Research On Computing Task Offloading Strategy Based On Internet Of Vehicle

With the advancement of urban traffic intelligence and communication technology,vehicles are now transforming from means of transportation to intelligent terminals,leading to many vehicular applications.However,the limited computing resources of today’s vehicles cannot meet the emerging vehicular applications’ computing requirements and latency constraints.The mobile edge computing paradigm effectively addresses the emerging vehicular applications’ ultrareliable and low latency communications requirments.By using the computing resources of vehicles on the road to perform collaborative computing to achieve low-latency network services,the problem of insufficient computing power of the vehicle itself can be compensated.Many studies have been done on the task offloading strategy utilizing the internet of vehicles for computation.However,the traditional method transforms the offloading decision problem into an iterative optimization problem but cannot cope with the large-scale and complex internet of vehicles’ environment.Existing methods using deep reinforcement learning usually have problems such as insufficient consideration of the vehicle’s mobility and cannot adapt well to the resource fluctuation characteristics of the task offloading environment.Considering the issues mentioned above,the contribution of this thesis is twofold:(1)For vehicular edge computing scenario of using nearby vehicles as edge servers,a deep reinforcement learning task offloading algorithm considering vehicle mobility is proposed.First,heterogeneous offoading tasks are prioritized by using entropy method.The offloading process of each prioritized task is then regarded as a Markov decision process,which is formulated by communication and computing models to maximize the system utility value.The mobility of the vehicle will affect the task offloading,vehicle sojourn time is added to the environment state.Assuming that vehicle sojourn in a time slot follow a normal distribution,Bayesian algorithm is used to estimate the posterior distribution of the current time slot based on the prior distribution of the historical sojourn time.Simulation results show that the proposed algorithm can achive a system utility value 11% higher than the deep Q network,and 45% higher than the greedy strategy.In addition,the proposed algorithm also has the highest task completion rate.(2)For task offloading and resource allocation problem in vehicular cloudlet environment,a computing resource allocation algorithm is proposed considering heterogeneous vehicles and heterogeneous tasks.First,M/M/1 queue model and computing model are formulated for arriving tasks,and then a utility function is defined to maximize the overall utility of the system.In order to deal with high fluctuations of computing resources in the environment,a secondary resource allocation mechanism based on dual-time-scales is proposed.In this mechanism,two difference time-scales are used for resource allocation decision-making actions,and constructing partially observable Markov decision processes respectively.The two decision-making actions are connected through the reward feedbacks obtained by their respective execution strategies,and the problem is modeled as a two-layered computing resource allocation problem.After that,a multi-agent reinforcement learning algorithm is proposed based on the secondary resource allocation mechanism to get an optimal strategy.Compared with the deep deterministic policy gradient algorithm,the simulation results show that the scheme based on the secondary resource allocation mechanism can significantly improve the overall utility value and task completion rate,and better adapt to the dynamic offloading environment of vehicular cloudlets.