research article intelligent testing of traffic light...

Research ArticleIntelligent Testing of Traffic Light ProgramsValidation in Smart Mobility Scenarios

Javier Ferrer Joseacute Garciacutea-Nieto Enrique Alba and Francisco Chicano

Departamento de Lenguajes y Ciencias de la Computacion Universidad de Malaga ETS Ingenieria InformaticaBulevar Louis Pasteur 35 29071 Malaga Spain

Correspondence should be addressed to Javier Ferrer ferrerlccumaes

Received 1 July 2015 Revised 24 December 2015 Accepted 12 January 2016

Academic Editor Dan Simon

Copyright copy 2016 Javier Ferrer et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

In smart cities the use of intelligent automatic techniques to find efficient cycle programs of traffic lights is becoming an innovativefront for traffic flow management However this automatic programming of traffic lights requires a validation process of thegenerated solutions since they can affect the mobility (and security) of millions of citizens In this paper we propose a validationstrategy based on genetic algorithms and feature models for the automatic generation of different traffic scenarios checking therobustness of traffic light cycle programsWe have concentrated on an extensive urban area in the city ofMalaga (in Spain) in whichwe validate a set of candidate cycle programs generated bymeans of four optimization algorithms Particle SwarmOptimization forTraffic Lights Differential Evolution for Traffic Lights random search and Sumo Cycle Program Generator We can test the cyclesof traffic lights considering the different states of the city weather congestion driver expertise vehiclersquos features and so forth butprioritizing the most relevant scenarios among a large and varied set of them The improvement achieved in solution quality isremarkable especially for CO

2emissions in which we have obtained a reduction of 12699 compared with the expertsrsquo solutions

1 Introduction

Nowadays all initiatives for the development of the SmartCity [1ndash3] focus onpublic institutions for the impact they haveon society in general A Smart City is a holistic initiative tomanage the city where actions and services based on infor-mation technology are optimally conceived and deployed forsustainable living Traffic flowmanagement is one of themostimportant aspects in the context of smart cities due to thelarge number of vehicles to be managed in the metropolitanarea An optimal management of traffic might be beneficialto minimize journey times and reduce fuel consumption andharmful emissions For this purpose the cycle programmingof traffic lights constitutes a key task [4] Nevertheless thelarge number of cycle combinations that appear in currenttraffic light schedules require automatic intelligent tools to beused by the experts in this field

In this regard several research studies on automatictraffic light scheduling exist which use different intelligenttechniques such as fuzzy logic controllers [5 6] multia-gent systems [7] neural networks [8] and metaheuristic

algorithms [9ndash11] However these solutions entail a highdependency on the scenario instances under examinationwith specific conditions and limited variability Furthermorethe variability of the traffic system is due among other factorsof varying importance to different weather conditions thedaily traffic versus the weekendrsquos rush hours changingenvironment and so forth Therefore our main objective inthis paper is to offer a validation strategy for the generatedcycle programs of different and variable traffic scenariossince they could determine the robustness of the proposedcycle programs of traffic lights for the experts

With the aim of representing high variability systems(like vehicular traffic scenarios) feature models (FMs) [12]have emerged in the last decade as a standard strategy formodeling common and variable features of a system andtheir relationships FMs have been applied to test SoftwareProduct Lines (SPL) [13ndash15] with great success although theyhave still not been considered for representing features inother domains Mindful of this our motivation is thereforeto develop a model for traffic management with the targetbeing to represent all feature variability for the generation

Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2016 Article ID 3871046 19 pageshttpdxdoiorg10115520163871046

2 Mathematical Problems in Engineering

of different scenarios of urban mobility In this regard weaim to extract validated information about which traffic lightsprogram (from among a number of them) is best adaptedfor most traffic scenarios with different feature combinationsand which fails to manage the traffic flow with other specificfeature combinations

In fact among the main features that affect traffic flowwe can enumerate the following different meteorology con-ditions type of vehicle driver expertise simulation timeand so forth However the analysis of all possible trafficscenarios is inviable due to the large number of feature com-binations to take into account For this reason we propose anevolutionary approach called the Prioritized Genetic FeatureModel (PGFM) algorithm designed for the generation ofa minimal prioritized test suite of traffic scenarios whichmeet a coverage criterion (pairwise) The pairwise coveragecriterion [16] is the most popular in the literature Sinceall scenarios are not equally important our feature modelhas been extended to provide priorities for the featuresTherefore in the case there are time or cost requirements thisprioritization allows us to first generate the most importantscenarios in which we test the traffic lights programs

In this paper we focus on an extensive urban area ofMalaga city (in Spain) for the case study We use the well-known SUMO (Simulator of Urban Mobility) [17] trafficsimulator which offers a continuous source of informationabout the vehiclersquos flow velocity fuel consumption emis-sions journey time and so forth giving rise to configurationsof realistic scenarios according to real patterns of mobility Inparticular we validate a large number of traffic light programspreviously generated by four optimization algorithms twometaheuristic techniques PSOTL [9] and DETL [9] a blindstochastic search algorithm random search (RS) and adeterministic procedure following human expertsrsquo rules fortraffic configuration SCPG [17] Our present work will helpdetermine the suitability of the computed configurations fora wider set of city scenarios for which they have not beentrainedThemain contributions of this paper are summarizedas follows

(i) Design and application of a prioritized feature model(PFM) for the representation of traffic light man-agement systems as far as we know this is the firstapplication of a PFM to traffic light management sys-tems It constitutes the actual use of cross-fertilizingtechniques since we take advantage of prioritizedfeature models typically applied to software testingto leverage traffic and environmental solutions incurrent smart cities (like Malaga) This model rep-resents traffic variability through different featuresweighted according to their probability of occurrenceand importance This testing technique is useful hereto identify which cycle programs fail to manage thetraffic flow in different scenarios

(ii) Development of Prioritized Genetic algorithm forfeature models (PGFM) for the automatic generationof weighted traffic scenarios PGFM is shown here tocomply with the pairwise covering criterion [16] Inthis way we are able to work with a reduced set of

scenarios that cover all the different traffic featuresconsidered in our PFM

(iii) Thorough analysis and comparison of the differenttraffic light programs for the Malaga city case studythese cycle programs were generated as candidatesolutions by means of four optimization algorithmsPSOTL DETL RS and SCPG (human expertsrsquo solu-tions) Although these solutions were obtained for aspecific traffic scenario they have been validated formultiple scenarios covering our PFM

The remainder of the paper is organized as followsSection 2 presents an overview of related work in the lit-erature In Section 3 basic concepts are explained for thesake of a better understanding of the work presented hereAfter this Section 4 details the validation strategy Section 5describes how traffic scenario features are represented bymeans of our feature model Section 6 outlines the PGFMalgorithm and presents the generated test suite of trafficscenariosThen Section 7 is devoted to presenting theMalagacity case study In Section 8 the experimental procedureand the analysis of results are addressed In Section 9 weprovide a deeper analysis focusing on the prioritization andthe ranking of the solutions Section 10 discuses the threats tovalidity Finally Section 11 outlines some concluding remarksand future work

2 Related Work

This section presents an overview of related work consideringa two-pronged approach combinatorial testing for featuremodels and optimization of traffic light cycle programs Theformer focuses on intelligent techniques for combinatorialinteraction testing (CIT) using covering arrays The latterconsiders the use of metaheuristics for traffic optimization

CIT [16] is an effective testing approach for detectingfailures caused by certain combinations of components orinput values Generally this task consists of generating atleast all possible combinations of the parametersrsquo values (thistask is NP-hard [18]) A large number of CIT approaches havealready been published A good overview and classificationof approaches can be found in [19 20] and more recentlyin [21] In addition an extensive survey that focuses onCIT with constraints is also given in [22] There are onlytwo approaches (to the best of our knowledge) that supportprioritized test data generation the Deterministic DensityAlgorithm (DDA) presented in [23] and an approach basedon Binary Decision Diagrams (BDD) [24] However neitherof them have been applied to feature models prioritizationSo here we use the priorities of the traffic features to generateprioritized scenarios for testing traffic light programs

Except for a few efforts the application of bioinspiredtechniques to combinatorial interaction testing with featuremodels remains largely unexplored Garvin et al appliedsimulated annealing to CIT for computing n-wise coveragefor feature models [25] Ensan et al also propose a geneticalgorithm approach for test case generation [26] In contrastwith the proposed work here they use as fitness function avariation of cyclomatic complexity metric adapted to feature

Mathematical Problems in Engineering 3

models their goal is not n-wise coverage and they do notconsider priorities Henard et al propose an approach thatuses a dissimilarity metric that favors individuals whose n-wise coverage varies the most from the current populationthus increasing the chances of wider coverage [27] A keydifference with this work is that the prioritization of them isnot based on assigned weights that have a real value Recentsurveys on feature models [28 29] attest to the increasingrelevance of the topic within the community but also confirmthat the latent potential of search based testing techniquesremains largely untapped

The generation of test suites from feature models hasrecently been examined in several studies Oster et al pro-posed MoSo-PoLiTe [30] an approach that translates featuremodels and their constraints into binary constraint solverproblems (CSP) fromwhich they compute pairwise coveringarrays Hervieu et al developed a tool called PACOGENthat also relies on constraint programming for computingpairwise coverage from feature models [31] Johansen et al[15] proposed a greedy approach to generate n-wise test suitesthat adapts Chvatalrsquos with the aim of solving the set coverproblem In the work presented here a test suite is generatedalthough it is composed of the different traffic scenarioswhere traffic light programs are tested

In terms of traffic optimization metaheuristic algorithms[32] have become very popular for solving traffic light stagingproblems A first attempt was proposed by Rouphail et al[33] where a genetic algorithm (GA) was coupled with theCORSIM [34] microsimulator for the timing optimization ofnine intersections in the city of Chicago (USA) Followingthe model proposed in Brockfeld et al [35] Sanchez et al[10] designed a GA with the objective of optimizing the cycleprogramming of traffic lights in a commercial area in thecity of Santa Cruz de Tenerife (Spain) In that approach thecomputation of valid states was done before the algorithmbegan and it highly depended on the scenario instancetackledAGAwas also used byTurky et al [36] to improve theperformance of traffic lights and pedestrian crossing controlin a single four-way two-lane intersection

