Towards Task Number Adaptive Offloading in MEC Systems: A Transformer-based DRL Approach

In a practical Mobile Edge Computing (MEC) system, the stochastic arrival and departure of heterogeneous Mobile Devices (MDs) can cause fluctuations in the number of generated tasks to be scheduled over time, with the generated tasks differing in data size, complexity, and delay constraint. In this work, we consider a dynamic Task Offloading (TO) problem aiming at maximizing the long-term average system utility jointly defined by task completion and energy consumption in the MEC system, where the number of MDs to be served varies over time, and the processing of generated tasks may span multiple time slots. Particularly, we first transform the TO problem into a Markov Decision Process (MDP) with time-varying state and action spaces, which cannot be effectively solved by conventional Deep Reinforcement Learning (DRL) algorithms. Then, we propose a state and action spaces adaptive DRL algorithm to efficiently solve the formulated MDP by leveraging the Transformer model. Finally, simulation results demonstrate the superiority of our proposed algorithm over baseline algorithms and emphasize the limitation of the conventional DRL algorithm in handling time-varying state and action spaces.