DRL-based Resource Allocation for Multi-Semantic Tasks over Air-Ground MEC Networks

This paper investigates unmanned aerial vehicles (UAV)-assisted air-ground multi-access edge computing (MEC) networks, with particular emphasis on semantic communication. We present a comprehensive air–ground MEC framework in which MEC servers mounted on millimeter wave (mmWave) UAVs and small-cell base stations (SBSs) will provide efficient services for multi-semantic tasks generated by internet of things (IoT) devices. A joint resource allocation problem is formulated to systematically optimize four key factors, aiming to maximize the average system utility while balancing task revenue and cost. Considering the complexity and the mixed-timescale characteristic of the formulated problem, we decompose the original challenging problem into two subproblems and propose a two-stage solution. Specifically, we transfer the first short-timescale resource allocation problem into Markov decision process (MDP) coupled with user association, bandwidth allocation, and computational resource allocation. Then we complete the second long-timescale user admission stage by Lyapunov optimization. Accordingly, an attention-enhanced proximal policy optimization (APPO) algorithm based on deep reinforcement learning (DRL) and simulated annealing (SA) algorithm are developed to achieve intelligent resource adaptation. Simulation results demonstrate the fast convergence and superior performance of our proposed algorithm compared to other algorithms.