Expert Systems with Applications

Transcription

1 Expert Systems with pplications 37 (2010) Contents lists available at ScienceDirect Expert Systems with pplications jornal homepage: On optimal decision for QoS-aware composite service selection Ping Wang a, *, Ko-Ming Chao b, Chi-Chn Lo c a Department of MIS, Kn Shan University, No. 949, Da Wan Rd., Yng-Kang City, Tainan Hsien 710, Taiwan b DSM Research Grop, School of MIS, Coventry University, UK c Institte of Information Management, National Chiao Tng University, Taiwan article info abstract Keywords: Web services Services selection QoS LINMP Fzzy linear programming The increasing poplarity of employing web services for distribted systems contribtes to the significance of service discovery. However, dplicated and similar fnctional featres existing among services reqire service consmers to inclde additional aspects to evalate the services. Generally, the service consmers wold have different view on the qality of service (QoS) of service attribtes. How to select the best composite service in theory among available service (WS) candidates for consmers is an interesting practical isse. This work proposes a QoS-aware service selection model based on fzzy linear programming (FLP) technologies, in order to identify their dissimilarity on service alternatives, assist service consmers in selecting most sitable services with consideration of their expectations and preferences. This approach can obtain the optimal soltion of consensal weight of QoS attribte and fzzy positive ideal soltion (FPIS) by extending LINMP method, developed by Srinivasan and Shocker. Finally, two nmerical examples are given to demonstrate the process of QoS-aware web service selection. The experimental reslts demonstrated that it is a feasible and spplementary manner in selecting the of web services. Ó 2009 Elsevier Ltd. ll rights reserved. 1. Introdction Web services enable bsiness applications rnning on distinct platforms and exchanging data over the Internet, to be applied in bsiness and daily life regardless of the platforms or locations. It has created nprecedented opportnities for organizations to shorten software development time by composing existing services across Internet. Effective mechanisms for spporting service discovery have considerable contribtion to the sccess of web service composition. n efficient web service can bring a serios competitive advantage to the service providers as well as carry social welfare to the consmers. n application assisting in service selection based on certified QoS, cost and trst can bring essential benefits to the service consmers. Practically, the service providers are spposed to garantee QoS of WS, which are advertised on the Internet for service consmers. When service providers annonce their available services, crrent advertising approaches of web services create a WSDL or OWL-S docment to sbscribe the web service profile and service gronding, then promote it throgh UDDI registration, or other web services registries sch as ebxml. For emerging e-commerce bsiness, the selected services are aggregated to form composite services. The composite service is * Corresponding athor. Tel.: +886 (6) ; fax: +886 (6) addresses: pingwang@mail.ks.ed.tw (P. Wang), k.chao@coventry.ac.k (K.-M. Chao), cclo@faclty.nct.ed.tw (C.-C. Lo). a service prodced by a composition of other services to complete the desired service activities (nane, Chao, & Li, 2005). For example, Google research application are accepted as a web service and integrated with other services, sch as Gmail, dwords, Yo Tbe and Google Maps service, to provide an integrated environment for service consmers. Microsoft and Yahoo also provide the services analogos to that of Google for bsiness competition. The other example, consmer likes to discover the composite service, sch as flight booking, restarant reservation, and rent a car at a time, as illstrated in Fig. 1. What is the optimal approach of linking each service reqest to an approximate service? This problem may be nontrivial if the ser reqests mltiple services at one time. nmber of works on composite service discovery and selection have been carried ot to locate the reqired services and compose them to meet reqirements sing ontology (Zho, Chin, & Lee, 2004, 2005) or service matchmaking techniqes (Chao, Yonas, Lo, & Tan, 2005; Hang et al., 2005a; Hang, Chao, & Lo, 2005b). Ontology technology is developed to answer the semantic confsion problem which cold be effectively solved by semantic registration and discovery, by defining the appropriate meaning of the service s fnctionality. Part of researches (nkolenkar, Brstein, & Hobbs, 2002; Borenstein & Fox, 2003; Jorge & mit, 2006; Zho, Chia, & Lee, 2005) on semantic service discovery were investigated via Semantics Web Service (SWS) technologies to locate the reqired services and compose them to meet reqirements, as illstrated in Fig /$ - see front matter Ó 2009 Elsevier Ltd. ll rights reserved. doi: /j.eswa

2 P. Wang et al. / Expert Systems with pplications 37 (2010) Cstomer #1 Reqirements S 1 S 3 S 4 S 7 S 13 Flight Services (S 1 S 2 ) Insrance Services (S 3 S 6 ) Cstomer #2 Reqirements S 2 S Hotel Services (S 7 S 12 ) S 5 14 S 11 S 1 Crisine Services (S 13 S 15 ) S 6 S 12 S 15 Composite Service List Fig. 1. The discovery and selection of composite services. Reqests Individal Preferences Reqirements. Price. Qality. Consenss Reaching Fzzy Moderator Grop Consenss Service Discovery & Selection Registry Exection() Internet Flight Booking Semantic Web Services Consmers Discovery(),Selection() Semantic Service Profile What it does QoS Ranking UDDI OWL-S Ontologies Service Gronding How it access Reports Service Model Services How it works Fig. 2. The moderated fzzy discovery, selection and exection sing ontology. The other approaches focs on fzzy matchmaking techniqe that remains fzzy semantics on terms and handle this problem via fzzy theory. For example, moderated fzzy discovery method (Hang et al., 2005a, 2005b), measres the similarity between services in terms of capability, syntax and semantics throgh a moderator initiates to minimize the differences among service consmers and providers. For the consmer consenss of WS selection isse, service consmers and providers may have different expectations, experiences, and preferences abot the services. Frthermore, consmer preferences often remain imprecise, ncertain or ambigos on service QoS terms; the preferences over the QoS attribtes are hard to be qantified especially in distingishing the importance among these service attribtes. Therefore, the adoption of fzzy terms sch as reasonable price, reliable service, and comfortable feeling in the reqests becomes inevitable. Moreover, consmers sally have distinct view with providers for service terms, sch as cheap flight ticket, comfortable leg-room or delicios food, simply becase they have divergent perception of these terms. From the consmers point of view WS providers sally advertise on the Internet exaggerating the featres of web services for appealing to cstomers, which might lead to misnderstanding or confsing abot some service terms for WS consmers. In addition, the providers prefer to advertise their services to cstomers in sbjective terms, which might be short of considering the consmers expectations and preferences. Hence, it is imperative to reach the consenss for service consmers on the specific specification terms (i.e., QoS), where they find and search WSDL docment in the service discovery process. Based on these reqirements, W3C working grop has defined varios QoS attribtes for WS. That docment comprises a nmber of generic and specific items for cross-referencing between the possible needs of service consmers and the fnctions spported by web services. lthogh reglar QoS attribtes have been listed, some nclear problems are yet to be clarified on selection of WS processes. For example, QoS attribtes perception of importance is generally different from consmers and providers preferences. It is widely accepted that the consmers have been taking an active role in the expansion of e-commerce. In this paper we consider optimal service selection based on a given set of service reqests interacted with a set of service candidates sing fzzy linear programming (FLP) model (Li & Yang, 2004). This investigation leads to a need of developing a grop consenss-centric approach to investigate QoS attribte preferences and determine the ranking order of service alternatives according to the distance from the positive ideal soltion nder grop consenss. Conseqently, service consmer is able to redce redndancy in search, and service provider can improve the qality of services. The remainder of the paper is organized as follows. Section 2 describes the existing QoS-aware selection of web service methods. Section 3 describes proposed method. Section 4 reports on two illstrational examples of selection of service alternatives. Finally, Section 5 illstrates the conclsion and the ftre work.

