RIS-Aided Tracking: Neuroevolution and Supervised Learning

A Dual-Agent deep learning framework jointly optimizes discrete Reconfigurable Intelligent Surface (RIS) phase profiles and a user equipment's transmit power in real time. The approach, submitted to arXiv as arXiv:2607.00056, was posted 30 Jun 2026 (v1) and revised 2 Jul 2026 (v2) by George Stamatelis, Hui Chen, Henk Henk Wymeersch and George C. Alexandropoulos; the manuscript is 15 pages long.

What did the authors build?

The paper presents a DA (Dual-Agent) active sensing framework that co-designs RIS phase settings and uplink pilot power with a low-overhead feedback link from the Base Station (BS) to the user. The authors introduce a single-bit feedback channel to enable dynamic uplink power control, and the framework works with both single- and multi-antenna BSs, the latter handled by adding an extra NN output branch that selects a valid digital combiner from a finite set.

After that core design, the authors emphasize energy efficiency: because localization pilot transmissions dominate the energy budget of power-limited devices, the single-bit feedback and joint control aim to reduce that overhead while preserving tracking accuracy.

How does the Dual-Agent approach work?

The DA framework uses two cooperating neural agents: one that selects discrete RIS phase profiles and another that controls the UE's transmit power; training combines neuroevolution with supervised learning to handle discreteness and tight feedback. The paper states the neuroevolution paradigm overcomes the non-differentiability of discrete RIS unit-element phase responses, while supervised learning addresses the information bottleneck imposed by single-bit feedback messages for pilot power control.

The method is designed for real-time operation and requires only minor structural changes to support a multi-antenna BS: an extra output branch in one neural network picks a digital combiner from a finite set. The authors position this hybrid training pipeline as a way to avoid relying solely on backpropagation through nondifferentiable components.

How does it perform versus established methods?

Extensive numerical simulations reported by the authors show the proposed scheme delivers highly accurate and robust tracking across diverse target motion models, outperforming extended Kalman filters, particle filters, and machine learning-based trackers. For static localization tasks the paper claims the DA framework significantly outperforms traditional fingerprinting schemes, deep reinforcement learning baselines, and standard backpropagation-based estimators.

Those performance statements are based on the simulations in the manuscript; the arXiv entry lists the submission details as 30 Jun 2026 (v1) and a revision on 2 Jul 2026 (v2).

Why it matters

Jointly optimizing RIS configurations and UE transmit power targets two practical barriers for RIS deployment: discrete phase control and constrained device energy. The paper addresses both directly by combining a low-overhead, single-bit feedback link and a hybrid neuroevolution-plus-supervised training approach. Systems that must localize or track power-limited devices could reduce pilot energy while keeping or improving accuracy if the simulation claims translate to hardware.

What to watch

Look for peer-reviewed publication of this manuscript and for open-source code or reproducible experiment artifacts linked to the arXiv entry; the paper lists the arXiv DOI arXiv:2607.00056 and shows it was submitted on 30 Jun 2026 and revised on 2 Jul 2026. Follow-up signals to watch are hardware experiments validating energy savings and any public release of the Dual-Agent training code or the neural combiner set used for multi-antenna BSs.