Taguchi's T-method with nearest integer-based binary bat algorithm for prediction

Abstract

Taguchi’s T-method is a new prediction technique under the Mahalanobis-Taguchi system to predict unknown output or future states based on available historical information. Conventionally, in optimizing the T-method prediction accuracy, Taguchi’s orthogonal array is utilized to determine a subset of significant features to be used in formulating the optimal prediction model. This, however, resulted in a sub-optimal prediction accuracy due to its fixed and limited feature combination offered for evaluation and lack of higher-order feature interaction. In this paper, a swarm-based binary bat optimization algorithm with a nearest integer discretization approach is integrated with the Taguchi’s T-method. A comparative study is conducted by comparing the performance of the proposed method against the conventional approach using mean absolute error as the performance measure on four benchmark case studies. The results from experimental studies show a significant improvement in the T-method prediction accuracy. A reduction in the total number of features results in a less complex model. Based on the general observation, the nearest integer-based binary bat algorithm successfully optimized the selection of significant features due to recursive and repetitive searchability, in addition to its adaptive element in response to the current best solution in guiding the search process towards optimality.

Enhanced Taguchi’s T-method using angle modulated Bat algorithm for prediction

“…The factorial effect chart reveals the relative importance of features based on the deterioration of the average integrated estimate SNR (db) of features when it is not used in combination. The detail of the steps is elaborated in [16].…”

“…Studies by [14], [15] that employ an artificial bee colony algorithm and a modified artificial bee colony algorithm to optimize the feature selection process show promising improvement in the T-method prediction accuracy over the conventional OA approach in terms of error rate. In the latest development, [16] reported that integrating the binary Bat algorithm based on the nearest integer binarization scheme with the T-method enhanced the prediction accuracy ranging from 7.2% to 55%, dependent on case studies. In a continuous quest for optimality, the ideal approach is through an exhaustive search approach where every possible combination is assessed and evaluated.…”