Abstract
This paper describes numerical simulations of a moving virtual robot. The virtual robot moves on a two-dimensional plane by the law based on the Newton's equations. The motion of the virtual robot is controlled by an artificial neural network (ANN). In order to control the virtual robot moving for a given target place, parameters of ANNs are optimized by particle swarm optimization (PSO). This type of control and learning method is called evolutionary ANN (EANN). It is shown by many numerical simulations that difficulties of learning rely on a position at which the target is placed. The phenomena is also found that a generalization ability of an ANN, that is, the ability to go toward not learned targets, decreases as learning of EANN progresses. Additional simulations show that learning simultaneously for two of more target places brings the high generalization ability.
