Efficient Computation Offloading With Reinforcement Learning Assistance In Vehicular Networks

With the continuous development of wireless communication technology,mobile internet has begun to enter people’s lives,and people have begun to enjoy the timeliness,convenience and accuracy of internet services.However,mobile internet technology is limited by the insufficient computing and storage capabil-ities of mobile terminals,and computing offloading technology can effectively solve this problem.Wireless communication technology is also driving the con-tinuous development of the vehicular networks.Vehicles can be connected to each other through advanced communication technology for resource sharing.Cloud computing technology can make full use of communication,computing and storage resources in the network,and vehicles can play two roles.One role is the computing resource provider,and the mobile terminal offloads the computing task to the vehicle calculation;the other is the computing resource requester,the vehicle offloads the computing task to the connected mobile ter-minal.Conventional computing tasks are mainly random splitting tasks and fixed splitting tasks.This thesis conducts computing offloading research for these two computing tasks.The proposed computing offloading algorithm can minimize its corresponding objective function.At first,the offloading problem of random splitting tasks is studied.The objective function of the problem is to minimize the average delay of the system.In order to achieve this goal,vehicles can selectively offload some tasks to the edge cloud.The decision-making process of the system is modeled as a Markov dynamic decision-making process,and an offloading algorithm based on Asynchronous Advantage Actor-Critic(A3C)algorithm is proposed.The simulation results indicate that the performance of the A3C algorithm is better than the greedy algorithm and the other two comparison strategies.Then the problem of computing offloading of fixed splitting tasks is stud-ied.Also,with the objective function of minimizing the average system de-lay,vehicles can offload some subtasks to the car cloud or edge cloud.The decision-making process of the system can be regarded as a Markov dynamic decision-making process,and an A3C-based offloading algorithm is proposed.Computational complexity analysis indicates that the computational complex-ity of the A3C algorithm is significantly better than the deep Q-learning(DQN)algorithm.The simulation results reveal that the performance of the A3C algo-rithm is better than the DQN algorithm,the greedy algorithm and an another offloading strategy.