Abstract
In recent years, multi-agent systems that perform tasks by using multiple autonomous robots have been researched. We try to acquire cooperative behavior by using a method called Graph Structured Program Evolution (GRAPE) which uses graph structure as a representation of programs. GRAPE was applied to test problems such as making a sort program and demonstrate that the optimum solution in each problem. In this paper, we apply GRAPE to tasks which need cooperative behaviors among agents and introduce a new method in genetic operators focusing a using rate of graph structure. We treat a box pushing task and verify the effectiveness.
