Description Logic Based Composition of Web Services

Transcription

1 Description Logic Based Composition of Web Services Fen Lin 1,2 Lirong Qiu 1,2 He Huang 1,2 Qing Yu 2 Zhongzhi Shi 1 1 Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, China 2 Graduate University of the Chinese Academy of Sciences {linf,qiulr,huangh,yuq,shizz}@ics.ict.ac.cn Abstract. Automatic service composition may dramatically improve development efficiency of Web Service applications. This paper proposes an approach to automatically process semantic and dynamic service composition using Description Logics(DLs) and AI planning techniques. Services and service composition problems are formalized with DLs to provide well-defined semantics. Four relationships among services as well as two combined service expressions are defined, with which AI planning techniques can be used to reason about how to compose services to achieve user-defined goals. We present an algorithm for automatic service composition, which uses backward-chaining search of potential services, and automatically eliminates irrelevant services while selecting, thus guarantee the execution efficiency. We also make some performance optimization of the algorithm such as removing redundant services and reusing previously achieved goals. All the composition steps could be done dynamically and automatically. Finally, we present an example to show how the algorithm works. Key Words: Semantic Web Services(SWS), Web Services Composition(WSC), Description Logic(DL). 1 Introduction Web services provide a standard means of interoperating between different software applications, running on a variety of platforms and/or frameworks [1]. In a Web service application, if there is no single Web service that can achieve the goal required by the user, there should be a software agent which can automatically compose existing services together in order to fulfill the request. Automatic service composition has the potential to reduce development time and effort for new applications. However, it is a hard problem, and the first step toward which is to give a well-defined formal semantics of the services. DLs are a family of knowledge representation languages that are able to represent structural knowledge in a formal and well-understood way [2], which play an important role in the Semantic Web since they are the basis of the W3Crecommended Web ontology language OWL [3, 4]. However, automatic service

2 composition not only needs static information, but also needs dynamic information to allow the users to effect changes in the world. In this paper, we use DLs to formalize the services in order to provide a well-defined semantics based on [5], with the description of the preconditions and effects of a service. We also define four relationships among the services as well as two combined service expressions, with which we can reason about how to compose services to achieve the user-defined goals. Recently there has been a lot of work applying AI planning techniques to the service composition problem such as [6, 7]. Despite the work done, composition is still largely unexplored. The combination of DLs and AI planning techniques becomes directly relevant to service composition. In this paper, we use DLs to formalize the composition problem in a general and formal plan execution task. We also propose an algorithm for automatic service composition. The algorithm uses backward-chaining search of potential services, and automatically eliminates irrelevant services while selecting, thus guarantee the execution efficiency. We also make some performance optimization of the algorithm such as removing redundant services and reusing previously achieved goals. All the composition steps could be done dynamically and automatically. The arrangement of the rest of this paper is as follows. Section 2 gives a brief survey of the DLs. In Section 3, we introduce the definition of the semantic web service and its formal semantics. Then we define four relationships among the services. Besides, we define two combined service expressions: sequence and parallel, which are used to composite the services. In Section 4, we present the composition algorithm in detail including: the service composition problem, removing redundant services, composition algorithm, performance evaluation and a travel example. Finally, Section 5 contains a brief conclusion and describes the future plans. 2 Preliminaries This section gives a short and formal introduction to DLs. See [2, 8] for a comprehensive introduction or details. DLs are a family of knowledge representation languages that are able to represent structural knowledge in a formal and wellunderstood way. A description logic system consists of four parts: constructors which represent concept and role, Terminological assertion (Tbox) subsumption assertion, Assertions about individual (Abox) instance assertion, and reasoning mechanism of Tbox and ABox. The constructors determine the expressive power of the DL. Given mutually disjoint sets N C of concept names, N R of role names, and N I of individual names, concept and role constructors in ALCN R, can be defined using the following syntax: C, D A C D C D C R.C R.C ( nr) ( nr) R P 1... P m

