| dc.description.abstract | With the increasing number of resource-rich handsets equipped with diverse wireless communication technologies, users within a limited geographical area can share the services deployed on their mobile devices to form service-sharing communities. By leveraging the computing resources on nearby devices, new service-based applications can be developed to expand users' service options. Many applications from multiple domains have the potential for improvement with flexible, dynamic service composition, including automotive (e.g., real-time hazard warnings), energy demand-side management (e.g., communities maximising use of renewable energy while catering to individual home needs), and FinTech (e.g., fast insurance response).
Automatic planning, with adaptive composition recovery mechanisms, has been used to tackle complex service provisioning in such dynamic environments. Existing service composition proposals generate a service dependency graph based on the interoperable relationships between available services, and use goal-driven techniques to discover paths that can functionally satisfy user requests. Apart from functional requirements, though, Quality of Service (QoS) such as execution reliability and latency are also major concerns that impact users' satisfaction. Finding paths in this graph that can functionally satisfy a user's request while simultaneously guaranteeing user-acceptable QoS levels is difficult in mobile environments. Given mobile devices' limited communication ranges, the frequent network topology changes, and services with time-dependent QoS, the existing proposals for QoS-optimal service composition trade-off computational efficiency for optimality to provide only best-effort QoS. The existing mechanisms also use a-priori articulation of the QoS objectives' weights, which does not allow for the exploration of different QoS trade-offs. These relative weights might differ at runtime, and the constant enquiry for user's preferences can inhibit the development of automatic, planning-based service composition.
This thesis presents SBOTI (Stigmergic-Based OpTImisation), a decentralised, QoS optimisation mechanism for automatic, planning-based service composition. SBOTI uses a community of homogeneous, mobile software agents, which share the same goal, to effectively and efficiently approximate the set of QoS-optimal service composition configurations available in a geographically-limited, mobile environment. The proposed mechanism uses an iterative, reinforcement-based approach to control the trade-off between computational efficiency and the optimality of the identified service composition solutions. SBOTI incorporates a non-dominated sorting technique to identify the Pareto-optimal set solutions, which allows the user to explore various QoS trade-offs. To control the diversity of the solutions in this set, SBOTI globally updates both dominated and non-dominated solutions using digital pheromones. To allow for exploration of new service composition configurations that may emerge as a result of providers' mobility, SBOTI uses an adaptation procedure that limits the amount of pheromone on previously identified solutions. SBOTI also engages multiple communities, with diverse properties, to collaboratively address the computational efficiency and optimality concerns introduced by a single community of homogeneous agents.
SBOTI is evaluated using simulations under various dynamic conditions. The evaluation metrics are the size of the dominated space, which indicates the optimality of the identified set of solutions, and communication overhead. Baselines for comparison are SimDijkstra, GoCoMo and a Random approach. Also, an utility metric is used to compare the performance of SBOTI with the baselines that require a-priori articulation of user preferences. The evaluation results illustrate both the strengths and the limitations of SBOTI in a mobile environment, under different network densities and mobility speeds. | en |
