Training a Feed-Forward Neural Network Using Particle Swarm Optimizer with Autonomous Groups for Sonar Target Classification

Feed-Forward Neural Networks (FFNNs), as one of the wide-spreading Artificial NNs, has been used to solve many practical problems such as classification of the sonar dataset. Improper selection of the training method, which is an important part of the design process, results in slow convergence rate, entrapment in local minima, and sensitivity to initial conditions. To overcome these issues, the recently proposed method known as Particle Swarm Optimizer with Autonomous Groups (AGPSO) has been used in this paper. It is known that the FNNs are very sensitive to the problems dimension, so clearly applying it to a dataset with large dimension results in poor performance. However, the combination of FNN and AGPSO solves this problem because of the ability of AGPSO in the optimization of high dimension problems. To evaluate the performance of the proposed classifier, it is applied to other datasets and the results are compared to the standard PSO, modified PSO (IPSO), Gravitational Search Algorithm (GSA) and Gray Wolf Optimizer (GWO). Simulation results show that the AGPSO classifier provides better performance in terms of convergence speed, entrapment in local minima and classification accuracy compared to the other algorithms.