Abstract
Electroencephalogram recordings provide insightful information concerning the diagnosis and prognosis of human thinking and memory-related processes, aiding researchers and physicians during Brain-Computer Interface systems development. In electroencephalogram memory pattern identification, feature extraction, and feature selection are determining factors for an impartial data description and an accurate classification. The electroencephalogram signals analyzed in this study are collected from sixteen electrodes split into four frequency bands during specific working memory-related tasks on different reasoning scenarios.
Although most genetic algorithm based optimization procedures tackle the minimization of a classifier’s error rate and the number of selected features, they are independent of how feature selection procedures are configured, either in single or multi-objective optimization manners, the major problem is multidimensionality and quantity of redundant and noisy electroencephalogram recordings. Since single objective optimization applied separately for two objectives: the minimization of the misclassification rate and the minimization of the number of selected features bias the final results to a specific objective direction, all these limited explorations ground the use of multi-objective optimization procedures for better and sound results.
Regarding all the multi-objective optimization procedures, the compared Pareto ranking schemes are meant for the selection of parents and survivors in evolutionary multi-objective optimization. Usually, Pareto methods use only the dominance analysis for providing the partial sorting of solutions without considering the specific strength of the conflict between them. The methods compared in this paper assign the ranks by combining the search with the decisional mechanism. The decision is implemented through adaptive grouping schemes meant to guide the search towards the middle of the first Pareto fronts, enabling the progressive rejection of profitless solutions. The population is split into several groups to preserve its diversity, and a supplementary objective is added to control the variety of the most valuable genetic information. Finally, the layout of the available solutions in the objective space is examined based on clustering procedures and by individually ranking of the resulting clusters of solutions to counteract the inherent disadvantages of Pareto methods. All compared ranking schemes demonstrate their effectiveness during the evolutionary selection of features. Furthermore, various classifiers distinctively address the problem at hand, illustrating different decisional mechanisms.