The practice of composing web services has received an increasing interest with the emerging application development architecture called Service Oriented Architecture (SOA). A web service composition can be done either manually or (semi-) automatically. Doing composition (semi-) automatically minimizes runtime problems that arise due to dynamic nature of runtime environments. However, the implementation of (semi-) automatic composition demands for the automation of a process model or a composition plan generation process. In addition, creating a composite service or applications from component services, that are developed and meant to work independently, causes unavoidable dependencies among the services involved. Consequently, in a composite service development, understanding, analyzing and tracking of such dependencies becomes important. This thesis views the process model generation sub-task of a service composition as a service dependency identifification and analysis problem. In this thesis, we propose a dependency based automatic process model generation methods. For this purpose, the following issues are explored. First, a top layer architecture with a composition engine is developed. The architecture gives a complete picture of dependency based automatic service composition. Second, the process model generation sub-task is formulated as a service dependency identification and analysis problem. Third, a two-stepped method for automatic process model generation, given a set of candidate web service descriptions, is proposed. The first step of the proposed approach deals with the identifification of potential direct and indirect dependencies between abstract services. The direct dependency extraction is done by assuming a semantic I/O matching of service parameters. The extraction of indirect dependency from direct dependency is done using a recursive algorithm derived from the transitive closure property. Alternatively the Warshall algorithm is used. The second step of the proposed approach deals with analysis of dependency information and generation of process model (PM) automatically. To execute this step, we propose two approaches: matrix based and graph based approaches. The matrix based approach utilizes both direct and indirect dependencies. This approach represents dependencies using matrix and takes advantages of a sorting algorithm. The matrix representation facilitates a simplistic mathematical dependency analysis for generating important indicators during automatic process model creation. The process model is generated using a sorting algorithm that uses the analysis result obtained from the dependency matrix as sorting criterion. The graph based approach uses only direct dependency among candidate services. As its name indicates, in this approach the extracted I/O dependencies are represented using a directed graph. A modifified topological sorting algorithm is used for generating a process model that shows the execution order of candidate services. Both of the proposed approaches (matrix and graph based approaches) recognize the existence of cyclic dependencies and provide ways of dealing with them. The resulting process model or composition plan from both approaches has a sequential, concurrent and loop control flows. Finally, the performance of the proposed approaches is studied theoretically as well as experimentally. For the experimental validation and evaluation purpose, the approaches are implemented in a prototype that facilitates the validation and evaluation of the approaches at a larger scale. An extensive experimental performance evaluation is done fifirst on each proposed approach. The two approaches are then compared and their pros and cons under difffferent scenarios are assessed.

Dans le but d’exploiter au mieux les grandes masses de données hétérogènes produites par les instruments scientifiques modernes de l’astrophysique, les scientifiques ont développé le concept d’Observatoire Virtuel (OV). Il s’agit d’une architecture orientée services, qui a pour objectif de faciliter l’identification et l’interopérabilité des données astrophysiques. Malgré le développement et les avancées permises par l’OV dans l’exploitation de ces données, certains objectifs sont partiellement atteints notamment l’interopérabilité, la sélection de services et l’identification de services connexes, etc. Par ailleurs, l’ergonomie des outils à la disposition de l’utilisateur final reste perfectible. De même l’utilisation actuelle des ressources de l’OV, s’appuyant sur des compétences humaines, gagnerait à être automatisée. Les services de données astrophysiques n’étant pas tous inscrits dans l’OV, il serait aussi souhaitable pour permettre une utilisation plus large de ces outils, qu’ils s’appuient également sur des services disponibles en-dehors de l’OV. En vue d’automatiser l’utilisation des ressources en ligne, les sciences de l’information travaillent depuis 2001 à l’élaboration du Web sémantique. Cette évolution apporte au Web des capacités de raisonnement automatiques, basées sur des algorithmes utilisant une nouvelle forme de description des contenus. Cette nouvelle forme de description sémantique se trouve exprimée dans des représentations informatiques appelées ontologies. Malheureusement, les méthodes actuelles d’élaboration du Web sémantique ne sont pas complètement compatibles avec les services OV qui utilisent des modèles de données, des formats et des protocoles d’accès aux services qui s’éloignent de ceux rencontrés habituellement dans les sciences de l’information. Dans ce contexte, cette thèse décrit une méthodologie générique de composition de services sans état, basée sur la description des services par une ontologie dont la définition est proposée dans ce document. Cette ontologie représente aussi bien des services Web que des services non accessibles par le Web. Elle prend en compte certaines spécificités qui peuvent être rencontrées dans les infrastructures de services préexistantes. L’enrichissement de l’ontologie par des concepts issus de domaines d’application spécifiques pour lesquels il n’existe que peu de représentations ontologiques est également pris en compte. La population de cette ontologie, par des services éventuellement éloignés des standards utilisés habituellement dans les sciences de l’information, est aussi traitée. La méthodologie a