Hybrid bootstrap-based approach with binary artificial bee colony and particle swarm optimization in Taguchi's T-Method

Abstract

Taguchi's T-Method is one of the Mahalanobis Taguchi System (MTS)-ruled prediction techniques that has been established specifically but not limited to small, multivariate sample data. When evaluating data using a system such as the Taguchi's T-Method, bias issues often appear due to inconsistencies induced by model complexity, variations between parameters that are not thoroughly configured, and generalization aspects. In Taguchi's T-Method, the unit space determination is too reliant on the characteristics of the dependent variables with no appropriate procedures designed. Similarly, the least square-proportional coefficient is well known not to be robust to the effect of the outliers, which indirectly affects the accuracy of the weightage of SNR that relies on the model-fit accuracy. The small effect of the outliers in the data analysis may influence the overall performance of the predictive model unless more development is incorporated into the current framework. In this research, the mechanism of improved unit space determination was explicitly designed by implementing the minimum-based error with the leave-one-out method, which was further enhanced by embedding strategies that aim to minimize the impact of variance within each parameter estimator using the leave-one-out bootstrap (LOOB) and 0.632 estimates approaches. The complexity aspect of the prediction model was further enhanced by removing features that did not provide valuable information on the overall prediction. In order to accomplish this, a matrix called Orthogonal Array (OA) was used within the existing Taguchi's T-Method. However, OA's fixed-scheme matrix, as well as its drawback in coping with the high-dimensionality factor, leads to a sub- optimal solution. On the other hand, the usage of SNR, decibel (dB) as its objective function proved to be a reliable measure. The architecture of a Hybrid Binary Artificial Bee Colony and Particle Swarm Optimization (Hybrid Binary ABC-PSO), including the Binary Bitwise ABC (BitABC) and Probability Binary PSO (PBPSO), has been developed as a novel search engine that helps to cater the limitation of OA. The SNR (dB) and mean absolute error (MAE) were the main part of the performance measure used in this research. The generalization aspect was a fundamental addition incorporated into this research to control the effect of overfitting in the analysis. The proposed enhanced parameter estimators with feature selection optimization in this analysis had been tested on 10 case studies and had improved predictive accuracy by an average of 46.21% depending on the cases. The average standard deviation of MAE, which describes the variability impact of the optimized method in all 10 case studies, displayed an improved trend relative to the Taguchi’s T-Method. The need for standardization and a robust approach to outliers is recommended for future research. This study proved that the developed architecture of Hybrid Binary ABC-PSO with Bootstrap and minimum-based error using leave-one-out as the proposed parameter estimators enhanced techniques in the methodology of Taguchi's T-Method by effectively improving its prediction accuracy.