Mobility-Aware Robust Resource Allocation and Computation Migration in Heterogeneous Virtualized Vehicular Networks by Low Overhead Multi-Agent DRL-based Approach

In emerging vehicular networks, delay-sensitive tasks can be processed in real time by offloading tothe edge computing servers. Unlike the legacy scenarios, in this paper a novel data offloading/deliverydecision making framework is proposed, where users have the option to divide their task into severalportions and partially offload their data to a complex multi-access edge computing (MEC) environment,consisting of several MEC servers located on road side units (RSUs), base stations (BSs), and unusedcomputing power in other vehicles and pedestrians. This can increase the reliability and the spectralefficiency, and decrease delay. However, due to high mobility of vehicles, the vehicle-mounted MECservers act as a uncertain payoff for the users, while the local computation server (e.g., RSUs, and BSs)alternatives constitute safe and guaranteed options. Therefore, the offloaded tasks may be migrated toother MEC servers to provide continuous execution. In addition, users’ high mobility and traffic variationcan result in task offloading demand uncertainty and the imperfect channel state information (I-CSI)estimation that lead to difficulties in guaranteeing the quality of service (QoS) for each task. In thispaper, we consider a two-time scale resource management scheme for task offloading with fluctuatedtraffic demands and I-CSI, aiming at minimizing total network cost and delay. We consider a hierarchicalcontrol architecture in our proposed scheme including large time-scale RSU activation/deactivation andshort time-scale resource allocation, and making task offloading decisions. We extend our problem as multi-time scale Markov decision process. Simulation results validate that our proposed task offloadingscheme can increase the task completion rate by 27% and 98% when considering task partitioning andtask migration, respectively. Moreover, considering demand uncertainty, we can achieve more robustsolution by increasing the task completion rate by 42.8%. Finally, we show that considering MADDPG can increase task offloading cost by 45% due to communication overhead between agents.