3 442 P. Wang et al. / Expert Systems with pplications 37 (2010) vailable soltions for web service selection nmber of stdies for web service selection have been carried ot. One of the most well known techniqes is matchmaking. It is employed in the sitation where services with semantic descriptions for their fnctional attribtes are available in the Internet search system. How to ensre the QoS of web services for service consmers. Ran (2003) proposed a new QoS-based service registration and discovery model to explore the possibility of QoS involving into UDDI registry information. In this model, service providers have to send QoS claims to service QoS certifiers, responding to third party or form web services, for certification. The service cstomer is responsible to verify QoS claims. The QoS information finally will be registered in the UDDI registry associated with fnction description, if QoS claims have passed QoS certifier verification. The new UDDI registration mechanisms help cstomers to discover and locate the reqired service by looking p WSDL docment as well as certified QoS. Moreover, consenss of service consmers on QoS attribtes has to be considered for web service QoS certifier in the QoS comptation process. Balke and Wagner (2003) introdced the cooperative discovery concept for evalating web services in detail which composes three phases of interaction with services, i.e., (i) service discovery, (ii) service selection, and (iii) service exection. Based on Fig. 2 we reorganized three phases as Table 1, which specifies the extensive definition for selection of QoS-aware web services provisioning. Several service matchmaking techniqes have been developed to meet the needs of both consmers and providers. Zeng et al. (2004) addressed this isse of selecting web services by maximizing ser satisfaction expressed as tility fnctions over QoS attribtes; Kafmann and Gpta (1991) and Sirin et al. (2004) developed a goal-oriented and interactive composition approach that ses matchmaking algorithms to help sers filter and select services while bilding their composition service. Zho et al. (2005) discriminated between fnctional and non-fnctional QoS properties of web services, where fnctional properties can be measred in terms of throghpt, latency, response time; where non-fnctional properties are addressing of varios isses inclding integrity, reliability, availability and secrity of web services. The crrent techniqes and tools for measrement are more sitable to qantify fnctional QoS properties (for example, network throghpt, latency, and response time) than non-fnctional properties. Basically, non-fnctional QoS properties rely heavily on the perceptions of service providers and consmers that are not easy to assess de to the fact of complexity and the involvement of illstrctred information. Li, Ng, and Zeng (2004) treated the selection of QoS-driven web service with dynamic composition as a fzzy constraint satisfaction problem and applied an optimal search approach with adjstments to service composition; The matchmaking approach, however, relies on the advertisements from service providers sbjective views that cold lead to divergent perception between consmers and providers. Consmer expectations and their common preferences (i.e., consenss) on QoS shold be considered in the process of service selection. The aforementioned three major approaches are compared as illstrated in Table 2. To see in detail, composite service search approaches solved by nmerical methods can be generally divided into two categories: Mltiple ttribte Decision Making (MDM) and mathematical programming. MDM methods (Hang et al., 2005a; Li et al., 2004; Zho et al., 2004) concentrates on that QoS attribtes be collected and enforced objectively, then MDM theory can be applied to obtain a consistent ranking of service alternatives. Mathematical programming methods (Sirin, Parsia, & Hendler, 2004; Zeng et al., 2004) comprise linear programming (a single objective fnction) and mltiple goal programming. It concerns abot interactive composition selection that se a preset planning to optimally select component services dring the exection of a composite service. Those methods advanced the knowledge in QoS-aware service discovery and selection, bt nevertheless, remained the following significant isses for debate: (i) The perception of QoS attribtes needs to adjst according to consmer s preferences, (ii) How to objectively determine weights (importance) of QoS attribtes and Table 1 QoS-aware web services discovery and selection. Phases Operation Tasks Task description Phase I Service registry Fnction definition Specify the terms of WS fnctionalities sing ontology langage or WSDL Service registry Register and receive a official ID for applied service to pblish to the Internet Phase II Service discovery QoS certification ccept and certify the application of service QoS attribtes Service advertisement nnonce the featres of WS Service discovery Perform and find the related services based on a ser s reqest Service selection Select one of the desired service Phase III Service exection Service exection Carry ot service binding and exection QoS monitoring Collect cstomer opinions to QoS certifier for