Federated Reinforcement Learning-Based Dynamic Resource Allocation and Task Scheduling in Edge for IoT Applications.

Area of Science:

• Edge Computing
• Internet of Things (IoT)
• Machine Learning
• Federated Learning

Background:

• Optimizing resource allocation and energy consumption in edge computing environments for IoT applications is crucial.
• Existing task offloading algorithms often struggle with dynamic resource availability and data distribution.
• Ensuring privacy and fairness in federated learning models for edge devices presents significant challenges.

Purpose of the Study:

• To develop an efficient task offloading algorithm for IoT edge computing that enhances performance and reduces energy consumption.
• To propose a hybrid forecasting model combining Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) with attention for resource usage prediction.
• To introduce a Deep Deterministic Policy Gradient (D4PG) based Federated Learning algorithm for dynamic user equipment participation, focusing on accuracy, efficiency, and privacy.

Main Methods:

• Utilized Google cluster traces for algorithm development and EdgeSimPy for simulations.
• Developed a hybrid forecasting model integrating BiLSTM, GRU layers, and an attention mechanism.
• Implemented and compared a D4PG-based Federated Learning algorithm against DQN, DDQN, Dueling DQN, and Dueling DDQN on EMNIST and Crop Prediction datasets.

Main Results:

• The proposed task offloading algorithm outperformed best-fit, first-fit, and worst-fit algorithms, ensuring stable edge server power consumption.
• The D4PG-based Federated Learning achieved 92.86% accuracy on the Crop Prediction dataset and high F1-scores (0.9192 on Non-IID EMNIST, 0.82 on IID EMNIST).
• The hybrid offloading algorithm demonstrated reduced power consumption fluctuations across edge nodes compared to existing methods.

Conclusions:

• The developed hybrid forecasting model and task offloading algorithm significantly improve performance and energy efficiency in IoT edge computing.
• The D4PG-based Federated Learning approach offers superior accuracy, efficiency, and privacy preservation in dynamic edge environments.
• The research provides a robust solution for stable energy usage and enhanced data handling in distributed edge systems.