Study on evolution of the artificial flying creature controlled by neuro-evolution

摘要

The objective of this study is to realize efficient learning for the high generalization ability of evolutionary artificial neural network (EANN). In order to achieve this objective, the evolutionary process of behavior acquisition is analyzed, and then an efficient evaluation function is led by the analysis. An artificial flying creature (AFC) is controlled to fly towards a given target point by EANN. The three-dimensional motion of the AFC is calculated by the physical engine PhysX and a numerical expression of the simple drag force. To evolve ANNs and to have the AFC flight suitably for given target points, particle swarm optimization optimizes parameters of ANNs. The results of evolutionary simulation show that generalization ability of ANNs does not always increase as evolution progresses, and it depends on given tasks of the AFC. It is also shown that diversity of input signals about target points, which the AFC goes through in flight, has positive correlation with generalization ability.