3 DL commonly have a set-theoretic semantics. The semantics of a DL knowledge base are given via interpretation I = I, I, where I is a non-empty set of objects, and I maps each individual name a to an element a I I ; each concept name C to a subset of I, i.e., C I I ; each role name R to a binary relation on I, i.e., R I I I. Let T be an ancylic Tbox,and A denotes an Abox. We say that I satisfies T, A, written as I = T, A, if I satisfies every sentence in T, A. We define the binary relations and on models of I, I in A w.r.t. T, I I, this means the two models are equal, that is to say, a I = a I and C I = C I and R I = R I ; I I, this means one model is inferior to the other, that is to say, a I = a I and C I C I and R I R I. In a DL reasoning system, several different kinds of reasoning problems are typically supported w.r.t. a knowledge base T, A. For the purpose of this paper, it suffices to introduce concept satisfiability and ABox consistency: Concept satisfiability: the concept C is satisfiable w.r.t. the TBox T iff there exists a model I of T such that C I ABox consistency: the ABox A is consistent w.r.t. the TBox T iff there exists an interpretation I that is a model of both T and A. All other inferential services can be realized with satisfiability tests of knowledge bases. For example, the problem whether I = C D holds, is equivalent to the problem whether I {a : C, a : D} is satisfiable(consistent). 3 Service Descriptions In this section, to begin with we introduce the definition of Semantic Web services and its formal semantics based on [9]. Then we define serval relationships among them, which are useful to remove redundant services in the following section. Furthermore, we define two combined service expressions: sequence and parallel, which are used to composite the services to achieve the user-defined goal. Finally, we present some theorems about these operators. 3.1 Service Definition Definition 1 (Semantic Web Services) Let T be an acyclic Tbox. An atomic Semantic Web Service S for T is defined in the form of S = (P, E), where: S is the service name.

4 P is a finite set of Abox assertions, the pre-conditions, which must be satisfied before the service is executed. E is a finite set of conditional post-conditions, which denote the effects of the service, E is a set of pair ϕ/ψ, where ϕ is an Abox assertions for T, ψ is a primitive literal for T, i.e., C(p), C(p), R(p, q), R(p, q) with C a primitive concept in T and R a role name. Each Semantic Web Service can be specified by its preconditions and effects in the planning context. Precondition presents logical conditions that should be satisfied prior to the service being requested. Effects are the result of the successful execution of the service. To illustrate the definition of services, consider a travel website, whose business is providing services for tourists including traffic service, hotel service, and taxi service as defined below. T rafficservice = ({ P erson(a) }, { bookt icket(a, b), T icket(b) }) M odesthotelservice = ({ P erson(a), hast elepone.t elephone(a) }, { bookhotel(a, c), M odesthotel(c) }) P alacehotelservice = ({ P erson(a), hasm oney.enoughm oney(a) }, { bookhotel(a, d), P alacehotel(d) }) T axiservice = ({ P erson(a), bookt icket.t icket(a) }, { bookt axi(a, e), bookhotel.m odesthotel(a)/m odestt axi(e), bookhotel.p alacehotel(a)/p alacet axi(e) }) The traffic service provides ticket booking for the customer and everybody can book a ticket for himself. The modest hotel service provides a modest hotel booking for the customer by telephone. The palace hotel service provides a palace hotel booking for the customer if the customer has a great amount of money. Which hotel to choose is decided by the economic ability of the customer. The taxi service provides the transportation for the daily routines of the customer if the customer has booked the ticket. Which taxi to choose is decided by the choice of the hotel, if the customer has booked a modest hotel, it will book a modest taxi, otherwise a palace taxi. The meaning of the concepts used in these services are defined in the following acyclic Tbox T : T = { EnoughMoney Money 10 hast housandrmb, Hotel M odesthotel P alacehotel, T elephone F ixedlinep hone MobileP hone, T axi ModestT axi P alacet axi.} The formal semantics of services can be defined by means of a transition relation on interpretations. The service S may transform I to I (I T,A S I ), if C is a primitive concept and R a role name, then C I := (C I {c I ϕ/c(c) E, and I = ϕ}\{c I ϕ/ C(c) E, and I = ϕ}) R I := (R I {(a I, b I ) ϕ/r(a, b) E, and I = ϕ}\{(a I, b I ) ϕ/ R(a, b) E, and I = ϕ}) Then we present executability and projection of services as follows [9]:

