![]() | Only 14 pages are availabe for public view |
Abstract The Internet of Things (IoT), real-time media streaming, and other related technologies have increased due to the rapid development of wireless communication technologies and the enormous growth of computation, storage, and data transmission tasks. Edge-Cloud Computing (ECC) is a technology that can be used to better meet the diverse needs of IoT users, it combines the benefits of Mobile Cloud Computing (MCC) and Mobile Edge Computing (MEC) to meet energy consumption and delay requirements, and achieve more stable and affordable task execution. The most significant challenge in ECC is making real-time task offloading decisions. The main purpose of task offloading in ECC is to decide which tasks are offloaded from user devices and which edge or cloud an offloaded task is assigned. For edge/clouds, the task offloading problem is much more complex than the cloud or edge problem because it not only involves all the challenges of offloading tasks on both mobile clouds and edges but also introduces new challenges such as [1]; the heterogeneity between edges and clouds in terms of various resources, a more complex network, and the decision of which edge or cloud each offloaded task is assigned. Achieving optimal offloading decisions usually requires accurate information about network resources and, in turn, increases the overhead and operating costs of Mobile Devices (MDs). Also, cloud environments don’t need a dynamic approach all the time, as most of the traffic on the cloud can be predicted; Therefore, Machine Learning (ML) methods are incorporated into the task offloading algorithm to achieve the offloading decision based on workload information only rather than all information about the system. In order to generate offloading decisions in ECC environments in an efficient and near optimal manner, a Deep Reinforcement Learning (DRL)-based Distributed task Offloading (DRL-DO) framework is proposed. The Keras ML library is used to implement and evaluate the proposed DRL-DO and other offloading algorithms in Python experiments. Experimental results demonstrate the accuracy of the DRL-DO framework, it achieves high Gain Ratio (GR) and greatly reduces the energy consumption, response time, while attaining moderate time cost compared with other offloading algorithms. The work in this thesiscomprises of three stages: 1- The proposed DRL-DO aims to offload the incoming dynamic workloads into the edge or cloud servers, generate offloading decision in distributed manner which preserve generalizability and quick adaptability to new environments is based on learning automata (DRL) as a decision-maker It utilizes a three-dimension Convolutional Neural Network (CNN), while considering multiple tasks with multiple features, and the current workload of the total system as input to CNN to generate offload decisions. 2- The primary goal of the proposed DRL-DO framework is to create a multi-class offload action that can be used to execute independent tasks locally and offload them to the central Cloud Server (CS), or Edge Servers (ES), while taking into account the workload that the Mobile Devices (MD), MEC, and MCC are currently handling to maintain availability, preserve generalizability, and enable quick adaptation to new environments. 3- The Keras Machine Learning (ML) library is used to implement and evaluate the proposed DRL-DO and other offloading algorithms in Python experiments. 4- Experimental results demonstrate the accuracy of the DRL-DO framework, it achieves high Gain Ratio (GR) and greatly reduces the energy consumption, response time, while attaining moderate time cost compared with other offloading algorithms. According to experimental results, the DRL-DO algorithm is accurate; in comparison to DDOA and TOLBO offloading algorithms, it achieves high performance, significantly lowers energy consumption by about 7.6%, and 3.8%, by about 43%, and 34% for response time, and increases the GR by about 28.4 %, and 16.2% in contrast of DDOA and TOLBO, respectively, while achieving acceptable time cost. |