Resource Allocation for Multi-Modal Semantic–Bit Coexisting Communication in Hybrid Vehicular Networks
Abstract
Semantic communication is a promising enabler for future vehicle-to-everything (V2X) networks. Nevertheless, for a long period of time, connected autonomous vehicles (CAVs) equipped with semantic transceivers will coexist with conventional human-driven vehicles (HDVs) that still rely on bit-level communications. In this paper, we consider the resource allocation for multi-modal semantic–bit coexisting communication in the hybrid vehicular network where CAVs receive multi-modal semantic information, including images and point clouds from the roadside base station, while HDVs receive conventional bit-level data streams. We formulate a joint optimization problem that maximizes a weighted sum of the semantic rates of CAVs and the bit rates of HDVs by jointly optimizing downlink beamforming, power allocation, and the number of transmitted semantic symbols. To capture the bursty communication demand of vehicles about to change lanes, lane-changing probabilities inferred from a long short-term memory (LSTM)-based predictor are embedded as weights in the objective function, enabling resource prioritization for such vehicles. To address the highly coupled and non-convex optimization problem arising from the coexistence of discrete and continuous variables, we propose a graph reinforcement learning framework based on a multi-head policy network (GRLMP). Simulation results show that the proposed GRLMP framework can proactively enhance the communication performance of lane-changing vehicles while maintaining the quality of service of other vehicles and achieving significant gains in overall semantic and bit rates compared with several baselines.