TY - JOUR
T1 - Self-Learning Based Model Predictive Control for Nonlinear Systems With State-Dependent Uncertainty
AU - He, Ning
AU - Li, Wenzhuo
AU - Ma, Kai
AU - Li, Huiping
N1 - Publisher Copyright:
© 1982-2012 IEEE.
PY - 2026
Y1 - 2026
N2 - To address the state-dependent uncertainty issues in nonlinear systems, this article presents a self-learning based model predictive control (SLMPC) method. Unlike existing online learning MPC architectures where control and data transmission are tightly coupled, this study develops an active intermittent learning strategy. Data transmission is activated only when the existing model cannot accurately describe the current dynamics, and incorporates the data collected during this period into the training dataset for model updating. First, the model uncertainty is learned by Gaussian process regression (GPR), and the cost function is constructed by combining it with the nominal model to ensure the stability of the system. Then, the reproducing kernel Hilbert space (RKHS) theory is introduced to construct a deterministic learning error bound, and the self-learning mechanism is proposed to reduce the frequency of online learning updates and the consumption of communication resources. In addition, the theoretical properties of the predictive control method based on the self-learning mechanism are rigorously analyzed, including the analysis of Zeno-free behavior, recursive feasibility, and closed-loop stability. Finally, the effectiveness of the proposed method is verified through simulations and experiments on a robotic system.
AB - To address the state-dependent uncertainty issues in nonlinear systems, this article presents a self-learning based model predictive control (SLMPC) method. Unlike existing online learning MPC architectures where control and data transmission are tightly coupled, this study develops an active intermittent learning strategy. Data transmission is activated only when the existing model cannot accurately describe the current dynamics, and incorporates the data collected during this period into the training dataset for model updating. First, the model uncertainty is learned by Gaussian process regression (GPR), and the cost function is constructed by combining it with the nominal model to ensure the stability of the system. Then, the reproducing kernel Hilbert space (RKHS) theory is introduced to construct a deterministic learning error bound, and the self-learning mechanism is proposed to reduce the frequency of online learning updates and the consumption of communication resources. In addition, the theoretical properties of the predictive control method based on the self-learning mechanism are rigorously analyzed, including the analysis of Zeno-free behavior, recursive feasibility, and closed-loop stability. Finally, the effectiveness of the proposed method is verified through simulations and experiments on a robotic system.
KW - Gaussian process regression (GPR)
KW - learning-based model predictive control (MPC)
KW - self-learning mechanism (SLM)
KW - uncertain nonlinear systems
UR - https://www.scopus.com/pages/publications/105050237240
U2 - 10.1109/TIE.2026.3720726
DO - 10.1109/TIE.2026.3720726
M3 - 文章
AN - SCOPUS:105050237240
SN - 0278-0046
JO - IEEE Transactions on Industrial Electronics
JF - IEEE Transactions on Industrial Electronics
ER -