5 Executability: S is executable in A w.r.t. T iff I = P in all models I of A w.r.t. T and I T S I, I with no conflict. Projection: an assertion a is a consequence of applying S in A w.r.t. T iff for all models I of A and T, and all I with I T S I, we have I = a. 3.2 Service Relationship Among semantic web services, there exist relationships, which are the important features of services for composition. Let T be an acyclic Tbox, and A be an Abox, and S i denotes services. Formally, we define the following relationships about services. Identical Service: S 1 = S 2, this means the two services can provide the same function in spite of the fact that they may have different service names. Conditionally Identical Service: S 1 = S2, this means S 1 can provide the same function as S 2 in some situation. For instance, there is a traffic service that provides ticket booking just for students, StudentT raf f icservice = ({Student(a)}, {bookt icket(a, b), T icket(b)}), and add Student P erson to Tbox(T ). If all the people in the system are students, then we have that, StudentT rafficservice = T rafficservice in the system. Substitute Service: S 1 S 2 or S 2 S 1, this means a service S 1 can be substituted by service S 2 in any case, if S 1 is possible, S 2 is also possible and has exactly the same effects as S 1 or more. For instance, as described above, StudentT rafficservice is a sub-service of T rafficservice. Of course, the former ticket service can be alternated by the other service. Conditionally Substitute Service: S 1 S 2 or S 2 S 1, this means a service S 1 can be substituted by service S 2 in some situation. For instance, there is a traffic service that provides train ticket booking for everybody, T raint raf f icservice = ({P erson(a)}, {bookt icket(a, b), T raint icket(b)}), and add T icket P lanet icket T raint icket to Tbox(T ). If all the people in the system are students, then we have that, T raint rafficservice StudentT rafficservice in the system. Obviously, we have the following theorems about the relationships on services. Theorem 1 If S 1, S 2 are executable in A w.r.t. T, then I T,A S 1 I, I T,A S 2 I, in all models I of A w.r.t. T, if we have I I hold, then S 1 = S2 in the case of A w.r.t. T. Theorem 2 If S 1, S 2 are executable in A w.r.t. T, then I T,A S 1 I, I T,A S 2 I, in all models I of A w.r.t. T, if we have I I hold, then S 1 S 2 in the case of A w.r.t. T. 3.3 Combined Service A combined service is an aggregation of some independent and interactive web services (individual or combined web services). We define two combined service expressions: sequence and parallel as follows. Let T be an acyclic Tbox, A denotes an Abox and S i denotes services for T.

6 Definition 2 (Sequence Services) S = S 1 ; S 2 ; ; S k. A sequence service S is a service that is achieved by an order of services in a sequence, a following of one service after another. Executability: S = S 1 ; S 2 ; ; S k is executable in A w.r.t. T iff the following conditions are true in all models I of A and T : I = P 1 and For all i with 1 i < k and all interpretations I with I T S 1;S 2; ;S i I, we have I = P (i+1) and I T S I, I with no conflict. Definition 3 (Parallel Services) S = S 1 S 2 S k. A parallel service S is a service that is achieved by a set of services of equal rank, all services executing at the same time. Executability: S = S 1 S 2 S k is executable in A w.r.t. T iff each service in S is executable. A combined service can be composed by sequence and parallel services. For example, we can compose a combined service of the foregoing services such as T ravelservice = (T raf f icservice HotelServices); T axiservices. By executing the travel service, the customer will have the suitable ticket, hotel and taxi for his trip. After the definitions, we have some theorems of the operators as follows, which is used to reduce the complexity of the combined services. Theorem 3 S 1 (S 1 ; S 2 ) in the case of A w.r.t. T, if S 1 ; S 2 are executable in A w.r.t. T. Proof. If S 1 is possible, S 1 ; S 2 is also possible, and all the effects S 1 produced obviously can be produced by S 1 ; S 2. Theorem 4 S 1 S 2 = S 2 iff S 1 S 2 in the case of A w.r.t. T, if S 1 S 2 are executable in A w.r.t. T. Proof. If S 1 can be substituted by S 2, executing S 1 S 2 has the same effects as executing S 2, or vice versa. Theorem 5 S 1 S 2 S 3 = (S 1 S 2 ) S 3 = S 1 (S 2 S 3 ), that is to say, the parallel services are irrelevant with the order of the services. 4 Planning for Service Composition With the results of Section 2 and Section 3, we have addressed a fundamental barrier to automated service composition. This section aims to give a specific approach to automated composition of web services. Most important of all we use DLs to formalize the composition problem and present its formal definition. Moreover we remove the redundant services, which is desirable because it reduces the plan search space. After the preprocessing, we propose the planning algorithm with an analysis of the algorithm. Finally we provide an example to show how the algorithm works.