reflecting ser expectation Table 2 The comparison of three approaches for selection of web services. Matchmaking techniqe Composite service search method Consenss moderation approach ssmptions service description may be booked in registry Service specification is described with standard ontology tool sch as WSDL and DML-S QoS recognized by consmer s expectations and grop consenss are considered Featres The research engine always makes se of a matching algorithm to retrieve some services The selection of web service is an iterative process to discover and compose the services nder some constraints The fzzy grop decision-making methods is employed to assist service consmers in discovering appropriate services Sitable for The definition of QoS criteria are clear between consmers and providers The definition of QoS criteria are clear The sccessive selection of web services to satisfy the reqirements QoS criteria may be vage between consmers and providers

4 P. Wang et al. / Expert Systems with pplications 37 (2010) (iii) The ranking order of service alternatives shold be decided on the basis of grop consenss. To enable effective QoS-aware composite service selection, a new web service model is proposed, which inclded the following important aspects: Vage preference. This model shold be able to handle vage preferences or lingistic opinions for QoS attribtes expressed by service consmers in the process of selecting web services. Weighting of QoS attribtes. Be able to explore the optimal soltion of weighting of QoS attribtes. Service ranking. The approach shold be capable of realistically gaining a consensal ranking on web service alternatives according to consistence and inconsistence measrement between individal ratings and ideal performance soltion. To flfill these reqirements, we extend or previos work (Wang, Chao, Lo, Hang, & Li, 2006), to select QoS-aware composite services sing fzzy linear programming techniqes by minimizing the inconsistency measrement. More detailed information abot this model is described in the next section. 3. n QoS-aware services selection model based on LINMP In this section, we introdce a new fzzy grop consenssaware service selection model, which extends LINMP (LINear programming techniqes for Mltidimensional nalysis of Preference) method, developed by Srinivasan and Shocker (1973). In the LIN- MP, decision maker gives the performance ratings matrix of alternatives with a pair wise comparison form to obtain the best soltion that has the shortest distance to positive ideal soltion (PIS) (Hwang & Yoon, 1981). The aim of the model is to find the optimal weighting of QoS attribte for a set of web services and locate fzzy positive ideal soltion (FIPS) considering grop consenss, and determine a rational ranking order of web service alternatives Basic definitions and notations In this section, we review some arithmetic operations on fzzy nmbers for the prpose of representing the proposed algorithm in Section 3.2 (Kafmann & Gpta, 1991). Definition 1 (Trianglar fzzy nmber (TFN)). trianglar fzzy nmber e can be defined by ða; b; cþ. The membership fnction is defined as 8 0 for x > a >< x a for a 6 x < b b a ðxþ ¼ ð1þ c x for b < x 6 c c b >: 0 for x < c Definition 2 (Fzzy arithmetic operations). The arithmetic operations of the positive fzzy nmbers described by the interval of confidence are expressed below: ddition : ða 1 ; b 1 ; c 1 Þða 2 ; b 2 ; c 2 Þ¼ða 1 þ a 2 ; b 1 þ b 2 ; c 1 þ c 2 Þ Sbtraction : ða 1 ; b 1 ; c 1 Þ ða 2 ; b 2 ; c 2 Þ¼ða 1 a 2 ; b 1 b 2 ; c 1 a 2 Þ Mltiplication : B ¼ða 1 ; b 1 ; c 1 Þða 2 ; b 2 ; c 2 Þ¼ða 1 a 2 ; b 1 b 2 ; c 1 c 2 Þ B ¼ða 1 ; b 1 ; c 1 Þða 2 ; b 2 ; c 2 Þ¼ða 1 a 2 ; b 1 b 2 ; c 1 c 2 Þ k ¼ k ða 1 ; b 1 ; c 1 Þ¼ðka 1 ; kb 1 ; kc 1 Þ 8k 2 R Division = : = B ¼ða 1 ; b 1 ; c 1 Þ=ða 2 ; b 2 ; c 2 Þ¼ðða 1 =c 2 ; b 1 =b 2 ; c 1 =a 2 ÞÞ: ð2þ Definition 3 (The normalized Eclidean distance between two trianglar fzzy nmbers). If and B are two TFNs, then the normalized Eclidean distance between and B can be calclated as eð ; BÞ¼ 1 1=2 3 ½ða 1 b 1 Þ 2 þða 2 b 2 Þ 2 þða 3 b 3 Þ 2 Š ð3þ Definition 4 (The weighted sqare distance from positive ideal soltion, r j ). If r ij is an individal rating, then the weighted sqare distance between r ij and r j can be calclated as (Hwang & Yoon, 1981) d i ¼ Xn 2; w j r ij r j i ¼ 1; 2;...; m ð4þ 3.2. Consistence and inconsistence measrements Consider the problem of ranking WS alternatives a i ði ¼ 1;...; mþ. grop of decision makers ðd p ; p ¼ 1;...; qþ is formed to identify n QoS attribtes, say c j ðj ¼ 1;...; nþ. Each decision maker has to assign performance rating x ij ðd p Þ to service alternatives, x ij ðd p Þ represents the rating of web service s i with respect to criterion c j evalated by d p.ifx ij ðd p Þ is a fzzy data expressed by lingistic terms, then it mst be converted to a trianglar fzzy nmber (TFN) of the form ða ij ; b ij ; c ij Þ defined in Definition 1, where a ij ; b ij ; c ij are real nmbers and a ij 6 b ij 6 c ij. The performance rating matrix e X assessed by decision maker d p is shown as Eq. (4) where and represent fzzy additive and mltiplication operation that defined in Definition 2, respectively. x ij ðd p Þ might be crisp (nonfzzy) or fzzy form