Co-evolutionary Fitness Switching: Learning Complex Collective Behaviors Using Genetic Programming

摘要

Genetic programming provides a useful paradigm for developing multiagent systems in the domains where human programming alone is not sufficient to take into account all the details of possible situations. However, existing GP methods attempt to evolve collective behavior immediately from primitive actions. More realistic tasks require several emergent behaviors and a proper coordination of these is essential for success. We have recently proposed a framework, called fitness switching, to facilitate learning to coordinate composite emergent behaviors using genetic programming. Coevolutionary fitness switching described in this chapter extends our previous work by introducing the concept of coevolution for more effective implementation of fitness switching. Performance of the presented method is evaluated on the table transport problem and a simple version of simulated robot soccer problem. Simulation results show that coevolutionary fitness switching provides an effective mechanism for learning complex collective behaviors which may not be evolved by simple genetic programming. Evolving complex collective behaviors is an interesting problem for distributed intelligence and artificial life. Some tasks can be done faster or more easily by dividing them up among