Inverse Reinforcement Learning for Disturbed Networked Nonlinear Systems With Data Dropouts

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Networked control systems (NCSs) face challenges from data dropouts and external disturbances.
• Mimicking target system trajectories is crucial for NCS performance but difficult with unknown dynamics.

Purpose of the Study:

• To develop inverse reinforcement learning (IRL) control algorithms for nonlinear NCSs.
• To address challenges of data dropouts and external disturbances in trajectory imitation.
• To enable effective control under partial model knowledge.

Main Methods:

• Developed a model-based IRL algorithm integrating $H_{\infty}$ control for disturbance rejection and uncertainty management.
• Proposed a neural-network-based, data-driven IRL algorithm to infer cost functions and control policies from available data.
• Addressed data dropouts in trajectory, state feedback, and control input data transmission.

Main Results:

• Demonstrated effective trajectory imitation capabilities of the proposed IRL algorithms.
• Showcased robustness against random data dropouts and external disturbances.
• Validated the performance through comprehensive simulation studies.

Conclusions:

• The developed IRL algorithms enable robust trajectory imitation in nonlinear NCSs.
• The methods reduce reliance on complete system models, offering practical advantages.
• Successful trajectory mimicry is achieved despite significant operational uncertainties.