Training multilayer perceptrons in neural network using interior search algorithm

Abstract

Multilayer perceptron (MLP) is the most popular neural network method and it has been widely used for many practical applications. In this paper, recently developed interior search algorithm (ISA) is proposed for training MLP. Five of most important standard classification datasets (balloon, XOR, Iris, heart, and breast cancer) are employed to evaluate the proposed algorithm performance. The obtained results from ISA-based are compared with five well-known algorithms including ant colony optimization (ACO), genetic algorithm (GA), particle swarm optimization (PSO), population-based incremental learning (PBIL), and evolution strategy (ES). The statistical results reflect that the performance of the proposed algorithm can train MLPs with a very high degree of accuracy and it is capable of outperforming the well-known algorithms. The results also show that the high convergence rate of the ISA and it is potential to avoid local minima.