Automatic fuzzy parameter tuning for neural network models

The area of machine learning in practical applications is constantly growing. One of the main problems with implementing such solutions is related to the research phase. This is the phase where a given machine learning architecture is modeled and tested to obtain the best model for the processed data. The main problem most often turns out to be the selection of architecture and model parameters. Hence, in this paper, we propose the use of the Takagi-Sugeno-Kang controller with simulated annealing for tuning parameters of neural networks. Based on the rules scheme, the controller will analyze various parameters provided by the network and evaluate them to create the best selector for input parameters. The proposed model is an adaptive approach that selects input parameters during analysis to achieve the best accuracy. After modeling the controller, a simulated annealing algorithm is used for finding the best values of input parameters according to the fuzzy model. The solution has been subjected to performance tests based on classic and convolutional neural networks with two databases. Based on the obtained results, the proposed method can be determined as efficient and enables quick analysis of parameters in real-time.