Peng et al [37] presented a particle swarm optimization(PSO)with isolation niches for the scheduling of traffic lightsIn that approach a purely academic instancewith a restrictiveone-way road with two intersections was used to test theproposal Kachroudi andBhouri [38] applied amultiobjectiveversion of PSO for optimizing cycle programs using a predic-tive control model based on a public transport progressionmodel In that approach private and public vehicle modelswere used to carry out simulations on a virtual urban roadnetwork made up of 16 intersections and 51 links Kesur [39]used the nondominated sorting genetic algorithm (NSGA-II) to evaluate two confront objectives delay and number ofstopsThiswork focuses on evaluating differentmodificationsof the NSGA-II algorithm In addition the trafficmodel usedwas the microscopic stochastic traffic network simulationproposed by him Garcıa-Nieto et al [9 40] proposed a PSOapproach with the SUMO microsimulator for traffic lightprogramming in large realistic urban areas like Malaga andSeville (Spain) and Bahıa Blanca (Argentina) For all thescenarios this approach considered hundreds of traffic lights

and circulating vehicles In addition these last works [9 40]evaluated other tools like SCOOT that uses different trafficlight schemes of scheduling on different instances adapted tovery specific use cases Performing comparative studies withthese tools is out of the scope of the proposed work herewhere the main goal is the validation of traffic light plans incertain urban area under different environmental conditions

In summary the aforementioned approaches havefocused on different aspects of traffic light programmingHowever a common limitation in all of them is that in theoptimization phase they considered just one scenario withspecific traffic conditions This means that the resulting cycleprograms could be strongly biased to the optimized scenarioinstances so even a slight change in the traffic conditions(eg weather change) might suppose that the cycle programdoes not work properly

In contrast our initiative stands for cross-fertilization ofthe two domains involved here Feature Modeling and TrafficLights Optimization Therefore we consider a number ofdifferent traffic conditions in scenarios covering all possiblefeatures according to our designed feature model so fromnow on bymeans of our approach the cycle programs will beempirically validated thereby providing the experts in trafficmanagement with robust and generalist solutions

3 Background

In this work we do not focus on the simulationmodel hencea comparison of simulation algorithms is out of the scopeof this study although we do use one simulator (SUMOsimulator) tomeasure the quality of the traffic light programsWe focus on the optimization and validation models In thisregard the validation model has been specified as general aspossible in order to be used with multiple general purposeoptimization algorithms With this aim we have tested twometaheuristic algorithms PSO and DE a stochastic randomsearch algorithm and a deterministic traffic scheduler SCPG

Before describing our validation strategy several requiredconcepts are introduced the SUMO traffic simulator whichis the simulator used to measure the quality of the solutionproposed the cycle program optimization of traffic lightsprogram the optimization algorithms used to generate trafficlights programs and the feature models proposed here

31 SUMO Traffic Simulator SUMO (Simulator of UrbanMobility) [17] is a well-known traffic simulator that providesan open source highly portable and microscopic road trafficsimulation tool designed to handle large road scenariosSUMO requires several input files that contain informationabout the traffic and the streets to be simulated A journeyis a vehicle movement from one location to another definedby the starting edge (street) the destination edge and thedeparture time A route is an extended journey meaning thata route definition contains not only the first and last edgesbut also all the edges the vehicle will pass through Additionalfiles can be added to SUMOcontaining themap or traffic lightpositions and cycles


Phase duration = lsquolsquo36rsquorsquo

Solution vector of integer variables

Current state in intersection

Intersection id = lsquolsquoirsquorsquo Intersection id =

40 5 40 10 36 6 22

i = lsquolsquoGrr GGGr rrG GGrsquorsquo

middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot middot

ldquoi + 1rdquo

Figure 1 Cycle program (phase duration) of traffic lights within intersections Dashed circles indicate where traffic lights are located in thereal intersection

It is important to note that SUMO by default providesthe valid combination of states that the traffic light controllercan go through inside the map specification file and anapproximation of interval times for these states [41] Thismeans that SUMO already incorporates a solver (SCPG) forthe cycle program of traffic lights based on greedy and humanknowledge

The output of a SUMO simulation is registered in ajourney information file that contains information about eachvehiclersquos departure time the time the vehicle waited to setoff (offset) the time the vehicle arrived the duration of itsjourney and the number of steps in which the vehicle speedwas below 01ms (temporal stops in driving) Other outputfiles gather information about emission traces in vehicles(CO2 NO119909 PM etc) and hydrocarbon consumption This

information is used to evaluate the quality of alternativetraffic light cycle programs

32 Cycle Program Optimization of Traffic Lights The objec-tive in this problem is to find optimized cycle (timing)programs for all the traffic lights located in a given urban areawith the aim of reducing the global journey time emissionsand fuel consumption Specifically cycle programs are thetime span that a set of traffic lights (at an intersection)keep their color states This is the first step of our workobtaining configurations of the city traffic lights optimizingthese metrics the second is to find the best of them in termsof generalization for different city conditions (weather trafficdriver and vehicle conditions)

An example of this mechanism can be observed inFigure 1 where the intersection with id = ldquoirdquo contains sevenphases with durations of 40 5 40 10 36 6 and 22 secondsIn these phases the states have twelve signals (colors) eachone of them corresponding to one of the twelve signallights located in the intersection being studied (dashedcircles indicate where traffic lights are located in the realintersection) These states are the valid ones generated bySUMO [42] obeying real traffic rules In this instance the fifth

phase contains the state ldquoGrr GGGr rrG GGrdquo meaning thatseven traffic lights are green (119866) and the other five are red (119903)for 36 seconds The following phase changes the state of thetraffic lights to another valid combination for example ldquoyGGyyyG GGy yyrdquo (119910means yellow) for 6 seconds The last phaseis followed by the first one and this cycle is repeated for theentire analysis All the intersections in the complete scenarioperform their own cycles of phases at the same time therebycomprising the global schedule of signal lights Asmentionedbefore computing the Signal Light Timing Program (SLTP)is based on optimizing the combination of phase durationsof all traffic lights (in all intersections) with the intention ofimproving the global flow of vehicles

A formal definition of the optimal SLTP is as followsLet 119875 = 119868

1 119868

119899 be a set of intersections where each

one has a different set of phases 119868119894= 1205931 120593

119898 and each

120593119895represents the timespan that the set of traffic lights in

intersection 119868119894keep one valid state of light colors (eg ldquoGrr

GGGr rrG GGrdquo) find a program 1198751015840 that minimizes a scoring

function Θ Γ rarr 119901 such that

1198751015840= argmin119901subΓ

Θ (119901) (1)

where Γ is the space of all possible combinations of cycleprograms and 119901 a given program of Γ

Since the timespans of phase durations are calculatedin seconds (as done in real traffic lights) Γ can be repre-sented with a tuple of positive integer numbers Z+ Thenthe number of possible program combinations that is thesolution space size can be calculated as 119879sum

119899

119894=1|119868119894| 119879 being

the value in the interval of possible timespans This way asintervals were set to 119879 = [5 60] and the worked instanceshad 304 phases (at least) the problem solution space wouldconsist of 55304 = 118 sdot 10

529 combinations Thereforeefficient automated approaches are required to tackle it Thenumber of phases (304) is automatically computed accordingto the traffic model by the SUMO simulator when a probleminstance is generated


Input A scenario instance 120601(119888) of a given city 119888Output Best found solution 119887 encoding a cycle program(1) 119878 larr initializeSwarm()(2) while 119892 ltMAXIMUM

119892do

(3) for each particle 119909119894119892in 119878 do

(4) 119909119894

119892+1larr update(119909119894

119892 V119894119892 119901119894

119892 119887119892) update velocity and particlersquos position as in Standard PSO 2011

(5) 119902119894

119892+1larr Q(119909119894

119892+1) Mid-Thread quantisation adaptation to cycle programs encoding

(6) ekaluate(119902119894119892+1 120601(119888)) solution evaluation by micro-simulation on scenario instance 120601(119888)

(7) end for(8) 119887

119892+1larr updateLeader(119887

119892 119902119894

119892+1) if better

(9) end while

Algorithm 1 Pseudocode of PSOTL

33 Optimization Algorithms for Cycle Programming In thissubsection we briefly describe the algorithms proposed forthe optimization of cycle programs of traffic lights

331 PSOTL This is a particle swarm optimization algo-rithm for finding quasi-optimal cycle programs for trafficlights In PSOTL the initial swarm is composed of a numberof particles (solutions) initialized with a set of random valuesrepresenting the phase durationsThese values are within thetime interval [5 60] sube Z+ and constitute the range of possibletime spans (in seconds) a traffic light can be kept on a signalcolor (only green or red the time for yellow is a constantvalue) This interval is specified to follow several examples ofreal traffic light programs provided by Malagarsquos City Council(Spain)

Since the optimal SLTP requires solutions encoded with avector of integers (representing phase durations in seconds)we have used the quantisation method provided in thestandard specification of PSO 2011 [43] This quantisation isapplied to each new generated particle and transforms thecontinuous values of particles to discrete ones It consistsof a Mid-Thread uniform quantiser as specified in (2) Thequantum step is set here to Δ = 05 Consider

119876 (119909) = Δ sdot lfloor119909

Δ+ 05rfloor (2)

Algorithm 1 describes the pseudocode of PSOTL Thealgorithm starts by initializing the swarm (Line (1)) whichincludes both positions and velocities of the particles Thecorresponding personal best 119901119894 of each particle is randomlyinitialized and the leader 119887 is computed as the best particleof the swarm Then for a maximum number of iterationseach particle is updated (Line (4)) quantised (Line (5)) andevaluated (Line (6)) according to a scenario instance 120601(city)At the end of each iteration leader 119887 is also updated (Line(8)) Finally the best solution (cycle program in individual 119887)found so far is returned

332 DETL This algorithm also performs a population-based search but following in this case the operation schemeas specified by the Differential Evolution algorithm (versionDErand1) [44]

As shown in Algorithm 2 after initializing the populationin (Line (1)) the individuals evolve for a number of itera-tions performing differential operators (Line (4)) After thissolutions are quantised (Line (5)) and evaluated (Line (6))according to a scenario instance 120601(119888) At the end of eachiteration each particle is either selected or not (Line (7))depending on whether it outperforms the previous one in theevolution procedure Finally the best solution (cycle programin particle V) found so far is returned

333 RS Random search is included here not as a seriouscompetitor but as a sanity check to find out whether anintelligent algorithm is actually needed In RS at eachiteration step a new solution vector of integer variables israndomly generated (uniformly) in the range of [5 60] Thenew individual replaces the previous one if it is better

334 SCPG This is a deterministic algorithm provided bythe SUMO [42] package for generating cycle programs Thistechnique incorporates actual scheduling information usedby human experts in the domain Cycle programs generatedby SCPG are actually running our city traffic light systemsat present This algorithm consists of assigning fresh valuesin the range [6 31] to the phase durations according to threedifferent factors

