Abstract: Multi-Agent System (MAS) designing has faced some challenging work such as cooperation among agents which are vital to the performance of this system. A much advanced agent role-value method based upon Q-learning is proposed in this paper to avoid the unstabilizing factors and the loss of efficiency caused by possibility of too frequent role switching between robots. Other new methods based on this role model are suggested to solve the problems associated with system designing and implementation. Application to Federation of International Robot-Soccer Association (FIRA) simulation system proves that this method is effective, and reduces the possibility that the robots loss ball, fumble ball and nonfeasance, and remedies the shortage that roles are assigned according to fixed regions.