Towards Intelligent Antenna Positioning: Leveraging DRL for FAS-Aided ISAC Systems

Shunxing Yang, Junteng Yao, Jingyu Tang, Tuo Wu, Maged Elkashlan, Chau Yuen, Mérouane Debbah, Hyundong Shin, Matthew C. ValentiarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a novel framework for 6G networks that uses Fluid Antenna Systems (FAS)—antennas that can physically move within a defined space—to optimize both communication and multi-target radar sensing. By combining traditional convex optimization with Deep Reinforcement Learning (DRL), the researchers demonstrate that moving antennas dynamically can significantly outperform traditional fixed-position setups, even in complex scenarios with multiple sensing targets and users.

What Problem Does This Paper Address?

In the push toward 6G, researchers are focused on Integrated Sensing and Communication (ISAC), where the same hardware and spectrum are used to both transmit data and perform radar-like sensing of the environment. However, traditional ISAC systems use Fixed-Position Antennas (FPAs). These fixed arrays often struggle to balance the competing needs of high-speed data transmission and precise sensing, especially when there are multiple targets to track.

Optimizing an ISAC system involves two main "knobs": beamforming (how you shape the signal) and antenna positioning (where the antennas are). When antennas can move (FAS), the math becomes "highly non-convex"—essentially a computational nightmare that traditional optimization methods can't solve efficiently as the number of antennas or targets grows. This paper aims to solve this joint optimization problem to make FAS-aided ISAC practical and scalable.

Methods

The authors propose a Block Coordinate Descent (BCD) framework that splits the problem into two alternating steps:

  1. Beamforming Optimization: For a given set of antenna positions, they optimize the transmit covariance matrix (beamforming) using standard convex optimization tools (CVX).
  2. Antenna Positioning via DRL: To handle the messy, non-convex task of moving the antennas, they use a Deep Reinforcement Learning algorithm called Deep Deterministic Policy Gradient (DDPG).

Key Technical Details: * State Space: The DRL agent looks at current antenna coordinates and "summary statistics" of the beamforming matrix (like its trace and eigenvalues) to understand the system's power and directionality. * Action Space: The agent decides the incremental movement (displacement) for each antenna at the Base Station and the User Terminal. * Reward Function: The agent is "rewarded" for high communication rates but "penalized" if it fails to meet sensing gain requirements, exceeds power limits, or moves the antennas too much (to ensure stability).

Datasets

The study relies on numerical simulations rather than a static dataset. The researchers modeled a 2D Cartesian environment where:

  • Channel paths (AoA/AoD) are randomly distributed.
  • The system uses a far-field channel model.
  • The path response matrix accounts for both Line-of-Sight (LoS) and non-Line-of-Sight (NLoS) components.
  • The environment is restricted to a movement area defined by the wavelength ($\lambda$).

Key Insights and Ideas

  • The "FAS Advantage": The core "aha" moment is that FAS provides a "spatial degree of freedom" that fixed antennas lack. By moving just a few centimeters, an antenna can exit a "fading hole" or reposition itself to create a better null/peak for sensing.
  • Continuous vs. Discrete: Unlike previous studies that simply switched between pre-defined "ports," this paper treats antenna position as a continuous variable, allowing for much finer optimization.
  • Feature Engineering for DRL: Instead of feeding the raw, high-dimensional beamforming matrix into the DRL agent, the authors extracted key features (eigenvalues). This significantly reduces the complexity of the neural network and speeds up learning.
  • Scalability: The DRL approach proved remarkably robust. As the number of sensing targets ($K$) increased, the FAS system maintained a performance edge that traditional fixed systems couldn't match.

Key Results

  • FAS vs. FPA: FAS consistently outperformed fixed-position antennas. Remarkably, an FAS system sensing three targets still achieved a higher communication rate than a fixed antenna system sensing only one target.
  • Multi-User Gains: In a scenario with 12 antennas at the base station, moving from 1 user to 3 users resulted in a 27.6% improvement in the maximum communication rate (reaching 14.84 bps/Hz at 30 dB SNR).
  • Constraint Satisfaction: The DRL agent successfully learned to satisfy the "sensing gain" requirement ($\Gamma$), ensuring that radar performance wasn't sacrificed for data speed.

Limitations and Future Work

  • Hardware Realism: The paper assumes antennas can move freely via "integrated waveguides or flexible cables." In reality, the mechanical latency of moving an antenna or the hardware complexity of a high-speed switching fabric remains a significant implementation challenge.
  • Training Overhead: While DRL is fast once trained (enabling real-time decision-making), the initial training phase is computationally expensive.
  • Future Directions: The authors suggest extending this to more complex, "cell-free" massive MIMO environments where many distributed base stations must coordinate their antenna positions simultaneously.

Summary updated

Abstract

Fluid antenna systems (FAS) enable dynamic antenna positioning, offering new opportunities to enhance integrated sensing and communication (ISAC) performance. However, existing studies primarily focus on communication enhancement or single-target sensing, leaving multi-target scenarios underexplored. Additionally, the joint optimization of beamforming and antenna positions poses a highly non-convex problem, with traditional methods becoming impractical as the number of fluid antennas increases. To address these challenges, this letter proposes a block coordinate descent (BCD) framework integrated with a deep reinforcement learning (DRL)-based approach for intelligent antenna positioning. By leveraging the deep deterministic policy gradient (DDPG) algorithm, the proposed framework efficiently balances sensing and communication performance. Simulation results demonstrate the scalability and effectiveness of the proposed approach.
BibTeX
@misc{yang2025towards,
  author = {Shunxing Yang and Junteng Yao and Jingyu Tang and Tuo Wu and Maged Elkashlan and Chau Yuen and Mérouane Debbah and Hyundong Shin and Matthew C. Valenti},
  title = {{Towards Intelligent Antenna Positioning: Leveraging DRL for FAS-Aided ISAC Systems}},
  year = {2025},
  eprint = {2501.01281},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.01281},
}
Figure