7 4.1 Service Composition Problem A Web services composition problem can be described as: Given a set of web services and a description of some tasks or goals to be achieved(e.g., Make the travel arrangements for my trip ), find a composition of services that achieves the task. And it can be viewed as a planning problem, depending upon how we represent our services. In this paper, we use DLs to formalize the services as described in Section 3. Then the definition of the composition problem is as follows: Definition 4 (Service Composition Problem) the composition problem can be described as a four-tuple < T, A, G, S >, where: T describes the vocabulary of the application domain. A contains assertions about named individuals in terms of this vocabulary and also denotes the initial state of the world. G is a set of assertions, which represent the goal attempting to reach. S is the set of services as described in Section 3.1. A composition problem is to find a combined service performed to reach the goal. 4.2 Removing Redundant Services Removing Weaker Services: In section 3.2, we define substitute relationship among services. If a service S 1 can be substituted by service S 2 in any case, S 1 may be redundant in the sense that in any situation, if S 1 is possible, S 2 is also possible and has exactly the same effects as S 1. More generally, we define the notion that service S 2 is stronger than service S 1. So we can remove service S 1 to reduce the plan search space. Each weaker services, which can be substituted by another service, could be removed. The pseudo-code for removing weaker services is in Algorithm 1. Algorithm 1 Removing Weaker Services 1: T is the Tbox 2: A is the ABox 3: S = (S 1,, S n) 4: for i = 1 to n do 5: if there exists S k S, k! = i, and S i S k in the case of A w.r.t. T then 6: remove S i from S 7: end if 8: end for Removing Useless Services: If the preconditions of the service S can never be satisfied in any case, then S is useless and can be removable. The pseudo-code for removing useless services is in Algorithm 2. For simplicity, we just consider those services whose preconditions is inconsistent with G in A w.r.t. T. These

8 Algorithm 2 Removing useless Services 1: T is the Tbox 2: A is the ABox 3: G is the Goal 4: S = (S 1,, S n) 5: for i = 1 to n do 6: S i = (P i, E i) 7: if P i is inconsistent with G in A w.r.t. T. then 8: remove S i from S 9: end if 10: end for services are not possible, otherwise there will be a conflict in the system. Removing irrelevant Services with respect to a Goal: If the effects of the service S can never make the goal true and if S can not directly achieve the preconditions of any services, then S is irrelevant with respect to the goal and can be removed. The composition algorithm described in Section 4.3 is goaldriven, thus the services irrelevant to the goal are eliminated automatically while selecting. 4.3 Composition Algorithm The algorithm is a simple backward-chaining algorithm, which executes as follows. To begin with the user initializes the algorithm by specifying the goals. Then the algorithm divides the goals into two parts: a sub goal(subgoal) and the rest goals(restgoal). Furthermore the algorithm checks if the subgoal has already been satisfied previously. If so, the algorithm includes subservices in the solution. Now the algorithm tries to satisfy the rest goals. This corresponds to a new iteration which involves executing the same steps and include restservices in the solution, thus return subservices; restservices. If not, the algorithm checks if the subgoal could be satisfied by the initial assertions. If so, the algorithm tries to satisfy the rest goals as described above and return restservices. If not, the algorithm looks for service(s) whose effect(s) could satisfy the subgoal and include subservices in the solution. Now the preconditions of the newly included service(s) are treated as unsatisfied goals. And the algorithm tries to satisfy the unsatisfied goals. This corresponds to a new iteration and includes preservices in the solution. And then the algorithm tries to satisfy the rest goals as described above and include restservices in the solution, thus return (preservices; subservices) restservices. Note that the algorithm terminates when failing to find any services that satisfy any of the unsatisfied goals. If all unsatisfied goals correspond to the initial assertions then the algorithm could be successful. The pseudo-code for automatic service composition is in Algorithm 3.

