DRL-OS: A Deep Reinforcement Learning-Based Offloading Scheduler in Mobile Edge Computing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Mobile Computing

Background:

• Smart device performance is limited by hardware bottlenecks for intensive applications.
• Task offloading in Mobile Edge Computing (MEC) can mitigate these bottlenecks but faces challenges with delays and expired deadlines.
• Determining optimal offloading decisions (local, remote, or drop) is complex for smart devices due to uncertain task processing.

Purpose of the Study:

• To propose a novel deep-reinforcement-learning-based offloading scheduler (DRL-OS) for smart devices.
• To address the challenge of optimal task offloading decisions considering energy balance, task characteristics, and device constraints.
• To enhance the performance of computation-intensive and delay-sensitive applications on smart devices.

Main Methods:

• Developed a DRL-OS utilizing the double dueling deep Q-network (D3QN) algorithm.
• The DRL-OS learns to make offloading decisions based on task size, deadline, queue status, and residual battery charge.
• Simulations were conducted to evaluate the scheduler's performance against existing schemes.

Main Results:

• The DRL-OS demonstrated a significant reduction in average battery level consumption (up to 54%).
• A notable decrease in the task drop rate was observed (up to 42.5%).
• The scheduler achieved lower average latency, ranging from 0.01 to over 0.25 seconds, outperforming existing methods.

Conclusions:

• The proposed DRL-OS effectively balances energy consumption and task execution for smart devices.
• DRL-OS provides a robust solution for optimizing task offloading in mobile edge computing environments.
• The scheduler enhances overall smart device performance by intelligently managing computation-intensive and delay-sensitive tasks.