield recharging which limits the coverage and applicability.
An energy consumption model is presented for automated drone deliveryservices in [20]. They assumed that drones can perform multi-package deliveriesin a predefined service area. The drone fleet size is optimized by analyzing theimpact of payload weight and flight range considering battery capacity. Theyexplore the relationship between four variables (working period, drone speed,demand density of service area, and battery capacity) to minimize the totalcosts of the drone delivery system. The study indicated that the long hours ofoperation would benefit both service providers and customers. They found thatdrone deliveries are more cost-effective in areas with high demand densities.
This study does not consider the dynamic congestion conditions at rechargingstations and uncertain weather conditions .A scheduling model is presented to support persistent drone delivery servicesin [21]. The relationship between the intrinsic factors such as payload and flightrange is considered for the effective use of drone delivery services. Multipleservice stations are assumed to replenish batteries of drones during the deliveryoperation. A MILP formulation is presented to model the problem and solvedusing a heuristic approach. An exact solution through MILP and a heuristicalgorithm are provided. It is assumed that the recharging time at the service sta-tion is constant, which is not realistic in practical applications. The flight timeis assumed as a function of payload weight. In real-world problems, the flighttime depends upon the payload weight, drone speed, and environmental weatherconditions such as wind speed and temperature.
The proposed solutions do nottake into account the extrinsic factors such as dynamic operating environment,recharging constraints at each station, the influence of one drones recharging onother drones, congestion conditions at each station, uncertain weather condi-tions, and failures in drone delivery services.
Hence, a heuristic-based approachis required which incorporates the aforementioned real-world aspects of dronedelivery services. 9 .2. Failure Detection and Recovery in Composite Services
Many research works discuss the problem of failure detection and recoveryin composite services [22, 23, 24, 25, 26, 27]. In [22], a service failure recoveryapproach is presented using subgraph replacement of services containing a failedservice. They first represent the composite services as directed graphs. Theypre-calculate the subgraphs and then rank them to speed up the recovery processat the time of failure. The subgraph calculation is time-expensive as it considersall possible compositions of all the component services. The subgraph of a failedservice is replaced by the best-ranked alternative subgraph. The replacementpatterns simply consider the functional and non-functional differences betweenthe new subgraph and replaced subgraph containing the failed service. Theproposed approach is highly time-consuming and limited to considering onlythe sequential digraphs.A region-based service reconfiguration approach is proposed to repair multi-ple failed services and satisfy the original end-to-end QoS constraints in [25]. Areconfiguration region is composed of one or more failed services. When one ormore services in a service composition fail at runtime, they try to replace onlythose failed services. The proposed approach uses Mixed Integer Programming(MIP) to recompose each region until all regions have a satisfactory composition.Generally, MIP methods are very effective when the size of the problem is small.However, these methods suffer from poor scalability due to the exponential timecomplexity of the applied search algorithms.Yu and Lin [27] proposed two algorithms to solve service failures. The pro-posed algorithms compose offline backup service paths for each component ser-vice. When a component service incurs a failure, the predecessor of the failedservice quickly switches to a backup path to skip the failed service. However, theproposed approach does not consider the QoS in the execution of the compositeservice. Also, the approach presented can only handle a single point of failure.Because of the dynamic nature of services, the availability of the backing upprocesses may not be guaranteed when failure happens.A two-phase approach is proposed for the recovery of failed composite ser-10ices in [26]. The two proposed phases are the offline phase and the onlinephase. In the offline phase, the subgraphs of services are calculated and addedto a composite service registry. The offline phase pre-calculations can quickenthe replacement. The online phase refers to the execution of composite ser-vices. Found subgraphs are ranked according to the semantic description oftheir component services. The online phase comprises forward and backwardapproaches. Forward recovery approach attempts to reach the original goal ofthe composite service by retrying or replacing components and continuing theprocess. If the forward approach fails to accomplish, the backward approach isapplied. The proposed recovery approach does not consider the QoS-awarenesscapabilities and the dynamism of the execution context environment to adaptthe most appropriate recovery strategy.Recomposition is a naive solution to handle the problem of service executiontime failures [28]. However, it is extremely time-consuming which is undesired.A repair approach based on planning graphs is proposed as an alternative torecomposition in [24]. Repair is a form of heuristic and guided partial recompo-sition. Repair is time-efficient compared to recomposition while generates solu-tions of similar quality. The proposed approach is restricted to the compositionof deterministic services with simple composition requirements. The presentedtechnique does not consider the QoS criteria, which simplifies the problem. In[29], the service composition problem is transformed into a non-deterministicplanning problem for creating workflows with contingency plans. The before-hand planning for failures saves execution time. However, the generation of allpossible alternative composition plans is a time-consuming process.A constraint-aware failure recovery approach is proposed to explore the re-liability of service composition in [30]. Existing approaches do not consider theconstraint verification failures in composite services. They predict failures in-side a composite service to reduce the number of service rollbacks upon failurerecovery. The proposed solution includes a planning-based service compositionapproach and a constraint-processing method. The planning-based algorithmconstructs constraint-aware composite service plans. The constraint-processing11ethod proceeds with constraint verification in constructed composite service.The proposed approach is restricted to only a small number of possible solutions,i.e., inefficient for a very large number of plans.An adaptive composition approach is proposed to handle the service changesoccurring at runtime, for both repair and optimisation purposes [17]. The pro-posed approach adapts to changes as soon as possible in parallel to the executionprocess. In this way, the interruption time reduces, the chances of a successfulrecovery increase, and the most optimal solution is produced according to thecurrent state of the environment. The results show that the proposed approachmanages to recover from unexpected situations with minimal interruption, evenwith frequent changes or in the cases where interference with execution is non-preventable.The service paradigm is leveraged to abstract the line segment as a ser-vice (e.g., a bus service) for multi-modal travel purposes in [31, 32]. A servicecomposition framework is proposed for composing spatio-temporal line segmentservices. A novel spatio-temporal A*-based algorithm is proposed to composethe services. It is assumed that the services are deterministic, i.e., time andavailability are unknown in advance. A failure-proof composition approach forSensor-Cloud services is presented in [16] considering the dynamic features suchas position and time. The proposed approach is based on D*Lite algorithm todeal with the changes in QoS of Sensor-Cloud services at runtime.A spatio-temporal service model is proposed for drone services in [7]. Adrone delivery function over a line segment in a skyway network is abstractedas a service. A spatio-temporal service model is also proposed for drone de-livery services. A spatio-temporal service selection and composition algorithmis proposed to compose line segment services considering QoS properties. Thebattery capacities and recharging constraints are not considered in the proposedmodel. A constraint-aware deterministic drone service composition approach isproposed in [15]. The proposed approach considers the recharging constraintsat each station. A skyline approach is presented to select an optimal set ofdrone services. The drone service composition problem is formulated as a multi-12rmed bandit tree exploration problem. A lookahead heuristic-based algorithmis developed to compose the selected services. However, the dynamic serviceenvironment, the uncertain weather conditions, and the failure in drone servicesat runtime are not considered in the proposed approach.
To the best of ourknowledge, this paper is the first attempt to model the influence of rechargingconstraints in a drone service environment and resilient composition of dronedelivery services .
3. Constraint-Aware System Model for Drone Services
We propose a constraint-aware system model for drone delivery services.The proposed model includes four main parts: (1) Skyway Network, (2) DroneServices, (3) Effects of Wind Speed and Direction in DaaS, and (4) Constraint-Aware Model for Drone Delivery Services.