depending on the natre of QoS attribtes. When x ij ðd p Þ is a nonfzzy datm, it shold be converted from the distinct scales of ratings to a nmerically comparable scale. In contrast, if x ij ðd p Þ is a fzzy form then it has to be normalized by sing Eq. (5) to rank the web services compatibly between evalation QoS attribtes. For QoS attribtes, two types simltaneosly exist: benefit-oriented and cost-oriented. Both are mtally conflict and inconsistent and needs to be trade-off. To avoid generating an otbond condition, when r ij exceeds the vale 1, it needs to be constrained by pper bond 1. The linear scale transformation is sed for forming the normalized fzzy matrix R as (Chen, 2000) R ¼½r ij Š mxn! r ij ¼ x ij ¼ x j a ij c j ; b ij c j ; c ij ^ 1 c j r ij ¼ x i ¼ a j ; b j ; c j ^ 1 x ij c j c j c j where 8 j; xj 2 B! 8j; xj 2 B a j ¼ max i a ij ; b j ¼ max i b ij ; c j ¼ max i c ij ; if j 2 B a j ¼ min i a ij ; b j ¼ min i b ij ; c j ¼ min i c ij ; if j 2 C where U, C represent a set of benefit-based and cost-based QoS attribtes, respectively. Stdies regarding distance-based consenss methods have been carried ot (Cook, 2006; Cook, Kress, & Seiford, 1997), with focs on solving the nonfzzy ranking order problems. Cook, Kress, and Seiford (1997) investigated two specific cases (i.e., s ¼ 1; s ¼ 2) to solve the consenss degree of grop on ordinal rankings; the general form of consenss measrement fnction is constrcted by minimizing a normalized weighted metric distance between individal opinions and positive ideal soltion (PIS), D i, that is, ð5þ ð6þ

5 444 P. Wang et al. / Expert Systems with pplications 37 (2010) Min D i ¼ Xn ðw j jr ij r j js Þ 1=s ; i ¼ 1; 2;...; m ð7þ where w j is the weighting of QoS attribte j; ðjr ij r j js Þ 1=s is Minkowski metric, s is metric nmber. For example, s = 2, then D i becomes as d i ¼ Xn ðw j jr ij r j j2 Þ 1=2 ð8þ In this paper, we address the consistence measrement of service cstomers by aggregated difference between fzzy performance ratings of each alternative and fzzy positive ideal soltion (FPIS). Then the sqare distance, s i, defined in LNMP, is sed for assessing the weights of QoS attribtes, that is, s i ¼ Xn w j ðr ij r j Þ2 ; i ¼ 1; 2;...; m: ð9þ s X ¼fðk; lþja k Pa l ; k; l ¼ 1;...; mg denotes a set of preference relations which is composed of the ordered pairs ðk; lþ for service alternatives, where P represents a preference relation given by decision maker. There are nðn 1Þ=2 elements in X. Member a k Pa l represents that decision maker prefers a k to a l. Frthermore, analogos to s i, the fzzy form of sqare distance between a pair of alternative ðk; lþ; S k and S l, is defined by sqare distance sing the normalized Eclidean distance, defined in Definition 3, as follows: S k ¼ Xn w j ½eðr kj ; r j ÞŠ2 S l ¼ Xn w j ½eðr lj ; r j ÞŠ2 ð10þ By definition of inconsistence measrement, inconsistence index, ðs l S k Þ, measring the discrepancy between S l and S k, is given by ðs l S k Þ ¼ 0; if ðs l P S k Þ ¼ maxf0; ðs k S l Þg ð11þ S k S l ; if ðs l < S k Þ Then, the inconsistence measrement for all the ordered pairs ðk; lþ for all service alternatives in X can be compted by B ¼ X ðs l S k Þ ¼ X maxf0; ðs k S l Þg ð12þ Similar to Eq. (11), the consistence measrement between S l and S k ; ðs l S k Þ þ, is given by ðs l S k Þ þ ¼ S l S k ; if ðs l P S k Þ ð13þ 0; if ðs l < S k Þ The consistence measrement for all the ordered pairs ðk; lþ in X is given by G ¼ X ðs l S k Þ þ ð14þ 3.3. Problem formlation To avoid obtaining a trivial soltion with w j ¼ 0, we add two additional constraints, G B ¼ h, where h is also an arbitrary positive nmber, and w j P d, where d may be zero or a sfficient positive nmber. Or goal is to obtain the optimal soltion of weight of QoS attribte and fzzy positive ideal soltion (FIPS), ðw; r Þ in term of minimizing the inconsistence measrement B. The constraint, G B ¼ h, is needed to ensre the tolerance (hþ between G and B. The problem of finding the optimal consensal weights and positive ideal vales of soltion can be formlated as a linear programming model as follows: Min s:t: B G B ¼ h w j P d; j ¼ 1;...; n ð15þ By the definition of Eqs. (11) and (13), we have ðs l S k Þ þ ðs l S k Þ ¼ S l S k ð16þ Sbstitting Eq. (16) into Eq. (15), then it can be rewritten as G B ¼ X ðs l S k Þ¼h: ð17þ ðk;l2xþ Therefore, the optimal soltion ðw; r Þ can be obtained by solving the constrained optimized problem of 8 9 < X = Min maxf0; ðs k S l Þ : ; 8 P ð18þ < ðs l S k Þ¼h s:t: ðk;l2xþ : w j P d; j ¼ 1;...; n Let z kl ¼ maxf0; ðs k S l Þg, we have z kl P 0 and z kl P ðs k S l Þ, then the third and the forth constraints are obtained. z kl P ðs k S l Þ can be rewritten as z kl þðs l S k Þ P 0 ð19þ dding two constraints, then Eq. (18) is obtained as X Min z kl 8 P ðs l S k Þ¼h for ðk; lþ 2X ðk;l2xþ >< s:t: z kl þðs l S k Þ P 0 for ðk; lþ 2X z kl P 0 for ðk; lþ 2X >: w j P d; j ¼ 1;...; n In the following, sbstitting Eqs. (10) (20), then we have X Min s:t: 8 P z kl P n ðk;l2xþ w j ½eðr lj ; r j Þ eðr kj; r j ÞŠ ¼ h >< z kl þ Pn w j ½eðr lj ; r j Þ eðr kj; r j ÞŠ P 0 z kl P 0 >: w j P d Obviosly, item r 2 j for ðk; lþ 2X for ðk; lþ 2X for ðk; lþ 2X j ¼ 1;...; n ð20þ ð21þ will be omitted in the comptation process of the first constraint. Hence a new variable v j is introdced to replace w j r j for simplification of comptation, that is, v j ¼ w j r j ¼½a vj ; b vj ; c vj Š ð22þ By sing Definition 3, Eq. (21) can be transformed into the following form X Min z kl 8 P P n w j ½ða 2 r lj a 2 r kj Þþðb 2 r lj b 2 r kj Þþðc 2 r lj c 2 r kj ÞŠ 1 3 ðk;l2xþ P P 2 n 3 ðk;l2xþ w j ½a vj ða rlj a rkj Þþb vj ðb rlj b rkj Þ þc vj ðc rlj c rkj ÞŠ ¼ h for ðk; lþ 2X >< P P z kl þ 1 n w 3 j ½ða 2 r s:t: 8 lj a 2 r kj Þþðb 2 r lj b 2 r kj Þþðc 2 r lj c 2 r kj ÞŠ ðk;l2xþ P P n w j ½a vj ða rlj a rkj Þþb vj ðb rlj b rkj Þ 2 3 ðk;l2xþ þc vj ðc rlj c rkj ÞŠ P 0 for ðk; lþ 2X z kl P 0 for ðk; lþ 2X w j P d j ¼ 1;...; n >: 0 6 a vj 6 b vj 6 c vj 6 1 j ¼ 1;...; n ð23þ

