Characterization of bees algorithm into the Mahalanobis-Taguchi system for classification

Abstract

Mahalanobis-Taguchi System (MTS) is a pattern recognition tool employing Mahalanobis Distance (MD) and Taguchi Robust Engineering philosophy to explore and exploit data in multidimensional systems. In order to improve recognition accuracy of the MTS, features that do not provide useful and beneficial information to the recognition function is removed. A matrix called Orthogonal Array (OA) to search for the useful features is utilized by MTS to accomplished the search. However, the deployment of OA as the feature selection search method is seen as ineffective. The fixed-scheme structure of the OA provides a non-heuristic search nature which leads to suboptimal solution. Therefore, it is the objective of this research to develop an algorithm utilizing Bees Algorithm (BA) to replace the OA. It will act as the alternative feature selection search strategy in order to enhance the search mechanism in a more heuristic manner. To understand the mechanism of the Bees Algorithm, the characteristics of the algorithmic nature of the algorithm is determined. Unlike other research reported in the literature, the proposed characterization framework is similar to Taguchi-sound approach because Larger the Better (LTB) type of signal-to-noise formulation is used as the algorithm’s objective function. The Smallest Position Value (SPV) discretization method is adopted by which the combinations of features are indexed in an enumeration list consisting of all possible feature combinations. The list formed a search landscape for the bee agents in exploring the potential solution. The proposed characterization framework is validated by comparing it against three different case studies, all focused on performance in terms of Signal-to-Noise Ratio gain (SNR gain), classification accuracy and computational speed against the OA. The results from the case studies showed that the characterization of the BA into the MTS framework improved the performance of the MTS particularly on the SNR gain. It recorded more than 50% improvement (on average) and nearly 4% improvement on the classification accuracy (on average) in comparison to the OA. However, the OA on average was found to be 30 times faster than the BA in terms of computational speed. Future research on improving the computational speed aspect of the BA is suggested. This study concludes that the characterization of BA into the MTS optimization methodology effectively improved the performances of the MTS, particularly with respect of the SNR gain performance and the classification accuracy when compared to the OA.