(1) the proportion of green states in the phases(2) the number of incoming lanes into the intersection(3) the braking time of the vehicles approaching the

traffic lights

34 Feature Models Feature models (FMs) provide a way tomake a compact representation for modeling the commonand variable features of a system their relationships andthe constraints between them FM can be understood as atree of concepts where the nodes are the features (whichare depicted as labeled boxes) and the edges represent therelationships between them Thus an FM denotes the set offeature combinations In an FM each feature (except the root)has one parent feature and can have a set of child featuresNote here that a child feature can only be included in a featurecombination of a valid product if its parent is included as


Input A scenario instance 120601(119888) of a given city 119888Output Best found solution V encoding a cycle program(1) 119875 larr initializePopulation()(2) while 119892 lt MAXIMUM

119892do

(3) for each individual V119894119892do

(4) 119908119894

119892+1larr differentialOps(V119892 119865) differential mutation and crossover as in DErand1

(5) 119902119894

119892+1larr Q(119908119894



(7) V119894119892+1

larr selection(V119894119892 ℎ119894

119892+1) differential selection of new vector ℎ119894

119892+1if better

(8) end for(9) end while

Algorithm 2 Pseudocode of DETL

Basic car (Bc)

Transmission (Ts) Car Body (Cb) Air Conditioning (Ac) 90 Engine (En) GPS (Gp) 15

Manual (Ma) 75 Automatic (Au) 5Electric (El) 25 Gasoline (Gs) 90

Figure 2 Basic car feature model

well The root feature is always included In order to illustratethese concepts in Figure 2 we show an instance from theSPLOT repository [45] of a basic car This instance defines24 different configurations combinations and four kinds offeature relationships

(i) Mandatory features are depicted as filled circlesA mandatory feature is selected whenever itsrespective parent feature is selected Featurescalled Transmission (Ts) Car Body (Cb) andEngine (Eg) are mandatory

(ii) Optional features are depicted with an empty circleAn optional feature may or may not be selectedif its respective parent feature is selected Fea-tures Air Conditioning (Ac) and GPS (Gp) areoptional

(iii) Exclusive-or relationships are depicted as empty circlescrossed by a set of lines connecting a parent featurewith its child features They indicate that exactly oneof the features in the exclusive-or group must beselected whenever the parent feature is selected Iffeature Transmission (Ts) is selected then eitherfeature Manual (Ma) or feature Automatic (Au)must be selected

(iv) Inclusive-or relationships are depicted as filled circlescrossed by a set of lines connecting a parent featurewith its child featuresThey indicate that at least one ofthe features in the inclusive-or groupmust be selectedif the parent is selected If feature Engine is selected

then at least one of the features Electric (El) orGasoline (Ga)must be selected

Besides the parent-child relationships features can alsorelate across different branches of the FM with the so-calledCross-Tree Constraints (CTC) These constraints as well asthose implied by the hierarchical relationships between fea-tures are usually expressed and checked using propositionallogic [46] These FMs could be extended to take into accountfeature priority In our example we indicate the weightassociated with the optional features as a percentage themandatory features should always be present (do not need tobe weighted) In the following paragraphs we summarize thebasic terminology that we use in this paper related to featuremodels

A configuration is a pair [sel sel] where sel and sel arethe sets of selected and unselected features respectively Inthis paper a configurationwill be a traffic scenario Regardingpriorities each feature 119891 has a weight 119891

119908in the range [0 1]

Then theweight119891119908is 1minus119891

119908 A prioritized pair pc is composed

of two features 1198911and 119891

2and a weight such that the pair

weight is calculated as the product of the weight of 1198911and 1198912

119901119888119908= 1198911119908lowast 1198912119908 Consequently the configuration priority is

calculated as the sum of the prioritized feature pairs whichare present in the configuration It should be noted that aconfiguration is valid in FM iff it does not contradict anyimplicit or explicit constraints introduced by the FM

Combinatorial interaction testing (CIT) is a constructiveapproach that builds test suites in order to systematicallytest all configurations of a system [16] In this paper we aregoing to use the pairwise coverage criterion which is the


Validation

Modeling

(i) Traffic model design(ii) Cross-Tree Constraints

definition

Generation

(i) PGFM design(ii) Scenarios generation

(i) Cycle programssimulation on thegenerated scenarios

(ii) Experimentalanalysis

(iii) Validated cycleprograms selection

Optimization

(i) Selection of algorithms PSOTL DETL RS and SCPG

(ii) Multiple cycle programs generation

Figure 3 Modeling optimization generation and validation phases of our proposed strategy

most popular method in CIT and is based on the assumptionthat most errors originating in a parameter are caused by theinteraction of two values [47] This criterion is satisfied ifall feature pairs ([(119891

1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are

present in at least one configuration of the test suite Whenthis technique is applied to FMs the idea is to select a setof valid configurations where the errors could be presentwith higher probability In addition the prioritization of thefeatures makes it possible to first test the more frequent ormore important features

4 Validation Strategy

At this point once we have explained some preliminaryrequired concepts we describe the main steps involved inour validation strategy for the sake of a better understandingof this study Figure 3 illustrates the general scheme ofthis procedure which consists of four main phases cycleprograms optimization Feature Model Design ScenariosGeneration and Cycle Programs Validation

A brief explanation of how the four phases of ourvalidation strategy are carried out is given in the following

(1) Optimization As described in Section 3 given anurban scenario related to a certain area in a real citythe cycle programs of traffic lights operating in thisarea are optimized by means of several specializedalgorithms PSOTL DETL RS and SCPG describedin Section 33 As most of these algorithms are basedon stochastic procedures a number of different can-didate solutions (representing traffic light programs)appear that must be thoroughly analyzed with theaim of selecting the most accurate one for a widerange of traffic conditions Note that we have used

a neutral scenario for the optimization of the solu-tions (traffic light programs) In a later step in phase 4(Validation) we validate these solutions withmultiplescenarios with different characteristics generated inphase 3 (Generation)

(2) Modeling This phase entails the feature model(Section 34) design for our traffic scenarios At thisstage the human expert has to select the features andconstraints and decide how the priorities are assignedto the featuresMore details about the trafficmodelingare given in Section 5

(3) Generation The generation of scenarios is automati-cally carried out by means of our Prioritized Geneticalgorithm for feature models (detailed in Section 6)In this way we can obtain a test suite of scenarios thatfulfill the pairwise coverage criterion In addition thegenerated scenarios are ordered by priority accordingto the combination of features that are involved in thespecific urban area Note that the scenarios generatedare only used for the validation and not for creatingbetter schedules by the optimizers

(4) Validation Finally the optimized cycle programs arethen validated with regard to the generated scenariosfollowing our traffic FMThe experimental procedureleading up to accomplishing this phase is described inthe experimental section (Section 8) and the result-ing valid cycle programs are analyzed according to thestakeholder interests

In addition there exist some dependencies betweenthese phases that have to be considered for example thetraffic scenarios cannot be generated unless the traffic featuremodel has first been defined whereas the cycle programs


Traffic management (T)

Weather (W)

Rainy (Ra) 16 Sunny (Su) 84 Windy (Wd) 30 Foggy (Fg) 7

Stormy (St) 3

Emergency (Em)

Yes (Ys) 50 No (No) 50

Vehicles amount (Va) Driver imperfection (DI) Driver reaction (DR) Vehicle type (Vt)

Few (Af) 50 Many (Am) 50

Low (Il) 30 Medium (Im) 30 High (Ih) 30

Slow (Rs) 20 Medium(Rm) 50 Fast (Rf) 80

Light (Li) 90 Heavy (Hv) 10

Exclusive-or Exclusive-or

Exclusive-orInclusive-or

Exclusive-or

Exclusive-or

Vehicles (V)

Cross-tree constraints(1) Su rarr Ra(2) St rarr Ih and Rf(3) Su rarr Il and Rs

Figure 4 Traffic management feature model with priorities

optimization phase could be done in parallel to the definitionof the generation of the feature model and scenarios Theoptimization phase is independent from the definition ofthe traffic FM because we optimise the cycle programs on aneutral scenario which is not affected by any feature definedby the traffic FM Finally the cycle program validationdepends on the generated traffic light programs and the trafficscenarios

After applying these steps we are now able to numericallyevaluate a set of cycle programs of traffic lights in order toobjectively select the overall best taking into account severalscenarios With this goal we use some traffic measures inorder to evaluate a cycle program of traffic lights accordingto the desirable behavior of the whole traffic system In otherwords we select the best cycle program depending on thestakeholder interest such as saving fuel reducing emissionsor avoiding traffic jams

5 Modeling Traffic Representation withFeature Models

Trafficmanagement is a hard task that involves lots of varyingfeatures Traffic is a highly variable system that could provokea wide range of possible scenarios Therefore it could berepresented by a feature model which could be used formodeling the common and variable features of a system andtheir relationships In addition not every traffic scenario hasthe same importance or probability of occurrence So wehave extended the FM to prioritize features in order to firsttest the most important scenarios

Figure 4 shows the generated FM for representing thetraffic management system Among the main features thatcould describe a traffic scenario we have taken into accountseveral conditions affecting the traffic such as the weather if

there exists an emergency situation and the different vehiclersquoscharacteristics Among these vehicle features we have thenumber of vehicles type of vehicle driver imperfection anddriver reaction time In the FM the features are defined asfuzzy however in Section 7 we set concrete values corre-sponding to the traffic characteristics available in the SUMOsimulator For some of these features we select the weightsaccording to some real-world data that we describe in thefollowing When no data is available to justify a weight wesimply assign equal weights to all the possible values In whatfollows we are going to justify the chosen features

(i) Weather The weather types we have considered areRainy Stormy Sunny Windy and Foggy They arerelated with a inclusive-or so at least one featureshould be selected In order to assign the weightswe have consulted the data of the Spanish NationalInstitute of Meteorology concretely the data of 2012from the Malaga airport weather station The weightassigned is the percentage of days of the year that theassociated condition held

(ii) Emergency We have considered whether an emer-gency situation has occurred or not because it couldbe important to know how traffic flow evolves in anemergency situation Emergency values consideredare Yes and NoWhen there is an emergency situationin the whole network the reaction time is shorter andthe traffic light programs have less time to help thedriverrsquos arrive at their destination They are related toan exclusive-or so only one feature can be selectedThey are equally weighted

(iii) Vehicle We represent the main vehicular character-istics under this mandatory feature For some childnodes of the Vehicle feature it was not possible


to obtain reliable public data so in those cases weconsidered equal weight for the features All childnodes of the Vehicle feature are also mandatory andare as follows

(1) Vehicles Amount The number of vehicles con-sidered is Few or Many They are related to anexclusive-or so only one feature can be selectedThey are equally weighted

(2) Driver ImperfectionThis feature represents howbad a driver is (low medium or high) forexample a high driver imperfection rate is abad driver In many traffic situations the drivermakes the difference So we have consideredthree levels of expertise They are related to anexclusive-or so only one feature can be selectedThey are equally weighted

(3) Driver Reaction Time Time spent by the driverto carry out an action in the vehicle We haveconsidered three levels of reaction time (SlowMedium and Fast) in which the feature Fasthas a larger priority than Medium Medium haslarger priority than Slow They are related to anexclusive-or so only one feature can be selected

(4) Vehicle Type We have taken into account twotypes of vehicles Light and Heavy In orderto assign the weights we have consulted publicdata of the traffic control center (city hall) ofMalaga Most of the vehicles are light so wehave assigned a weight of 90 to them Theyare related to an exclusive-or so only one featurecan be selected

Besides the parent-child relationships features can alsobe related across different branches of the feature model withthe so-called Cross-Tree Constraints (CTC) In this model wehave generated three CTCs that are shown in the upper rightcorner of Figure 4 and are as follows

(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908

Let us explain our interpretation of the CTCs includedin our traffic feature model It is impossible for the weatherto be sunny and rainy at the same time We think this firstconstraint does not need any further explanation On theone hand when the weather is stormy we consider that thedriverrsquos imperfection must be high and the driverrsquos reactionmust be fast because the driver is payingmuchmore attentionwhen the weather is bad On the other hand when theweather is sunny we consider that the driverrsquos imperfectionmust be low and the driverrsquos reaction must be slow becausethe driver is much more relaxed

Finally this model represents 960 valid scenarios withdifferent levels of importance and possibility of occurrenceThe ideal situation is to generate an optimized traffic light

program for each scenario but this could be costly Consider-ing our previous results in programoptimization for synchro-nizing traffic lights [9] the generation of 960 different cycleprograms could take around 19 years of computation timeMoreover here we have applied four different algorithmsso the computation time is multiplied by four resulting ina total of 76 years of computation time For this reasonthe use of automatic intelligent algorithms to reduce thetesting scenarios is mandatory for this task Therefore inthe next section we explain how we reduce the number oftesting scenarios without loosing the capacity of measuringthe quality of the cycle programs

6 Generation PGFM Algorithm forScenarios Generation

The Prioritized Genetic Feature Model (PGFM) algorithmis an evolutionary approach that constructs an entire testsuite taking into account the feature model the constraintsbetween the features and their priorities in the generation ofthe test suite of traffic scenarios It is a constructive algorithmthat adds one new valid scenario to the partial solution ineach iteration until all pairwise combinations of features arecovered In each iteration the algorithm tries to find the validscenario that adds more coverage to the partial solution

Algorithm 3 sketches the pseudocode of PGFMAs inputit receives a feature model for generating the test suiteInitially the test suite is initialized with an empty list (Line(1)) and the set of remaining pairs (RP) is initialized withall the valid weighted pairwise combinations of features(Line (2)) In each iteration of the external loop (Lines(3)ndash(21)) the algorithm creates a random initial populationof individuals (Line (5)) and enters an inner loop whichapplies the traditional steps of a generational evolutionaryalgorithm (Lines (6)ndash(18)) That is some individuals (trafficscenarios in our case) are selected from the population 119875(119905)recombined mutated evaluated and finally inserted in theoffspring population 119860 An individual only contains theselected features so the operators only affect these featuresThe nonselected features are those which are not contained inthe individual If a generated offspring individual is not a validscenario (ie it violates a constraint derived from the featuremodel) it is transformed into a valid scenario by applying aFix operation (Line (12)) provided by the FAMA tool [48]

The fitness value of an offspring individual (Line (13)) isthe sumof the weights of the weighted pairwise combinationsthat would still to be covered after adding the offspringsolution to the test suite This is a minimization fitnessfunction that promotes the generation of scenarios that coverthe pairs of features with higher weights Note that as thesearch procedure advances the cost of computing the fitnessfunction decreases since each time fewer weighted pairwisecombinations remain uncovered

The best individuals of 119875(119905) and 119860 are kept for the nextgeneration in119875(119905+1) (Line (16))The internal loop is executeduntil the maximum number of evaluations is reached Afterthis the best individual found is included in the test suite(Line (19)) and RP is updated by removing the weighted


Input Traffic Feature Model (tfm)Output Prioritized Test Suite(1) TSlarr 0 Initialize the test suite(2) RPlarr weighted pairs to cover (tfmfeatures)(3) while not empty (RP) do(4) 119905 = 0

(5) 119875(119905) larr Create Population() 119875 = population(6) while 119890V119886119897119904 lt 119905119900119905119886119897119864V119886119897119904 do(7) 119860 larr 0 119860 = auxiliary population(8) for 119894 larr 1 to (1199011199001199011199061198971198861199051198941199001198991198781198941199111198902) do(9) parentslarr Selection(119875(119905))(10) offspring larr Recombination(119901119886119903119890119899119905119904)(11) offspring larr Mutation(119900119891119891119904119901119903119894119899119892)(12) Fix(119900119891119891119904119901119903119894119899119892)(13) Ekaluate Fitness(119900119891119891119904119901119903119894119899119892)(14) Insert(119900119891119891119904119901119903119894119899119892 119860)(15) end for(16) 119875(119905 + 1) fl Replace(119860 119875(119905))(17) 119905 = 119905 + 1

(18) end while computing the best test case(19) TSlarr TS cup 119887119890119904119905119878119900119897119906119905119894119900119899(119875(119905))

(20) RemokePairs(RP 119887119890119904119905119878119900119897119906119905119894119900119899(119875(119905))) Remove the covered pairs(21) end while computing the best test suite

Algorithm 3 Pseudocode of PGFM

pairwise combinations covered by the selected best solution(Line (20))Then the external loop starts again until there areno weighted pairs left in the RP set

We set the configuration parameters of PGFMwith valuesfrequently used for genetic algorithms one-point crossoverstrategy with a probability of 08 selection strategy binarytournament population size of 10 individuals mutationthat iterates over all selected features of an individual andreplaces a feature by another randomly chosen feature with aprobability of 01 and termination condition of 1000 fitnessevaluations and full weight coverage in the external loop

As a result of the execution of PGFM Table 1 shows a listof 10 prioritized scenarios for testingThis list of scenarios (119878

119894

with 119894 isin [1 10]) fulfills the pairwise coverage criterion whichensures that all distinct traffic conditions are considered inour validation model In this table feature abbreviations inthe first row are defined in Figure 4 corresponding to labelsin the FM tree A feature is included in the scenario if itis marked in the scenariorsquos row The last column indicatesthe total weight of the scenario according to the priorityof the features included In fact we have to highlight thereduction achieved by considering only 10 scenarios frominitial 960 valid scenarios although all the possible scenariosthat need to be explored are 225This is a successful reductionespecially if we consider the time spent for an exhaustiveexperimentation (57 years on one single CPU) compared tothe actual 83 days spent

In this way shown in Table 1 the use of priorities led usto only execute the six most important scenarios since theyguarantee 99 coverageThis indicates that several scenariosshould be considered but some of them are more importantthan others and so should be tested first that is 119878

1scenario

adds 6359 coverage then 1198781covers 6359 of the total

weight of feature pairsThe optional features that are selectedin this scenario (119878

1) are as follows it is sunny the simulation is

long there are a lot of vehicles the driver skills are mediumthe driver reaction is fast and there is a high percentage oflight vehicles If we test under the conditions defined by 119878

1we

know that the results provided of our optimization algorithmfor finding the traffic light cycleswill cover 6359of city totalconditions which is a lot Doing so we add completeness toour study and weight to its robustness in a holistic scientificanalysis of the whole city

7 Generation of Traffic Light ProgramsMalaga City Case Study

As we are interested in developing an optimization solvercapable of dealing with close-to-reality and generic urbanareas we have generated an instance by extracting actualinformation from real digital maps From this instanceconsidering the scenarios computed with PGFM algorithmextracted fromour FM(different trafficdensity weather etc)the optimized cycle programs were generated The selectedurban area covers approximately 750000m2 and is physicallylocated in the city of Malaga in Spain The information usedis all real and concerns traffic rules traffic element locationsbuildings road directions streets intersections and so forthMoreover we have set the number of vehicles circulatingas well as their speeds by following current specificationsavailable from the Mobility Delegation of the Malagarsquos CityCouncil This information was collected from sensorizedpoints in certain streets obtaining a measure of traffic densityat different time intervals


Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001


Table 2 Relationship between features and simulator parameters

Features Ra St Su Wd Fg Ys No Af Am

Parameters 120590+ = 01 120590+ = 01 120590minus = 01 120590+ = 01 120590+ = 01 500 s 1000 s 250 u 500 u120591+ = 1 120591+ = 1 120591minus = 1 120591+ = 1 120591+ = 1

Features Il Im Ih Rs Rm Rf Li Hv mdashParameters 120590minus = 01 minus 120590+ = 01 120591+ = 1 minus 120591minus = 1 95 vl 85 vl mdash

Maacutel

aga

Google map OpenStreetMap SUMO

Figure 5 Malaga scenario instance Selection from OpenStreetMaps and exportation to SUMO format

In Figure 5 the selected area of Malaga city is shownwith its corresponding captured views ofOpenStreetMap andSUMO (as explained in Section 31) In the zone between thecity center and the harbor this instance comprises streetswith different widths and lengths and several roundaboutsThe main streets found in this area are Alameda PrincipalAndalucıa Manuel Agustın Heredia Colon and AuroraThearea contains 70 intersections with 4 to 16 traffic lights at eachone adding up to a total number of 312 traffic lightsThere arebetween 250 and 500 vehicles circulating during the analysisperiod each one of the vehicles completes its own route fromorigin to destination circulating with a maximum speed of50 kmh (typical in urban areas) The traffic flow is generatedby means of the DUARouter tool [42] that computes vehicleroutes used by SUMO using shortest path computation anddynamic user assignment (DUA) algorithms

With regard to the driverrsquos features there are two mainaspects to characterize the driving style imperfection andreaction represented in SUMObymeans of parameters120590 and120591 respectivelyThefirst one is a real number in the range [0 1]for weighting the driverrsquos skills A high value of 120590means thatthe driver commits many driving errors In contrast a low120590 means good driving The second parameter (120591) measuresthe driverrsquos reaction time in seconds In addition we considerother parameters such as the simulation time the numberof vehicles and the percentage of light and heavy vehiclesA heavy vehicle has lower acceleration deceleration andmaximum velocity than a light vehicle The optional featuresselected from our traffic model are the source of variabilityleading to different traffic scenarios

