Distributed Multi-Objective DRL for Edge Offloading in Time-Varying and Heterogeneous AGI–MEC Environments

The integration of air-ground integrated (AGI) networks and mobile edge computing (MEC) presents a promising direction to provide flexible computing services for diverse mobile terminals (MTs). However, a key challenge for AGI-MEC is how to cope with the time-varying and heterogeneous performance preferences of MTs while ensuring data security, which indicates a complex and multi-constrained multi-objective optimization problem. To address this problem, we propose an adaptive distributed multi-objective deep reinforcement learning (ADMODRL) algorithm. Firstly, an edge offloading training framework based on federated learning is developed, enabling agents to implicitly acquire global knowledge by sharing model parameters rather than local raw data. As a result, the framework ensures solving accuracy while preserving data privacy. Secondly, a locally trained continuous optimization algorithm, based on actor-critic networks, is proposed to achieve multi-objective optimization specific to local preferences. Additionally, an adaptive gradient fusion (AGF) method is introduced to limit the differences between local models and the global model through adaptive gradient aggregation, so as to alleviate the impact of heterogeneous preferences on co-training. Finally, the convergence of the ADMODRL algorithm is analyzed, and its effectiveness is verified by simulation experiments.