Dynamic Genotype Reduction for Narrowing the Feature Selection Search Space

Large search space optimization problems can pose a challenge to any optimization approach, as there are many chances for the algorithm to become stuck in the local optima. In this paper, we present a novel dynamic representation of genotype, where the search areas with low potential are discarded from the search, and areas of the more significant potential for global optimum are kept in focus. The validity of the proposed method is tested on a large scale feature selection problem, using an extensive dataset with a large number of features compared to instances. An arrhythmia dataset was specifically chosen for a case study, and a self-adaptive differential evolution algorithm with the dynamic genotype reduction was implemented. The results of the experiments are promising for the future since the proposed method achieves better results than standard feature selection with stochastic population-based nature-inspired algorithms.