Deep Reinforcement Learning for Resource Allocation in RIS-Assisted NOMA-MEC Vehicular Networks

Mobile edge computing (MEC) enables efficient computation offloading for mission-critical applications in resource-constrained vehicles, while reconfigurable intelligent surface (RIS) help address connectivity challenges for vehicles in urban environments with severe signal blockages. Non-orthogonal multiple access (NOMA) is an appealing technique that improves spectral efficiency while mitigating multi-user interference. This work proposes the RIS-assisted NOMA-MEC in vehicular networks, considering dynamic challenges such as heterogeneous vehicle processing capability, time-varying channel from high-mobility and dynamic task workloads. We formulate a system latency minimization problem by jointly optimizing the task offloading ratio, edge server resource allocation and RIS passive beamforming, while satisfying the task deadline and Signal to Interference plus Noise Ratio (SINR) requirements. To overcome the limitations of conventional optimization methods in such dynamic environments, we propose a soft actor critic (SAC)-based deep reinforcement learning (DRL) framework, which dynamically adapts to real-time channel state information (CSI), task workload and vehicle processing capability of all vehicles. Simulation results demonstrate that our approach achieves lower latency performance compared with the Deep Deterministic Policy Gradient (DDPG) baselines. Moreover, the proposed SAC method exhibits robustness and adaptivity to various levels of uncertainty in the CSI.