Training Multi-layer Perceptron Using Hybridization of Chaotic Gravitational Search Algorithm and Particle Swarm Optimization

Abstract

A novel amalgamation strategy, namely, chaotic gravitational search algorithm (CGSA) and particle swarm optimization (PSO), has been employed for training multi-layer perceptron (MLP) neural network. It is called CGSAPSO. In CGSAPSO, exploration is carried out by CGSA, and exploitation is performed using PSO. The sigmoid activation function is utilized for training MLP. Besides, a matrix encoding strategy has been used for providing a synergy between neural biases, weights, and CGSAPSO searcher agents. To validate the effectiveness of the hybrid framework, CGSAPSO is applied to three different classification datasets, namely, XOR, Iris, and Balloon. The investigation of results is carried out through various performance metrics like average, standard deviation, median, convergence speed, execution time, and classification rate analysis. Besides, a pair-wise non-parametric signed Wilcoxon rank-sum test has also been conducted for statistical verification of simulation results. In addition, the numerical outcomes of CGSAPSO are also compared with standard GSA, PSO, and hybrid PSOGSA. The experimental results indicate that CGSAPSO provides better results in the form of recognition accuracy and global optima as compared to competing algorithms.