Training Multilayer Perceptron: A Comparison of Back-propagation and Enhanced Jaya Algorithm for Data Classification

Choosing the most suitable weights and bias for feed-forward neural networks plays a critical role in the process of data classification. The back-propagation algorithm, commonly used in the multilayer perceptron, can get stuck in the local minima, resulting in poor generalization for data classification. To address this issue and improve accuracy, researchers have explored metaheuristic algorithms as replacements for back-propagation in the training of multilayer perceptron. However, many metaheuristic algorithms suffer from unsatisfactory performance due to the need for tuning a large number of control parameters and potential premature convergence. This paper introduces an effective Jaya global optimization algorithm for training multilayer perceptron, which is both efficient and does not require manual adjustment of control parameters. To strike a balance between exploiting and exploring the solution space in the Jaya algorithm, we present an enhanced version that offers an optimal configuration of connection weights to enhance the accuracy of the multilayer perceptron. Through extensive experiments using five well-known standard medical datasets, we demonstrate that the proposed method outperforms the Jaya algorithm and traditional back-propagation algorithm in terms of performance accuracy.