Social behaviour in intelligent agent systems is often considered to be achieved by deliberative, in-depth reasoning techniques. This paper shows, how a purely reactive multi-agent system can learn to evolve cooperative behaviour, by means of learning from previous experiences. In particular, we describe a learning multi agent approach to the problem of controlling power flow in an electrical power-grid. The problem is formulated within the framework of dynamic programming. Via a global optimization goal, a set of individual agents is forced to autonomously learn to cooperate and communicate. The ability of the purely reactive distributed systems to solve the global problem by means of establishing a communication mechanism is shown on two prototypical network configurations.
展开▼
