Performance analysis of various training algorithms of deep learning based controller

Abstract

Abstract Advances in artificial neural networks (ANN), specifically deep learning (DL), have widened the application domain of process control. DL algorithms and models have become quite common these days. The training algorithm is the most important part of an ANN that affects the performance of the controller. Training algorithms optimize the weights and biases of the ANN according to the input-output patterns. In this paper, the performance of different training algorithms was evaluated, analysed, and compared in a feed-forward backpropagation architecture. The training algorithms were simulated on MATLAB R2021b with license number 1075356. Training data were generated using two benchmark problems of the process control system. The performance, gradient, training error, validation error, testing error, and regression of the different training algorithms were obtained and analysed. The data shows that the Levenberg-Marquardt (LM) algorithm produced the best validation performance with a value of 2.669*10−14 at 2000 epochs, while ‘traingd’ and ‘traingdm’ algorithms did not improve beyond their initial values. The LM algorithm tends to produce better results than other algorithms. These results indicate that the LM backpropagation best suits these types of benchmark problems. The results also suggest that the choice of training algorithm can significantly impact the performance of a neural network.