The Use of Radial Basis Function Networks Based on Leader Mass Gravitational Search Algorithm for Sonar Dataset Classification

Due to the very similar physical characteristics of active sonar echoes, which are related to true and false targets, the task of classification and distinguishing them from each other has become difficult and complex issues for the researchers and industrialists in this field. Radial Basis Function Neural Networks (RBF NN) is one of the most used artificial NNs in the classification of the real-world objects. Training is an important part of the development of this type of network that it has been highly regarded in recent years. For RBF NN training, the use of recursive and gradient descent methods has traditionally been common. However, poor classification accuracy, trapped in local minima, and low convergence speed are among the disadvantages of these methods. In recent years, the use of heuristic and meta-heuristic methods has been very common to overcome these disadvantages. To overcome the GSA’s weakness in the exploitation phase, this paper introduces and uses Leader Mass Gravitational Search algorithm (LMGSA) in the training of RBF NNs. The results show that the designed classifier provides better results than benchmark classifiers in all areas. To a comprehensive comparison, the designed classifier is compared to GSA, Gradient Descent (GD), Genetic Algorithm (GA), Kalman Filter (KF), and Extended Kalman Filter (EKF) algorithms through three benchmark datasets. The evaluation criteria are convergence rate, the probability of being caught in local minima, and classification accuracy. Finally, as a practical application, sonar dataset is classified by the designed classifier.