Reinforcement learning-based SDN routing scheme empowered by causality detection and GNN.

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• High-quality network service demands Quality-of-Service (QoS) routing, a critical technology driven by expanding network applications.
• Machine learning, especially reinforcement learning (RL) and graph neural networks (GNNs), shows promise for QoS routing.
• Existing RL methods neglect the causal impact of agent actions, and GNNs struggle to represent crucial link features for routing.

Purpose of the Study:

• To address limitations in current RL and GNN approaches for QoS routing.
• To quantify causal influence between intelligent agents and the network environment for improved action space exploration.
• To enhance GNNs for effective node and link feature representation in routing optimization.

Main Methods:

• Causal inference techniques are applied to quantify the causal impact of agent actions on the network environment.
• Graph neural networks (GNNs) are utilized to embed both node and link features, improving network representation.
• A centralized reinforcement learning (RL) method is proposed, incorporating a reward function that considers network performance and causality.

Main Results:

• The proposed method effectively achieves QoS-aware routing in Software-Defined Networking (SDN) environments.
• Experimental results demonstrate superior performance compared to baseline methods across key metrics.
• Significant improvements were observed in packet loss reduction, delay minimization, and throughput enhancement.

Conclusions:

• The integration of causal inference with GNNs offers a promising direction for advanced QoS routing.
• The developed centralized RL approach provides an effective solution for optimizing network performance.
• This research advances intelligent routing strategies by considering causal relationships and comprehensive network features.