In this section, we describe the structure of the skyway network in whichdrone delivery services operate. Let D = { d , d , . . . , d n } be a set of n dronesand T = { t , t , . . . , t m } be a set of m delivery targets. The skyway network isrepresented as an undirected graph G = ( V, E ), where V is a set of vertices (ornodes) each of which represents a target and E is a set of edges each of whichrepresents a skyway segment service joining any two vertices. We assume thateach vertex is also a recharging station. Each node is assumed to have a finitenumber of recharging pads . B is a set of battery capacities for all the drones.The travelling cost and battery consumed in travelling from node i to j arerepresented by c ij and b ij respectively. The battery consumption of the dronehas a proportional relationship with payload weight, the distance travelled bythe drone, and the wind speed and direction. We formally defined a model for drone services in our previous work [7].The proposed model includes the formal definitions of DaaS, DaaS composite13ervice, and DaaS composition problem as follows.
Definition 1: Drone-as-a-Service DaaS . A DaaS is defined as a deliveryfunction of a drone which takes a package from a pickup location to a deliverylocation (i.e., longitude and latitude) having a start time and an end time andmeeting a set of QoS attributes (e.g., flight range). A DaaS is a 3-tuple
DaaS i ∈ CS , i.e., CSID = concat ( DaaS i .id ) • CSF is a set of functions { f ( DaaS ) , f ( DaaS ) , . . . , f n ( DaaS n ) } , whereeach f i represents the function of corresponding component DaaS DaaS i ∈ CS • CSQ is an m-tuple < Q , Q , . . . , Q m > , where each Q j denotes an ag-gregated value of j th quality parameter of component DaaS DaaS i ∈ CS .14 efinition 3: DaaS Composition Problem . For a given set of DaaS S DaaS = { DaaS , DaaS , ..., DaaS n } services in a skyway network, the DaaScomposition problem is to compose the services for delivering a package from apickup location to a delivery location in minimum time. The wind is a major environmental factor affecting the drone’s performanceand flight behaviour [33]. The wind effect that causes the drone to drift in acertain direction is studied in [34]. They designed a method based on a modifiedaccelerated A* algorithm to take the wind effects into account and generatereachable states. It is assumed that the wind is constant which does not reflectthe real-world scenarios. A deadline-constrained routing scheme is presented fordelivery drones in [35]. The objective of this study is to minimize the energyconsumption under wind conditions.We consider the effects of wind speed and direction in dynamic weatherconditions. Highly random nature of wind speed and direction (i.e., headwindand tailwind) greatly influences the battery consumption rate and flight rangeof the drone [36] [37]. We present a model to determine the impact of windspeed and direction on the travel time of a drone. The travel time of a droneincreases with headwind and reduces with the tailwind. We calculate the effectsof wind speed and direction on travel time using a method in [38] for a dronetravelling from node i to j as follows. δ = θ ij − θ W S (1) A = W S. cos(180 − δ ) (2) C = W S. sin(180 − δ ) (3) B = (cid:112) AS − C (4) GS = A + B = W S. cos(180 − δ ) + (cid:113) AS − W S . sin (180 − δ ) (5) T ij = d ij GS (6)15here, • θ ij = bearing from node i to j • θ W S = wind bearing • δ = course correction angle • W S = wind speed • A = headwind/tailwind. When | δ | < A is negative and de-notes headwind. When 90 < | δ | ≤ A is positive and de-notes tailwind. • C = wind adjustment angle • B = wind adjustment angle • AS = air speed • GS = ground speed • d ij = distance between node i and j • T ij = travel time from node i to j In this section, we first present our previous constraint-aware DaaS com-position model for drone delivery services. The constraint-aware compositionmeans to compose the drone services knowing the availability of recharging padsat intermediate stations and the arrival of other drone services. In our previouswork [15], we assume all the drone services and service environment are deter-ministic, i.e., the QoS attributes of drone services, the availability of rechargingpads, and the trajectory of other drone services are all known beforehand. Ourobjective was to compose the drone services avoiding the congested rechargingstations and delivering the packages in the shortest time. However, such anassumption of the deterministic service environment is not realistic in practice.The QoS may fluctuate and fail due to the dynamic nature of drone services andchanging wind patterns. We relax the assumption of the deterministic serviceenvironment. We consider that the service environment is stochastic and flighttime may vary with the changing wind conditions and the arrival of other droneservices.We compose the drone services to generate an initial service compositionplan using our previous deterministic approach. Different types of drones have16arying payloads, flight ranges, and battery capacities. There is a constraintthat the same drone delivers the package from source to destination. A dronecan either recharge, wait, or travel from one station to the next station . Thedeterministic approach estimates the arrival time, waiting time, and rechargingtime of each drone at a specific recharging station. The initial composition planadapts to the failures dynamically occurred at runtime. Here, failure means thelate or early arrival of drones than the scheduled arrival in the initial plan. Thisfailure may have a cascading effect to the execution of subsequent drone services,thus affecting the initial composition. Therefore, a resilient DaaS compositionframework is required to ensure the on-time delivery of drone services.
4. Drone Service Selection using Skyline Approach
The first step to compose drone services is the selection of appropriate candi-date services. For this purpose, we consider several drone services from multipleservice providers. The QoS properties of drone services distinguish among func-tionally equivalent services. Some of the available drones may not carry thepackage because of its higher weight. Therefore, we use the difference betweenthe payload capacity of the drone and package weight to filter out the candidatedrone services. We use skyline approach [39][40] to further reduce the numberof candidate drone services by selecting only the non-dominated services. Sky-line computation speeds up the service selection process and selects the serviceswith best QoS attributes . Skyline approach is also used to deal with the uncer-tainty of service in the process of selection [41]. A multi-attribute optimizationtechnique, called service skyline computation, guarantees to provide the bestuser-desired service providers [42].For a given set
DaaS = { DaaS , DaaS , . . . , DaaS n } of functionally similardrone services and a set Q = { q , q , . . . , q m } of QoS attributes, we presentformal definitions of drone service domination and service skyline as follows. Definition 4: Drone Service Domination . The domination relationshipbetween a drone service
DaaS i ∈ DaaS and another drone service
DaaS j ∈ able 1: A set of functionally similar Drone services
Drone service Flight time (min) Flight range (km) Recharging time (hours) Is skyline?
DaaS
20 0.8 1.5 No
DaaS
20 56 2 Yes
DaaS
25 8 1 Yes
DaaS
30 7 1.5 No
DaaS
20 1.6 1.5 No
DaaS
18 0.8 1.5 Yes
DaaS
120 100 2 Yes
DaaS
20 3 1 Yes
DaaS
27 7 1 No
DaaS
40 1.9 1.5 No
DaaS
22 5 1.5 Yes
DaaS
24 8 1.5 Yes
DaaS is defined as
DaaS i ≺ DaaS j , if ∀ q k ∈ Q , q k ( DaaS i ) (cid:22) q k ( DaaS j ), and ∃ q l ∈ Q , q l ( DaaS i ) ≺ q l ( DaaS j ) where ≺ denotes better than and (cid:22) denotesbetter than or equal to relationship. Definition 5: Service Skyline . The service skyline comprises a set ofdrone services, denoted by
SKY DS , that are not dominated by any other droneservice, i.e., SKY DS = { DaaS i ∈ DaaS |¬∃
DaaS j ∈ DaaS : DaaS j ≺ DaaS i } .We compute the skyline using the following three QoS properties: (1) flighttime (in minutes) represents the time duration a drone can fly with batterycharged to its capacity, (2) flight range (in kilometres) represents the distancea drone can travel with full capacity charge, and (3) recharging time (in hours)for 0 to 100% recharge. We use Block Nested Loop (BNL) algorithm [39] forskyline computation. The non-dominated skyline services are obtained by repet-itive scanning of the candidate drone services. The
BNL algorithm can be usedfor any dimensionality without requiring any indexing or storage. It performswell most of the time for dealing with our low dimension and small domainrange data. Table 1 presents an example of skyline computation for function-ally similar drone services that are differed in QoS properties. For instance,a drone service
DaaS dominates another drone service DaaS according toaforementioned domination relationship. The “Is skyline?” column illustratesthe outcome of skyline computation. 18 ulti-armed Bandit Tree ExplorationInitial Offline Composition Resilient Online CompositionService Selection using Skyline
Candidate
Drone Services
Failure
DetectionFailure
AnalysisAdaptive Lookahead Tree ExplorationLocal Recomposition
Input
Figure 2:
Resilient Drone Service Composition Framework
5. Resilient Drone Service Composition Framework
We divide the resilient drone service composition framework into two cate-gories: (1) Constraint-Aware Drone Service Composition using Lookahead and(2) Resilient Drone Service Composition using Adaptive Lookahead. Fig 2presents an overview of the resilient drone service composition framework. Theinitial offline composition is provided by constraint-aware drone service com-position in a deterministic fashion. While the resilient online composition iscarried out to handle the dynamic failures in the initial offline composition atruntime.
