We have been studying a new, lean and agile manufacturing concept. In this paper, we focus on an Automated Guided Vehicle (AGV) dispatching problem which in most cases, is resolved by using a First Come First Service (FCFS) rule in existing systems. We have developed a cooperative dispatching strategy using a Stochastic Learning Automaton (SLA) reinforcement learning method within the framework of a multi-agent system. We have implemented the idea that if an agent can't learn for itself, it should imitate other agents. We think the learning property including an imitation should be acquired in accordance with an agent's characteristics, such as its speed and failure rate. We will present an approach in which two kinds of SLAs run in parallel and each agent can autonomically acquire both behavior and learning properties. This paper demonstrates that our strategy can achieve an approximate 30% improvement in handling capability, compared with a FCFS rule in a realistic system. Furthermore, we demonstrate that an AGV can obtain its desired behavior more quickly by composing double SLAs.