Reinforcement Learning-based Safe Optimal Control Of Nonlinear Discrete-Time Affine Systems Under Adversarial Inputs
Abstract
This paper presents a resilient reinforcement learning (RL)-based safety-aware optimal adaptive tracking framework for nonlinear discrete-time affine systems under state constraints and adversarial sensor, actuator, and reward attacks. Sensor and actuator attacks are modeled through networked communication, while reward attacks affect the reward function or temporal difference error (TDE). A multilayer neural network (MNN)-based actor-critic architecture estimates the cost and optimal policy, using a clipped TDE and an adaptive Gaussian-based forgetting factor to balance resilience and optimality. Sensor and actuator attack resilience is achieved via a fault-aware cost-to-go function and a control barrier function (CBF) constraint embedded in a quadratic program (QP). The effectiveness is validated on a rear-wheel-drive vehicle.
Recommended Citation
B. Farzanegan and S. Jagannathan, "Reinforcement Learning-based Safe Optimal Control Of Nonlinear Discrete-Time Affine Systems Under Adversarial Inputs," Proceedings of the American Control Conference, pp. 1173 - 1178, Institute of Electrical and Electronics Engineers, Jan 2026.
Department(s)
Electrical and Computer Engineering
Second Department
Computer Science
Keywords and Phrases
Approximate dynamic programming; CBF; Quadratic programming; RL
International Standard Book Number (ISBN)
979-8-3315-9381-0; 979-8-3315-9382-7
International Standard Serial Number (ISSN)
0743-1619
Document Type
Article - Conference proceedings
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2026 Institute of Electrical and Electronics Engineers, All rights reserved.
Publication Date
01 Jan 2026
Comments
Office of Naval Research, Grant N00014-24-1-2338