9 Algorithm 3 1: T is the Tbox 2: A is the ABox, add the initial state to A 3: G is the Goal 4: S = (S 1,, S n) 5: GoalServices = 6: if T, A = G then 7: GoalServices = 8: return GoalServices 9: end if 10: select an assertion subgoal from the unsatisfied goal G 11: restgoal = G subgoal 12: if the subgoal has already been processed by subservices then 13: executed recursively, and get restservices to reach restgoal. 14: GoalServices = subservices restservices 15: return GoalServices 16: else if T, A = subgoal then 17: executed recursively, and get restservices to reach restgoal. 18: GoalServices = restservices 19: return GoalServices 20: else if there exists no services to satisfy the subgoal subgoal then 21: return NULL 22: else 23: there exists a set of services S = {S 1,, S k} can satisfy the subgoal 24: loop 25: select a set of services subservices from S that can satisfy the subgoal 26: there exists a set of preconditions to satisfy the subgoal, G = {G 1,, G m}, where G i = P re(subservices) + P ost i(subservices, subgoal) 27: loop 28: select a pregoal from G that can satisfy the subgoal 29: executed recursively, and get preservices to reach pregoal. 30: executed recursively, and get restservices to reach restgoal. 31: if preservices! = NULL then 32: if restservices! = NULL then 33: GoalServices = (preservices; subservices) restservices 34: return GoalServices 35: end if 36: end if 37: remove pregoal from G 38: end loop 39: remove subservices from S 40: end loop 41: return NULL 42: end if

10 4.4 Performance Evaluation In the service composition problem, we make a closed world assumption, and the solution either exists or does not exist. If the solution exists, the algorithm will surely find one and return it. If not, the algorithm will return NULL. All the composition steps could be done dynamically and automatically. Most important of all we remove the redundant services including weaker services and useless services, which is desirable because it reduces the search space and improves efficiency in some cases. Note that removal of weaker primitive actions may result in removal of the optimal plan. We may lose the optimal plan with respect to the number of primitive actions in our initial domain. The search tree created by the above algorithm can be seen as a sort of AND-OR tree, where the OR branches represent different ways of satisfying the goal, and the AND branches represent combinations of service effects that together reach the goal. Leaves in the tree represent initial assertions to be available. Viewing the search and execution as an AND-OR tree allows us to investigate the use of different search techniques to improve execution performance. Moreover the algorithm described above uses backward-chaining search of potential services, and automatically eliminates irrelevant services while selecting, thus guarantee the execution eciency. Finally the algorithm reuses previously achieved goals, which makes the algorithm more efficient. 4.5 A Sample Case In this section we provide an example that illustrates how the algorithm works. For example, a tourist named Jim wants to have a trip to Beijing, and he wants the website to make the travel arrangements for his trip.we suppose that the tourist has a telephone but not enough money. The problem could be formalized as < T, A, G, S >,where: T, S defined in Section 3.1. A = {P erson(jim), hast elephone(jim, t), M obilep hone(t)}. G = { bookt icket.t icket(jim), bookhotel.hotel(jim), bookt axi.t axi(jim)}. The algorithm executes as the Figure 1 shown. To start with we divide the goals({g1, G2, G3}) into two parts:{g1} and {G2, G3}, as shown in the left tree in Figure 1. And there exists S1 to satisfy the first subgoal. Next we try to satisfy the rest goals, which leads to a new iteration. We find that G2 could be satisfied by S2 and S3. Obversely, the preconditions of the former service could be satisfiable in A w.r.t. T, while the other couldn t. In addition we try to reach the rest goal,{g3}, as shown in the right tree in Figure 1. There exist only one service to meet the requirement. And we have two pregoals of the service: {G1, G6, G7} and {G1, G6, G8}, each leads to a new iteration. Obviously, the subgoal,{g6}, is in the initial assertions. Now the left goals are treated as the unsatisfied goals. The subgoal {G1} has already been satisfied and could

