A Learning Algorithm For Service Oriented Agent

Service oriented architecture, such as Web service technology, and thesoftware agent technology has been widely applied to a variety of IT fields.The service agent technology, based on Web service and software agent, is aneffective combination of them. It contains not only the repeatability and low-coupling features of Web services, but also the persistent, autonomous andinteractive features of the software agent. As another significant feature of theintelligent software agent, it must also have the capability of learning throughthe knowledge of the past. Confronted with a totally new challenge, theservice agent should find an optimal solution independently and quickly.This thesis proposes a service agent model, the semantic architecture aswell as the working mechanisms for the Web service composition. A noveleffective algorithm for service agent learning is given. This algorithm is basedon the reinforcement learning and the team Markov game process. Moreover,this algorithm makes it possible to utilize the knowledge learned in the past.The experiments show that in complex circumstances, especially in multi- agent system, this algorithm can achieve convergence quickly. The algorithmtakes advantage of the theoretical basis of Q-learning. Both the multi-agentgame and negotiation mechanisms are considered. Meanwhile, the algorithmalso supports the dynamic storage and update of the knowledge model, so thatthe convergence can be achieved more efficiently.