A Multi-agent Learning Model for Service Composition

Agent technology has gained increasing popularity in service oriented architecture (SOA) because of its features of autonomy, initiative, interactivity, persistency and adaptability. There are already a plenty of implementations which integrate SOA with multi-agent systems (MAS). The ability of learning is a significant feature of MAS. This paper proposes a learning model of the service-oriented MAS for the service composition problem. It adopts the principle of reinforcement learning and is based on the Markov game and Q-learning. The reward of the learning procedure is determined by the QoS parameters such as responding time and cost. The mechanism of multi-agent leaning for service composition is introduced. The results of experiments and case study show that our multi-agent learning approach can reach convergence efficiently and it can also accelerate the service composition process based on the knowledge continuously learned from past composition experiences.