As a large amount of data is increasingly generated from edge devices,such as smart homes,mobile phones,and wearable devices,it becomes crucial for many applications to deploy machine learning modes across edge devices.The execution speed of the deployed model is a key element to ensure service quality.Considering a highly heterogeneous edge deployment scenario,deep learning compiling is a novel approach that aims to solve this problem.It defines models using certain DSLs and generates efficient code implementations on different hardware devices.However,there are still two aspects that are not yet thoroughly investigated yet.The first is the optimization of memory-intensive operations,and the second problem is the heterogeneity of the deployment target.To that end,in this work,we propose a system solution that optimizes memory-intensive operation,optimizes the subgraph distribution,and enables the compiling and deployment of DNN models on multiple targets.The evaluation results show the performance of our proposed system.

As a combination of edge computing and artificial intelligence,edge intelligence has become a promising technique and provided its users with a series of fast,precise,and customized services.In edge intelligence,when learning agents are deployed on the edge side,the data aggregation from the end side to the designated edge devices is an important research topic.Considering the various importance of end devices,this paper studies the weighted data aggregation problem in a single hop end-to-edge communication network.Firstly,to make sure all the end devices with various weights are fairly treated in data aggregation,a distributed end-to-edge cooperative scheme is proposed.Then,to handle the massive contention on the wireless channel caused by end devices,a multi-armed bandit (MAB) algorithm is designed to help the end devices find their most appropriate update rates.Diffe-rent from the traditional data aggregation works,combining the MAB enables our algorithm a higher efficiency in data aggregation.With a theoretical analysis,we show that the efficiency of our algorithm is asymptotically optimal.Comparative experiments with previous works are also conducted to show the strength of our algorithm.

The continuous changes of the external environment lead to the performance regression of neural networksbased on traditional deep learning methods.Therefore,continual learning(CL) area gradually attracts the attention of more researchers.For edge intelligence,the CL model not only needs to overcome catastrophic forgetting,but also needs to face the huge challenge of severely limited resources.This challenge is mainly reflected in the lack of labeled resources and powerful devices.However,the existing classic CL methods usually rely on a large number of labeled samples to maintain the plasticity and stability,and the lack of labeled resources will lead to a significant accuracy drop.Meanwhile,in order to deal with the problem of insufficient annotation resources,semi-supervised learning methods often need to pay a large computational and memory overhead for higher accuracy.In response to these problems,a low-cost semi-supervised CL method named edge hierarchicalmemory learner (EdgeHML) is proposed.EdgeHML can effectively utilize a large number of unlabeled samples and a small number of labeled samples.It is based on a hierarchical memory pool,leverage multi-level storage structure to store and replay samples.EdgeHML implements the interaction between different levels through a combination of online and offline strategies.In addition,in order to further reduce the computational overhead for unlabeled samples,EdgeHML leverages a progressive learning method.It reduces the computation cycles of unlabeled samples by controlling the learning process.Experimental results show that on three semi-supervised CL tasks,EdgeHML can improve the model accuracy by up to 16.35% compared with the classic CL method,and the training iterations time can be reduced by more than 50% compared with semi-supervised methods.EdgeHML achieves a semi-supervised CL process with high performance and low overhead for edge intelligence.

The existing research on resource allocation strategies of mobile edge computing mostly focus on delay and energy consumption,but relatively less consideration is given to the benefits of edge servers.When considering the benefits of edge servers,many studies ignore the optimization of delay.Therefore,a two-way update strategy based on game theory(TUSGT) is proposed.TUSGT transforms the task competition between servers into a non-cooperative game problem on the side of edge servers,and adopts a joint optimization strategy based on potential game,which allows every edge server to determine the task selection prefe-rence with the goal of maximizing its own profit.On the side of mobile devices,the EWA algorithm in online learning is used to update the parameters,which affects the task selection preference of the edge servers from a global perspective and improves the overall deadline hit rate.Simulation results show that,compared with BGTA,MILP,greedy strategy,random strategy,and ideal strategy,TUSGT can increase the deadline hit rate by up to 30%,and increase the average profit of edge servers by up to 65%.

In order to cope with the performance loss caused by temporal-spatial resource dispersion of edge service providers (ESPs) in edge intelligence-driven industrial Internet of Things system,this paper proposes a coalition game-based joint resource allocation scheme assisted by digital twin.Firstly,we design a transferable utility coalition game model consisting of a primary problem of utility maximization of edge devices and a sub-problem of utility maximization of ESPs under the constraints of ESPs' resource limitation including bandwidth,computation and cache capabilities.Then,the original multi-objective problem is transformed into one convex problem with linear constraints.Finally,an alternative optimization method is leveraged for solving the equivalent optimization problem.Simulation results show the effectiveness of the proposed coalition game-assisted scheme for improving system resource utilization,with greater promotion as the number of ESPs grows.This proves that the proposed scheme is more adaptable to large scale edge intelligence systems,compared with traditional Nash equilibrium and grand coalition method.

