Learning joint coordinated plans in Multi-Agent Systems

One important class of problems in Multi-Agent Systems (MASs) is planning, that is constructing an optimal policy for each agent with the objective of reaching some terminal goal state. The key aspect of multi-agent planning is coordinating the actions of the individual agents. This coordination may be done through communication, learning, or conventions imposed at design time. In this paper we present a new taxonomy of MASs that is based on the notions of optimality and rationality. A framework that describes the interactions between the agents and their environment is given, along with a reinforcement learning-based algorithm (Q-learning) for learning a joint optimal plan. Finally, we give some experimental results on grid games that show the convergence of this algorithm.