There is a growing demand for new mobile and pervasive environments where software services, devices, and resources can be selected and combined to build flexible and adaptive applications. Pervasive environments are characterized by richness of context, by the mobility of users and devices, and by the appearance and disappearance of resources over time. This paper presents a framework that deals with dynamic service composition and context awareness, two key aspects of application development for pervasive environments.
Dynamic service composition aims at composing complex services (or applications) on the fly, from primitive distributed components. Context awareness aims to explicitly link services to contexts and adapt services, as the context of the user changes. The framework presented in this paper offers an approach that takes a user’s context and preferences into account when determining which services to provide.
A key idea of the paper is the use of semantics, in the form of conceptual graphs annotated with ontologies, to model the contexts of users and express how to compose applications in specific contexts. The approach then uses semantic-based similarity of component models to adapt dynamic service composition to the user’s preferences and contexts. The framework assumes the availability of domain experts or component designers to provide the related ontologies in some appropriate language, such as the Web ontology language (OWL). For modeling contextual information, the framework can leverage existing ontologies about locations, device capabilities, and user profiles.
Another key idea is the use of machine learning to learn additional rules for composing context-aware applications, apart from those initially designed or input by the end user. The framework records the history of user contexts and workflow executed by the user, and then uses a decision tree construction algorithm to formalize the contextual conditions under which components are selected. Such learning can improve adaptation to the user and alleviate the need for an initially complete specification.
The style of the paper is rather informal and somewhat cumbersome to read. A formal description of the overall approach is lacking, and there is little coverage of the semantic modeling of components and services and the rule languages for stating context-awareness rules and user preferences. There are only two exceptions: how preference values of components are computed for the learning scheme (expressed in Equation 1) and how to determine semantic similarity between pairs of components (expressed in Equation 2). The paper also fails to provide a critical discussion of other approaches to dynamic composition, such as those based on artificial intelligence planning and semantic Web languages.
The paper presents some experiments based on the implementation of the approach, using an emulated environment that simulates components and contexts. The simulation experiments compute the number of synthesized workflows and the success rates, as metrics, in order to evaluate the proposed framework. The paper describes empirical results on use cases, comparing four schemes for selecting components: random, popularity, rule based, and learning based. The results of the experiments are interesting, but not surprising. The simulations show that “the rule-based and learning-based schemes ... are more adaptive,” more scalable, and “synthesize a [lower] number of workflows in dynamic environments than the random and popularity-based schemes.” The authors provide no experiments that compare the proposed framework with other semantic-based approaches for context-aware dynamic service composition. Although the simulated experiments show promise for the proposed framework, it would be premature to draw practical conclusions before deploying the framework on real components, in real-life environments.
