Optimal Feature Selection of Taguchi Character Recognition in the Mahalanobis-Taguchi System using Bees Algorithm

The Mahalanobis-Taguchi System (MTS) is a data mining method employing Mahalanobis distance (MD) and Taguchi′s Robust Engineering philosophy to explore and exploit data in a multidimensional system. The MD calculation provides a measurement scale to discriminate sample data and gives an approach of measuring the level of severity among them. One unique feature of MTS lies its robustness to assess variability among all levels of samples (noise) and ability to evaluate significant and insignificant factors which contributed to the system (optimization) by means of simplistic yet robust technique via orthogonal array (OA) and signal to noise ratio (SNR). The optimized system obtained is considered robust, since the SNR identifies the useful variables that are most insensitive to variation, and cost efficient, as it constitutes a smaller number of attributes with better system performance. In this paper, a novel useful variable selection (feature selection) approach using Bees Algorithm (BA) replacing conventional OA technique is presented. BA is a heuristic search technique that finds optimal (or near optimal) result which falls under the Swarm Intelligence field. The solution search strategy mimics social behaviour of animals or insects (bee colony in particular). MD is used as the result assessment metric while the larger-the-better type of SNR is deployed as the algorithm objective function. Character recognition based on Taguchi concepts (exploiting variation and abundance items) is used as the case study on which the comparison between BA and OA performances is made. The results show a promising discriminant power of the optimized system via BA as compared to OA, however, the OA approach outperforms BA in terms of optimization speed to a great extent. AMS subject classification: