Binary Particle Swarm Optimization for Variables Selection Optimization in Taguchi’s T-Method

Abstract

Prediction analysis has drawn significant interest in numerous field. Taguchi’s T-Method is a prediction tool that developed practically but not limited to small sample analysis. It was developed explicitly for multidimensional system prediction by relying on historical data as the baseline model and adapting the signal to noise ratio (SNR) as well as zero proportional concepts in strengthening its robustness. Orthogonal array (OA) in T-Method is a variable selection optimization technique in improving the prediction accuracy as well as help in eliminating variables that may deteriorate the overall performance. However, the limitation of OA in dealing with higher multidimensionality restraint the optimization accuracy. Binary particle swarm optimization used in this study helps to cater to the limitation of OA as well as optimizing the variable selection process to better prediction accuracy. The results show that if the historical data consist of samples with higher correlation of determination (R²) value for the model creation, the optimization process in reducing the number of variables would be much reliable and accurate.  Comparing between T-Method+OA and T-Method+BPSO in four different case study, it shows that T-Method+BPSO performing better with greater R² and means relative error (MRE) value compared to T-Method+OA.