Hybrid Adaptive Moth-Flame Optimizer and Opposition-Based Learning for Training Multilayer Perceptrons

Abstract

AbstractThis chapter is dedicated to improving the optimization capability of moth-flame optimizer (MFO), which works based on a spiral operation mimicking the moths’ navigation behavior and dynamic coefficients. However, such a basic version can be easily trapped in the local optima and is associated with an unstable balance between exploratory and exploitative cores. To mitigate the shortcomings of slow convergence and local stagnation, an adaptive moth flame optimization with opposition-based learning (AMFOOBL) is presented, employing a new adaptive structure and the opposition-based learning (OBL) strategy. This original adaptive tool is devised to reduce the number of flames around which agents update their positions for balancing the exploration and exploitation stages more effectively. The performance of AMFOOBL is evaluated in two experiments: First, the quantitative results of 23 benchmark function tests show that AMFOOBL outperforms AMFO, followed by MFO, validating the effectiveness of the proposed approach in terms of accuracy and convergence rate. Second, AMFOOBL is demonstrated on multilayer perceptron's (MLP) structural realization and training compared with nine advanced algorithms. The simulation on eight datasets for pattern classification and function approximation reveals outstanding performance in the AMFOOBL-based trainer concerning classification accuracy and test error. Our findings suggest that AMFOOBL is a superior algorithm, and the developed evolutionary-enhanced MLP can be considered a helpful tool.KeywordsSwarm-intelligenceMoth-flame optimizationMultilayer perceptronOpposite-based learningSoft computing