A Learning-Based V2V Task Offloading Algorithm for Multi-RAT Vehicular Networks

With the emergence of computing-intensive vehicular applications, limited on-board computing resources become insufficient to meet the computing requirement of many vehicular applications. Vehicle-to-vehicle (V2V) task offloading is a promising approach to addressing the limited on-board computing resources of vehicles. Meanwhile, the rapid development of radio access technologies (RATs) has enabled a vehicle network to incorporate multiple RATs for supporting data transmissions between vehicles. This article studies the V2V task offloading problem in a vehicular network with multiple RATs. The problem is formulated as a mixed integer nonlinear programming (MINP) problem with an objective to minimize the average offloading delay of all offloading tasks generated by vehicles in the network at a given time by optimizing offloading decisions and computing resource allocation, subject to the delay constraints of the tasks and the computing resource constraints of the vehicles. To solve the formulated problem, a log-sum-exp approximation is used to transform the MINP problem into a combinatorial optimization problem to obtain a probability distribution of all possible solutions to the problem. Based on the probability distribution, a Markov chain based method is used to obtain the transition probabilities between different solutions and based on the transition probabilities, a learning-based offloading decision and computing resource allocation (L-D&A) algorithm is further proposed to solve the formulated problem. The L-D&A algorithm introduces a learning process, in which the problem is decomposed into three sub-problems: SV and RAT selection (SRS), sub-channel selection (SCS), and computing resource allocation (CRA), and correspondingly a Markov chain-based SRS algorithm, an interference minimization based SCS algorithm, and a convex optimization based CRA algorithm are proposed to solve the three sub-problems, respectively. Simulation results show that the proposed L-D&A algorithm can efficiently reduce the average offloading delay as compared with three benchmark algorithms.