Evaluation of Convergence of Neural Network Regulators Training Methods

Neural network controllers are a promising replacement for classical controls in closed tracking systems, but for the correct operation of an artificial neural network (ANN) in this capacity, its correct configuration is necessary. One of the most frequently used algorithms for training neural netwo...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2023 7th International Conference on Information, Control, and Communication Technologies (ICCT) s. 1 - 4
Hlavní autor: Feofilov, Dmitry
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 02.10.2023
Témata:
Approximation algorithms
Artificial neural networks
automatic control systems
Backpropagation
error back propagation algorithm
gradient descent method
Learning (artificial intelligence)
neural network regulators
Process control
Regulators
Training
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Neural network controllers are a promising replacement for classical controls in closed tracking systems, but for the correct operation of an artificial neural network (ANN) in this capacity, its correct configuration is necessary. One of the most frequently used algorithms for training neural networks is the method of error back propagation, which is used when updating the weights of a multilayer neural network. This paper presents recommendations for analyzing the convergence of the error back propagation method in the training of artificial neural networks. The algorithm is considered as a variation of gradient descent, as a result of which the approaches used to analyze the convergence of numerical iterative methods are used. The final expression for evaluating the possibility of successful training of neural networks by the method under consideration is obtained, recommendations are given to increase the probability of convergence of the learning process, the most frequent problems that arise when configuring the ANN are indicated.
DOI:10.1109/ICCT58878.2023.10347103