Multi-agent learning by evolutionary subsumption

We present the emergence of cooperative behaviors of heterogeneous robots by means of evolutionary subsumption in both simulation and real world environments. The key idea of evolutionary subsumption is to apply GP to the design of subsumption architecture, thus hierarchically constructs the control architecture of robots, and enables us to build domain knowledge into the genetic programming system. We claim that this method can facilitate the transformation from simulation to real world. Our approach is evaluated with an "eye"-"hand" cooperation problem. The domain knowledge of this problem is that the "eye" is an observer and the "hand" is an actor, namely we let the "eye" to observe each action of the "hand" and give appraisement as reinforcement signal to the "hand", thus endow the "hand" with learning ability. Experimental results show that by applying this approach, multirobot system exhibits identical behaviors in simulation and real world.