Automatic design of neural network structures

Abstract

Artificial neural networks are being extensively applied in many fields of science and engineering. Despite their wide range of applications and their flexibility, there is still no general framework or procedure through which the appropriate neural network for a specific task can be designed. The design of neural networks is still very dependent upon the designer's experience. This is an obvious barrier to the wider application of neural networks. To mitigate this barrier methods have been developed to automate the design of neural networks. A new method for the auto-design of neural networks was developed, which is based on genetic algorithms (GA) and Lindenmayer Systems. The method is less computationally intensive than existing iterative design procedures, hence it can be applied to the automatic design of neural networks for complex processes. To evaluate the performance of the new design procedure, it was tested for the design of industry standard neural networks. The method was also applied to design neural networks to model the dynamics of a pH neutralization process and a CSTR reactor in which a set of nonlinear reactions takes place. The networks obtained by the new algorithm for these typical chemical processes was much simpler, yet more accurate than those designed by traditional methods.