Research On Optimization Of Service Composition Based On Partially Observable Environment

In a service-oriented architecture,web service composition provides a new model for software construc-tion by integrating multiple exiting services into a value-added composite service to satisfy more complex demands.As an efficient way to build complex software applications,it is widely recognized for its advan-tages of high efficiency and flexibility.The goal of web service composition is to find the optimal service composition strategy under the premise of meeting the specific functional requirements.In real application scenarios,there are many uncertainties and dynamic factors in the service composition system,such as the changes of the external environment,the evolution of the services,as well as the uncertain result of the behavior and attributes,which will affect the result and performance of service composition.Thus,this requires that the service composition method should have certain adaptive ability.In the meanwhile,considering that the operating environment of the system can not be fully observed in some cases,and that the internal state is opaque,some available information is incomplete.Therefore,service composition method should also provide an effective solution for this kind of partial observability environment.Based on the above discussion,this thesis introduces the Partially Observable Markov Decision Pro-cess(POMDP)into web service composition,the POMDP model regards the running system as partially observable,and uses the observable partial information to model and solve the problem.Therefore,the introduction of POMDP can solve the partial observability of the information in the service composition process.Combining the self-learning characteristic of reinforcement learning,we combine the POMDP ser-vice composition model and the reinforcement learning algorithm to realize the adaptive service composition technology.Considering the continue expansion of the scale of the service composition,we modify the tra-ditional reinforcement learning algorithm and implement an optimization algorithm based on kernel method.This improved algorithm can achieve better generalization ability through the kernel function approximation technique,and thus have good performance for large scale service composition problem.Finally,a series of simulation experiments are used to verify the effectiveness,scalability,adaptability and statistical significance of the proposed algorithm.