Service Provisioning in Edge Computing for IoT Applications via Intelligent Resource Optimization and Allocation

Abstract

The Internet of Things (IoT), as an emerging technology that connects a wide range of smart devices to the Internet, has significantly impacted our daily lives and will continue to grow and evolve in the coming years. To meet delay-sensitive user service requirements, mobile edge computing (MEC) provides computing resources by placing cloudlets in the proximity of users. Also, with the development of MEC, an era of rapid growth and expansion in the fusion of Artificial Intelligence (AI) and IoT has begun. The increasing computation power allows the shift of computation and services from the remote cloud to the edge of the core network, which provides low-latency services for IoT devices. As a result, more and more sensors and smart devices are being deployed. AI techniques help to manage limited computing, storage and communication resources in a dynamic MEC environment. To utilize the computation power of smart IoT devices while protecting user privacy, federated learning enables the training of deep neural networks at smart IoT devices and the aggregation of trained models at edge servers. Digital twins (DTs) as replicas of real objects, enable the provisioning of IoT services between the real world and the virtual digital world. In this thesis, we study service provisioning for IoT applications in mobile edge computing networks via intelligent resource allocation and optimization. This introduces several challenges: (1) How to maximize the data collection from IoT devices deployed in remote areas by using energy-constrained unmanned aerial vehicles (UAVs); (2) How to synchronize DTs in cloudlets with their IoT devices in a real-time manner to minimize the DT state staleness under the limited synchronization budget per update round; (3) How to maximize the profit of IoT service providers by utilizing deep learning techniques to manage service placement on edge servers and allocation of user requests; (4) How to maximize the performance of deep neural networks by selecting federated learning clients in a training round in a device-to-device communication enabled edge environment. Firstly, we study the sensing data collection of IoT devices in a sparse IoT-sensor network, using an energy-constrained UAV, where the sensory data is stored in IoT devices. We proposed a novel framework to enable the UAV to collect data from multiple devices simultaneously. We also devised efficient approximation and near-optimal algorithms to find the trajectory that maximizes the data collection. Secondly, we consider the DT state staleness minimization problem of objects, by a real-world application example: a UAV-enabled data collection in a renewable sensor network, and the collected data are then uploaded to their corresponding DTs in an MEC network for fidelity-aware user queries. We devise efficient algorithms and a data volume prediction mechanism for the problem. We also study on how to develop evaluation plans for queries on DT data. Thirdly, we investigate the online service placement and request assignment problem in MEC networks with the aim to maximize the profit of the service provider, by admitting as many service requests as possible for a given monitoring period. We implement a prediction mechanism for future request arrivals to pre-install service instances. We also develop an efficient request assignment algorithm. Fourthly, we study energy-aware DNN model training in an edge computing environment. We formulate a novel federated learning framework that enables device-to-device communication when model uploading to mitigate communication overhead. We also develop near-optimal algorithms to minimize the loss of global models trained in the formulated framework, where the energy resources of devices and the bandwidth of the server are limited. Finally, we summarise the work and discuss several potential research topics for future research based on the work in this thesis.