Evolutionary regional network modeling for efficient engineering optimization

This study presents a soft computing based optimization methodology, the evolutionary regional neural network modeling for engineering applications with sampling constraints. Engineering optimization often involves expensive experiment costs. Intelligent optimization advocates establishing a neural network model using small training samples such as orthogonal array to set up a surrogate model for the engineering system followed by an optimum search in the model to reduce optimization cost. However, scarce training samples might compromise modeling generality for a complex problem. Empirical rules suggest reliable predictions are likely restricted to the neighboring space of training samples, and interpolating designs are more reliable than extrapolating designs. To avoid imperfection of model due to small learning samples, an evolutionary regional network model is set up to confine the search of quasi-optimum using genetic algorithm. The constrained search in the regional network model provides a reliable quasi-optimum. The verification of the optimum is added to the learning samples to retrain the regional network model while the size and the distribution of reliable space will evolve intelligently during the optimization iteration using a fuzzy inference according to the prediction accuracy. An engineering case study, the optimal die gap parison programming of extrusion blow molding process for a uniform thickness, is presented to demonstrate the robustness and efficiency of the proposed methodology.