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

Technical Summary: Active Sensing for RIS-Aided Tracking and Power Control

Problem Definition
This paper addresses the challenge of energy-efficient tracking and localization for power-limited mobile User Equipments (UEs) in multipath environments, assisted by a Reconfigurable Intelligent Surface (RIS). The study focuses on two primary constraints inherent to practical 6G IoT deployments:

1. Hardware Limitations: Practical RIS implementations utilize discrete (often binary) phase shifts rather than continuous tuning, rendering traditional gradient-based optimization non-differentiable and computationally intractable (NP-hard).
2. Energy and Feedback Constraints: UEs operate under strict energy budgets where continuous high-power pilot transmissions are unsustainable. Furthermore, the feedback link from the Base Station (BS) to the UE for power control is restricted to a single-bit message (e.g., "increase" or "decrease" power) to accommodate ultra-low-power non-coherent receivers.

The objective is to jointly optimize the discrete RIS phase configuration and the UE's uplink transmit power in real-time to minimize the Mean Squared Error (MSE) of the position estimate while adhering to a cumulative power budget.

Methodology
The authors propose a Dual-Agent (DA) Deep Learning framework that integrates Neuroevolution (NE) with Supervised Learning to solve this non-differentiable, decentralized control problem. The system comprises two collaborating agents:

• BS Agent: A Recurrent Neural Network (RNN) utilizing Long Short-Term Memory (LSTM) units. It observes the received pilot signals and outputs two actions: the next discrete RIS phase profile and a single-bit feedback command to the UE.
• UE Agent: An LSTM-based RNN that processes the history of received single-bit feedback commands to dynamically adjust its pilot transmission power for the next time slot.

Training Procedure
The framework employs a three-stage hybrid training methodology to overcome the limitations of standard backpropagation:

1. Initial Estimator Training: An estimator network is trained via supervised learning on a dataset generated with random RIS profiles and power levels to establish a baseline for position estimation.
2. Policy Evolution (Neuroevolution): The BS and UE policy networks are jointly optimized using a gradient-free evolutionary strategy (specifically, a variant of CoSyNE). The fitness function penalizes violations of the power budget and rewards lower tracking errors. This stage bypasses the non-differentiability of the discrete RIS elements.
3. Final Estimator Fine-Tuning: Once the policies are evolved, a final estimator network is retrained on the data generated by the optimized policies to maximize localization accuracy.

The framework is designed to be scalable, supporting both single-antenna and multi-antenna BSs (with minor architectural modifications to select digital combiners from a finite codebook).

Key Contributions

• Novel Problem Formulation: The paper formulates active uplink tracking for power-constrained UEs with strict hardware constraints (discrete RIS phases) and information bottlenecks (single-bit feedback), extending beyond prior static localization studies.
• Hybrid DA Algorithm: The development of a Dual-Agent NE framework that effectively handles the non-differentiable nature of discrete RIS control and the sparse feedback loop, enabling efficient joint learning of RIS configuration and power control.
• Robustness to Constraints: The proposed scheme is shown to function effectively with single-bit feedback, where the UE learns to interpret the temporal history of binary commands to infer sensing urgency, thereby mitigating the information bottleneck.

Numerical Results
Extensive simulations demonstrate the proposed scheme's superiority over several baselines:

• Tracking Performance: The DA framework significantly outperforms classical filters (Extended Kalman Filter, Particle Filter) and machine learning-based trackers (Supervised RNNs, Deep Reinforcement Learning using A2C) across various motion models, including high turn rates and motion noise.
• Localization Performance: In static scenarios, the method outperforms traditional fingerprinting (including genetic algorithm-optimized variants), purely supervised learning, and backpropagation-based estimators that assume continuous RIS phases.
• Scalability and Robustness: The algorithm maintains high performance as RIS size increases, whereas gradient-based and DRL baselines degrade. It also remains robust under varying Ricean fading conditions (low $\kappa$) and high noise levels.
• Feedback Efficiency: Despite using only single-bit feedback, the proposed scheme achieves performance nearly identical to schemes utilizing full scalar or multi-bit feedback links, while consuming significantly less UE energy (adhering to the power budget).

Significance and Claims
The paper claims that active sensing, when combined with a hybrid neuroevolutionary and supervised learning approach, can effectively overcome the dual challenges of discrete hardware constraints and strict energy budgets in 6G networks. The authors assert that their learned collaborative protocol allows lightweight IoT devices to achieve high-accuracy tracking without the need for complex, power-hungry feedback mechanisms or continuous RIS tuning. The work establishes a new benchmark for energy-efficient, RIS-aided localization, demonstrating that intelligent adaptation of transmission power and RIS configuration can compensate for severe information bottlenecks.

Future work is identified as extending this framework to multi-RIS deployments, exploring location privacy guarantees, and integrating model-based NN architectures to further reduce computational overhead.