We formulate the constraint-aware drone service composition as the multi-armed bandit tree [43] exploration problem. In multi-armed bandits, an armdenotes an action or a choice which is initially unknown to the player. If thearms are deterministic, i.e., known beforehand, the problem would be reducedto the selection of arms with the highest reward. We assume that the droneservices and the services environment are initially deterministic. Our target is to19 tate = [1, t ]State = [2, t ] State = [3, t ]State = [4, t ] State = [5, t ] State = [4, t ] State = [5, t ] t = 02:00 pmt = 02:20 pm t = 02:25 pm t = 02:40 pmt = 02:35 pm t = 02:30 pm t = 02:35 pm State = [Node, Time]Source Node = 1
Destination Node = 5
Figure 3:
An example of a state tree maximize the reward by selecting optimal arms. A drone can take the followingset of actions at each station: recharge, wait, or travel from one station to thenext. These actions generate a large set of possible states. Fig. 3 presents anexample of a temporal state tree. We formally define a state as follows:
Definition 5: State . A state is a tuple of < N odeID, T imeStamp > ,where • N odeID is a unique node identifier, • T imeStamp represents the arrival time of drone at a certain node.For the sake of simplicity, we consider that the states are known beforehand.In case of immediate state selection, the temporal optimal neighbour state maylead to a non-optimal state, e.g., long waiting time due to congestion at thenext station.The selection and composition of optimal drone services from a large numberof candidate services is a challenging task. The uncertainty is the main issue ina DaaS composition. In many cases, an immediate optimal service may lead toa non-optimal service. For example, we have a skyway network where node 120 tate = [1, t ]State = [2, t ] State = [3, t ]State = [4, t ] State = [5, t4]
State = [4, t ] State = [5, t ] t = 02:00 pmt = 02:20 pm t = 02:25 pm t = 02:40 pmt = 02:35 pm t = 02:30 pm t = 02:35 pm State = [Node, Time]Source Node = 1
Destination Node = 5
Figure 4:
State selection without lookahead is the source node and node 5 is the destination node. Here we find a temporaloptimal neighbour leading to a non-optimal state. Temporal optimal meanstaking towards destination faster. As shown in Fig. 3, the service of state [2, t ]is optimal but the overall delivery time is more compared to state [3, t ]. Thisuncertainty can cause long delays for drones to deliver packages. Looking forall possible service compositions or deep tree exploration is not computationallyfeasible to find the best composition. The time complexity for such problemsis exponential. Hence, we need a heuristic-based solution to find the optimalcomposition of drone services.We propose a lookahead heuristic-based solution to the multi-armed bandittree exploration problem. The selection of optimal actions in a DaaS compo-sition is performed by looking ahead of neighbour services. We consider thecurrent waiting time, expected waiting time, and flight time to the destinationfor selection of optimal drone services. The term lookahead means consideringthe next-to-adjacent states while making the state selection decision. Fig. 4 and5 illustrate the difference between without lookahead and with one lookaheadbased service (state) selection. Without lookahead considers only the neighbour21 tate = [1, t ]State = [2, t ] State = [3, t ]State = [4, t ] State = [5, t ] State = [4, t ] State = [5, t ] t = 02:00 pmt = 02:20 pm t = 02:25 pm t = 02:40 pmt = 02:35 pm t = 02:30 pm t = 02:35 pm State = [Node, Time]Source Node = 1
Destination Node = 5
Figure 5:
State selection with one lookahead optimal states which leads to an overall non-optimal solution. Using lookaheadheuristic provides more information to select the overall optimal states. Webuild our initial composition plan using the aforementioned lookahead strategy.But, this approach does not take into account the runtime failures in the exe-cution of the initial composition plan such as uncertain weather conditions. Weneed a resilient composition approach for drone services to ensure the in-timepackage delivery.
The underlying initial DaaS composition approach is formulated as a multi-armed bandit problem [15]. Multi-armed bandits are a special type of sequentialdecision problems which demonstrate exploration and exploitation trade-offsand produce maximum rewards under uncertainty [44]. The exploration refersto trying each possible action to find an optimal reward. In contrast, exploita-tion refers to trying the actions that are believed to provide higher payoffs in thefuture. We focus on the constraints at recharging stations and dynamic weatherconditions. However, multi-arm bandits are generally proposed for tree-basedsearch exploration in the context of combinatorial optimization [45]. An exact22pproach such as MILP does not naturally fit to solve such exploratory opti-mization [46]. MILP approaches are usually applied in solving deterministiclinear optimization problems [47]. While heuristic-based lookahead [48], geneticalgorithm [49], and tabu search [50] are usually used for exploratory optimiza-tion problems. We focus on the adaptive lookahead heuristic-based approachwhich is typically used to solve combinatorial multi-armed bandit problems [51].The heuristics are widely used in multi-armed bandit literature and provide sub-stantially more efficient solutions than traditional optimization approaches [52].Therefore, we focus on exploring a heuristic-based solution for the compositionof drone services.The stochastic arrival of other drone services at intermediate stations and thechanges in wind pattern influence the initial composition plan. As a result, theestablished composition plan may become non-optimal and fail. Such failuresmay impact on the initial composition in two ways: (1) local impact (2) globalimpact. The term local impact means the effect of failure propagates to a certainnumber of recharging stations. The rest of the plan is still recoverable. The termglobal impact refers to the propagation of failure effect till the destination.We propose a resilient drone service composition using adaptive lookaheadheuristic-based approach. The resilient means that the delivery operation issuccessfully carried out even the established composition plan adapts to thefailures. We require a lookahead algorithm to handle time-varying constraintsand weather conditions. The adaptive lookahead performs lookahead accord-ing to the type of failure occurred rather than a fixed number of lookaheads.There is only one difference between an adaptive lookahead and a standardlookahead algorithm: the distance in adaptive lookahead is no longer a fixedlength but varies with the propagation effect of the failures. Once the adaptivelookahead finds the distance of failures, we locally recompose the drone servicesfrom the current station to the next failure-free station. The state selectionusing adaptive lookahead approach is shown in Fig. 6. The selective states ofinitial composition plan are represented by green colour. Failure occurs at node3 which requires the recomposition of services. We recompose the services until23 tate = [1, t ]State = [2, t ] State = [3, t ]State = [4, t ] State = [3, t ] State = [4, t ] State = [6, t ] t = 02:00 pmt = 02:20 pm t = 02:25 pm t = 02:40 pmt = 02:35 pm t = 02:45 pm t = 02:50 pmState = [5, t ] State = [6, t ] State = [7, t ] State = [8, t ]t = 02:45 pmt = 02:50 pm t = 02:30 pm t = 02:45 pmState = [7, t ] State = [8, t ]t = 02:45 pm t = 02:55 pm State = [7, t ] State = [8, t ]t = 02:30 pm t = 02:35 pm State = [Node, Time]
Source Node = 1Destination Node = 8
Initial Composition Plan Resilient Composition PlanFailure PointAffected NodeCongested Node
Figure 6:
State selection using adaptive lookahead the next state where no change is observed. Fig. 7 presents an example of initialservice composition, failed service, and its impact on other services. We firstcompute an initial offline composition plan from source to destination, denotedby a sequence of solid line arrows connecting green colour nodes. The initialcomposition plan avoids the congested recharging stations (yellow colour nodes)for faster delivery services. A failed service is represented by a red colour nodeand its impact on the next services in the initial plan is shown by orange colournodes. The formal definitions of failure, service failure, and resilient servicecomposition are given as follows.
