The repetitive optimization design strategy using neural network and hybrid algorithm

In this paper, a Bayesian learning technique, mapped into feed-forward artificial neural networks, is considered as a system approximation, which, for training, highly non-linear and implicit complex functions. This process is integrated with a hybrid algorithm (HA) in the proposed design optimization strategy. The combination of the back-propagation Levenberg-Marquardt (BPLM) algorithm and the Bayesian learning technique shows good and accurate generalization, which creates the meta-model, considered as the fitness and constraints function in the hybrid algorithm. Here, a genetic algorithm (GA), hybridized with a local gradient-based method, performs the effective and robust evolutionary search and reduces the computation cost. D-optimality is used to select the appropriate points in the design space, to obtain the significant responses. A numerical example, the design of a two-member frame and air intercept missile-AIM design optimization problem are presented to demonstrate the accuracy and feasibility of the process.