Research On Computation Offloading And Multi-channel Access Based On Reinforcement Learning

Driven by the Internet of Things and 5G communications,in recent years,mobile computing has gradually shifted from centralized mobile cloud computing to multi-access edge computing(MEC).The main feature of MEC is to push mobile computing,network control and storage to the edge of the network(for example,base stations and access points).MEC enables compute-intensive and delay-critical applications on resource-constrained mobile devices,and can significantly reduce latency and mobile energy consumption.There are two key issues in MEC,the computation offloading problem and the multi-channel access problem.The computation offloading problem is that the mobile device needs to develop an offload strategy to determine whether each computation task is offloaded to the edge server.The multi-channel access problem illustrates how to select channels to transmit data in the case of limited spectrum resources.Existing research only builds models from the perspective of a single device,without considering the cooperation and competition between multiple devices.Moreover,computation offloading and multi-channel access are usually considered independently,and have not been considered jointly in existing studies.Aiming at the joint problem of computation offloading and multi-channel access in MEC,and describes this problem in a multi-agent environment,a solution based on multi-agent reinforcement learning is proposed.This solution enables mobile devices to learn the best computation offloading and multi-channel access strategy,which can significantly reduce the calculation delay and improve the success rate of channel access,and does not require prior knowledge of system parameters such as channel state distribution.This solution learns the best strategy over time,provides a low-complexity learning solution for PSPACE-hard’s multi-channel access problem,and makes up for the deficiencies of existing research.In order to verify the effectiveness of the scheme based on multi-agent reinforcement learning algorithm,the simulation of mobile devices and MEC environment was implemented in the Tensor Flow environment,and the proposed multi-agent reinforcement learning algorithm is compared with the single agent algorithm DQN,Actor-Critic and the greedy algorithm with known system state.The greedy algorithm with known system state can be regarded as the theoretical optimal solution of the single agent algorithm.A large number of simulation results under different system parameters show that,compared with the traditional single agent reinforcement learning method,the proposed scheme can reduce the calculation delay by 33.38%,increase the channel access success rate by 14.88%,and increase the channel utilization by 3.24%.