11 Fig. 1. Example of Service Composition Algorithm Execution Tree be achieved by S1. the goals G7 and G8 could be achieved by S2 and S3 respectively. The pregoals : {G4, G6} can be satisfiable, but the other {G5, G6} couldn t. Finally we get the solution, (S1 S2) ((S1 S2); S4), which could be reduced to (S1 S2); S4) by Theorem 3 and 4. So the result service of the problem is (T raf f icservice M odesthotelservice); T axiservice). 5 Conclusions and Future Work In this paper, semantic web services and their composition are studied and a specific approach to web services composition is proposed. In the approach, we use DLs to formalize the services and the services composition problem in order to provide a well-defined semantics. And we define four relationships among the services as well as two combined service expressions including sequence and parallel, with which AI planning techniques can be used to reason about how to compose services to reach the user-defined goal. The algorithm uses a backward-chaining search with some performance optimizing including removing redundant services and reusing previously achieved goals. All the composition steps could be done dynamically and automatically. The idea presented in this paper can be extended in future from different points of view. One is the improvement of the composition algorithm. If there exist various composition solutions to achieve the user-defined goal, which one should be chosen to satisfy the goal best? Another is the improvement of the composition model. If the goals couldn t be satisfied, how can we find the largest

12 compatible subset of the goals to achieve. Another is the implementation of the algorithm. From a multi-agent systems perspective, a service composition software system can be viewed as a collection of sociable agents, representing individual services, which cooperate, coordinate and work together to achieve the goal [10]. So we can implement the system on MAGE [11], which is a Multi-AGent Environment developed by Intelligent Science Research Group, at Key Lab of IIP, ICT, CAS, China, using agent technology to analyze, design, implement and deploy multi-agent systems. Acknowledgements Our work is supported by the Natural National Science Foundation of China (No ), the National 973 Project of China (No.2003CB317004) and the Natural Science Foundation of Beijing (No ). References 1. David Booth et al. Web services architecture. Technical report, W3C Working Group Note, See 2. Franz Baader, Diego Calvanese, Deborah McGuinness, Daniele Nardi, and Peter F. Patel-Schneider. The Description Logic Handbook: Theory, Implementation, and Applications. Cambridge University Press, Franz Baader, Ian Horrocks, and Ulrike Sattler. Description logics as ontology languages for the semantic web. In In Festschrift in honor of Jorg Siekmann,LNAI, Springer-Verlag, Ian Horrocks, Patel-Schneider, and Frank van Harmelen. From shiq and rdf to owl: The making of a web ontology language. Journal Web Semantics, 1(1):7 26, F. Baader, M. Milicic, C. Lutz, U. Sattler, and F. Wolter. Integrating description logics and action formalisms for reasoning about web services. LTCS-Report LTCS-05-02, Chair for Automata Theory, Institute for Theoretical Computer Science, Dresden University of Technology, Germany, See 6. Marco Aiello, Mike P. Papazoglou, Jian Yang, M. Carman, Marco Pistore, Luciano Serafini, and Paolo Traverso. A request language for web-services based on planning and constraint satisfaction. In Proceedings of the Third International Workshop on Technologies for E-Services, pages 76 85, Springer-Verlag London, UK, S. McIlraith and T. Son. Adapting golog for composition of semantic web services. In Proceedings of the 8th International Conference on Principles of Knowledge Representation and Reasoning (KR 2002), pages , Description Logics website 9. F. Baader, C. Lutz, M. Milicic, U. Sattler, and F. Wolter. A description logic based approach to reasoning about web services. In Proceedings of the WWW 2005 Workshop on Web Service Semantics (WSS2005), Chiba City, Japan, Paul A. Buhler and José M. Vidal. Toward the synthesis of web services and agent behaviors. In Proceedings of the First International Workshop on Challenges in Open Agent Systems, pages 25 29, Zhongzhi Shi, Haijun Zhang, and Mingkai Dong. Mage: Multi-agent environment. In ICCNMC-03, pages , 2003.