Comparison of classical control algorithms and reinforcement learning methods using an industrial benchmark in the context of adaptive digital twins

Comparison of classical control algorithms and reinforcement learning methods using an industrial benchmark in the context of adaptive digital twins

Shchelov V.A. (MIPT FPMI, Dolgoprudny, Moscow Region, Russia)

Abstract

The results of a study of several architectural approaches to the creation of adaptive controllers based on reinforcement learning (RL) for the task of controlling a binary distillation column are presented. The analysis of the influence of the form of the reward function and the structure of the neural network on the effectiveness of learning, which previously had not received sufficient coverage in the literature, was carried out. It is shown that the class of neural networks based on radial basis functions (RBF) can be expanded by using fully connected neural networks with several layers and a significant number of neurons in the hidden layer. The results of the study indicate that architectures in which RL selects the parameters of a conventional regulator demonstrate better performance and require significantly fewer iterations to achieve optimal performance compared to using pure RL.

Keywords

reinforcement learning; adaptive controllers; hybrid control systems.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 161-176

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(6)-10

For citation

Shchelov V.A. Comparison of classical control algorithms and reinforcement learning methods using an industrial benchmark in the context of adaptive digital twins. Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 161-176 DOI: 10.15514/ISPRAS-2026-38(6)-10.

Full text of the paper in pdf (in Russian) Back to the contents of the volume