DQN-based Computation-Intensive Graph Task Offloading for Internet of Vehicles

A computation-intensive graph task comprises a set of tasks and corresponding data flows between adjacent tasks. In this paper, we consider a mobile edge computing (MEC) system based on vehicle-to-everything (V2X) communication in which each task initiator (TI) generates and offloads a set of correlated tasks to different task executors (TEs). We formulate the graph task assignment problem as a mixed-integer nonlinear programming problem (MINLP) to minimize the weighted sum of time-energy consumption (WETC). Due to the data-flow dependency and time-varying characteristics of the operating environment including channel gain, communication distance between TIs and TEs, available computing resources of TEs, traditional numerical optimization algorithms cannot solve such optimization problem efficiently, especially when the scale of the MEC system is quite large. To this end, we propose a graph task offloading mechanism named GT-DQN by integrating deep Q-Network (DQN) with breadth-first search technique. Firstly, DQN is trained to generate a near-optimal offloading strategy, through numerous interactions with the time-varying operating environment. Secondly, a breadth-first search algorithm is adopted to traverse the graph task, which can significantly reducing the computational complexity. Compared with existing algorithms, simulation results verify the superiority of GT-DQN.