Adaptive Inference Reinforcement Learning for Task Offloading in Vehicular Edge Computing Systems

Vehicular edge computing (VEC) is expected as a promising technology to improve the quality of innovative applications in vehicular networks through computation offloading. However, in VEC system, the characteristics of distributed computing resources and high mobility of vehicles bring a critical challenge, i.e., whether to execute computation task locally or in edge servers can obtain the least computation overhead. In this paper, we study the VEC system for a representative vehicle with multiple dependent tasks that need to be processed successively, where nearby vehicles with computing servers can be selected for offloading. Considering the migration cost incurred during position shift procedure, a sequential decision making problem is formulated to minimize the overall costs of delay and energy consumption. To tackle it effectively, we propose a deep Q network algorithm by introducing Bayesian inference taking advantage of priori distribution and statistical information, which adapts to the environmental dynamics in a smarter manner. Numerical results demonstrate our proposed learning-based algorithm achieve a significant improvement in overall cost of task execution compared with other baseline policies.