A Genetic Algorithm Based Adaptive Offloading Scheme for Domain Generalization in VEC

Tianyu Li, Liang Zhao, Chaojin Mao, Na Lin, Qiang He, Xingwei Wang

Abstract: Vehicular Edge Computing (VEC) is a computational paradigm which integrates Mobile Edge Computing (MEC) into vehicular environments. Task offloading in VEC enables tasks generated by vehicles to be offloaded to the nearby Roadside Units (RSU) or surrounding vehicles, thereby reducing computational latency and energy consumption. However, current task offloading schemes still face numerous challenges, such as varying scenarios, intricately highly dynamic network topologies, and demands for lower latency. In this paper, we introduce a vehicle-to-vehicle (V2V) offloading scheme, namely, a genetic algorithm Based adaptivE offloadiNg schEme For domaIn generalizaTion (BENEFIT). First, to deal with the problem of varying scenarios in VEC, we consider the Domain Generalization (DG) problem in trajectory prediction and employ causal inference to construct a generalized prediction model. This serves as a pre-processing step and simultaneously reduces the overall decision space of the system to assist offloading. Second, we put forth an adaptive model targeting the cumulative offloading cost to handle highly dynamic network topologies, which dynamically adjusts delay and energy coefficients in various scenarios to facilitate adaptive offloading. Finally, we utilize the enhanced Non-Uniform Genetic Algorithm (NUGA) for making offloading decisions with considering the limited resources and collaboration processes among vehicles. According to simulation results, our scheme effectively decreases the total cost of the system. The BENEFIT achieves a 3 percent cost savings with regard to other schemes.