Definition 6: Failure . A failure is defined as the deviation from expected(specified) behaviour. In some cases, the failure may result in the terminationof the ability to perform the required function.
Definition 7: Service Failure . Service failure is defined as an event thatoccurs when the delivered service deviates from the correct service. For example,a drone service
DaaS i is specified to reach a station Station j at 04 : 00 pm. If24 aaS Source DaaS DestinationDaaS Composition Congested NodeNode Affected NodeFailed Node Node ConnectionInitial Composition Plan Range of affected drone services by a service failure in initial composition plan
Figure 7:
An example of initial composition plan failure and its impact
DaaS i reaches at Station j before or after 04 : 00 pm, we say that the service isfailed. Definition 8: Resilient Service Composition . Resilience refers to theability or capacity of a system to adapt to dynamic changes (failures) withoutdeviating from the expected behaviour. Resilient service composition is a mech-anism for handling the failures occurred at runtime. When one or more servicesfail at runtime, the resilient service composition approach locally or partiallyrecomposes the failed services to deliver the expected behaviour.Fig 2 illustrates the process of resilient drone service composition. We firstexecute the initial offline composition plan. The failure detection module peri-odically checks for any failures at each station. Each drone service has a certaindeadline for each station defined in the initial offline plan. We compare thecurrent arrival time of a drone service with an expected arrival time given bythe established plan. In the case of the early or late arrival of a drone service,the failure is detected. The failure analysis module finds the number of services25 lgorithm 1
Resilient Drone Service Composition Algorithm procedure Execute Init Plan ( InitComp ) DaaS cur ← InitComp [ start ] DaaS dst ← InitComp [ end ] while DaaS cur (cid:54) = DaaS dst do Execute initial composition plan Monitor the execution to find the failed services f d ← failure detection ( DaaS cur .t e , curT ime ) if f d then DaaS affected ← failure analysis ( InitComp, DaaS cur ) LocalComp ← recompose ( InitComp, DaaS cur , DaaS affected , curT ime ) InitComp ← update plan ( InitComp, LocalComp, DaaS cur ) end if DaaS cur ← InitComp [ next DaaS ] end while end procedure affected due to failure. The failure may affect the execution of a couple of nextdrone services. The adaptive lookahead tree exploration module guarantees theexploration of all possible alternatives to the failed service. Finally, we locallyrecompose the explored alternatives to mitigate the effect of failure. The re-composition of drone services at the intermediate station obtains an optimalcomposite service in minimal computational time.Algorithm 1 provides the details of the proposed approach as follows. Thealgorithm generates a resilient composition of drone services using an initialcomposition plan as input. The first and last component services in the initialplan are the source and destination locations (Lines 2-3). We execute the initialplan and monitor periodically for any failure at runtime (Lines 4-6). The initialplan is executed smoothly until a failure is detected (Line 7). The actual arrivaltime at each station is compared with the expected arrival in the initial plan. Ifa failure is detected, the failure analysis algorithm computes the affected (i.e.,26 lgorithm 2 Failure Analysis procedure failure analysis ( InitComp, DaaS cur ) f ailedDaaS ← Find first congested node
CongN ode from
DaaS cur to the destination if CongN ode then DaaS affected ← compute number of services from DaaS cur to CongN ode end if for each DaaS ∈ InitComp from
DaaS cur do td ← compute time difference between actual and expected DaaS if td ≥ then f ailedDaaS ← f ailedDaaS + 1 else break end if end for return min( DaaS affected , f ailedDaaS ) end procedure failed) drone services (Line 9). Algorithm 2 presents the details of the failureanalysis algorithm. We find the first congested node in the initial plan fromthe failed service until the destination. If a congested node is found, we simplycompute the number of services from the current failed service to the congestednode. Moreover, we find the first unaffected service from the failed service untilthe destination. We calculate the difference between the first failed service andunaffected service in the initial plan. The minimum of distance (i.e., nodes)is selected from the congested node and unaffected service. We consider thecongested node for failure analysis because the delay of service failure resultsin less waiting time at a congested node. For example, a service failure causes15 minutes delay to the initial composition plan. Let’s assume that there is acongestion node in the initial plan ahead of failed service. The waiting time on27 lgorithm 3 Recomposition of Drone Services procedure recompose (( InitComp, DaaS cur , Ld adapt , curT ime )) srcLocal = DaaS cur dstLocal = InitComp [ DaaS cur .index + Ld adapt ] startT ime = curT ime newComp = find optimal comp ( G, RP, D, srcLocal, dstLocal, w,Ld adapt , W S, θ
W S , startT ime ) return newComp end procedure the congested node is 25 minutes for the availability of recharging pad. In sucha case, the waiting time will be reduced to 10 minutes because of 15 minutesdelay from failed service. The adaptive lookahead distance is equivalent to thenumber of affected drone services for exploration of all possible alternatives. Werecompose the services from the failed position to next unaffected drone serviceusing recompose algorithm (Line 10). The details of recompose algorithm aregiven in Algorithm 3. The recompose algorithm composes the services locallyby calling the f ind optimal comp function which is same as our drone serviceselection and composition algorithm in [15]. Finally, the new locally composeddrone services update the inconsistent affected services in our initial compositionplan. This process continues until the package is delivered to the destination.An alternative to the use of local recomposition is to replicate the delay in theinitial service composition till the destination. This alternative approach mayresult in longer delays, and in some cases, the package may not be delivered.
6. Experiments and Results
We evaluate the effectiveness of the proposed resilient drone service compo-sition approach in this section. A set of experiments are conducted to assessthe performance of the proposed approach. We compare the proposed approachwith a baseline (i.e., Brute-Force) approach and a without lookahead approach.28he most important features of the drone delivery services are the shorteningof the delivery time and cost reduction. The delivery cost is a function of dronetravelling distance. Therefore, we mainly focus on three evaluation metrics: (1) delivery time , (2) computation time , and (3) distance travelled . All the experi-ments are conducted on an Intel Core i9-9900X processor (3.50 GHz) with 32.0GB memory under Windows 10. Python is used to implement the algorithms.
