FMFO: Floating flame moth-flame optimization algorithm for training multi-layer perceptron classifier

As one of the most popular artificial neural networks, multi-layer perceptron (MLP) has been employed to solve classification problems in many applications. The main challenge in MLP application is finding the ideal set of network connection weights and biases in the training process, which minimizes the error of MLP in processing datasets. To efficiently address this challenge, numerous swarm intelligence (SI) algorithms with powerful search capabilities have been adopted for training MLP classifiers. However, these existing algorithms often suffer from problems of local optima stagnation, premature convergence, and inefficient search. In this study, a novel floating flame moth-flame optimization (FMFO) algorithm with remarkable exploitation and exploration search capabilities is proposed, offering an advantageous option for training MLP classifiers. To verify the performance of the proposed FMFO in training MLP classifiers, the FMFO-based MLP training approach (FMFO-MLP) is evaluated on eleven classification datasets that represent a wide range of the variable dimension scale. In addition, some recently developed well-known and state-of-the-art SI algorithms are applied to compare with the proposed FMFO. Experimental results demonstrate that the proposed FMFO outperforms the other competing algorithms in terms of approximating the optimal objective function value and achieving classification accuracy. Moreover, the proposed FMFO achieves a competitive computational efficiency in the experiment, confirming that it is an efficient optimizer for training MLP classifiers in practical applications.