A structure-improving adaptive fruit fly optimization algorithm (SFOA) for training artificial neural network was proposed. The new algorithm was successfully applied to pattern classification problems. The method of equal length coding based on link structure which map different network structure was proposed. The strategy of adaptive variable step size were adopted in the smell-based search procedure resulted in the dynamic balance between global and local optimizing capability. Finally

simulation results based on three benchmark testing shows that the new algorithm achieves simply

costs less time

eliminates effectively some redundant structure and improves the training efficiency and classification ability of the neural network.