Optimizing lightweight neural networks for efficient mobile edge computing | JoVE Visualize

Area of Science:

Computer Science
Artificial Intelligence
Network Engineering

Background:

Mobile Edge Computing (MEC) is crucial for latency-sensitive applications like IoT, autonomous driving, and smart cities.
Efficient resource allocation in dynamic MEC environments is challenging due to fluctuating workloads, network conditions, and diverse computational capabilities.
Traditional centralized and conventional machine learning methods struggle with scalability, adaptability, and computational overhead in MEC.

Purpose of the Study:

To propose an advanced Multi-Agent Reinforcement Learning (MARL) framework integrated with a lightweight neural network (LtNet) for optimizing task offloading and resource management in MEC.
To address the limitations of existing approaches in handling dynamic MEC environments, aiming for improved scalability, efficiency, and reduced complexity.
To enhance the performance of MEC systems through decentralized decision-making and adaptive learning strategies.

Main Methods:

Development of a Multi-Agent Reinforcement Learning (MARL) framework enabling decentralized decision-making for optimal task offloading.
Integration of a lightweight neural network (LtNet) featuring H-Swish activation and selective Squeeze-and-Excitation modules for reduced computational overhead.
Implementation of adaptive learning strategies allowing devices to dynamically adjust offloading and resource management in real-time.

Main Results:

Achieved a 12-22% reduction in task completion time compared to prior methods.
Demonstrated a 5-8% decrease in energy consumption.
Consistently maintained high resource utilization in dynamic MEC environments.

Conclusions:

The proposed MARL framework with LtNet effectively optimizes task offloading and resource management in dynamic MEC settings.
The approach offers significant improvements in efficiency, scalability, and reduced computational complexity over traditional and single-agent methods.
The findings highlight the potential of advanced AI techniques for enhancing the performance of critical MEC applications.