Pattern generation for multi-class LAD using iterative genetic algorithm with flexible chromosomes and multiple populations

In this paper, we consider a pattern generation method for multi-class classification using logical analysis of data (LAD). Specifically, we apply two decomposition approaches—one versus all, and one versus one—to multi-class classification problems, and develop an efficient iterative genetic algorithm with flexible chromosomes and multiple populations (IGA-FCMP). The suggested algorithm has two control parameters for improving the classification accuracy of the generated patterns: (i) the number of patterns to select at the termination of the genetic procedure; and (ii) the number of times that an observation is covered by some patterns until it is omitted from further consideration. By using six well-known datasets available from the UCI machine-learning repository, we performed a numerical experiment to show the superiority of the IGA-FCMP over existing multi-class LAD and other supervised learning algorithms, in terms of the classification accuracy.