A ternary bitwise calculator based genetic algorithm for improving error correcting output codes

This paper proposes a novel genetic algorithm (GA) for the error correction output coding (ECOC) framework. Different from other GA algorithms, a new individual structure is designed by setting a gene as the combination of three types of operators: (1) the column selector; (2) the ternary bitwise calculator; (3) the feature selector. In our GA algorithm, two column selectors first extract two columns to from a codematrix pool, then a Ternary bitwise Calculator (TC) transfers them to a new column through a ternary number calculation process. The feature selector selects a feature subset for training the associated dichotomizer. By doing so, an individual contains a set of genes to form an ECOC ensemble. The TCs deployed in our algorithm include both the traditional TCs and some newly proposed TCs, which aid to generate diverse codematrices. When the evolutionary process terminates, the best individual in the last generation is regarded as the final solution. The performance of our algorithm is verified on both the UCI and microarray data sets. Experiment results demonstrate that our GA based ECOC achieves promising performance comparing to other ECOC algorithms. Furthermore, results also confirm that various TCs contribute to the generation of discriminative individuals. (C) 2020 Elsevier Inc. All rights reserved.