Table 2 summarizes the parameter settings in terms ofdriving and simulation values with respect to each selectedfeature For example when selecting the feature Ra (rainy)parameters 120590 and 120591 are increased by 01 and 1 respectivelyRegarding the emergency feature if Yes is selected the timeto reach the destination is 500 seconds but when No isselected this time is 1000 seconds Note that the simulation

time is set according to the emergency feature So we havescenarios with different analysis times Finally when featureLi is selected there are 95 of light vehicles and 5 of heavyvehicles In contrast when Hv is selected there are 85 oflight vehicles and 15 of heavy vehicles

The trace information obtained after the simulation ofeach traffic light program for a given scenario is used tocompute a set of metrics vehicles arriving at their destination(VLL) vehicles not arriving at their destination (VNLL) CO

2

emissions (CO2) NO

119909emissions (NO

119909) fuel consumption

(FUEL) and global journey time (GJT) In this way it ispossible to numerically quantify how accurate a given cycleprogram is and compare it with all other existing solutionsfromhuman expertrsquos programs or from automatic optimizers

8 Experiments

This section describes a series of experiments and analysesof the results obtained so as to empirically test our approachThis allows us to put here into practice all initial specificationsof our strategy explained in Section 4 in order to validate theresulting traffic light programs in the ten scenarios generatedby means of the execution of the PGFM algorithm

81 Experimental Setting As stated in the introduction weinclude a varied analysis including four specialized algo-rithms for computing cycle programs of traffic lights PSOTLDETL RS and SCPGThe first three optimization algorithmsare nondeterministic so we have carried out 30 runs and con-sequently have obtained 30 different (best) cycle programs foreach algorithmThe fourth algorithm (SCPG) is a determinis-tic technique so it only computes one optimal cycle programAs our aim is to validate these cycle programs in differentscenarios in the case of nondeterministic algorithms wehave carried out 30 independent runs of the simulator pergenerated scenario (10) algorithm (3) and proposed best


FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx

Figure 6 Normalized star diagram containing traffic accuracymetrics for all algorithms compared

cycle program (30) adding up to a total number of 27000executions This accounts for the considerable effort wehave made to study the city in really different conditionsof the traffic flow to attain realistic results In the case ofSCPG we have performed 300 executions (10 scenarios times 30independent runs)

In order to check whether the differences betweenthe algorithms are statistically significant or just a matterof stochastic noise we have applied the nonparametricWilcoxon rank-sum test [49] The confidence level was setto 95 (119901 value below 005) In addition so as to properlyinterpret the results of statistical tests it is always advisableto report effect size measures For that purpose we havealso used the nonparametric effect size measure

12statistic

proposed by Vargha and Delaney [50] Effect size providesinformation about the magnitude of an effect which can beuseful in determining whether it is of practical significance ornot

All the executions were run in a cluster of 16 machineswith Intel Core2 Quad processors Q9400 (4 cores per proces-sor) at 266GHz and 4GB memory running Ubuntu 12041LTS managed by the HT Condor 784 cluster manager Inorder to offer an approximate idea of the required computa-tional effort to solve this realistic task we have to note that interms of a single coremachine our complete experimentationwith 27300 independent runs would require about 8 monthsof computation In contrast in the used parallel platform thecomputation of the valid scenarios required 83 days (2626seconds per run)

82 Experimental Results In order to illustrate the manyresults obtained at a glance Figure 6 shows the performanceof algorithms for the selected metrics (VLL CO

2 NO119909 Fuel

and GJT) for each vehicle and all scenarios These metricsare normalized for a proper visualization In this figure afirst observation is that parameters CO

2 NO119909 and Fuel are

Table 3 Best average VLL for each algorithm and scenario Thespecific cycle program (out of 30) that generates the best result forthis algorithm and scenario is indicated in parenthesis

PSOTL DETL RS SCPG1198781

9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928

correlated in almost all algorithms so we could considerto deal with them as one single factor Nevertheless as oneof our main focus is environmental resources consump-tionemission saving we decided to use these parameters asnonaggregate values The fact of getting correlated results isdue to the simulation strategy that indeed follows a realisticmodel (HBEFA)

A second observation in Figure 6 is that PSOTL obtainsthe maximum number of vehicles arriving at its destinationwith the lowest fuel consumption and also with the lowestemissions of CO

2and NO

119909 However for this algorithm the

global journey time (GJT) is the highest oneThe explanationof this counter-intuitive result is that a vehicle which does notarrive at its destination in the analysis period is not used forthe computation of the metrics Thus several vehicles do notarrive at their remote destinations so the global journey timewill not be increased by the stats of the vehicles with remotedestinations In Figure 6 we also note that SCPG is the worstalgorithm in terms of fuel consumption and CO

2and NO

119909

emissions even though it had a low number of vehicles thatarrive at their destination This means that the problem ishighly difficult and lacks clear patterns for experts when wego to a large scale optimisation scenario

In fact as increasing the number of vehicles that arriveat their destination during the study timeframe (VLL) is animportant metric for evaluating how the system preventstraffic jams we have organized in Table 3 the average VLLfor each scenario and algorithmWe have also highlighted theparticular cycle program that obtains the best average of VLLin the 30 independent executions Therefore we can easilysee that the cycle programs generated by PSOTL are alwaysthe best at preventing traffic jams since they guarantee thatalmost all vehicles reach their destination within the analysistime However the cycle program that obtains the best resultis not always the same in all scenarios For example forPSOTL programs 3 and 20 (see the numbers in parenthesis inTable 3) obtain the best results in 3 out of 10 scenarios Thismeans that there is no single cycle program that shows thebest results in all possible traffic scenarios as expectedTheseresults lead us to consider in Section 9 the priority assignedto each scenario in order to choose the best cycle program


Table 4 Number of scenarios where significant differences exist for VLL between the algorithms (120588) and Vargha and Delaneyrsquos statisticaltest results (

12)


12120588

12120588

12120588

12

PSOTL mdash 8 07901 10 07129 4 08303DETL 8 02099 mdash 6 03831 0 05503RS 10 02871 6 06169 mdash 3 06657SCPG 4 01697 0 04497 3 03343 mdash

from all of them In this way an unexpected change in trafficconditions will not end in an enormous traffic jam becausethe overall best cycle program is in some way more adaptiveto a greater number of traffic conditions than an overfittedcycle program

In order to check whether the differences between thealgorithms are statistically significant or not we have appliedthe nonparametric Wilcoxon rank-sum test In Table 4 weshow the number of scenarios where significant differencesexist between the algorithms In this regard we can see thatPSOTL is significantly different to RS for all the scenarios(thus an intelligent algorithm is needed) whereas solutionsprovided by DETL are not statistically different to the onesprovided by the human experts represented in SCPG (soDETL is just equivalent to the expertrsquos solutions not better)RS is significantly better than SCPG and DETL by 3 and 6times respectively

Table 4 also reports the effect size measure 12

for theVLL metric for all scenarios We interpret the

12statistic

as follows given a performance measure 119872 12

measuresthe probability that running algorithm 119860 yields higher 119872values than running another algorithm 119861 In this regard119860 represents algorithms in the columns and 119861 representsalgorithms in the rows If these two algorithms are equivalentthen

12= 05 A value of

12= 03 entails that one would

obtain higher values for119872with algorithm119860 30 of the timeAs shown in this table PSOTLrsquos solutions are the best for allscenarios and their differences are of actual importance inpractical cases of the city Numerically speaking these cycleprograms (the ones generated by PSOTL) are better than theones provided by DETL RS and SCPG by 7901 7129and 8303 respectively In fact these results are labeled aslarge effect size according to Cohenrsquos 119889 scale [51]These resultsprovide us with some insights into the successful applicationof PSOTLrsquos cycle programs to real traffic light schedules

83 Focusing on Scenarios From the point of view of thedifferent generated scenarios in Figure 7 we can observethe box plots of resulting traces in terms of metrics forall algorithms Box plots display variation in samples of astatistical population without making any assumptions ofthe underlying statistical distribution Box plots show themean the interquartile range (in color) and the maximumand minimum value on the extremes In the box plots ofFigure 7 there are no outliers which are observation pointsthat are distant from other observations Four metrics areused two of them measured per vehicle (CO

2and GJT)

Specifically the second subfigure shows the percentage oftime that the journey takes compared to the total analysistime for example 40 value means that the average vehicletakes 40 of the analysis time to reach to its destinationTheother twomeasures used are VLL andVNLL Note that we donot show the resulting box plots for VNLL because they arethe complementary results obtained for VLLThese diagramsfocus on the scenarios in order to show the variability oftrafficmeasures while at the same time we have tried to showthe results without the specific bias introduced by a particularsolving technique

An interesting observation in Figure 7 lies in the fact thatScenario 6 (119878

6) is the most favorable for the GJT and VLL

indicators whereas for the CO2indicator the third best result

is obtainedThe 1198786scenario induces a successful performance

of cycle programs In fact the main features of 1198786are ideal for

enhancing the traffic flow sunny weather (Su) no emergency(No) few vehicles (Af) and more drivers with a high level ofexpertise (Il) In contrast the most unfavorable scenario is1198785 since only a few vehicles arrive at their destination and

the CO2thrown up into the atmosphere is the highest for

each vehicle in our study This last scenario has unfavorableweather conditions rainy (Ra) stormy (St) and foggy (Fg) Inaddition there is an emergency (Ys) the number of vehiclesis higher (Am) and there are more heavy vehicles (Hv) Allthese features induce adverse conditions in this scenario (119878

5)

which negatively influence the traffic flowIn light of these results we claim that providing experts

with better general programs is possible by using ourapproach We consider different traffic conditions in a setof modeled scenarios which provides the experts with unbi-ased traffic light programs In contrast the aforementionedliterature (see Section 2) just offers ideal traffic conditionsIn addition here we go one step beyond by weighting thosefeatures with more probability of occurrence but withoutmissing those traffic situations that are more extreme

9 Prioritized Analysis

In this section as far as we know we are the first to applyprioritization techniques that consider all possible trafficscenarios at a time We analyze how each generated cycleprogram works for all scenarios as a whole Since not allscenarios are equally important or frequent the computedmetrics are weighted according to the scenariorsquos priority

91 Analysis of Best Performing Cycle Programs Designingrobust traffic light programs is a mandatory task when


S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL

Figure 7 Boxplots of the resulting distribution traces of our studied metrics CO2 GJT and VLL

tackling vehicular urban environments In this regard con-sidering all modeled scenarios as a whole helps us to give arobustness to our solutions so we have therefore weightedthese scenarios according to their priority

In Table 5 we report our top ten cycle programs for thefollowing metrics VLL FUEL CO

2 and GJT We have to

highlight that the PSOTL13 cycle program is the best solutionin terms of VLL for the most prioritized scenario (119878