Simulation tools offer a faster, cost-effective, and safe approach to assess theperformance of possible solutions before physical testing. There exists severalsimulators for drones, e.g., AirSim [53], Gazebo [54], and JMavSim [55]. Thesesimulators are not specifically designed for drone delivery services over skywaynetworks with recharging stations. For example, AirSim does not model droneenergy consumption in dynamic environments [56]. Energy is a scarce resourcein drones that affects the entire delivery operation. The AirSim platform doesnot implement payload effects on the power consumption of the drone. Thefailures in delivery services are not considered in AirSim. The skyway networkfor drone delivery services is also not a part of the AirSim platform. As thecentre of our paper is the drone-based delivery platform, we implement a customdrone-based delivery simulation model for the experiments. In future, we planto deploy a skyway network and delivery management framework on AirSim forgreater reachability to the research community.We design a custom drone simulation model using tools from drone energyconsumption model [57], weather model [58], operations research, i.e., deliveryservice management [59], and 2D path planning [60]. The simulation modelconsists of the following modules (as shown in Fig. 8): (a) controller , (b) energymodule , (c) flight path module , (d) weather module , (e) request dispatcher , (f) failure detection module , (g) failure recovery module , and (h) skyway network .The controller module ensures the long-term stability of drone delivery services.It keeps track of all types of manoeuvres in the dynamic environment. The con-troller realizes the desired composition objectives by handling more and more29 ontroller
Skyway
Network
Flight Path Module
Weather
Module
Request
Dispatcher
Energy
Module
Compute EnergyInitial Offline
Plan
Weather Data
Failure Detection ModuleFailure Recovery Module
Resilient Online
Plan
Figure 8:
Structure of drone simulation model services at each step. The energy module simulates the energy consumption ofa drone service travelling from one recharging station to the next station. Theenergy consumption is calculated using the method in [57]. An initial flightpath is generated using our existing deterministic offline composition approach.The flight path module contains the composed services and position informa-tion of the drone services operating in the skyway network. The flight path isupdated to maintain the resilience of composite services under dynamic weatherconditions. The changing weather conditions influence the initial compositionplan. The weather module is in charge of generating weather data for the wholesimulation. The request dispatcher module takes care of receiving drone servicerequests from users. The current implementation of drone simulation modeldeals with single package delivery service request. The failure detection modulemonitors the execution of the initial composition plan. If a failure occurs dueto dynamic weather conditions or stochastic arrival of other drone services, itnotifies the failure recovery module. The failure recovery module is responsiblefor the execution of two main actions: (1) estimation of the failure impact and(2) local recomposition of affected drone services. As there is no 3D graphicsinvolved (to simulate 3d drones), we do not require high capacity GPU. The30 able 2:
Dataset Description
Attribute Description Examplevalue
Dronename Represents the manufacturer of the drone DJI M200 V2 Payload Represents the weight a drone can carry (inkilograms) 1.45 kgFlight time Represents the time a drone can fly with fullpayload capacity (in minutes) 24 minRange Represents the distance a drone can coverwith full payload capacity (in kilometres) 32.4 kmSpeed Represents the flying speed of a drone withfull payload capacity (in kilometres perhour) 81 km/hRechargingtime Represents the time required by a drone forrecharging from 0% to 100% (in hours) 2.24 hourssimulation environment and composition algorithms are written in Python.We use NetworkX [60] python library to construct the topology of the sky-way network. We model the multiple delivery drones operating in the sameskyway network. We evaluate the proposed approach using a real drone dataset[61]. The dataset contains the trajectories of drones, which include data for co-ordinates, altitude, and timestamps. We augment a dataset for different typesof drones considering the flight range, payload, battery capacity, speed, andrecharging time. The details of the dataset are given in Table 2. The efficiencyof the proposed framework depends on the values of the environmental vari-ables. Table 3 describes the environmental variables used in the experiments.The number of drones varies from 50-80 for varying sizes of the skyway net- able 3: Experiment Variables
Variable Value
Number of drones [50, 80]Number of nodes (or recharging stations) [10, 60]Number of recharging pads at each station 5Number of DaaS services [500, 2500]Number of generated requests 1500Average battery consumption rate with 1 kg package 25%/10 kmNumber of sources 1 (random)Number of destinations 1 (random)Frequency of failures (% times the total nodes) [10, 50]Experiment run (% times the total nodes) 10work. We assume that each node is a recharging station. The number of nodes(i.e., recharging stations) varies from 10-60 for all approaches. Each rechargingstation has a finite number of recharging pads. The number of skyway segmentDaaS services depends upon the size of the skyway network and the number ofinterconnected nodes. The proposed approach focuses on the single package de-livery services from a given source to a destination. Each experiment starts witha random source and a destination point. The service failures occur randomlyat runtime. The frequency of failures in each experiment varies from 10-50%times the total number of nodes. The effect of each failure varies from a coupleof subsequent nodes to the destination node. We conducted the experiments for10% times the total number of nodes and computed the average results.
To the best of our knowledge, this paper is the first attempt for a resilientdrone service selection and composition in dynamic weather conditions. Toevaluate our proposed approach, we compare the resilient drone service compo-sition algorithm with Brute-Force algorithm. The Brute-Force approach is an32ll-paths search method. We apply the Brute-Force approach as a baseline togenerate the ground truth of optimal compositions. We use Brute-Force in twophases of experiments to find optimal service compositions. In the first phase ofexperiments, Brute-Force approach finds all the possible compositions of droneservices from a given source to a destination. We then select an optimal com-position based on the QoS parameters of drone services. In the second phaseof experiments, Brute-Force approach is used for the global recomposition ofservices to handle the service failures at runtime. Global recomposition refersto composing services from the failed point until the destination. Whenever aservice failure occurs, Brute-Force approach finds all the possible compositionsfrom current failed service until the destination. Finding all possible composi-tions of drone services is time exponential which is undesired. This significantlyreduces the performance of Brute-Force approach to find optimal drone servicecomposition and limits its use for large-scale problems.
We use without lookahead approach in comparison to the proposed looka-head heuristic-based approach. The without lookahead approach behaves simi-lar to a greedy shortest path algorithm. It always selects the least travel distanceservices leading towards the destination. The without lookahead approach hasa higher probability to fail under dynamic weather conditions. For example, theinitial composition plan and expected delivery time may be highly affected byadverse wind. Because of its greedy nature, the without lookahead is fast com-pared to baseline Brute-Force approach and the proposed lookahead approach.Sometimes, the selection of least travel distance services leads to the congestednodes which may result in longer delays for the availability of recharging pads.
The proposed approach performs composition of selective services based oncertain parameters to reach the destination faster. We first generate an initialservice composition plan and compare the Brute-Force, without lookahead, and33
Avg. Computation Time (ms)
N u m b e r o f N o d e s
W i t h o u t L o o k a h e a d A p p r o a c h
L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
Figure 9:
Average computation time
Avg. Delivery Time (min)
N u m b e r o f N o d e s
W i t h o u t L o o k a h e a d A p p r o a c h
L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
Figure 10:
Average delivery time lookahead approaches. We consider three evaluation parameters for comparisonas follows: (1) average computation time, (2) average delivery time, and (3)average distance travelled. We then compare the Brute-Force and adaptivelookahead heuristic-based approaches dealing with the runtime service failures.
1) Average Computation Time:
The baseline Brute-Force approach isnot time-efficient. The computational time for drone service composition usingBrute-Force approach is very high in comparison to without lookahead and pro-posed lookahead heuristic-based approaches. The computation time increasesdue to the increasing number of possible compositions for drone services. Fig. 9compares the average computation time for Brute-Force, without lookahead, andproposed heuristic-based approaches. We observe that the proposed approachsignificantly outperforms the Brute-Force approach by drastically reducing thecomputational time. This is because the proposed approach avoids expensivecomputations by looking ahead once per neighbour state. The computation timevaries for composing drone services depending upon the number of lookaheads.The higher the number of lookaheads we have, the more computational time isrequired to compose drone services.
2) Average Delivery Time:
The delivery time of a drone service includesthe flight time, recharging time, and waiting time. The selection of a right drone34
Avg. Distance Travelled (km)
N u m b e r o f N o d e s
W i t h o u t L o o k a h e a d A p p r o a c h
L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
Figure 11:
Average distance travelled
Avg. Computation Time (ms)
N u m b e r o f N o d e s
A d a p t i v e L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
Figure 12:
Average computation time service is of paramount importance as it ensures the availability of rechargingpads ahead of time minimizing the overall delivery time. Fig. 10 shows theefficiency of the proposed lookahead approach compared to Brute-Force andwithout lookahead approaches. The Brute-Force provides the exact solution asit finds all possible compositions. Our proposed approach obtains a near-optimalsolution compared to Brute-Force approach. However, the time complexity ofthe proposed approach is much better than the baseline Brute-Force approach.Our proposed approach delivers the package 36% faster than without lookaheadapproach. The without lookahead approach selects the services without antic-ipating the congestion conditions ahead which results in higher delivery timecompared to our proposed approach. Our proposed approach uses a lookaheadsearch strategy to reduce recharging and waiting times.
