Photonic spiking DQN-GAT architecture for intelligent routing

Area of Science:

• Photonics
• Artificial Intelligence
• Network Engineering

Background:

• Intelligent routing is vital for data centers, IoT, and satellite networks.
• Reinforcement learning (RL) shows promise but faces challenges in energy consumption and latency.
• Current routing algorithms like OSPF and ECMP have limitations in performance.

Purpose of the Study:

• To propose a novel photonic spiking architecture integrating deep Q-network (DQN) and graph attention network (GAT) for efficient intelligent routing.
• To enable low-latency and energy-efficient routing decisions in communication networks.
• To demonstrate the superiority of the proposed architecture over traditional routing algorithms.

Main Methods:

• Developed a photonic spiking-DQN-GAT architecture for intelligent routing.
• Evaluated performance on the Waxman network topology.
• Implemented a hardware-software co-design for the activation layer using a DFB-SA laser.

Main Results:

• Achieved 100% routing accuracy, significantly outperforming OSPF/ECMP.
• Reduced latency by 82.34%/81.86% and increased bandwidth by 66.64%/24.96% compared to OSPF/ECMP.
• Demonstrated low power consumption (1.685 μJ/inf) and latency (191.20 ps/inf) in hardware implementation.

Conclusions:

• The photonic spiking-DQN-GAT architecture offers a pathway to low-latency, energy-efficient intelligent routing.
• The proposed system significantly enhances routing performance metrics.
• Hardware implementation validates the practical feasibility and efficiency of the photonic approach.