| Under such environment that the global commercialization of the fifth generation communication technology,the popularization of a large number of smart wearable devices,including smart phones,smart watches,smart bracelets,and the deployment of a large number of edge sensors in various scenarios,massive data have been generated at the edge side of the network.By conservative estimating that the amount of data produced in one day is roughly equivalent to the amount produced in one year compared with a decade ago.Therefore,the processing efficiency of these data will drop dramatically under the traditional cloud computing framework.To overcome such situation,Mobile Edge Computing(MEC)framework has been proposed.The MEC framework reduces the transmission cost of data or tasks and improves the quality of computing services by placing serval small servers closer to the data generating side.The key problems are divided into two points:1)For the whole system,how to minimize the average delay of task completion;2)For the situation that users could communicate with each other,how to improve the performance of the time cost when offloading tasks?To address the above challenges,this article designs effective algorithms based on the Multi-armed Bandits(MAB)related problems.For the first type of problem,the BMSE algorithm is proposed to optimize the task processing delay for all users in the system.The first sub-algorithm,BMSE-UL,is proposed to optimizes the server selection for each user based on the upper confidence bound principle at the user level,with a regret upper bound of(?).Then,the second sub-algorithm,BMSE-SL,combining batch learning and the upper confidence bound principle at the system level is proposed to optimize the task deployment strategy for the whole system.with a regret upper bound of(?).For the second type of problem,the APU algorithm is proposed to achieve faster convergence by adjusting probabilities and using probability sampling to select servers for individual users,with a regret upper bound of(?).Rigorous regret analysis and convergence proof are conducted for these algorithms,and reasonable simulation experiments are designed to demonstrate their excellent performance,compared with traditional MAB algorithms in terms of average task processing delay,optimal server selection rate,and total task processing delay for the whole system. |