6 P. Wang et al. / Expert Systems with pplications 37 (2010) We solve the linear programming sing Simplex method, and then the optimal soltion of ðw; r Þ of linear programming is yielded. Once the optimal weights of QoS attribtes ðw j ; j ¼ 1...nÞ and fzzy positive ideal soltion (FIPS) of web service i are obtained, one can easily decide a ranking order by distance from FPIS. It means that the shortest distance from FPIS is the best soltion. 4. Cases stdy In this section, two illstrational examples for selecting an appropriate web service are sed as the application of the proposed model. To examine the process of soltions respectively, the former example is a case which emphases the selection of composite service alternatives sing the traditional LINMP method, wherein QoS attribtes are crisp data assessed by a single decision maker; whereas, the latter example is regarded as a grop decision problem where rating format is fzzy form given by a set of decision makers Nmerical Case I set of composite service for traveller comprises for primitive services flight, hotel, insrance and car rental serves. For flight service, there are for service alternatives, a i ði ¼ 1;...; 4Þ, are assessed by ser regarding QoS attribtes c j ðj ¼ 1;...; 6Þ maximm baggage allowance ðc 1 Þ, check-in efficiency ðc 2 Þ, flight safety ranking ðc 3 Þ, payment for baggage lost ðc 4 Þ, ticket price ðc 5 Þ, and satisfaction on food qality ðc 6 Þ. The check-in efficiency is a scale for the passenger and baggage service in airport regarding specific airliner, its rating range resides, 1 5. For rating of flight safety, International Civil viation Organization (ICO) se formlate to cont the following aircraft accident items: (1) million flights, (2) fatal events, (3) adjst fatal events, (4) last fatal accident, (5) accident rate, to decide the ranks of flight safety in overall rank The increasing score means higher flight risk to take. The decision makers assign the performance ratings to all service candidates and determine the ranking order of for candidates. The proposed model is applied to solve this problem according to the following procedres: Step 1. Three sers assess the performance rating of each service candidate and generate the individal performance rating matrix as shown in Table 3. In addition, service consmer gives the paired comparison jdgments among for service candidates as follows: XðdÞ ¼fð1; 2Þ; ð1; 3Þ; ð4; 1Þ; ð3; 2Þ; ð2; 4Þ; ð3; 4Þg Step 2. The performance ratings matrix has converted the distinct scales of ratings to a nmerically comparable scale in [10, 10] for comparison of relative importance of each attribte as 2 3 4:50 1:10 3:00 1:20 4:50 4:00 5:00 1:30 5:00 1:50 5:00 5:00 R ¼½r ij Š¼ :00 0:90 4:00 1:00 3:80 2:00 5 4:00 1:00 3:00 0:90 3:20 4:00 Step 3. Let h ¼ 1, and d ¼ 0:01. We can model the linear programming problem sing the crisp form v j to replace v j ¼½a vj ; b vj ; c vj Š and sbstitting r ij to r ij in Eq. (23), as follows Min s:t: fz 12 þ z 13 þ z 41 þ z 32 þ z 24 þ z 34 g 8 45:75w 1 þ 0:64w 2 þ 9:0w 3 þ 0:62w 4 þ 1:0w 5 þ 21:0w 6 þ7:0v 1 0:60v 2 2:0v 3 0:4v 4 6:0v 6 ¼ 1 z 12 þ 14:25w 1 þ 0:353w 2 þ 16:0w 3 þ 0:81w 4 þ 4:25w 5 þ9:0w 6 3:0v 1 0:3v 2 4:0v 3 0:6v 4 þ 1:0v 5 2:0v 6 P 0 z 13 þ 56:25w 1 0:4w 2 þ 7:0w 3 0:44w 4 3:75w 5 12:0w 6 9:0v 1 þ 0:4v 2 2:0v 3 þ 0:4v 4 1:0v 5 þ 4:0v 6 P 0 z 41 þ 3:75w 1 þ 0:108w 2 þ 0:63w 4 þ 7w 5 1:0v 1 0:10v 2 0:60v 4 2:0v 5 P 0 >< z 32 42:0w 1 þ 0:750w 2 þ 9:0w 3 þ 1:25w 4 þ 8:0w 5 þ21:0w 6 þ 6:0v 1 0:70v 2 2:0v 3 1:0v 4 þ 2:0v 5 6:0v 6 P 0 z 24 18:0w 1 0:46w 2 16:0w 3 1:44w 4 11:25w 5 9:0w 6 þ 4:0v 1 þ 0:4v 2 þ 4:0v 3 þ 1:2v 4 3:0v 5 þ 2:0v 6 P 0 z 34 60:0w 1 þ 0294w 2 7:0w 3 0:19w 4 3:25w 5 þ12:0w 6 þ 10:0v 1 0:30v 2 þ 2:0v 3 þ 0:2v 4 1:0v 5 4:0v 6 P 0 z 12 P 0; z 13 P 0; z 41 P 0 z 32 P 0; z 24 P 0; z 34 P 0 >: w j P 0; j ¼ 1;...; 6 Step 4. Solve the optimal soltion sing Simplex method. This step prodces ðw; v Þ as Table 4. The fzzy positive ideal soltion (FIPS) is located sing Eq. (22) r i ¼ v w ¼ðr 1;...; r 6 Þ¼ð6:71; 0:0; 0:0; 0:0; 0:0; 0:0Þ Table 4 The optimal soltion of Simplex method. Paired comparison jdgments QoS attribtes ðc j Þ Variable v j ¼ w j r j z ¼ðz 12 ; z 13 ; z 41 ; z 32 ; z 24 ; z 34 Þ w ¼ðw 1 ;...; w 6 Þ T v ¼ðv 1 ;...; v 6 Þ (0.0, 0.0, 0.5, 0.0, 0.0,0.0) ð0:17; 0:0; 0:09; 0:0; 0:0; 0:00Þ T (1.14, 0.0, 0.0, 0.0, 0.0, 0.0) a 3 s a3 a 4 s a4 ( a * ) s a2 * a a 2 (Fzzy Positive Ideal Soltion ) * * 2 s a = w1 ( r11 r1 ), 1 (Sqare distance from FPIS) a 1 Fig. 3. The sqare distance of all service alternatives from FPIS a. Table 3 Performance ratings of service alternatives. Flight services ða i Þ QoS attribtes ðc j Þ Maximm baggage Flight safety Check-in Payment for baggage Ticket Satisfaction on food allowance ranking efficiency lost price qality c 1 (ponds) c 3 (1 90) c 2 (1 5) c 4 ð$þ10 2 c 5 ð$þ10 2 c 6 (1 5) a a a a

