Fast Adaptive Optimization for AAV-Assisted MEC: A Constrained Decision Transformer Approach

Deep reinforcement learning (DRL) has been widely applied to dynamic resource allocation for the autonomous aerial vehicle (AAV)-assisted mobile edge computing (MEC) networks. However, it is often challenging to generalize a well-trained DRL model to new MEC scenarios once the system constraints change, since retraining DRL models from scratch is time and energy consuming. In this letter, we jointly optimize the AAV’s trajectory and computing resource allocation for maximizing the fairness-based throughput under the battery capacity and quality of service (QoS) constraints. The sequential optimization problem is formulated as a constrained Markov decision process (CMDP) and solved via the constrained DRL algorithms. To generalize the optimized resource allocation policies across various energy and QoS constraints, we propose an offline pre-training and online fine-tuning based constrained Decision Transformer (CDT) framework. In particular, the CDT is first pre-trained on the training samples collected by the constrained DRL algorithm offline, and then fine-tuned online for rapid adaptation to the unseen constraint thresholds. Simulation results show that compared with the benchmark DRL algorithms, the CDT is capable of effectively improving the fairness-based throughput under the battery capacity and QoS constraints, and demonstrates rapid convergence when constraints change.