An enhancement of constraint feasibility in BPN based approximate optimization

Abstract

Back-propagation neural networks (BPN) have been extensively used as global approximation tools in the context of approximate optimization. A traditional BPN is normally trained by minimizing the absolute difference between target outputs and approximate outputs. When BPN is used as a meta-model for inequality constraint function, approximate optimal solutions are sometimes actually infeasible in a case where they are active at the constraint boundary. The paper explores the development of the efficient BPN based meta-model that enhances the constraint feasibility of approximate optimal solution. The BPN based meta-model is optimized via exterior penalty method to optimally determine interconnection weights between layers in the network. The proposed approach is verified through a simple mathematical function and a ten-bar planar truss problem. For constrained approximate optimization, design of rotor blade is conducted to support the proposed strategies.