7 446 P. Wang et al. / Expert Systems with pplications 37 (2010) Step 5. Otrank the ranking of flight services.the sqare distance of service alternatives from PIS can be calclated sing Eq. (9): s i ¼ w 1 ðr i1 r 1 Þ2 ; i ¼ 1;...; 4; j ¼ 1 s 1 ¼ 1:25; s 2 ¼ 0:249; s 3 ¼ 0:545; s 4 ¼ 1:752: From Fig. 3, we can jdge the ranking order of service alternatives by the sqare distance of service alternatives from FPIS ða Þ, that is, a 2 is the best soltion. So the ranking order of flight service alternatives is generated as following: a 2 > a 3 > a 1 > a 4. Step 6. Similarly way, we examine the rest of services: hotel, insrance, car rental services, obtain the complete seqence of composite web services, a 2 a 6 a 9 a 13, as the recommendation which is expressed in Fig Nmerical Case II Consider to atomatically look and compose the delicacies from the available web services in Seattle. Focsing on the cisine or fine food where we se the Google Map to specify the context, as illstrated in Fig. 5. ser reqests Chinese dish service with service a 1 a 3 a 4 a 8 a 9 a 10 a 12 a 2 a 5 a 13 a 6 Flight services(a 1 a 4 ) a 11 a 7 Hotel services(a 5 a 7 ) Insrance services(a 8 a 11 ) Car rental services(a 12 a 13 ) Fig. 4. Search reslt of a composite web service. reqests {PM 6:30 7:30, $60 $100, Chinese dish, distance: within 15KM}. If someone discovers no appropriate service to select, one may relax the constraints to enlarge the discovery scope, that is, sia restarants (inclding of Japan, Thai, Indian and Vietnamese cisine). To relax constraints, it show sers have some possible service alternatives and improves the qality of service provisioning from Fig. 5. Three sers d p ðp ¼ 1;...; 3Þ offer their reqests to select five possible service alternatives a i ði ¼ 1;...; 5Þ based on QoS attribtes c j ðj ¼ 1;...; 3Þ acceptable price of ticket ðc 1 Þ, taste of food ðc 2 Þ, service of crew ðc 3 Þ. This case stdy focses on the satisfaction evalation on restarant services for different service consmer s reqests. The service cstomers have their different sbjective preferences on the definition of the index on satisfaction. The QoS term, satisfaction, is defined to illstrate the preference of consmer. Step 1. It is assmed that QoS term: satisfaction denoted as satisfaction ðq e Þ combining from the following three primitive fzzy terms, (i) cceptable price: s the cisine price always varies on different seasons, an acceptable price range is jdged by perception of cstomer, denoted as e for short. For example, ða; e b; cþ ¼ð100; 150; 180Þ represent the interval of acceptable price range for cstomer d i by TFN. (ii) Taste of food: it represents the satisfaction degree of food taste, qality and diversity service in the restarant, denoted as e F. (iii) Service of crew: it means the satisfaction degree of cisine service of crew represented as e S. So, the degree of satisfaction can be formlated by fzzy simple additive weighting rle (Chen & Hwang, 1992), i.e. Q ¼½ðw1 Þ e ð w 2 e SÞðw 3 e FÞŠ= ð w 1 w 2 w 3 Þ. The weightings will be evolved to reflect the sitation for a nmber of consmers preferences. Then, acceptable price, expressed by TFN, r ij ¼½ ei Š¼ða ij ; b ij ; c ij Þ, can be denoted as Fig. 6. Fig. 5. Restarant services selection.

8 P. Wang et al. / Expert Systems with pplications 37 (2010) Table 8 Ratings assigned by service cstomer d 3. Flight service ða i Þ QoS attribtes ðc j Þ cceptable price c 1 ð$þ Taste of food c 2 Service of crew c 3 a 1 ð80; 110; 140Þ Good Fair a 2 ð100; 120; 150Þ Good Good a 3 ð110; 130; 160Þ Poor Good a 4 ð70; 90; 110Þ Good Poor a 5 ð90; 100; 130Þ Fair Good Fig. 6. cceptable price expressed by fzzy sets. X USD ($) Step 2. In the following, each decision maker has to assign performance rating x ij ðd p Þ on service of food and service of crew with lingistic terms, defined in Table 5, to service alternatives. In Table 5, the membership fnction of lingistic terms for the rating of each service alternative is given by ðx 2; x; x þ 2Þ for x ij ¼ð3; 5; 7Þ except (0,1,3) for x ij is very poor and (7,9,10) for x ij is very good. The decision makers assign acceptable price to five service candidates as the first colmn of Tables 5 7. The individal fzzy performance rating matrix is shown as Tables 6 8. Three service consmers ðd 1 d 3 Þ give their paired comparison jdgments among five service candidates as: Table 5 Lingistic terms for the rating of service alternatives. Lingistic term Very poor (VP) Poor (P) Fair (F) Good (G) Very good (VG) Table 6 Ratings assigned by service cstomer. Flight service ða i Þ QoS attribtes ðc j Þ Trianglar fzzy nmber a ij ; b ij ; c ij ð0; 1; 3Þ ð1; 3; 5Þ ð3; 5; 7Þ ð5; 7; 9Þ ð7; 9; 10Þ cceptable price c 1 ð$þ Taste of food c 2 Service of crew c 3 a 1 ð100; 150; 180Þ Poor Good a 2 ð110; 150; 200Þ Good Very good a 3 ð90; 140; 200Þ Very good Poor a 4 ð60; 90; 120Þ Good Fair a 5 ð80; 