3) Average Distance Travelled:
Some studies investigate the costs as-sociated with drone delivery [4]. The drone delivery cost for a package of 2 kgwithin a 10 km range is estimated at 10 cents in [5]. For simplicity, we use thedistance travelled by a drone as a cost function. Due to dynamic rechargingconstraints and wind conditions, the immediate drone services with least traveldistance cost may lead to congested nodes. Fig. 11 shows the average travel dis-tances chosen by Brute-Force, without lookahead, and proposed heuristic-basedapproaches. The without lookahead approach always selects the least travel35istance services, therefore, ends in higher delivery time. The Brute-Force ap-proach always considers the least delivery time services leading towards thedestination. Our proposed lookahead approach makes a decision based on next-to-adjacent node congestion information which results in 6% improvement indelivery cost than the baseline approach.
When a service failure occurs at any point during the execution, we recom-pose the services to meet the delivery demands. We use the Brute-Force ap-proach for the global recomposition of drone service. While we propose adaptivelookahead heuristic-based local recomposition of affected drone services. In thiscontext, global recomposition refers to the recomposition of services from failurepoint until the destination. The local recomposition refers to the recompositionof only the affected services in the initial composition plan.
1) Average Computation Time:
The Brute-Force approach is highlytime-consuming which is undesired. Whenever a failure occurs, it finds allpossible compositions from failure point until the destination. The adaptivelookahead approach finds the best alternative composition from failure pointuntil the next unaffected drone service in the initial composition or the nextcongested node. In the case of congested node selection, the failure effect ondelivery time is compensated by subtracting it from waiting time at that node.Fig. 12 plots the computation times of the baseline Brute-Force approach andthe proposed heuristic-based approach. The computation time increases alongwith the number of services, which is an expected result. The computationalcomplexity of our proposed approach is more consistent over time and less de-pendent on the network size. It is impractical to use the baseline approach inreal-world scenarios as it is exhausted for large scale problems.
2) Average Delivery Time:
The delivery time for drone services is highlyuncertain when a single drone service cannot fulfil the user’s requirements. Theinter-dependencies on recharging constraints by other drones affect the overalldelivery time of a drone service. At each station, the number of recharging pads36
Avg. Delivery Time (min)
N u m b e r o f N o d e s
A d a p t i v e L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
I n i t i a l P l a n
Figure 13:
Average delivery time
Avg. Distance Travelled (km)
N u m b e r o f N o d e s
A d a p t i v e L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
I n i t i a l P l a n
Figure 14:
Average distance travelled are limited which can be occupied by other drones for long time periods. Fig. 13shows the comparison of the Brute-Force approach and the proposed approachcompared to the initial plan. It shows that the local recomposition providesa near-optimal solution in a significantly shorter period of time compared tothe Brute-Force approach. In some cases, only a single composition is possiblefrom a failed point until the destination. In such cases, we simply replicate thedelay effect of failures to the subsequent services. We observe that sometimesthe Brute-Force approach finds better alternate composition than the originalinitial plan.
3) Average Distance Travelled:
When a service failure occurs, the re-composition approach finds alternate routes to ensure the resilient delivery ofdrone services. In some cases, the travel distances may vary significantly com-pared to the original plan. Fig. 14 plots the average travel distances chosen byBrute-Force and the proposed heuristic-based approaches on top of the initialcomposition plan. We observe that the performance of our proposed approachis almost linear up to 40 nodes in terms of travelling distance and maintains anotable trend even for a higher number of nodes. Our proposed approach savesa substantial amount of time to generate near-optimal solutions.
4) Effects of Failure Rate:
We analyze the effects of the increasing num-ber of failure rates on the resilience of delivery time and travel distance. Fig. 1537
Avg. Delivery Time (min)
F a i l u r e R a t e
A d a p t i v e L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
I n i t i a l P l a n
Figure 15:
Average delivery time
Avg. Distance Travelled (km)
F a i l u r e R a t e
A d a p t i v e L o o k a h e a d A p p r o a c h
B r u t e - F o r c e
I n i t i a l P l a n
Figure 16:
Average distance travelled and 16 plots the effects of different rates of failures on the average delivery timeand distance travelled for the baseline Brute-Force approach and the proposedadaptive lookahead heuristic-based approach. We observe that the proposedapproach finds optimal or near-optimal solutions for the increasing number offailure rates. The performance of our proposed approach is close to the Brute-Force approach even when the failure rate is high. Experiments based on thedifferent failure rates demonstrate the effectiveness of our proposed approach interms of delivery time and distance travelled (i.e., delivery cost).
We observed several unique features from our experiments with resilientcomposition during drone delivery operations. First, drones are vulnerable toweather conditions such as wind. The dynamic changes in the service environ-ment may significantly influence the initial composition plan. Second, we mayhave a high failure rate of drone services due to dynamic weather conditions.The proposed adaptive recomposition algorithm can provide computationallyefficient and near-optimal solutions in the dynamic environment. Moreover, theadaptive recomposition algorithm provides significantly better solutions whenthe number of services is small. Third, the computational complexity of theadaptive recomposition algorithm remains consistent even when the networksize becomes large. Fourth, the use of local recomposition techniques over global38ecomposition techniques provides better practical solutions especially in termsof computational complexity. Finally, the use of global recomposition is imprac-tical in real-world scenarios as it exhausts for large scale delivery networks.
7. Conclusion
We propose a resilient service composition framework for drone-based deliv-ery considering the recharging constraints and dynamic weather conditions. Anoptimal set of candidate drone services is selected using the skyline approachat the source node in a skyway network. We present a formal model to rep-resent constraint-aware drone services. We propose a deterministic lookaheadalgorithm to build an initial offline composition plan. We develop a heuristic-based resilient service composition algorithm that adapts to changes in servicebehaviour at runtime. We run several experiments to illustrate the performanceof the proposed approach in comparison to Brute-Force and without lookaheadapproaches. We found that the proposed approach is runtime efficient and pro-duces significantly better results than the Brute-Force and without lookaheadapproaches. Moreover, the proposed approach guarantees the resilience of de-livery services for the increasing number of failure rates. Hence, it is a morepractical solution in real-world applications of drone delivery services.A key limitation of the proposed approach is that the proposed approachdoes not take into account the handover of packages among different drones atintermediate recharging stations. The handover of packages to spare drones atintermediate recharging stations may assist in minimizing the overall deliverytime. We plan to apply new optimization techniques for the handover initiation,the selection of the optimal drone service, and the handover management amongdifferent drones. The behaviour of a drone depends on wind patterns such astailwinds and headwinds in different geographical areas. Another limitationof the proposed approach is that it does not incorporate the changing windpatterns into the drone service model. The proposed approach only focuses onthe effects of wind speed and direction on the drone service composition plan.39here are several weather conditions that can affect a drone’s performance suchas precipitation (e.g., rain, snow, hail, and sleet), temperature, cloud cover, andvisibility. We intend to consider the effects of different weather conditions onthe performance of a drone and exploit deep learning techniques for predictingand forecasting weather patterns. A single drone can deliver multiple smallpackages from a warehouse to desired destinations in one trip. The proposedresilient composition approach is limited to generate solutions for single packagedelivery by a drone from a given source to a destination. As future work, weplan to explore different adaptive techniques to extend the proposed approachfor multi-package deliveries in a dynamic environment.
Acknowledgment
This research was partly made possible by DP160103595 and LE180100158grants from the Australian Research Council. The statements made herein aresolely the responsibility of the authors.
