A new approach to find optimum architecture of ANN and tuning it´s weights using krill-herd algorithm

Data classification is an important branch of data mining and there are different methods for its implementation. Neural networks are one of the best ways for classification in machine learning. Structure and weights of neural network are most important in their precision. In recent years, due to the defects in gradient-based search algorithms in neural network training algorithms, metahuristic algorithms have been of interest for researchers. Due to the random nature of these algorithms, the defects of trapped in local minimum can be largely resolved but Since training the weights of the neural network was done on specific network architecture, there were no guarantees for selecting the best architecture. So, in our work krill herd algorithm was used to improve the structure addition of network weights. Task of optimizing the network structure was on the three components of this algorithm (movement induced by the other krill, random diffusion, and foraging motion) along with a genetic operator; also dimensions of krill showed the desired structure for the neural network In this paper, the performance of the proposed method was tested on five UCI data sets and the results compared with the previous methods showed that the classification accuracy of the proposed method was considerably higher and the mean square error was low.