100; 120Þ Fair Good Xðd 1 Þ¼fð1; 2Þ; ð1; 4Þ; ð2; 3Þ; ð2; 4Þ; ð2; 5Þ; ð3; 1Þ; ð3; 5Þ; ð4; 3Þ; ð4; 5Þ; ð5; 1Þg; Xðd 2 Þ¼fð1; 5Þ; ð2; 1Þ; ð3; 1Þ; ð2; 3Þ; ð2; 4Þ; ð2; 5Þð4; 3Þ; ð5; 3Þ; ð5; 4Þg; Xðd 3 Þ¼fð1; 2Þ; ð1; 3Þ; ð1; 5Þ; ð2; 3Þ; ð2; 4Þ; ð2; 5Þ; ð3; 5Þð3; 4Þ; ð5; 4Þg: Step 3. By applying Eq. (6), the normalized performance ratings matrix by decision maker d 1 is formed as 2 3 ð0:50; 0:60; 1:00Þ ð0:10; 0:30; 0:50Þ ð0:50; 0:70; 0:90Þ ð0:45; 0:60; 1:00Þ ð0:30; 0:50; 0:70Þ ð0:70; 0:90; 1:00Þ erðd 1 Þ¼½r ij Š¼ ð0:45; 0:65; 1:00Þ ð0:70; 0:90; 1:00Þ ð0:10; 0:30; 0:50Þ ð0:75; 1:00; 1:00Þ ð0:50; 0:70; 0:90Þ ð0:30; 0:70; 0:70Þ 5 ð0:75; 0:95; 1:00Þ ð0:30; 0:50; 0:70Þ ð0:50; 0:70; 0:90Þ Similarly, the normalized performance ratings matrix by decision makers d 2 and d 3 can be obtained, respectively. Step 4. Let h ¼ 1, and d ¼ 0:01, we can formlate the linear programming problem sing Eq. (23). Step 5. Solve the optimal soltion sing Simplex method. This step prodces ðw; v Þ as follows z ¼ðz 12 ; z 14 ; z 23 ; z 24 ; z 25 ; z 31 ; z 35 ; z 43 ; z 45 ; z 51 Þ ¼ð0:37; 0:0; 0:0; 0:0; 0:0; 0:0; 0:0; 0:0; 0:0; 0:0Þ w ¼ðw 1 ; w 2 ; w 3 Þ T ¼ð0:00; 0:0; 1:0Þ v ¼ðv 1 ; v 2 ; v 3 Þ¼ðð0:0; 0:0; 0:0Þ; ð0:0; 0:0; 0:0Þ; ð0:89; 0:89; 0:89ÞÞ The fzzy positive ideal soltion (FIPS) is located sing Eq. (22) r ðd 1 Þ¼v =w ¼ðr 1 ;...; r 3 Þ T ¼ðð0:0; 0:0; 0:0Þ; ð0:0; 0:0; 0:0Þ; ð0:89; 0:89; 0:89ÞÞ Step 6. Otrank the ranking of web services.the sqare distance of service alternatives from FPIS can be compted by sing Eq. (10): S 1 ðd 1 Þ¼0:275; S 2 ðd 1 Þ¼0:026; S 3 ðd 1 Þ¼0:675; S 4 ðd 1 Þ ¼ 0:515; S 5 ðd 1 Þ¼0:275 So the ranking order of service alternatives is generated as follows: a 2 > a 1 ffi a 5 > a 4 > a 3 : Similarly, the optimal soltion of ðw; v Þ is obtained as Tables 9 and 10 sing Steps (1) (5) for the ratings of decision makers d 2 and d 3, respectively.the fzzy positive ideal soltion (FIPS) is compted by Table 7 Ratings assigned by service cstomer d 2. Flight service ða i Þ QoS ttribtes ðc j Þ cceptable price c 1 ð$þ Taste of food c 2 Service of crew c 3 a 1 ð70; 90; 110Þ Good Fair a 2 ð100; 120; 160Þ Fair Fair a 3 ð80; 100; 120Þ Good Fair a 4 ð100; 150; 180Þ Good Fair a 5 ð120; 160; 200Þ Good Fair Table 9 The optimal soltion of Simplex method for rating of decision maker d 2. Paired comparison jdgments Weight of QoS attribtes Variable v j ¼ w j r j z ¼ z 12 ; z 14 ; z 23 ; z 24 ; z 25 ; z 31 ; z 35 ; z 43 ; z 45 ; z 51 ðw 1 ; w 2 ; w 3 Þ T v ¼ðv 1 ;...;v 6 Þ (0.33, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0) (0.0, 1.0, 0.0) ((0.0,0.0,0.0) (0.23,0.23,0.23), (0.0,0.0,0.0))

9 448 P. Wang et al. / Expert Systems with pplications 37 (2010) Table 10 The optimal soltion of Simplex method for rating of decision maker d 3. Paired comparison jdgments Weight of QoS attribtes Variable v j ¼ w j r j z ¼ z 12 ; z 14 ; z 23 ; z 24 ; z 25 ; z 31 ; z 35 ; z 43 ; z 45 ; z 51 ðw 1 ; w 2 ; w 3 Þ T v ¼ðv 1 ;...;v 6 Þ (0.0, 0.0, 0.0, 0.0, 0.0, 0.0,0.0, 0.0,0.0,0.0) ( ) ((0.0,0.1,0.1), (0.17,0.17,0 17), (0.0,0.0,0.0)) this method will obtain a well-fitting soltion. Moreover, this method may try to set w j =d, intending to stimlate the feasible soltions for obtaining a non-trivial soltion of weight of QoS attribte by adjsting d. The target of consensal weight is to yield a compromise soltion of weights among items of QoS attribtes. From two distinct cases and trial and error examples, we knew that the optimal soltion sometimes tends towards converging to a single item of weight vale of QoS attribte in the resoltion process of LNMP method. Table 11 Synthetic Jdgment matrix. Decision maker r ðd 2 Þ¼ðr 1 ; r 2 ; r 3 Þ T ¼ðð0:0; 0:0; 0:0Þ; ð0:23; 0:23; 0:23Þ; ð0:0; 0:0; 0:0ÞÞ; The fzzy positive ideal soltion (FIPS) is calclated as r ðd 2 Þ¼ðr 1 ; r 2 ; r 3 Þ T ¼ðð0:0; 0:53; 0:53Þ; ð0:21; 0:21; 0:21Þ; ð0:0; 0:0; 0:0ÞÞ; The sqare distance of service alternatives from FPIS for rating of decision makers, d 2 and d 3 can be calclated: S 1 ðd 2 Þ¼0:064; S 2 ðd 2 Þ¼0:223; S 3 ðd 2 Þ¼0:714; S 4 ðd 2 Þ ¼ 0:464; S 5 ðd 2 Þ¼0:223; S 1 ðd 3 Þ¼0:171; S 2 ðd 3 Þ ¼ 0:298; S 3 ðd 3 Þ¼0:699; S 4 ðd 3 Þ¼0:555; S 5 ðd 2 Þ ¼ 0:348 Comparing the distance from FPIS sing Eq. (10), the ranking order of five service alternatives for three decision makers is shown as, respectively. a 2 > a 1 ffi a 5 > a 4 > a 3 ; a 1 > a 2 ffi a 5 > a 4 > a 3 ; a 1 > a 2 > a 5 > a 4 > a 3 Step 7. The prominent ranking approaches of grop decision making problems inclde HP, Borda cont, and entropy method (Hwang & Lin, 1987). Here Borda cont is selected for its eqity, scores of all service alternatives assessed by three decision makers and the aggregation score are listed as the forth row in Table 11. From Table 11, the complete ranking order of service alternatives is decided as a 1 > a 2 > a 5 > a 4 > a 3 : 5. Discssion Service alternatives ða i Þ a 1 a 2 a 3 a 4 a 5 d d d Borda s scores Even if the optimal soltion is obtained, a significant isse of LINMP is remained for debate, i.e., consistent soltion. From the soltion process of two cases, one cold note that (i) LINMP method need not reqire the complete paired comparison, it can gain a transitive ranking order for service alternatives, after compted, (ii) when the nmber of service alternatives ðiþ exceeds the nmber of attribtes ðjþ, then it is easy to yield a reliable soltion of weight by the LIMP, for example, six attribtes are sed for assessing for alternatives in case I, as well as three attribtes are sed for assessing five alternatives that is, i ¼ 5 and j ¼ 3 in case II, then the proposed model can gain a consistent soltion when i > j, 6. Conclsion This paper presents a fzzy grop decision model to solve the selection of QoS-aware web services provisioning sing a fzzy linear program. The proposed model has the following featres. 