2013 年 79 巻 6 号 p. 552-558
This paper describes a novel method of combining artificial neural networks (ANNs), the composite artificial neural network (CANN), to improve performance of evolving ANNs (EANNs) in dynamic system control problems. Methods of combining ANNs by majority voting or averaging are not effective in controlling dynamic systems by EANNs. Unsupervised learning of EANNs is mathematically described, and then it is shown that the reason for ineffectiveness depends on the mechanism of evaluating ANNs indirectly by states. To avoid this problem, the CANN selects the suitable ANN by a high-level ANN to combine some ANNs. In numerical experiments, a flapping flight model is controlled by a common EANN and the CANN. The model motion is calculated by physical engine PhysX, and a common EANN and the CANN are optimized by the particle swarm optimization (PSO) respectively. The experiments show that the average evaluation of the CANN is 6.46% higher than that of a common EANN for the same computational time.