été appliquée avec succès dans le cadre de l’astrophysique, et a permis de développer une application Web permettant la composition automatique de services utilisable par un public non averti. / Scientists have developed the Virtual Observatory (VO) concept in order to make the most of the large masses of heterogeneous data produced by the modern scientific instruments of astrophysics. It is a service-oriented architecture, aiming to facilitate the identification and interoperability of astrophysical data. Despite the development and advances made by VO in the exploitation of these data, some objectives are partially such as interoperability, service selection and identification of related services, etc. In addition, the ergonomics of the tools available to the end user can be improved. Similarly, the current use of VO resources, based on human skills, would benefit from being automated. As not all the astrophysical data services are included in the VO, it would also be desirable to allow a wider use of these tools, as they also rely on services available outside the VO. In order to automate the use of online resources, information sciences have been working since 2001 on the development of the Semantic Web. This evolution provides the Web with automatic reasoning abilities, based on algorithms using a new form of content description. This new form of semantic description is expressed in computer representations called ontologies. Unfortunately, the current semantic Web development methods are not fully compatible with VO services that use data models, formats and protocols for accessing services that differ from those typically encountered in information sciences. In this context, this thesis describes a generic methodology for the composition of stateless services, based on the description of services by a global ontology, the definition of which is proposed in this document. This ontology represents both Web services and services that are not accessible via the Web. It takes into account certain specificities that may be encountered in preexisting service infrastructures. The enrichment of the ontology by concepts derived from specific fields of application for which there are only a few ontological representations is also taken into account. The population of this ontology, by services possibly distant from the standards usually used in the information sciences, is also treated. The methodology was applied successfully in the framework of astrophysics, and allowed to develop a Web application allowing the automatic composition of services usable by an uninformed public.

The practice of composing web services has received an increasing interest with the emerging application development architecture called Service Oriented Architecture (SOA). A web service composition can be done either manually or (semi-) automatically. Doing composition (semi-) automatically minimizes runtime problems that arise due to dynamic nature of runtime environments. However, the implementation of (semi-) automatic composition demands for the automation of a process model or a composition plan generation process. In addition, creating a composite service or applications from component services, that are developed and meant to work independently, causes unavoidable dependencies among the services involved. Consequently, in a composite service development, understanding, analyzing and tracking of such dependencies becomes important. This thesis views the process model generation sub-task of a service composition as a service dependency identifification and analysis problem. In this thesis, we propose a dependency based automatic process model generation methods. For this purpose, the following issues are explored. First, a top layer architecture with a composition engine is developed. The architecture gives a complete picture of dependency based automatic service composition. Second, the process model generation sub-task is formulated as a service dependency identification and analysis problem. Third, a two-stepped method for automatic process model generation, given a set of candidate web service descriptions, is proposed. The first step of the proposed approach deals with the identifification of potential direct and indirect dependencies between abstract services. The direct dependency extraction is done by assuming a semantic I/O matching of service parameters. The extraction of indirect dependency from direct dependency is done using a recursive algorithm derived from the transitive closure property. Alternatively the Warshall algorithm is used. The second step of the proposed approach deals with analysis of dependency information and generation of process model (PM) automatically. To execute this step, we propose two approaches: matrix based and graph based approaches. The matrix based approach utilizes both direct and indirect dependencies. This approach represents dependencies using matrix and takes advantages of a sorting algorithm. The matrix representation facilitates a simplistic mathematical dependency analysis for generating important indicators during automatic process model creation. The process model is generated using a sorting algorithm that uses the analysis result obtained from the dependency matrix as sorting criterion. The graph based approach uses only direct dependency among candidate services. As its name indicates, in this approach the extracted I/O dependencies are represented using a directed graph. A modifified topological sorting algorithm is used for generating a process model that shows the execution order of candidate services. Both of the proposed approaches (matrix and graph based approaches) recognize the existence of cyclic dependencies and provide ways of dealing with them. The resulting process model or composition plan from both approaches has a sequential, concurrent and loop control flows. Finally, the performance of the proposed approaches is studied theoretically as well as experimentally. For the experimental validation and evaluation purpose, the approaches are implemented in a prototype that facilitates the validation and evaluation of the approaches at a larger scale. An extensive experimental performance evaluation is done fifirst on each proposed approach. The two approaches are then compared and their pros and cons under difffferent scenarios are assessed.