References [1] B. Liu, X. Guo, Y. Yu, Q. Zhou, Minimizing the total completion time ofan urban delivery problem with uncertain assembly time, Transp. Res. ELogist. Transp. Rev. 132 (2019) 163–182. doi:https://doi . org/10 . . tre . . . .[2] H. Shakhatreh, A. H. Sawalmeh, A. Al-Fuqaha, Z. Dou, E. Almaita,I. Khalil, N. S. Othman, A. Khreishah, M. Guizani, Unmanned aerial vehi-cles (uavs): A survey on civil applications and key research challenges,IEEE Access 7 (2019) 48572–48634. doi:https://doi . org/10 . . . .[3] G. Chmaj, H. Selvaraj, Distributed processing applications for uav/drones:A survey, in: Progress in Systems Engineering, Springer, Cham, 2015, pp.449–454. doi:https://doi . org/10 . .404] D. Bamburry, Drones: Designed for product delivery, Design ManagementReview 26 (1) (2015) 40–48. doi:https://doi . org/10 . . .[5] R. D’Andrea, Guest editorial can drones deliver?, Trans. Autom.Sci. Eng. 11 (3) (2014) 647–648. doi:https://doi . org/10 . . . .[6] A. Bouguettaya, et al., A service computing manifesto: The next 10 years,Commun. ACM 60 (4) (2017) 64–72. doi:https://doi . org/10 . .[7] B. Shahzaad, A. Bouguettaya, S. Mistry, A. G. Neiat, Composing drone-as-a-service (daas) for delivery, in: IEEE International Conference on Web Ser-vices, 2019, pp. 28–32. doi:https://doi . org/10 . . . .[8] M. J. Corbett, F. Xie, D. Levinson, Evolution of the second-story city:The minneapolis skyway system, Environ. Plan. B Plan. Des. 36 (4) (2009)711–724. doi:https://doi . org/10 . .[9] S. Park, L. Zhang, S. Chakraborty, Design space exploration of drone in-frastructure for large-scale delivery services, in: Proceedings of the 35thInternational Conference on Computer-Aided Design, Austin, Texas, USA,2016, pp. 1–7. doi:https://doi . org/10 . . .[10] J. Kim, S. Kim, J. Jeong, H. Kim, J. Park, T. Kim, Cbdn: Cloud-based drone navigation for efficient battery charging in drone networks,Trans. Intell. Transp. Syst. (2018) 1–18 doi:https://doi . org/10 . . . .[11] K. Dorling, J. Heinrichs, G. G. Messier, S. Magierowski, Vehicle routingproblems for drone delivery, Trans. Syst., Man, Cybern. 47 (1) (2017) 70–85. doi:https://doi . org/10 . . . .[12] S. Kim, I. Moon, Traveling salesman problem with a drone station,IEEE Trans. Syst. Man Cybern. Syst. 49 (1) (2019) 42–52. doi:https://doi . org/10 . . . .4113] K. Sundar, S. Rathinam, Algorithms for routing an unmanned aerial vehiclein the presence of refueling depots, IEEE Trans. Autom. Sci. Eng. 11 (1)(2014) 287–294. doi:https://doi . org/10 . . . .[14] S. Venkatachalam, K. Sundar, S. Rathinam, A two-stage approach for rout-ing multiple unmanned aerial vehicles with stochastic fuel consumption,Sensors 18 (11) (2018) 3756. doi:https://doi . org/10 . .[15] B. Shahzaad, A. Bouguettaya, S. Mistry, A. G. Neiat, Constraint-awaredrone-as-a-service composition, in: International Conference on Service-Oriented Computing, Springer, Cham, 2019, pp. 369–382. doi:https://doi . org/10 . .[16] A. G. Neiat, A. Bouguettaya, T. Sellis, H. Dong, Failure-proof spatio-temporal composition of sensor cloud services, in: International Conferenceon Service-Oriented Computing, Springer Berlin Heidelberg, Berlin, Hei-delberg, 2014, pp. 368–377. doi:https://doi . org/10 . .[17] L. Barakat, S. Miles, M. Luck, Adaptive composition in dynamic service en-vironments, Future Gener. Comput. Syst. 80 (2018) 215–228. doi:https://doi . org/10 . . future . . . .[18] N. B. Mabrouk, S. Beauche, E. Kuznetsova, N. Georgantas, V. Issarny,Qos-aware service composition in dynamic service oriented environments,in: Middleware 2009, Springer Berlin Heidelberg, Berlin, Heidelberg, 2009,pp. 123–142. doi:https://doi . org/10 . .[19] C. C. Murray, A. G. Chu, The flying sidekick traveling salesman prob-lem: Optimization of drone-assisted parcel delivery, Transp. Res. PartC Emerg. Technol. 54 (2015) 86–109. doi:https://doi . org/10 . . trc . . . .[20] Y. Choi, P. M. Schonfeld, Optimization of multi-package drone deliveriesconsidering battery capacity, in: 96th Annual Meeting of the Transporta-42ion Research Board, Washington, DC, USA, 2017, pp. 8–12.URL https://trid . trb . org/view/1439294 [21] B. D. Song, K. Park, J. Kim, Persistent uav delivery logistics: Milp for-mulation and efficient heuristic, Comput. Ind. Eng. 120 (2018) 418–428. doi:https://doi . org/10 . . cie . . . .[22] H. Saboohi, A. Amini, H. Abolhassani, Failure recovery of composite se-mantic web services using subgraph replacement, in: International Con-ference on Computer and Communication Engineering, 2008, pp. 489–493. doi:https://doi . org/10 . . . .[23] G. Canfora, M. Di Penta, R. Esposito, M. L. Villani, Qos-aware replanningof composite web services, in: IEEE International Conference on Web Ser-vices, 2005, pp. 121–129. doi:https://doi . org/10 . . . .[24] Y. Yan, P. Poizat, L. Zhao, Repair vs. recomposition for broken service com-positions, in: P. P. Maglio, M. Weske, J. Yang, M. Fantinato (Eds.), Inter-national Conference on Service-Oriented Computing, Springer, Berlin, Hei-delberg, 2010, pp. 152–166. doi:https://doi . org/10 . .[25] Y. Zhai, J. Zhang, K. Lin, Soa middleware support for service processreconfiguration with end-to-end qos constraints, in: IEEE InternationalConference on Web Services, 2009, pp. 815–822. doi:https://doi . org/10 . . . .[26] H. Saboohi, S. A. Kareem, Failure recovery of world-altering compositesemantic services - a two phase approach, in: Proceedings of the 14thInternational Conference on Information Integration and Web-based Ap-plications & Services, ACM, New York, NY, USA, 2012, pp. 299–302. doi:https://doi . org/10 . . .[27] T. Yu, K. J. Lin, Adaptive algorithms for finding replacement services inautonomic distributed business processes, in: Proceedings Autonomous De-43entralized Systems, 2005, pp. 427–434. doi:https://doi . org/10 . . . .[28] H. Saboohi, An automatic failure recovery method for world-alteringcomposite semantic web services, Ph.D. thesis, University of Malaya(2013).URL http://studentsrepo . um . edu . my/5747/2/Hadi_Saboohi_PhD_Thesis . pdf [29] J. Alves, J. Marchi, R. Fileto, M. A. R. Dantas, Resilient compositionof web services through nondeterministic planning, in: IEEE Sympo-sium on Computers and Communication, 2016, pp. 895–900. doi:https://doi . org/10 . . . .[30] T. Laleh, J. Paquet, S. Mokhov, Y. Yan, Constraint verification failurerecovery in web service composition, Future Gener. Comput. Syst. 89 (2018)387–401. doi:https://doi . org/10 . . future . . . .[31] A. G. Neiat, A. Bouguettaya, T. Sellis, Z. Ye, Spatio-temporal compositionof sensor cloud services, in: IEEE International Conference on Web Ser-vices, Anchorage, AK, 2014, pp. 241–248. doi:https://doi . org/10 . . . .