Battle royale optimizer for training multi-layer perceptron

Abstract

Artificial neural network (ANN) is one of the most successful tools in machine learning. The success of ANN mostly depends on its architecture and learning procedure. Multi-layer perceptron (MLP) is a popular form of ANN. Moreover, backpropagation is a well-known gradient-based approach for training MLP. Gradient-based search approaches have a low convergence rate; therefore, they may get stuck in local minima, which may lead to performance degradation. Training the MLP is accomplished based on minimizing the total network error, which can be considered as an optimization problem. Stochastic optimization algorithms are proven to be effective when dealing with such problems. Battle royale optimization (BRO) is a recently proposed population-based metaheuristic algorithm which can be applied to single-objective optimization over continuous problem spaces. The proposed method has been compared with backpropagation (Generalized learning delta rule) and six well-known optimization algorithms on ten classification benchmark datasets. Experiments confirm that, according to error rate, accuracy, and convergence, the proposed approach yields promising results and outperforms its competitors.