Abstract
This paper describes a behavior acquisition by a composite artificial neural network (CANN) and a region divided artificial neural network (RDANN). The CANN includes a high-level ANN, which is used for deciding which low-level ANN is activated to calculate output signals. The RDANN is a special case of the CANN, in which the structure or the selection function of the high-level ANN is given by a designer. To examine the relation between parameters of the high-level ANN and learning effectiveness, a CANN is applied for pole balancing problems. The controller, a CANN, a RDANN or a common ANN, is optimized by the particle swarm optimization. Experimental results show that the angle of the pole is important for selection by the high-level ANN. Results of additional experiment show that the RDANN that is designed to select low-level ANNs by the angle of pole is the most effective for pole balancing.
