Learning-Based Symbol-Level Precoding for Flexible Tradeoffs in MIMO-ISAC Systems

Abstract

Integrated sensing and communication (ISAC) has emerged as one promising candidate for 6 G wireless networks, yet it encounters difficulties in alleviating mutual interference in complex multi-user and multi-target scenarios. To address this issue, one feasible methodology is symbol-level precoding (SLP) that exploits the constructive interference (CI), which converts multi-user interference into beneficial signals. However, most existing SLP techniques are computationally expensive, which fails to realize low-latency implementation required for real-time symbol-level processing. Moreover, ISAC inherently requires a flexible tradeoff between communication and sensing, which is not explicitly considered in existing frameworks. To overcome these limitations, this paper formulates the SLP design as a multi-objective problem for communication and sensing, and proposes an adaptive gradient-refined learning engine (AGILE) to solve it efficiently. Specifically, the proposed AGILE incorporates a safe deep deterministic policy gradient (safe-DDPG) policy with proportional-integral (PI) control and gradient rescaling, in order to generate robust initial solutions. Subsequently, an online refiner is employed to approach the Pareto optimality under CI constraints. Numerical results demonstrate that AGILE can outperform its counterparts in terms of both communication and sensing performance, while significantly reducing the computational complexity.