AUTOMATIC DESIGN OF NEURAL NETWORKS BASED ON GENETIC ALGORITHMS

Abstract

Nowadays, the theoies modelm methods and tools concerning neural networks are approaching complete and mature. Nevertheless, there still exists a main difficulty for industrial applications. That is how to design optimal network architecture according to every specific problem. The task includes optimization of network size, network topology, connection weights between neurons etc. This paper proposes an automatic design methodology of neural networks based on generic algorithms. We analyze firstly the building blocks of neural networks in ordre to obtain design specifications. Then, we develop a genetic encoding method, after which the evolution process is elaborated for finding the optimal neural networks. The results of our experiments reveal that out methodology is superrior to the error back-propagation algorithm both for its executing efficiency and performance.