A metaheuristic algorithm based on a radial basis function neural networks

Metaheuristic methods are optimization techniques generally based on natural or physical phenomena. Although these processes inspire, the underlying mechanisms of these phenomena still need to be fully understood. Translating these mechanisms into algorithmic solutions may lead to unclear strategies, whose performance can exhibit unpredictable behavior. On the other hand, a radial basis function neural network (RBFNN) is a powerful method for approximating highly complex functions through a training process. Its key advantage lies in its interpretable structure, which allows us to understand the underlying relationships among the input–output data. This paper proposes a new metaheuristic algorithm based on the architecture of an RBFNN. Under this approach, an RBFNN is trained to approximate the objective function based on initial input data distributed from the search space. After training, the radial basis functions and weights associated with the neurons that maintain a strong influence or contribution to obtain the highest values of the cost function are determined. The parameters of these radial basis functions define the promissory regions of the input space that are used to produce the individuals of the new population through the Latin hypercube sampling technique. Therefore, during the optimization process, different regions of the search space represented by radial basis functions are explored and exploited as the accuracy of the cost function approximation increases. The experimental results demonstrated the success of our approach, as it outperformed the established metaheuristic algorithms on various benchmark functions. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.