Loading...

Applying Deep Reinforcement Learning for Task Offloading and Resource Allocation in Fog-Based Infrastructures

Fallah Mir Mousavi Ajdad, Zahra | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58188 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Izadi, Mohammad
  7. Abstract:
  8. In cloud computing, serving applications is often done by sending data to data centers and servers. This centralized processing model, especially in the field of the Internet of Things (IoT), where the number of devices is very large and their geographical distribution is significant, brings challenges such as limited bandwidth, high latency, and privacy issues. In this context, fog computing, as an efficient solution, organizes devices in a layered manner from the proximity of users to cloud processing centers. This layered structure allows for the distribution of resources and data processing across different regions of the network, tailored to the needs of users and their proximity to them, creating a balance between processing speed efficiency and reducing latency. In many real-world problems, it is possible to model them as workflows with interrelated structures. Workflows are an effective method for modeling and managing complex, distributed, interactive, and scientific applications. A key feature of these applications is the distribution of computations and data in execution environments that require dynamic interaction to provide preliminary results and continue the process. In this study, the simulated infrastructure consists of three layers: IoT devices, fog nodes, and cloud servers. The decision-making model is defined as a Markov Decision Process (MDP), and to solve it, a combination of two independent deep reinforcement learning (Deep Q-Learning) agents is used; one focuses on reducing processing costs and the other aims to reduce latency. The priority of these two agents is determined by user-configurable weightings. The main distinction of this approach lies in its high configurability to adapt to different user conditions and priorities. Furthermore, unlike most similar methods that focus only on independent tasks, this approach considers the interdependencies between tasks in the form of workflows, enhancing the accuracy of decision-making in resource allocation. The proposed framework is implemented and evaluated in a SimPy-based simulation environment. In this framework, the average response time and total processing cost are considered as evaluation parameters. Simulation results show that the proposed approach significantly improves resource utilization and reduces latency compared to the baseline random offloading method, and is capable of dynamically adapting to changing network conditions and workflow diversity
  9. Keywords:
  10. Workflow ; Fog Computing ; Offloading ; Resources Allocation ; Reinforcement Learning ; Cloud Computing

 Digital Object List

 Bookmark

...see more