Research On Computational Offloading Algorithms In Fog-radio Access Networks

Limited battery,higher wireless transmission delay and lower computing power of the user devices have limited the development of mobile computing for a long time.In recent years,the rapid development of the 5th generation mobile networks(5G)and fog-radio access networks(F-RANs)has made it possible to break above limitations and achieve sustainable operation of devices,which allows the user to process the highly loaded tasks in the local side.In the F-RANs,the system compensates for the lack of computing power of devices by offloading tasks to remote servers.However,if a large number of mobile devices offload tasks to the fog nodes(FNs),the FNs will be under a higher load.Meanwhile,the large number of offloaded tasks will congest the transport networks.Furthermore,these offloaded tasks may suffer from an additional waiting delay or even be discarded when the deadline comes.Hence,it is always an extremely challenge for each user to efficiently determine the offloading decision for the tasks.In this paper,we focus on the offloading strategies in the F-RANs with binary computing,which mainly contains the offloading strategies for single-user scenarios and multi-user scenarios.First,this work has investigated the offloading strategy in the single-user scenario.This paper has proposed a randomization-based dynamic programming offloading algorithm,DPOA,based on the randomization theory,to solve the offloading decision generation problem in mobile fog computing.The algorithm innovatively designs a dynamic programming(DP)table filling approach,which will iteratively generate a set of randomized offloading decisions.If some in these sets improve the decisions in the DP table,then they will be merged into the table.Besides,the iterated DP table is also used to improve the set of decisions,which is generated in the iteration,to obtain the optimal offloading solution.Extensive simulations show that the proposed DPOA algorithm can generate offloading decisions within 3ms with different parameter settings.Meanwhile,the time cost of generating decisions does not increase drastically with the number of tasks.Moreover,in this paper,the performance of offloading in the multi-user scenario is further studied,which aims at obtaining an intelligent algorithm that can generate the offloading decision and allocate the upload channel resource in an optimal way.For evaluating the offloading strategy intuitively,a system utility metric,defined as the sum of delay and energy cost for offloading all tasks,is designed in this work.Then,the problems of the offloading policy and the upload channel resource allocation are modeled as a joint optimal offloading problem.With the objective of minimizing the system utility,we transform the above problem into a mixed integer nonlinear programming(MINLP)problem.To solve this problem,a learning offloading algorithm based on distributed deep neural networks(DNNs)is proposed,DDOA,which uses multiple parallel DNNs to generate offloading decisions.Besides,the newly generated optimal offloading decisions are stored as a public training set to further train and improve the DNNs in the algorithm.Extensive simulations indicate that the DDOA is able to achieve significant reduction on the total offloading cost.Besides,the algorithm has better convergence and higher accuracy while offloading.