AI Optimizes Active Sensing for RIS-Aided Mobile Tracking

The research addresses the challenge of energy-efficiently tracking mobile users, particularly those with limited power, by leveraging Reconfigurable Intelligent Surfaces (RIS). Traditional localization methods often consume significant energy due to pilot transmissions. To mitigate this, the authors propose a novel approach that incorporates a low-overhead feedback mechanism from the base station to the user, enabling dynamic control over the user's uplink transmit power. The core of their solution is a Dual-Agent (DA) deep learning framework. This framework is designed to jointly optimize the discrete phase profiles of the RIS and the user equipment's transmit power in real-time. It employs a hybrid training methodology, combining neuroevolution with supervised learning, which helps overcome the difficulties associated with non-differentiable discrete phase responses and the strict information bottleneck of single-bit feedback for power control. Extensive simulations demonstrate that this DA active sensing framework achieves superior accuracy and robustness in tracking across various target motion models, outperforming established methods like extended Kalman filters, particle filters, and other machine learning-based trackers. It also significantly improves static localization compared to traditional fingerprinting and deep reinforcement learning baselines.