Abstract
High dimensionality of data represents a major problem that affects the accuracy of the classification. This problem related with classification is mainly resulted from the availability of irrelevant features. Feature selection represents a solution to a problem by selecting the most informative features and discard the irrelevant features. Generalized normal distribution optimization (GNDO) represents a newly developed optimization that confirmed its outperformance in comparison with well-known optimization algorithms on parameter extraction for photovoltaic models. As an optimization algorithm, however, GNDO suffers from degraded performance when dealing with a problem with a high dimensionality. The main problems of GNDO include exploitation problem by falling into local optima problem. Also, GNDO has solutions diversity problem when it deals with data with high dimensionality. To alleviate the drawbacks of this algorithm and solve feature selection problems, a local search algorithm (LSA) is used. The new algorithm is called dynamic generalized normal distribution optimization (DGNDO), which includes the following main improvements to GNDO: it can improve the best solution to solve the local optima problem, it can improve solution diversity by improving the randomly selected solution, and it can improve both exploration and exploitation combined. To confirm the outperformance and efficiency of the new DGNDO algorithm, DGNDO algorithm is applied on 20 benchmarked datasets from UCI repository of data. In addition, DGNDO algorithm results are compared with seven well-known optimization algorithms using number of evaluation metrics including classification, accuracy, fitness, the number of selected features, statistical results using Wilcoxon test and convergence curves. The obtained results reveal the superiority of DGNDO algorithm over all other competing algorithms.