Optimizing Weights And Biases in MLP Using Whale Optimization Algorithm

Abstract

Artificial Neural Networks are intelligent and non-parametric mathematical models inspired by the human nervous system. They have been widely studied and applied for classification, pattern recognition and forecasting problems. The main challenge of training an Artificial Neural network is its learning process, the nonlinear nature and the unknown best set of main controlling parameters (weights and biases). When the Artificial Neural Networks are trained using the conventional training algorithm, they get caught in the local optima stagnation and slow convergence speed; this makes the stochastic optimization algorithm a definitive alternative to alleviate the drawbacks. This thesis proposes an algorithm based on the recently proposed Whale Optimization Algorithm(WOA). The algorithm has proven to solve a wide range of optimization problems and outperform existing algorithms. The successful implementation of this algorithm motivated our attempts to benchmark its performance in training feed-forward neural networks. We have taken a set of 20 datasets with different difficulty levels and tested the proposed WOA-MLP based trainer. Further, the results are verified by comparing WOA-MLP with the back propagation algorithms and six evolutionary techniques. The results have proved that the proposed trainer can outperform the current algorithms on the majority of datasets in terms of local optima avoidance and convergence speed.