Research On Intelligent Optimization Of Offloading Decision And Resource Allocation In Multi-access Edge Computing Networks

Network artificial intelligence(AI)is an innovative technology that integrates connections,computing power,algorithms,and data,expanding the service scope of future networks and leading communication networks into a new domain.Concurrently,multi-access edge computing(MEC)is an emerging computing paradigm that brings cloud computing capabilities closer to the network edge,enhancing users’ quality of experience(QoE).Importantly,network AI and MEC are complementary rather than conflicting technologies.Their synergism can foster the swift advancement of future networks and the extensive deployment of intelligent applications.As one of the critical MEC use cases,computation offloading allows users to offload computation-intensive or latency-sensitive tasks to nearby edge nodes,thereby improving user computing capabilities and QoE.However,in a heterogeneous network,adaptive algorithms need to be proposed to formulate MEC offloading decisions and resource allocation strategies due to user mobility leading to base station handovers and scenario changes.This dissertation focuses on the offloading decision and resource allocation problems for MEC in heterogeneous networks.The aim is to propose intelligent optimization algorithms with adaptive capabilities and low complexity,to adapt to scenario changes and meet real-time decision-making needs in mobile environments.The main contributions of this dissertation are as follows:1)The influence of MEC offloading decisions and time scheduling on energy consumption performance in a heterogeneous network with multiple users is analyzed.Initially,a MEC offloading scheme is introduced,employing dual connectivity(DC)and non-orthogonal multiple access(NOMA)technologies to enable user equipment(UE)to offload tasks to both macro cell base stations(MBS)and small cell base stations(SBS)concurrently.The subsequent analysis explores the delay and energy consumption of each stage during the MEC offloading process,presenting the problem as an energy consumption minimization problem subject to delay constraints.The proposed algorithm employs the difference of convex programming(DCP)based on successive convex approximation(SCA)to achieve near-optimal task segmentation and time scheduling strategies.Moreover,a cyclic branching algorithm incorporating a branchy network(BranchyNet)and residual network(ResNet)is introduced to enhance inference efficiency.Finally,numerical results indicate that the proposed offloading scheme surpasses alternative schemes in energy consumption performance.Additionally,the deep learning(DL)-based algorithm demonstrates significantly decreasing inference time compared to traditional optimization algorithms while maintaining comparable energy consumption performance.2)The impact of MEC offloading decisions and power allocation on energy performance in single-user mobile scenarios within ultra-dense networks is investigated.Initially,a MEC offloading scheme is devised,enabling a single mobile user(MU)to offload computational tasks to multiple base stations simultaneously using downlink NOMA and DC technologies,and the energy consumption and delay of MEC offloading during user mobility are modeled.Subsequently,the optimization objectives are classified into two cases based on distinct considerations: solely the MU’s transmission energy consumption and the total energy consumption of MU’s transmission and computation on the MEC server(MES).As both optimization objectives constitute mixed-integer nonlinear programming(MINLP)problems with non-convex constraints,the problem is decomposed into two subproblems:task segmentation and power allocation.Then,a heuristic algorithm incorporating the interior-point approach and binary search is proposed to develop sub-optimal task segmentation and power allocation strategies.Furthermore,a deep learning(DL)algorithm employing the model pool and finetuning and a deep reinforcement learning(DRL)algorithm utilizing reward shaping and twin delayed deep deterministic policy gradient(TD3)are developed for the two objectives to satisfy the real-time demands of offloading decision.Numerical results demonstrate that the proposed machine learning(ML)-based algorithms confer advantages in enhancing inference efficiency relative to the heuristic algorithm but achieving an energy consumption performance commensurate with the heuristic algorithm.3)The impact of MEC offloading decisions and power allocation on energy efficiency performance is investigated in heterogeneous networks with multiple users.Initially,DC technology is introduced to enable MUs to connect to both an MBS and an SBS simultaneously,while NOMA technology is employed to facilitate multiple MUs within the same service cell sharing identical timefrequency resources.Subsequently,the analysis investigates the delay and energy consumption associated with MEC offloading during user mobility.The problem is formulated as minimizing total energy consumption with power and delay constraints.Then,the problem is partitioned into two subproblems: task segmentation and power allocation.An iterative optimization algorithm,integrating gradient descent and the interior-point method,is introduced for task segmentation and power allocation decisions.A DRL-based algorithm employing a model pool and TD3 is proposed to address model adaptability and real-time decision requirements.Numerical results suggest a superior inference efficiency of the DRL-based algorithm in comparison to both the iterative optimization algorithm and the Lingo algorithm.However,all three algorithms demonstrate comparable performance with respect to energy consumption.4)The impact of MEC offloading decisions,computing power control,and power allocation on computational delay is investigated in heterogeneous networks with multi-MUs.First,the delay and energy consumption during MEC offloading in different scenarios are examined,and the problem is constructed as a maximum delay minimization problem subject to energy consumption constraints.Subsequently,the maximum computational delay is relaxed in this problem,and closed-form solutions for computational capability control and power allocation are derived.The problem is then decomposed into three subproblems: power allocation,task segmentation,and achieving minimum delay.The proposed iterative optimization algorithm,which combines gradient descent and binary search,aims to achieve task segmentation and minimum delay.Concurrently,a multi-task DRL algorithm utilizing the model pool and TD3 is introduced to address the multi-task optimization problem and guarantee real-time deployment.Finally,simulation results reveal that the DRL-based algorithm significantly reduced inference time compared with the differential evolution(DE)and iterative optimization algorithms but is on par regarding average delay performance,rendering it more appropriate for mobility scenarios.