Deep Reinforcement Learning-Based Task Scheduling in IoT Edge Computing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Distributed Systems

Background:

• Edge computing (EC) enables low-latency services for Internet of Things (IoT) applications.
• Limited edge server capacity presents challenges for scheduling resource-intensive application tasks.
• Efficient task scheduling is crucial for maximizing service quality in EC environments.

Purpose of the Study:

• To address the challenge of task scheduling in edge computing environments.
• To maximize the long-term task satisfaction degree (LTSD) for diverse IoT applications.
• To develop an intelligent scheduling algorithm considering task diversity and resource heterogeneity.

Main Methods:

• Formulated the task scheduling problem as a Markov decision process (MDP).
• Leveraged deep reinforcement learning (DRL), specifically a policy-based REINFORCE algorithm.
• Utilized a fully-connected neural network (FCN) for feature extraction in task scheduling and resource allocation.

Main Results:

• The proposed DRL-based algorithm effectively handles both task execution order and VM assignment.
• Simulation results demonstrate superior performance compared to existing methods.
• Achieved higher average task satisfaction degree and success ratio.

Conclusions:

• Deep reinforcement learning offers a powerful solution for complex task scheduling in edge computing.
• The developed algorithm enhances the efficiency and reliability of edge computing services.
• This work provides a significant advancement in optimizing resource allocation for IoT applications on edge servers.