A Smart Agent Framework for Personalized Service Composition
End-users face a massive amount of services when selecting and composing the desired services to meet their personal preferences. For example, people frequently use the Internet to perform on-line activities (e.g., on-line shopping and banking) and control their IoT devices (e.g., turning off home light). The ever increasing amounts of services available could be overwhelming to end-users. To relieve end-users' cognitive overloading when performing on-line activities, agents can be designed and developed to act autonomously on end-users' behalf. However, the autonomous and cooperative natures of agents make it complex to design and implement by developers. General speaking, developers face the following challenges to adopt agents for service composition: 1) difficult to program agents; 2) lack of domain knowledge to identify necessary tasks for composite services; and 3) lack of a standard interface to integrate various services (e.g., IoT devices) with web services by agents. In this thesis, we present a smart agent framework that facilitates the design and implementation of agents for service composition. More specifically, our approach consists of 5 aspects: 1) we propose an easy-to-understand semi-natural language syntax that allows developers to specify the functionalities of agents; 2) we use natural language processing techniques and machine learning algorithms to identify service related tasks from on-line sources, e.g., on-line how-to instructions; 3) we propose an approach to transform functionalities of IoT devices to web services using the standard web service infrastructure; 4) we propose a deep learning based model (i.e., DeepCont) to predict end-users' ratings on services, using user reviews and service descriptions; and 5) we use a ranking algorithm (i.e., RankBoost) to automatically learn user preferences from the service usage history, and integrate the learned user preferences to recommend a collection of services for end-users. A series of case studies demonstrate that our approaches are effective to develop agents and recommend personalized services for end-users. Our approaches relieve the required efforts from developers to design agents for personalized service composition.