Bi-Directional Computation Offloading in Caching Enabled UAV-assisted MEC Systems

To alleviate the computation resource scarcity of terrestrial wireless devices (WDs), a mobile edge computing (MEC) system with the assistance of unmanned aerial vehicles (UAVs) is envisaged as a promising paradigm where UAVs, acting as aerial relays and mobile MEC servers can provide extensive coverage, line-of-sight (LoS) communication links and supplementary computation resources. However, the high-correlated WDs’ mobile applications will incur repeat transmission of programs, which induces a decline in the quality of service. Therefore, the paper considers a service caching enabled UAVs to assist MEC systems with a bi-directional computation offloading scheme, where UAVs can temporarily store the popular service programs from a cloud-like database and move to the vicinity of WDs to establish bi-directional connections for both data uploading and program downloading, thus can be beneficial for saving both the tasks’ local and remote execution latency. To fully exploit the potential of such a bi-directional offloading scheme, it is crucial to jointly optimize the offloading decision of multiple WDs, both the downlink and uplink channel allocation, the cache decision of UAVs, the computation resource allocation of each WD, as well as UAVs’ trajectory. Therefore, the paper proposes a clustering, heterogeneous multi-agent deep reinforcement learning (MADRL), and successive convex approximation (SCA) combined alternative optimization method to conquer the non-trivial feature of such a non-convex mixed integer non-linear programming (MINLP) problem. Numerical simulation results reveal its effectiveness and superiority in reducing the task latency and saving energy consumption compared to the uni-directional offloading and non-caching-equipped schemes.