Based on the idea of crowdsourcing,mobile crowdsensing(MCS) collects mobile sensing devices to sense the surroun-ding environment,which can make environment sensing and information collection more flexible,convenient and efficient.Whe-ther the task allocation strategy is reasonable or not directly affects the success of the sensing task.Therefore,formulating a reasonable task allocation strategy is a hotspot and focus in the research of MCS.At present,most of the task allocation methods in MCS systems are offline and targeted at single type tasks.However,in practice,online multi-type task allocation is more common.Therefore,this paper studies the task allocation method in MCS for multiple types of tasks,and proposes an online task allocation strategy oriented to system benefits combined with the characteristics of MCS technology in the military field.In this paper,a long-term,dynamic online task allocation system model is established,and the problem is solved based on Lyapunov optimization theory with the system benefit as the optimization goal,so that the online dynamic control of task admission strategy and task allocation scheme is realized.Experiment shows that the online task allocation algorithm proposed in this paper is effective and feasible.It can reasonably allocate the tasks arriving at the MCS system online,ensure the stability of the task queue,and increase the system utility by adjusting the parameter value.

Mobile CrowdSensing (MCS) is a promising sensing paradigm that recruits users to cooperatively perform sensing tasks.Recently,unmanned aerial vehicles (UAVs) as the powerful sensing devices are used to replace user participation and carry out some special tasks,such as epidemic monitoring and earthquakes rescue.In this paper,we focus on scheduling UAVs to sense the task Point-of-Interests (PoIs) with different frequency coverage requirements.To accomplish the sensing task,the scheduling strategy needs to consider the coverage requirement,geographic fairness and energy charging simultaneously.We consider the complex interaction among UAVs and propose a grouping multi-agent deep reinforcement learning approach (G-MADDPG) to schedule UAVs distributively.G-MADDPG groups all UAVs into some teams by a distance-based clustering algorithm (DCA),then it regards each team as an agent.In this way,G-MADDPG solves the problem that the training time of traditional MADDPG is too long to converge when the number of UAVs is large,and the trade-off between training time and result accuracy could be controlled flexibly by adjusting the number of teams.Extensive simulation results show that our scheduling strategy has better performance compared with three baselines and is flexible in balancing training time and result accuracy.

As a new architecture,the space-air-ground integrated communication technology can effectively improve the network service quality of ground terminal,and has attracted widespread attention in recent years.This paper studies a space-air-ground integrated mobile edge computing system,in which multiple UAVs provide low-latency edge computing services for ground devices,and low earth orbit satellites provide ubiquitous cloud computing services for ground devices.Since the deployment position of the UAVs and the scheduling scheme of computing tasks are the key factors affecting the performance of the system,the deployment position of the UAVs,the link relationship between the ground terminal and the UAVs,and the offloading ratio of computing tasks need to be jointly optimized to minimize the average task response delay of the system.Since the formally defined joint optimization problem is a mixed nonlinear programming problem,this paper designs a two-layer optimization algorithm.In the upper layer of the algorithm,a particle swarm optimization algorithm that combines genetic algorithm operators is proposed to optimize the deployment position of the UAVs,and the greedy algorithm is used in the lower layer of the algorithm to optimize the computing task offloading scheme.The extensive simulation experiments verify the feasibility and effectiveness of the proposed method.The results show that the proposed method can achieve lower average task response time,compared to other baseline methods.

Deep space exploration is a significant area of space missions in the modern world,and future large-scale deep space exploration will be greatly impacted by autonomous deep space exploration technologies.The autonomous technology of deep space exploration faces severe challenges because of the complicated and uncharted deep space environment,lengthy deep space communication time,and constrained on-board computing capacity.To address this issue,a cloud-edge-end cooperation architecture for deep space exploration tasks using digital twins is developed,which can offer more efficient resource services for deep space exploration autonomous technologies.Firstly,the complex deep space exploration task is decomposed into multiple sub-modules with dependencies.Secondly,the orbiter coverage time model,the collaborative computing model,and the task dependency model are established in the virtual space layer.Finally,based on the aforementioned models,the corresponding target optimization pro-blem is proposed.The optimization objective is to minimize the energy consumption and time of the landing rover for completing the deep space exploration mission under the constraints of module dependence,the effective communication service time of the orbiter and the transmit power control of the landing rover.In order to solve this optimization problem,an adaptive genetic algorithm is proposed,so that the optimal execution strategy for the landing rover in the physical space layer can be determined.Si-mulation results show that the proposed adaptive genetic algorithm can effectively reduce the mission completion time and energy consumption.Additionally,the proposed cloud-edge-end cooperation computing model is contrasted with the other three computing models,and the results reveal that,when it is used to achieve the same objective,the proposed cloud-edge-end cooperation framework has a greater resource utilization.