Composite Artificial Neural Network for Controlling Artificial Flying Creature

Abstract

This paper proposes a composite artificial neural network (CANN). The CANN is a method that contains concepts of an evolutionary artificial neural network, a neural network ensemble and subsumption architecture, and designed for efficient robot control. In the CANN, while low-level ANNs work as actual controllers for calculating outputs, a high-level work as a selector. The high-level ANN works up some optimized ANNs, which output real values, into a controller. In order to verify performance of the CANN, numerical experiments are carried out. An artificial flying creature (AFC) is controlled by the CANN for flying to a target point. Motions of the AFC is calculated by a virtual physics environment, which consists of functions of a physical engine PhysX and a simple drag force calculation. Experimental results show that performance of the CANN is higher than that of a simple ANN.Keywordsartificial lifeevolutionary artificial neural networkparticle swarm optimizationneural network ensemble