1in

Table 3) However whenwe consider the scenarios altogetherPSOTL13 is not the best for VLL so it appears in secondplace for this metric in this rankingThis fact was unexpectedbecause of the high weight of 119878

1 nevertheless the best one

for all scenarios as a whole (PSOTL20) is the best in threedifferent scenarios (119878

4 1198787 and 119878

10in Table 3)

In Table 5 we can observe the cycle program that is thebest for each metric in all scenarios together So we onlyhave to decide which metric is more important for us andthen set a particular configuration It is possible that differentstakeholders could be interested in setting up different cycleprograms For example the municipal authorities of the citymay be interested in avoiding traffic jams so the GJT metricshould be optimized However the national government

could impose a CO2emissions maximum rate so the CO

2

metric also ought to be minimized thus the best optionwould be to configure the traffic lights with the ProgramPSOTL20

92 Practical Benefits Our validation strategy providesexperts with many practical benefits In general we havealleviated the work of the experts with the cost reductionthis implies just by setting the priorities in the featuremodel From the featuremodel they obtain several validationscenarios that accomplish the pairwise coverage criterion andprovide us with some confidence for choosing the best trafficlights program In addition the priorities of scenarios allowus to select a good solution for most traffic conditions takinginto account theweight of the scenarios previously computedby means of the PGFM algorithm

Another practical benefit is the flexibility of the proposedapproach Since it is impossible to objectively decide whichcycle program of traffic lights is the best we provide severalsolutions according to the desired behavior of the systemAs we have previously said it is possible that differentstakeholders could be interested in setting up a different cycle


Table 5 Ordered best performing cycle programs after considering all weighted scenarios and the expertrsquos solution (SCPG) PSOTLs areoverwhelmingly the best in all cases DETLs just have a minor role regarding GJT

VLL-weighted Fuel-weighted CO2-weighted GJT-weighted

Program Value Program Value Program Value Program ValuePSOTL20 35920 PSOTL24 1825 PSOTL20 01477 DETL29 37586PSOTL13 35817 PSOTL23 1826 PSOTL13 01520 DETL28 37645PSOTL2 35445 PSOTL22 1828 PSOTL28 01525 DETL26 37847PSOTL5 34779 PSOTL25 1831 PSOTL14 01542 DETL16 38105PSOTL14 34472 PSOTL21 1833 PSOTL16 01552 DETL25 38133PSOTL1 34299 PSOTL29 1834 PSOTL2 01553 PSOTL26 38244PSOTL3 34269 PSOTL20 1835 PSOTL27 01554 DETL14 38605PSOTL27 34258 PSOTL28 1837 PSOTL17 01561 DETL11 38610PSOTL16 34167 PSOTL27 1840 PSOTL3 01571 PSOTL28 38617PSOTL17 34158 PSOTL26 1841 PSOTL23 01576 DETL13 38706SCPG 20017 SCPG 2444 SCPG 03353 SCPG 39927

program For example solution PSOTL20 is the best for themetrics of vehicles that arrive at their destination (VLL)however it is the seventh in fuel consumptionMoreover thissolution is not in the top ten ranking of GJT per vehicle sowe have obtained robust cycle programs for all scenarios buta different one if you are interested in a different metric

The benefits of comparing the cycle programs of trafficlights in the proposed scenarios can be now numericallymeasured We have compared the expertrsquos solution and thebest automatically generated solution for each metric (seevalues of PSOTL20 with regard to those of SCPG in Table 5)The improvement achieved in solution quality is remarkableespecially for CO

2emissions in which we have obtained

an improvement of 12699 Regarding the vehicles thatarrive at their destination we have obtained an improvementof 7945 with respect to the expertsrsquo solution The fuelconsumption per vehicle is reduced by 3387 which is also ahighly relevant practical result Nevertheless the reduction isslightly lower in the case of vehiclersquos journey time (GJT) butstill better than the expertrsquos solution

Since this may need revising when implemented in areal city (as do most scientific studies) we could expect achange in the percentages we here include However it is soimportant that we have strong evidence that a final practicalbenefit will come from using our approach

10 Threats to Validity

Threats to validity should be considered in any empiricalstudy So we have taken care of those that are applicable toour work

Threats to internal validity come from the fact that wehave used a single standard assignment for the parametervalues of the optimization algorithms based on the authorrsquosexperience A change in the values of these parameters couldhave an impact on the results of the algorithms Neverthelessthe default values found in the literature may already besufficient as analyzed in [52] We have taken into accountthe cost of the tuning phase which could become quite highin the case of this large study involving four metaheuristics

and more than 27000 independent runs This considerablecomputational cost also partially justifies not going furtherin our already large analysis

Regarding external threats we have identified threeof them the assignment of weights to the traffic featuremodel and the selection of the algorithms To address thefirst of these (weights) we have consulted the data of theMobility Delegation of Malaga Hall as well as the SpanishNational Institute of Meteorology more concretely thedata from 2012 (Spanish National Institute of MeteorologyhttpwwwtutiemponetclimaMalaga Aeropuerto201284820htm Mobility Delegation of Malaga Hall httpmovilidadmalagaeu) The weights were assigned accordingto the percentage of days of the year that each conditionoccurred The second external threat is that maybe wehave not chosen the best algorithms but we have includedfour distinct algorithms that represent a diverse varietyof optimization techniques and concepts even so theexperiments took several weeks using a computer clusterAlthough certainly there might be other algorithms thatcould potentially yield better results the ones included arehowever very popular in current research and in fact manyarticles just focus on only one or two of them

The third external threat concerns the generation ofdynamic traffic light programs Our aim here is not to gener-ate cycle programs dynamically during an isolated simulationas done in agent-based algorithms [4] but to obtain optimizedcycle programs for a given scenario and timetable In factwhat real traffic light schedulers actually demand are constantcycle programs for specific areas and for preestablished timeperiods (rush hours nocturne periods etc) which led us totake this focus Our approach is automatic and can be used inreal city traffic centers (in progress) It is an offline step usedto build or improve actual city traffic

11 Conclusions

In this paper we have proposed a validation strategyfor the traffic light programs to be used by the humanexperts of Smart Mobility We have carried out a thorough


experimentation aimed at testing our strategy on a largenumber of different cycle programs generated by automaticoptimizers aswell as by human expertrsquos proceduresThemainconclusions that we can draw are as follows

(i) We propose the use of a traffic feature model withpriorities which allows us not only to reduce thenumber of traffic scenarios to test the available cycleprograms but also to generate themost important sce-nariosThis can be carried out bymeans of the PGFMalgorithm proposed in Section 6 This algorithm isable to automatically generate a minimal prioritizedtest suite from a given feature model Experts cannowwork with a reduced set of scenarios with similarfault detection capacity which means a great costreduction

(ii) Our validation strategy allows the experts to numer-ically quantify the quality of a traffic light cycleprogram on different scenarios This quantification isperformed on different scoring metrics like vehiclesthat arrive at their destination CO

2emissions NO

119909

emissions global journey time and so forth Wecould choose a specific cycle program depending onthe traffic conditions and then the metric we wantto optimize In this way we offer a wide range ofsolutions according to the stakeholder interests

(iii) We have experimentally shown that there is no singlespecific cycle programwhich is the best for all scenar-ios with numerical models and results (not just withintuitive beliefs) Nevertheless for the sake of an easydecision-making process we also provide a methodto rank the cycle programs This is possible becausewe have taken into account the scenariorsquos priorityautomatically computed from our feature model

(iv) With regard to our case study we have analyzedhow the cycle programs generated by four algorithms(PSOTL DETL RS and SCPG) adapt to differenttraffic conditions (ten different scenarios) in Malagacity We have compared the generated cycle pro-grams with the solution provided by the experts(SCPG) Considering the prioritization of scenariosthe improvement achieved in solution quality isremarkable especially for CO

2emissions in which

we have obtained a reduction of 12699 comparedwith the expertsrsquo solutions

As a matter of future work we plan to consider sev-eral scenarios in a single fitness evaluation of solutions inoptimization algorithms Then we expect to enhance thesearch procedure of the algorithms in cycle programs forthe most influential traffic conditions Moreover we plan todeal with a multiobjective model of the problem accordingto stakeholders interests As a result we will provide a Paretofront the objectives of which are theminimization of theCO

2

emissions theminimization of the global journey time or theminimization of traffic jams

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research has been partially funded by the SpanishMinistry of Economy and Competitiveness (MINECO) andthe European Regional Development Fund (FEDER) undercontract TIN2014-57341-R moveON Project (httpmoveonlccumaes) It has been also partially funded by GrantTIN2014-58304 (Spanish Ministry of Sciences and Inno-vation) Regional Project P11-TIC-7529P12-TIC-1519 andby the University of Malaga under contract UMAFEDERFC14-TIC36 Finally Javier Ferrer acknowledges the Grantwith code BES-2012-055967 fromMINECO

References

[1] A Caragliu C Del Bo and P Nijkamp ldquoSmart cities in EuroperdquoJournal of Urban Technology vol 18 no 2 pp 65ndash82 2011

[2] C Harrison B Eckman R Hamilton et al ldquoFoundations forsmarter citiesrdquo IBM Journal of Research and Development vol54 no 4 pp 1ndash16 2010

[3] R G Hollands ldquoWill the real smart city please stand uprdquo Cityvol 12 no 3 pp 303ndash320 2008

[4] S Lammer and D Helbing ldquoSelf-control of traffic lights andvehicle flows in urban road networksrdquo Journal of StatisticalMechanics Theory and Experiment vol 2008 no 4 Article IDP04019 2008

[5] G Lim J J Kang and Y Hong ldquoThe optimization of trafficsignal light using artificial intelligencerdquo in Proceedings of the10th IEEE International Conference on Fuzzy Systems pp 1279ndash1282 Melbourne Australia December 2001

[6] C Karakuzu and O Demirci ldquoFuzzy logic based smart trafficlight simulator design and hardware implementationrdquo AppliedSoft Computing vol 10 no 1 pp 66ndash73 2010

[7] R-S Chen D-K Chen and S-Y Lin ldquoACTAM cooperativemulti-agent system architecture for urban traffic signal controlrdquoIEICE Transactions on Information and Systems vol 88-D no1 pp 119ndash126 2005

[8] J C Spall and D C Chin ldquoTraffic-responsive signal timingfor system-wide traffic controlrdquo Transportation Research Part CEmerging Technologies vol 5 no 3-4 pp 153ndash163 1997

[9] J Garcia-Nieto A C Olivera and E Alba ldquoOptimal cycleprogram of traffic lights with particle swarm optimizationrdquoIEEE Transactions on Evolutionary Computation vol 17 no 6pp 823ndash839 2013

