Trust-Preserving Distributed Learning Control Architecture for Safety-Constrained Cyber-Physical Systems
Year: 2025
Type: article
Abstract: Safety-critical cyber-physical systems (CPS) distributed control is increasingly based on learning-enabled agents that operate on cloud-edge systems, where loss of trust, mal-adaptations and safety-violence are of major danger. Available distributed learning-based control methods do not pay much attention to considering trust preservation and safety enforcement joint in cases of decentralised execution, which restricts their applied scope in adversarial and dynamic CPS scenarios. The current paper suggests trust preserving distributed learning control architecture, which is based on centralised training and decentralised execution (CTDE) multi-agent deep reinforcement learning (MARL) of safety-constrained CPS. The trust-aware learning is proposed to dynamically determine the reliability of agents based on their behavioural consistency and constraint-violation information, which are then introduced in the centralised critic and reward shaping process. Formulating the control problem as a constrained Markov decision process is necessary to enforce the operational safety and to accomplish the safety enforcement through policy shielding of the execution process. The suggested framework allows a strong distributed control as well as removes the effect of unreliable agents without affecting the scalability. Extensive simulations show the proposed approach is far more effective in bringing an accuracy of control, minimising safety transgression, and resisting the tendency to diminish trust as compared with non-trust-aware and centralised baselines. Those findings point to the success of the offered architecture in ensuring security, scalability and safety-critical cyber-physical systems which are capable of running across cloud-enabled distributed environment.
Author: Gaurav Tamrakar
Institution: Kalinga University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7220166960