Deep Learning-Based Dynamic Time Division ISAC Beamforming for Vehicular Networks

Integrated sensing and communications (ISAC) is a promising key technology for vehicular networks, because it allows roadside units to support both data transmission and radar-like sensing over the same spectrum and hardware platform. In conventional time division ISAC systems, each frame is divided into sensing and communication phases with a fixed ratio, which determines the tradeoff between the sensing accuracy and the communication throughput. However, in high-mobility vehicular environments, a fixed sensing-communication split is often suboptimal due to time-varying channel and intervehicle interference variations. In this paper, we propose a dynamic sensing-communication time division and ISAC beamforming scheme that minimizes the Cramér-Rao lower bound while satisfying the minimum effective communication sum rate. We further develop a deep reinforcement learning framework based on proximal policy optimization to find the optimal time division ratio and beamforming vectors. Simulation results show that the proposed dynamic time division beamforming scheme significantly outperforms the conventional fixed time division beamforming schemes in terms of sensing accuracy and the communication sum rate.