1. This proposed approach not only deals with the decision maker s imprecise perceptions nder incomplete information, bt also objectively determines the importance weights of QoS criteria. The weightings are based on grop preferences for a grop of participants and realistically attain a QoS-based ranking of a list of web services. 2. The proposed approach enables decision makers to select OoSaware services from the marketplace. In the mltiple attribte decision-making applications, or approach is a complement way to the extension works of Srinivasan and Shocker (1973) and Li and Sn (2007). cknowledgements This research is partially spported by Grant No. NSC H and TWISC@NCKU Project: NSC E from the National Science Concil of the Repblic of China. References nane, R., Chao, K.-M., & Li, Y. (2005). Hybrid composition of web services and grid services. In Proceedings of 2005 IEEE international conference on e-technology, e-commerce and e-service (EEE-05) (pp ). Hong Kong: IEEE CS. nkolenkar,., Brstein, M., Hobbs, J., Lassila, O., Martin, D., McDermatt, D., et al. (2002). DMLS: Web service description for the semantic web. In Proceedings of the international semantic web conference (ISWC 02), Sardinia, Italy. LNCS (Vol. 2342). Springer. Balke, W.T., & Wagner, M. (2003). Cooperative discovery for ser-centered web service provisioning. In International conference on web services (ICWS 03), Las Vegas, US. Borenstein, J., & Fox, J. (2003). Semantic discovery for web services. Web Services Jornal, 3(4). Chao, K.-M., Yonas, M., Lo, C.-C., & Tan, T.-H. (2005). Fzzy matchmaking for web services. In Proceedings of 19 IEEE conference on advanced network and information applications (pp ). IEEE CS. Chen, C. T. (2000). Extensions of the TOPSIS for grop decision-making nder fzzy environment. Fzzy Set and Systems, 114, 1 9. Chen, S.-H., & Hwang, C. L. (1992). Fzzy mltiple attribte decision making methods and applications. Berlin/New York: Springer-Verlag. Cook, W. D. (2006). Distance-based and ad hoc consenss models in ordinal preference ranking. Eropean Jornal of Operational Research, 172, Cook, W. D., Kress, M., & Seiford, L. M. (1997). general framework for distancebased consenss in ordinal ranking models. Eropean Jornal of Operational Research, 96(2), Hang, C.-L., Lo, C.-C., et al. (2005a). Service discovery throgh mlti-agent consenss. In Proceedings of ieee international workshop on service-oriented system engineering (SOSE 2005) (pp ). Hang, C.-L., Chao, K.-M., & Lo, C.-C. (2005b). moderated fzzy matchmaking for web services. In Proceedings of the fifth international conference on compter and information technology (CIT 05) (pp ). IEEE CS. Hwang, C.-J., & Lin, M. J. (1987). Grop decision making nder mltiple criteria: Methods and applications. Lectres Notes in Economics and Mathematical Systems, 281, Hwang, C. L., & Yoon, K. (1981). Mltiple attribte decision making: Methods and applications. Springer-Verlag.

10 P. Wang et al. / Expert Systems with pplications 37 (2010) Jorge, C., & mit, S. (Eds.). (2006). Semantic web services. Processes and applications. Springer. Kafmann,., & Gpta, M. M. (1991). Introdction to fzzy arithmetic theory and application. New York: Van Nostrand Reinhold. Li, D. F., & Sn, T. (2007). Fzzy LINMP method for mlti-attribte grop decision making with lingistic variables and incomplete information. International Jornal of Uncertainty, Fzziness and Knowledge-Based Systems, 15(2), Li, Y., Ng, H. H., & Zeng, L. (2004). QoS comptation and policing in dynamic web service selection. In Proceedings of 13th international conference on world wide web (pp ). Li, D. F., & Yang, J. B. (2004). Fzzy linear programming techniqe for mltiattribte grop decision making in fzzy environments. Information Sciences, 158, Ran, S. (2003). model for web services discovery with QoS. CM SIGecom Exchanges, 4, Sirin, E., Parsia, B., & Hendler, J. (2004). Filtering and selecting semantic web services with interactive composition techniqes. IEEE Intelligent Systems, Srinivasan, V., & Shocker,. D. (1973). Linear programming techniqes for mltidimensional analysis of preference. Psychometrika, 38, Wang, P., Chao, K.-M., Lo, C.-C., Hang, C.-L., & Li, Y. (2006). fzzy model for selection of QoS-aware web services. In ICEBE 06 (pp ). Zeng, L., Benatallah, B., et al. (2004). Qos-aware middleware for web service composition. IEEE Transactions on Software Engineering, 30(5), Zho, C., Chia, L.-T., & Lee, B.-S. (2005). Semantics in service discovery and QoS measrement. IT Professional, 7, Zho, C., Chin, L. T., & Lee, B. S. (2004). DML-QoS ontology for web services. In International conference on web services (ICWS 2004) (pp ).