Ibots: Learning Real Team Solutions

This paper presents \Ibots" (Integrating roBOTS), a computer experiment in team robotics designed on an arti cial mission. Our aim is to understand how to use reinforcement learning to program automatically a team of robots with a shared mission. Moreover, we are interested in learning real team solutions. These are programs whose form strongly depends on the number of robots composing the team, on their individual skills and limitations, and on any other mission boundary condition which makes it worth to prefer \at a team level" certain solutions to others. The Ibots mission is speci ed implicitly by means of a single reinforcement signal which measures the team performance as a whole. This form of payo leads to real team solutions. Bene ts and drawbacks of using team reinforcement as opposed to individual robot reinforcement are discussed.