[10] J Sanchez M Galan and E Rubio ldquoApplying a traffic lightsevolutionary optimization technique to a real case lsquoLas Ram-blasrsquo area in Santa Cruz de Teneriferdquo IEEE Transactions onEvolutionary Computation vol 12 no 1 pp 25ndash40 2008

[11] T Nagatani ldquoEffect of speed fluctuations on a green-lightpath in a 2D traffic network controlled by signalsrdquo Physica AStatistical Mechanics and Its Applications vol 389 no 19 pp4105ndash4115 2010

[12] K Kang S Cohen J HessWNovak andA Peterson ldquoFeature-oriented domain analysis (FODA) feasibility studyrdquo Tech RepCMUSEI-90-TR-21 Software Engineering Institute CarnegieMellon University 1990


[13] M Mendonca and D Cowan ldquoDecision-making coordinationand efficient reasoning techniques for feature-based configura-tionrdquo Science of Computer Programming vol 75 no 5 pp 311ndash332 2010

[14] M Johansen O Haugen and F Fleurey ldquoProperties of realisticfeature models make combinatorial testing of product linesfeasiblerdquo inModel Driven Engineering Languages and Systems JWhittle T Clark and T Kuhne Eds vol 6981 of Lecture Notesin Computer Science pp 638ndash652 Springer Berlin Germany2011

[15] M F Johansen Oslash Y Haugen and F Fleurey ldquoAn algorithm forgenerating t-wise covering arrays from large feature modelsrdquoin Proceedings of the 16th International Software Product LineConference (SPLC rsquo12) pp 46ndash55 ACM September 2012

[16] M B CohenM B Dwyer and J Shi ldquoConstructing interactiontest suites for highly-configurable systems in the presence ofconstraints a greedy approachrdquo IEEE Transactions on SoftwareEngineering vol 34 no 5 pp 633ndash650 2008

[17] D KrajzewiczM Bonert and PWagner ldquoThe open source traf-fic simulation package SUMOrdquo in Proceedings of the Infrastruc-ture Simulation Competition (RoboCup rsquo06) Bremen Germany2006

[18] A W Williams and R L Probert ldquoA measure for componentinteraction test coveragerdquo in Proceedings of the ACSIEEEInternational Conference onComputer Systems andApplicationsvol 30 pp 304ndash311 IEEE Beirut Lebanon June 2001

[19] M Grindal J Offutt and S F Andler ldquoCombination testingstrategies a surveyrdquo Software Testing Verification and Reliabilityvol 15 no 3 pp 167ndash199 2005

[20] R Kuhn Y Lei and R Kacker ldquoPractical combinatorial testingbeyond pairwiserdquo IT Professional vol 10 no 3 pp 19ndash23 2008

[21] C Nie and H Leung ldquoA survey of combinatorial testingrdquo ACMComputing Surveys vol 43 no 2 article 11 2011

[22] M B Cohen M B Dwyer and J Shi ldquoInteraction testing ofhighly-configurable systems in the presence of constraintsrdquo inProceedings of the International Symposium on Software Testingand Analysis (ISSTA rsquo07) pp 129ndash139 ACM 2007

[23] R C Bryce and C J Colbourn ldquoOne-test-at-a-time heuristicsearch for interaction test suitesrdquo in Proceedings of the 9thAnnual Conference on Genetic and Evolutionary Computation(GECCO rsquo07) pp 1082ndash1089 ACM London UK July 2007

[24] E Salecker R Reicherdt and S Glesner ldquoCalculating pri-oritized interaction test sets with constraints using binarydecision diagramsrdquo in Proceedings of the 4th IEEE InternationalConference on Software Testing Verification and ValidationWorkshops (ICSTW rsquo11) pp 278ndash285 IEEE Berlin GermanyMarch 2011

[25] B J Garvin M B Cohen and M B Dwyer ldquoEvaluatingimprovements to a meta-heuristic search for constrained inter-action testingrdquo Empirical Software Engineering vol 16 no 1 pp61ndash102 2011

[26] F Ensan E Bagheri and D Gasevic ldquoEvolutionary search-based test generation for software product line feature modelsrdquoin Proceedings of the 24th International Conference on AdvancedInformation Systems Engineering (CAiSE rsquo12) pp 613ndash628Gdansk Poland June 2012

[27] C Henard M Papadakis G Perrouin J Klein P Heymansand Y Le Traon ldquoBypassing the combinatorial explosion usingsimilarity to generate and prioritize t-wise test configurationsfor software product linesrdquo IEEE Transactions on SoftwareEngineering vol 40 no 7 pp 650ndash670 2014

[28] E Engstrom and P Runeson ldquoSoftware product line testingmdashasystematic mapping studyrdquo Information amp Software Technologyvol 53 no 1 pp 2ndash13 2011

[29] P A da Mota Silveira Neto I do Carmo Machado J DMcGregor E S de Almeida and S R de Lemos Meira ldquoAsystematic mapping study of software product lines testingrdquoInformation and Software Technology vol 53 no 5 pp 407ndash4232011

[30] S Oster F Markert and P Ritter ldquoAutomated incrementalpairwise testing of software product linesrdquo in Software ProductLines Going Beyond 14th International Conference SPLC 2010Jeju Island South Korea September 13ndash17 2010 Proceedings JBosch and J Lee Eds vol 6287 of Lecture Notes in ComputerScience pp 196ndash210 Springer Berlin Germany 2010

[31] A Hervieu B Baudry and A Gotlieb ldquoPACOGEN automaticgeneration of pairwise test configurations from featuremodelsrdquoin Proceedings of the 22nd IEEE International Symposium onSoftware Reliability Engineering (ISSRE rsquo11) pp 120ndash129 IEEEHiroshima Japan December 2011

[32] C Blum and A Roli ldquoMetaheuristics in combinatorial opti-mization overview and conceptual comparisonrdquo ACM Com-puting Surveys vol 35 no 3 pp 268ndash308 2003

[33] N M Rouphail B B Park and J Sacks ldquoDirect signal timingoptimization strategy development and resultsrdquo in Proceedingsof the 11th Pan American Conference in Traffic and Transporta-tion Engineering Gramado Brazil January 2000

[34] P Holm D Tomich J Sloboden and C Lowrance ldquoTraf-fic analysis toolbox volume IV guidelines for applying cor-sim microsimulation modeling softwarerdquo Tech Rep NationalTechnical Information Service Springfield Va USA 2007

[35] E Brockfeld R Barlovic A Schadschneider and M Schreck-enberg ldquoOptimizing traffic lights in a cellular automatonmodelfor city trafficrdquo Physical Review E vol 64 no 5 Article ID056132 12 pages 2001

[36] A M Turky M S Ahmad M Z Yusoff and B T HammadldquoUsing genetic algorithm for traffic light control system with apedestrian crossingrdquo in Rough Sets and Knowledge Technology4th International Conference RSKT 2009 Gold Coast AustraliaJuly 14ndash16 2009 Proceedings vol 5589 of Lecture Notes inComputer Science pp 512ndash519 Springer Berlin Germany 2009

[37] L Peng M-H Wang J-P Du and G Luo ldquoIsolation nichesparticle swarm optimization applied to traffic lights control-lingrdquo inProceedings of the 48th IEEEConference onDecision andControl Held Jointly with the 28th Chinese Control Conference(CDCCCC rsquo09) pp 3318ndash3322 IEEE Shanghai China Decem-ber 2009

[38] S Kachroudi and N Bhouri ldquoA multimodal traffic responsivestrategy using particle swarm optimizationrdquo in Proceedings ofthe 12th IFAC Symposium on Transpotaton Systems (CT rsquo09) pp531ndash537 Redondo Beach Calif USA September 2009

[39] K B KesurMetaheuristics inWater Geotechnical and TransportEngineering Elsevier Philadelphia Pa USA 2013

[40] J Garcıa-Nieto E Alba and A C Olivera ldquoSwarm intelligencefor traffic light scheduling application to real urban areasrdquoEngineering Applications of Artificial Intelligence vol 25 no 2pp 274ndash283 2012

[41] J Leung L Kelly and J H Anderson Handbook of SchedulingAlgorithmsModels and Performance Analysis CRC Press BocaRaton Fla USA 2004

[42] M Behrisch L Bieker J Erdmann and D KrajzewiczldquoSUMOmdashsimulation of urban mobility an overviewrdquo in Pro-ceedings of the 3rd International Conference on Advances in


System Simulation (SIMUL rsquo11) pp 63ndash68 Barcelona SpainOctober 2011

[43] M Clerc ldquoStandard PSO 2011rdquo Tech Rep Particle SwarmCentral 2011 httpwwwparticleswarminfo

[44] K V Price R M Storn and J A Lampinen DifferentialEvolution A Practical Approach to Global Optimization NaturalComputing Series Springer Berlin Germany 2005

[45] SPLOT ldquoSoftware Product Line Online Toolsrdquo 2013 httpwwwsplot-researchorg

[46] D Benavides S Segura and A Ruiz-Cortes ldquoAutomatedanalysis of feature models 20 years later a literature reviewrdquoInformation Systems vol 35 no 6 pp 615ndash636 2010

[47] B Stevens and E Mendelsohn ldquoEfficient software testingprotocolsrdquo in Proceedings of the Conference of the Centre forAdvanced Studies on Collaborative Research (CASCON rsquo98) p22 IBM Press 1998

[48] P Trinidad D Benavides A Ruiz-Cortes S Segura andA Jimenez ldquoFAMA frameworkrdquo in Proceedings of the 12thInternational Software Product Line Conference (SPLC rsquo08) p359 Limerick Irland September 2008

[49] D J Sheskin Handbook of Parametric and NonparametricStatistical Procedures Chapman amp HallCRC 4th edition 2007

[50] A Vargha and H D Delaney ldquoA critique and improvement ofthe CL common language effect size statistics of McGraw andWongrdquo Journal of Educational and Behavioral Statistics vol 25no 2 pp 101ndash132 2000

[51] R J Grissom ldquoProbability of the superior outcome of onetreatment over anotherrdquo Journal of Applied Psychology vol 79no 2 pp 314ndash316 1994

[52] A Arcuri and G Fraser ldquoParameter tuning or default valuesAn empirical investigation in search-based software engineer-ingrdquo Empirical Software Engineering vol 18 no 3 pp 594ndash6232013

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttpwwwhindawicom

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Discrete MathematicsJournal of


Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Stochastic AnalysisInternational Journal of


of different scenarios of urban mobility In this regard weaim to extract validated information about which traffic lightsprogram (from among a number of them) is best adaptedfor most traffic scenarios with different feature combinationsand which fails to manage the traffic flow with other specificfeature combinations

