Active Sensing for RIS-Aided Tracking and Power Control: A Hybrid Neuroevolution and Supervised Learning Approach

How can power-constrained mobile devices be tracked efficiently in multipath environments using a Reconfigurable Intelligent Surface (RIS)? The authors aim to address the challenge of balancing accurate location estimation with the strict energy limitations of IoT devices, specifically by optimizing both the RIS phase configuration and the user's uplink transmit power in real time.

The authors introduce a Dual-Agent (DA) deep learning framework consisting of two collaborating Recurrent Neural Networks (RNNs): one at the Base Station (BS) and one at the User Equipment (UE). Because practical RIS hardware uses discrete phase shifts, the optimization problem is non-differentiable, rendering standard backpropagation-based training ineffective. To solve this, the authors employ a hybrid training methodology that integrates neuroevolution (a gradient-free global search) with supervised learning. The BS agent manages the RIS configuration and sends single-bit feedback to the UE, while the UE agent interprets this history of bits to adaptively control its pilot transmission power.

Numerical simulations demonstrate that the proposed DA framework achieves superior tracking accuracy and robustness compared to traditional methods like Extended Kalman Filters (EKF) and Particle Filters (PF), as well as existing machine learning benchmarks. The study highlights that the learned collaborative protocol effectively mitigates the information bottleneck of 1-bit feedback, allowing the system to achieve performance nearly identical to schemes using high-capacity scalar control links. Furthermore, the framework scales effectively to multi-antenna BS configurations and maintains high performance across diverse scattering conditions and target motion models.

This research provides a practical, energy-efficient solution for 6G localization and tracking. By enabling lightweight IoT devices to operate with minimal pilot transmission power while still benefiting from RIS-assisted signal enhancement, the proposed framework addresses a critical bottleneck in the deployment of large-scale, power-constrained wireless sensor networks.