Structural learning of neural networks by coevolutionary genetic algorithm with degeneration
Tetsuyuki Takahama, Setsuko Sakai · 2005
Structural learning, in which the structure of estimation systems are optimized, has been actively studied in research on supervised learning of fuzzy rules and neural networks. The Ga/sup d/ (genetic algorithm with degeneration) is one of the structural learning methods, which is modeled on genetic damage and degeneration. In GA/sup d/, degeneration pressure must be controlled properly to get the better structure of the estimation systems. But it is difficult to tune the degeneration pressure manually. In this paper, the idea of coevolution is introduced into GA/sup d/ and a coevolutionary genetic algorithm with degeneration (CGA/sup d/) is proposed to control the degeneration pressure adoptively. Coevolution is an evolution model, where two types of individuals evolve cooperatively or competitively each other, in CGA/sup d/, the learning individuals that learn the estimation systems and the control individuals that control the degeneration of the learning individuals evolve cooperatively. To show the advantage of CGA/sup d/ it is applied to the structural learning of neural networks. It is shown that CGA/sup d/ can find better structures than GA/sup d/.