Achieving Coordination through Combining Joint Planning and Joint Learning

There are two major approaches to activity coordination in multiagent systems. First, by endowing the agents with the capability to jointly plan, that is, to jointly generate hypothetical activity sequences. Second, by endowing the agents with the capability to jointly learn, that is, to jointly choose the actions to be executed on the basis of what they know from experience about the interdependencies of their actions. This paper describes a new algorithm called JPJL (“Joint Planning and Joint Learning”) that combines both approaches. The primary motivation behind this algorithm is to bring together the advantages of joint planning and joint learning while avoiding their disadvantages. Experimental results are provided that illustrate the potential benefits and shortcomings of the JPJL algorithm.