Cloud-Edge Resource Scheduling and Offloading Optimization Based on Deep Reinforcement Learning

Area of Science:

• Smart Manufacturing
• Industrial Internet of Things (IoT)
• Edge Computing

Background:

• Smart manufacturing relies on Industrial Internet of Things (IoT) devices, generating numerous latency-sensitive tasks requiring real-time processing.
• Dynamic changes in edge node load cause increased latency and task dropouts, posing challenges for cloud-edge-end collaboration.
• Existing task offloading strategies struggle with unknown edge node loads and dynamic system states.

Purpose of the Study:

• To propose a distributed algorithm for effective task offloading in smart manufacturing environments.
• To address the challenges of unknown edge node loads and dynamic system state changes.
• To optimize task allocation and execution order for latency-sensitive tasks.

Main Methods:

• A distributed algorithm based on deep reinforcement learning, incorporating Convolutional Neural Networks (CNN) and the Informer architecture.
• CNN extracts local features of edge node loads; Informer's self-attention captures long-term load trends.
• Integration of Dueling Deep Q-Network (DQN) and Double DQN for precise state-action value function approximation.

Main Results:

• The proposed algorithm reduces task dropout rates by 82.3-94%.
• Average latency is decreased by 28-39.2% compared to existing algorithms.
• The method demonstrates significant advantages in high-load, latency-sensitive manufacturing scenarios.

Conclusions:

• The developed deep reinforcement learning algorithm effectively handles dynamic edge node loads and system uncertainties.
• Independent task offloading decisions by mobile devices enable dynamic task allocation and optimized execution.
• The algorithm offers a robust solution for real-time processing in smart manufacturing with Industrial IoT.