In fact among the main features that affect traffic flowwe can enumerate the following different meteorology con-ditions type of vehicle driver expertise simulation timeand so forth However the analysis of all possible trafficscenarios is inviable due to the large number of feature com-binations to take into account For this reason we propose anevolutionary approach called the Prioritized Genetic FeatureModel (PGFM) algorithm designed for the generation ofa minimal prioritized test suite of traffic scenarios whichmeet a coverage criterion (pairwise) The pairwise coveragecriterion [16] is the most popular in the literature Sinceall scenarios are not equally important our feature modelhas been extended to provide priorities for the featuresTherefore in the case there are time or cost requirements thisprioritization allows us to first generate the most importantscenarios in which we test the traffic lights programs

In this paper we focus on an extensive urban area ofMalaga city (in Spain) for the case study We use the well-known SUMO (Simulator of Urban Mobility) [17] trafficsimulator which offers a continuous source of informationabout the vehiclersquos flow velocity fuel consumption emis-sions journey time and so forth giving rise to configurationsof realistic scenarios according to real patterns of mobility Inparticular we validate a large number of traffic light programspreviously generated by four optimization algorithms twometaheuristic techniques PSOTL [9] and DETL [9] a blindstochastic search algorithm random search (RS) and adeterministic procedure following human expertsrsquo rules fortraffic configuration SCPG [17] Our present work will helpdetermine the suitability of the computed configurations fora wider set of city scenarios for which they have not beentrainedThemain contributions of this paper are summarizedas follows

(i) Design and application of a prioritized feature model(PFM) for the representation of traffic light man-agement systems as far as we know this is the firstapplication of a PFM to traffic light management sys-tems It constitutes the actual use of cross-fertilizingtechniques since we take advantage of prioritizedfeature models typically applied to software testingto leverage traffic and environmental solutions incurrent smart cities (like Malaga) This model rep-resents traffic variability through different featuresweighted according to their probability of occurrenceand importance This testing technique is useful hereto identify which cycle programs fail to manage thetraffic flow in different scenarios

(ii) Development of Prioritized Genetic algorithm forfeature models (PGFM) for the automatic generationof weighted traffic scenarios PGFM is shown here tocomply with the pairwise covering criterion [16] Inthis way we are able to work with a reduced set of

scenarios that cover all the different traffic featuresconsidered in our PFM

(iii) Thorough analysis and comparison of the differenttraffic light programs for the Malaga city case studythese cycle programs were generated as candidatesolutions by means of four optimization algorithmsPSOTL DETL RS and SCPG (human expertsrsquo solu-tions) Although these solutions were obtained for aspecific traffic scenario they have been validated formultiple scenarios covering our PFM

The remainder of the paper is organized as followsSection 2 presents an overview of related work in the lit-erature In Section 3 basic concepts are explained for thesake of a better understanding of the work presented hereAfter this Section 4 details the validation strategy Section 5describes how traffic scenario features are represented bymeans of our feature model Section 6 outlines the PGFMalgorithm and presents the generated test suite of trafficscenariosThen Section 7 is devoted to presenting theMalagacity case study In Section 8 the experimental procedureand the analysis of results are addressed In Section 9 weprovide a deeper analysis focusing on the prioritization andthe ranking of the solutions Section 10 discuses the threats tovalidity Finally Section 11 outlines some concluding remarksand future work

2 Related Work

This section presents an overview of related work consideringa two-pronged approach combinatorial testing for featuremodels and optimization of traffic light cycle programs Theformer focuses on intelligent techniques for combinatorialinteraction testing (CIT) using covering arrays The latterconsiders the use of metaheuristics for traffic optimization

CIT [16] is an effective testing approach for detectingfailures caused by certain combinations of components orinput values Generally this task consists of generating atleast all possible combinations of the parametersrsquo values (thistask is NP-hard [18]) A large number of CIT approaches havealready been published A good overview and classificationof approaches can be found in [19 20] and more recentlyin [21] In addition an extensive survey that focuses onCIT with constraints is also given in [22] There are onlytwo approaches (to the best of our knowledge) that supportprioritized test data generation the Deterministic DensityAlgorithm (DDA) presented in [23] and an approach basedon Binary Decision Diagrams (BDD) [24] However neitherof them have been applied to feature models prioritizationSo here we use the priorities of the traffic features to generateprioritized scenarios for testing traffic light programs

Except for a few efforts the application of bioinspiredtechniques to combinatorial interaction testing with featuremodels remains largely unexplored Garvin et al appliedsimulated annealing to CIT for computing n-wise coveragefor feature models [25] Ensan et al also propose a geneticalgorithm approach for test case generation [26] In contrastwith the proposed work here they use as fitness function avariation of cyclomatic complexity metric adapted to feature









3 Background









40 5 40 10 36 6 22



ldquoi + 1rdquo









1 119868



119898 and each





1198751015840= argmin119901subΓ

Θ (119901) (1)



119899

119894=1|119868119894| 119879 being





119892do


(4) 119909119894

119892+1larr update(119909119894

119892 V119894119892 119901119894


(5) 119902119894

119892+1larr Q(119909119894



(7) end for(8) 119887


119892 119902119894

119892+1) if better

(9) end while





119876 (119909) = Δ sdot lfloor119909

Δ+ 05rfloor (2)







traffic lights




119892do


(4) 119908119894


(5) 119902119894

119892+1larr Q(119908119894



(7) V119894119892+1



119892+1if better



Basic car (Bc)






















Validation

Modeling


definition

Generation





Optimization





1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are













Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra

















3 Background









40 5 40 10 36 6 22



ldquoi + 1rdquo









1 119868



119898 and each





1198751015840= argmin119901subΓ

Θ (119901) (1)



119899

119894=1|119868119894| 119879 being





119892do


(4) 119909119894

119892+1larr update(119909119894

119892 V119894119892 119901119894


(5) 119902119894

119892+1larr Q(119909119894



(7) end for(8) 119887


119892 119902119894

119892+1) if better

(9) end while





119876 (119909) = Δ sdot lfloor119909

Δ+ 05rfloor (2)







traffic lights




119892do


(4) 119908119894


(5) 119902119894

119892+1larr Q(119908119894



(7) V119894119892+1



119892+1if better



Basic car (Bc)






















Validation

Modeling


definition

Generation





Optimization





1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are













Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra














40 5 40 10 36 6 22



ldquoi + 1rdquo









1 119868



119898 and each





1198751015840= argmin119901subΓ

Θ (119901) (1)



119899

119894=1|119868119894| 119879 being





119892do


(4) 119909119894

119892+1larr update(119909119894

119892 V119894119892 119901119894


(5) 119902119894

119892+1larr Q(119909119894



(7) end for(8) 119887


119892 119902119894

119892+1) if better

(9) end while





119876 (119909) = Δ sdot lfloor119909

Δ+ 05rfloor (2)







traffic lights




119892do


(4) 119908119894


(5) 119902119894

119892+1larr Q(119908119894



(7) V119894119892+1



119892+1if better



Basic car (Bc)






















Validation

Modeling


definition

Generation





Optimization





1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are













Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











119892do


(4) 119909119894

119892+1larr update(119909119894

119892 V119894119892 119901119894


(5) 119902119894

119892+1larr Q(119909119894



(7) end for(8) 119887


119892 119902119894

119892+1) if better

(9) end while





119876 (119909) = Δ sdot lfloor119909

Δ+ 05rfloor (2)







traffic lights




119892do


(4) 119908119894


(5) 119902119894

119892+1larr Q(119908119894



(7) V119894119892+1



119892+1if better



Basic car (Bc)






















Validation

Modeling


definition

Generation





Optimization





1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are













Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











119892do


(4) 119908119894


(5) 119902119894

119892+1larr Q(119908119894



(7) V119894119892+1



119892+1if better



Basic car (Bc)






















Validation

Modeling


definition

Generation





Optimization





1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are













Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










Validation

Modeling


definition

Generation





Optimization





1 1198912) (1198911 1198912) (1198911 1198912) and (119891

1 1198912)]) are













Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











Weather (W)


Stormy (St) 3

Emergency (Em)









Exclusive-or

Exclusive-or

Vehicles (V)



















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra
















(i) 119878119906119899119899119910 rarr 119877119886119894119899119910

(ii) 119878119905119900119903119898119910 rarr 119863119903119894119868119898119901119890119903119891119867119894119892ℎ and 119863119903119894119877119890119886119888119905119894119900119899119865119886119904119905

(iii) 119878119906119899119899119910 rarr 119863119903119894119868119898119901119890119903119891119871119900119908 and 119863119903119894119877119890119886119888119905119894119900119899119878119897119900119908


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra


















119894



1scenario





1we





Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










Table1Com

putedscenariosb

yPG

FMwith

pairw

isecoverage

criterio

nFeaturea

bbreviations

ared

efinedin

Figure

4

119878119894

TW

RaSt

SuWd

FgEm

YsNo

VVa

Af

Am

Di

IlIm

IhDr

RsRm

RfVt

LiHv

119882

1198781

6359

1198782

2237

1198783

70

91198784

332

1198785

141

1198786

12

11198787

043

1198788

034

1198789

023

11987810

001






Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra














Maacutel

aga








2

emissions (CO2) NO

119909emissions (NO



8 Experiments




FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










FUELVLL

GJTV

DETLPSOTL

RandomSCPG

CO2

NOx




12statistic




2 NO119909 Fuel





9100 (13) 5393 (4) 6663 (1) 49731198782

7900 (3) 5012 (26) 5688 (4) 45361198783

6898 (3) 4440 (26) 5000 (17) 39471198784

5206 (20) 3441 (26) 3757 (24) 29121198785

2668 (14) 1936 (26) 2000 (20) 15781198786

9676 (2) 7450 (22) 8583 (0) 68081198787

7378 (20) 4118 (4) 4928 (24) 38831198788

6828 (2) 4019 (4) 4706 (24) 36841198789

5902 (3) 3840 (11) 4315 (17) 365011987810

5446 (20) 3457 (26) 3827 (20) 2928



2and NO



2and NO

119909





12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











12)


12120588

12120588

12120588

12






12statistic



12= 05 A value of




2and GJT)










5)







S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

100

200

300

400

500

600

GJT

(s) p

er V

LL

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

0

20

40

60

80

100

VLL

()

S1 S3 S4 S5 S6 S7 S8 S9 S10S2

Scenario

000

020

040

060

080

100

CO2

(gs

) per

VLL






1in




4 1198787 and 119878

10in Table 3)



2



















11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra
























11 Conclusions






2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra













2emissions NO

119909




2emissions in which



2




Acknowledgments


References































































Volume 2014




Journal of











Journal of


Function Spaces






Algebra



























































Volume 2014




Journal of











Journal of


Function Spaces






Algebra









research article intelligent testing of traffic light...

Documents