[32] A. G. Neiat, A. Bouguettaya, T. Sellis, S. Mistry, Crowdsourced coverageas a service: Two-level composition of sensor cloud services, Trans. Knowl.Data Eng. 29 (7) (2017) 1384–1397. doi:https://doi . org/10 . . . .[33] A. Thibbotuwawa, P. Nielsen, B. Zbigniew, G. Bocewicz, Energy consump-tion in unmanned aerial vehicles: A review of energy consumption modelsand their relation to the uav routing, in: Proceedings of 39th Interna-tional Conference on Information Systems Architecture and Technology,Springer, Cham, 2019, pp. 173–184. doi:https://doi . org/10 . . 4434] M. Seleck`y, P. V´aˇna, M. Rollo, T. Meiser, Wind corrections in flightpath planning, Int. J. Adv. Robot. Syst. 10 (5) (2013) 248. doi:https://doi . org/10 . .[35] K. Takegami, S. Negoro, Y. Funabashi, I. Taniguchi, H. Tomiyama, Low-energy routing for deadline-constrained delivery drones under windy con-ditions, BNCSS 8 (1) (2019).URL . bncss . org/index . php/bncss/article/view/102 [36] J. Osborne, R. Rysdyk, Waypoint guidance for small uavs in wind, in:Proceedings of the AIAA Infotech@Aerospace, Arlington, Virginia, 2005,pp. 1–12. doi:https://doi . org/10 . . .[37] R. Vazquez, D. Rivas, A. Franco, Stochastic analysis of fuel consump-tion in aircraft cruise subject to along-track wind uncertainty, Aerosp.Sci. Technol. 66 (2017) 304 – 314. doi:https://doi . org/10 . . ast . . . .[38] K. P. O’Rourke, W. B. Carlton, T. G. Bailey, R. R. Hill, Dynamic routingof unmanned aerial vehicles using reactive tabu search, Mil. Oper. Res.6 (1) (2001) 5–30.URL . jstor . org/stable/43943662 [39] S. Borzsony, D. Kossmann, K. Stocker, The skyline operator, in: Proceed-ings 17th International Conference on Data Engineering, 2001, pp. 421–430. doi:https://doi . org/10 . . . .[40] Q. Yu, A. Bouguettaya, Computing service skylines over sets of services,in: IEEE International Conference on Web Services, 2010, pp. 481–488. doi:https://doi . org/10 . . . .[41] D. Papadias, Y. Tao, G. Fu, B. Seeger, An optimal and progressive al-gorithm for skyline queries, in: Proceedings of the 2003 ACM SIGMODInternational Conference on Management of Data, ACM, New York, NY,USA, 2003, pp. 467–478. doi:https://doi . org/10 . . .4542] Q. Yu, A. Bouguettaya, Multi-attribute optimization in service selection,World Wide Web 15 (1) (2012) 1–31. doi:https://doi . org/10 . .[43] P. A. Coquelin, R. Munos, Bandit algorithms for tree search, in: Proceed-ings of the Twenty-Third Conference on Uncertainty in Artificial Intelli-gence, Arlington, Virginia, USA, 2007, pp. 67–74.URL https://dl . acm . org/doi/10 . . [44] S. S. Villar, J. Bowden, J. Wason, Multi-armed bandit models for theoptimal design of clinical trials: benefits and challenges, Statist. Sci. 30 (2)(2015) 199–215. doi:https://doi . org/10 . .[45] A. S. Ferreira, R. A. Gonalves, A. Pozo, A multi-armed bandit se-lection strategy for hyper-heuristics, in: IEEE Congress on Evolution-ary Computation, 2017, pp. 525–532. doi:https://doi . org/10 . . . .[46] M. Radmanesh, M. Kumar, Flight formation of uavs in presence ofmoving obstacles using fast-dynamic mixed integer linear programming,Aerosp. Sci. Technol. 50 (2016) 149–160. doi:https://doi . org/10 . . ast . . . .[47] M. H. Lin, J. F. Tsai, C. S. Yu, A review of deterministic optimizationmethods in engineering and management, Math. Probl. Eng. 2012 (2012)1–15. doi:https://doi . org/10 . .[48] R. R. Lam, K. E. Willcox, Lookahead bayesian optimization with inequal-ity constraints, in: Advances in Neural Information Processing Systems30, Curran Associates, Inc., Red Hook, NY, USA, 2017, pp. 1890–1900.URL http://papers . nips . cc/paper/6785-lookahead-bayesian-optimization-with-inequality-constraints . pdf [49] A. Berro, S. L. M. Sainte, A. R. Gazen, Genetic algorithms and particleswarm optimization for exploratory projection pursuit, Ann. Math. Artif.46ntell. 60 (2010) 153–178. doi:https://doi . org/10 . .[50] B. Soykan, A hybrid tabu/scatter search algorithm for simulation-basedoptimization of multi-objective runway operations scheduling, Ph.D. thesis,Old Dominion University (2016).URL https://digitalcommons . odu . edu/emse_etds/4/ [51] S. Pandey, D. Chakrabarti, D. Agarwal, Multi-armed bandit problems withdependent arms, in: Proceedings of the 24th International Conference onMachine Learning, ACM, New York, NY, USA, 2007, p. 721728. doi:https://doi . org/10 . . .[52] R. Degenne, W. M. Koolen, P. M´enard, Non-asymptotic pure explorationby solving games, in: Advances in Neural Information Processing Systems32 (NIPS 2019), Curran Associates, Inc., 2019, pp. 14492–14501.URL http://papers . nips . cc/paper/9592-non-asymptotic-pure-exploration-by-solving-games [53] S. Shah, D. Dey, C. Lovett, A. Kapoor, Airsim: High-fidelity visualand physical simulation for autonomous vehicles, in: Field and ServiceRobotics, Springer, Cham, 2018, pp. 621–635. doi:https://doi . org/10 . .[54] N. Koenig, A. Howard, Design and use paradigms for gazebo, an open-source multi-robot simulator, in: IEEE/RSJ International Conference onIntelligent Robots and Systems, Vol. 3, 2004, pp. 2149–2154. doi:https://doi . org/10 . . . .[55] A. Babushkin, Jmavsim (2018).URL https://dev . px4 . io/v1 . . . html [56] S. Krishnan, B. Boroujerdian, W. Fu, A. Faust, V. J. Reddi, Air learn-ing: An AI research platform for algorithm-hardware benchmarking of au-47onomous aerial robots, CoRR abs/1906.00421 (2019).URL http://arxiv . org/abs/1906 . [57] P. M. Kornatowski, M. Feroskhan, W. Stewart, D. Floreano, Down-side up: Rethinking parcel position for aerial delivery, IEEE Robot.Autom. Lett. 5 (3) (2020) 4297–4304. doi:https://doi . org/10 . . . .[58] L. Roubeyrie, S. Celles, Windrose: A python matplotlib, numpy library tomanage wind and pollution data, draw windrose, J. Open Source Softw. 3(2018) 268. doi:https://doi . org/10 . . .[59] A. Hadjimichael, D. Gold, D. Hadka, P. Reed, Rhodium: Python library formany-objective robust decision making and exploratory modeling, J. OpenRes. Softw. 8 (1) (2020) 12. doi:http://doi . org/10 . . .[60] A. A. Hagberg, D. A. Schult, P. J. Swart, Exploring network structure,dynamics, and function using networkx, in: Proceedings of the 7th Pythonin Science Conference, Pasadena, CA, USA, 2008, pp. 11–15.URL http://conference . scipy . org/proceedings/SciPy2008/paper_2/ [61] G. V´as´arhelyi, C. Vir´agh, G. Somorjai, T. Nepusz, A. E. Eiben, T. Vic-sek, Flight logs of drone swarms (2018). doi:https://doi . org/10 . . mq85r61/1 .URL https://dryad . figshare . com/articles/Flight_logs_of_drone_swarms/6843977/1com/articles/Flight_logs_of_drone_swarms/6843977/1