On the optimum design of cluster structures by using a genetic algorithm

✍ Scribed by Honglan Jin; Yoshikazu Miyanaga; Koji Tochinai

• Publisher: John Wiley and Sons
• Year: 1998
• Tongue: English
• Weight: 164 KB
• Volume: 81
• Category: Article
• ISSN: 8756-663X
• DOI: 10.1002/(sici)1520-6432(199803)81:3<53::aid-ecjb7>3.0.co;2-7

✦ Synopsis

In discussing self-organizing neural networks, to some extent a large-scale network is assumed in order to achieve generality and adaptability. This paper discusses an optimal structurization method for a nonlinear network, based on a self-organizing algorithm with a two-layer structure. The basic structure of the network combines a selforganizing layer and a single-layer perceptron network. In the learning stage, both the self-organizing algorithm and the supervised learning algorithm are applied for each datum. Because of this structure, the network achieves highly precise signal processing based on learning, i.e., self-organization and supervised learning. A previous paper used this kind of network in the estimation of spectra. However, among problems that remained were the long processing time required in the learning stage due to the formation of unnecessary cluster nodes, and the fact that unnecessary nodes sometimes degrade estimation performance. From this perspective, it seems important in achieving a high-speed and highly precise system, to optimize cluster structure by eliminating unnecessary nodes. This paper presents a method for optimal network design based on a genetic algorithm that can attain a smaller scale network with higher precision than any conventional network. It is shown that performance is better than for a conventional network. The network is applied to the spectra estimation problem to demonstrate its effectiveness.