Deep Reinforcement Learning Multi-Agent System for Resource Allocation in Industrial Internet of Things

Area of Science:

• Computer Science
• Artificial Intelligence
• Industrial Engineering

Background:

• The Industrial Internet of Things (IIoT) faces challenges with devices having limited computational and communication resources.
• Industry 4.0 necessitates edge computing for data processing, further constrained by available resources.
• Deep Reinforcement Learning (DRL) and Multi-Agent Systems (MASs) show promise for industrial applications like robotics and scheduling.

Purpose of the Study:

• To apply DRL for intelligent resource allocation in industrial edge devices.
• To achieve optimal utilization of limited resources in IIoT devices.
• To leverage MASs for decentralized decision-making in complex IIoT environments.

Main Methods:

• A network of physical and virtualized IIoT devices was constructed.
• Deep Reinforcement Learning (DRL) was employed for resource allocation strategies.
• Multi-Agent Systems (MASs) were utilized for decentralized control and decision-making.
• Performance was evaluated based on MAS overhead, resource usage improvement, latency, and error rates.

Main Results:

• The proposed DRL-based MAS approach effectively managed dynamic system changes.
• MAS agents demonstrated very low resource consumption (traffic, computation, time).
• The system achieved significant improvements in device resource utilization.
• Low latency and error rates were observed during performance evaluation.

Conclusions:

• DRL-powered MASs provide an efficient solution for resource allocation in constrained IIoT environments.
• The developed approach is robust and adaptable to dynamic industrial settings.
• The learned resource allocation policies are transferable from simulation to real-world IIoT systems.