Research On Service Composition Using Multi-Objective Reinforcement Learning And Skyline Computing

In service computing, a simple atomic web service cannot meet increasing and complex users'require-ments. In this case, the combination of existing services to accomplish the needs of users has become a hot research topic, namely, service composition.With the rapidly development of the Internet, the growth of the number of web service with same func-tional properties but difference in QoS attributes, along with as the complexity of the business process, led to low efficiency problem in service composition. QoS-aware service composition method is aim to maximize meet customer needs and become a hot research topic. On the one hand, with the changing environment and online web services keep evolving, web service composition adjusts its behavior in response to changes in order to maintain its performance even reach a better performance. On the other hand, most of existing service composition methods synthesize multiple QoS attributes to one synthesized function regardless of whether these QoS attributes are conflicting or not.To handle these key challenges, we put forward a new service composition scheme, integrating Multi-Objective Reinforcement Learning and Skyline Computing for service composition. For web services with multidimensional QoS properties, we skyline Computing extracts no-dominated web service and reduces com-bination space to reduce the complexity. Multi-Objective Reinforcement Learning is cope with the dynamic changing environments and multiple conflicting objectives in service composition. Reinforcement learning (RL) is a major class of machine learning methods to solve sequential decision making problems. In an RL system, a learning agent aims to learn an optimal action policy via interactions with an uncertain and dynam-ic environment, thus, the RL is adaptive to changing environment. However, the traditional reinforcement learning method aims to learn an optimal strategy with maximum reward, which is reinforcement learning with single objective. For adaptive service composition with multiple objectives scenarios, the traditional reinforcement learning is no longer appropriate and we use Multi-Objective Reinforcement Learning to solve this problem. Finally, we conducted a set of experiments to demonstrate the effectiveness, scalability and self-adaptivity of the method.