Research On Mobile Edge Computing Oriented Multi-objective Computation Offloading Problem

Benefitting from the continuous development and maturity of technologies such as the Internet of Things and artificial intelligence,mobile devices(MDs),e.g.,connected vehicles and smartphones,have become the dominant computing platforms.Meanwhile,the emergence of computing-intensive applications like autonomous driving and augmented reality,poses significant challenges to the computing resources of MDs due to their physical size constraint.The contradiction between computing-intensive applications and resource-constrained MDs has been prominent recently.Mobile edge computing(MEC)brings abundant cloud computing resources to the edge of networks close to MDs and provides wireless access networks and computing offloading services.Therefore,MEC has been considered a promising technology to address the above contradiction.MDs can offload their computing-intensive applications to the edge cloud for processing via wireless networks,which shortens the end-to-end delay of mobile applications and alleviates the pressure on MDs in terms of computation and energy consumption.However,due to the complexity of offloading applications in the base station(BS)based MEC networks,the high dynamic of computing services in the unmanned aerial vehicles(UAV)assisted MEC networks,and the difficulty of balancing multiple performance indicators in computation offloading,serious challenges are posed to the optimization of MEC system performance and the design of computation offloading strategies.Focusing on BSbased and UAV-assisted MEC networks and measuring the system performance by multiple indicators,such as delay and energy consumption,this thesis studies the task graph(i.e.,application)offloading problem with multiple MDs in the static multi-cell network,task graph offloading problem with dynamic preferences in the dynamic multi-access network,and trajectory control and task offloading problem in the UAV-assisted network.The research on the above three issues can support computing-intensive mobile applications for future edge networks,providing important references for decision-makers.The main works are summarized as follows:This thesis proposes an improved multi-objective evolutionary algorithm based on decomposition to address the task graph offloading problem in the static multi-cell MEC network.This thesis studies the task graph offloading problem in the static multi-cell MEC network.The problem aims to minimize the average execution delay(AED)of applications and average energy consumption(AEC)of MDs by optimizing each application’s task execution location and order,subject to dependency constraints between tasks in each application.A task graph offloading algorithm based on the multi-objective evolutionary algorithm is proposed to strike a balance between AED and AEC,namely MOEA/D-STGO.First,based on the characteristics of each application,a chromosome encoding method is designed to include each application’s task execution location and order.Each chromosome represents a solution to the problem,which achieves a tradeoff between AED and AEC.Subsequently,a problemspecific population initialization scheme is used to obtain a set of high-quality initial solutions in the initialization phase,guiding the proposed algorithm to locate regions with excellent solutions quickly.The energy conservation based on the dynamic voltage frequency scaling(DVFS)scheme is proposed to reduce MDs’ energy consumption without prolonging application execution delays.The two schemes can enhance the global and local search abilities of MOEA/D-STGO,thus significantly improving its optimization performance.The experimental results show that MOEA/D-STGO can obtain better multi-objective offloading solutions than five multi-objective evolutionary algorithms and three heuristic algorithms.The proposed MOEA/D-STGO decreases applications’ AED and MDs’ AEC and extends their battery life.This thesis presents an improved multi-objective reinforcement learning algorithm with dynamic weight setting to deal with the task graph offloading problem in the dynamic multi-access MEC network.This thesis studies the task graph offloading problem in the dynamic multi-access MEC network,with the application execution delay(AED),energy consumption of the MD,and execution charge for edge computing minimized simultaneously.The wireless channel quality and preferences between objectives will change dynamically over time in the process of computation offloading.Therefore,when making the offloading decisions,it is necessary to pay attention to the channel quality and preferences changes.First,based on the MEC network’s environmental states and the MD’s application attributes,a multiobjective Markov decision process is built,where a vector-valued reward function related to the optimization objective is designed.Then,this thesis proposes a task graph offloading algorithm based on multi-objective reinforcement learning(namely MORL-DTGO)for single MD,aiming at adapting to the dynamics and uncertainty of the MEC network.A tournament selection method is devised to select important preferences to effectively maintain previously learned policies.The simulation results show that MORL-DTGO performs better than three multi-objective reinforcement learning algorithms in adapting to the dynamic preferences and responding to the user’s offloading requirements.The proposed MORL-DTGO also outperforms five algorithms regarding multiple offloading performance indicators and can simultaneously minimize the application execution delay,energy consumption,and execution charge,thus satisfying the offloading demands of end-users with dynamic preferences.This thesis proposes an improved evolutionary multi-objective reinforcement learning algorithm to solve the trajectory control and task offloading(TOCO)problems in the UAV-assisted MEC network.This thesis investigates the trajectory control and task offloading problems,where a UAV and a BS work together to provide MDs with computing services.The computation tasks arriving at each MD can be modeled as an independent and identically distributed sequence of Bernoulli random variables.The UAV collects computation tasks from the MDs within its coverage.Then it stores these tasks in the computing queue and makes task offloading decisions.This thesis formulates the TCTO problem as a multi-objective optimization problem,aiming to minimize the task execution delay and UAV’s energy consumption,and maximize the number of tasks collected by the UAV,simultaneously.First,based on the UAV’s environmental states and task profiles in the computing queue,the multi-objective Markov decision process is established,where a vector-valued reward function related to the optimization objective is designed.This thesis proposes a trajectory control and task offloading algorithm based on the evolutionary multi-objective reinforcement learning(EMORL),namely EMORL-TCTO.The multi-task multi-objective proximal policy optimization in the original EMORL is improved by retaining all new learning tasks in the offspring population,which can preserve promissing learning tasks.The simulation results show that EMORL-TCTO exceeds two multi-objective evolutionary algorithms and two multi-policy multi-objective reinforcement learning algorithms regarding two algorithm-related performance indicators.Thus,EMORL-TCTO can provide abundant and better solutions for decision-makers.EMORLTCTO also outperforms the above four algorithms in several offloading performance metrics.Therefore,it can minimize task execution delay and UAV’s energy consumption and maximize the number of tasks collected.