Intelligent learning-based routing algorithm for optical network-on-chips

Abstract

Optical network-on-chips (ONoCs) play a vital role in interconnecting chip cores. Currently, transmission loss is the major limiting factor that affects the size of interconnect networks. As such, reducing the transmission loss is very important for optimizing the operating efficiency and size of ONoCs. Despite already significant progress on transmission loss reduction, further work is still needed to enhance the performance of ONoCs. In this paper, we propose an intelligent deep Q-network (DQN)-based routing algorithm, for arriving at a low-loss optimal route. More specifically, we combine the power loss model with DQN, where the transmitted data packet is regarded as an agent. Through continuous interaction with the environment and iterative trial, the agent can learn a near-optimal routing strategy. Numerical results show that our proposed routing algorithm has better performance and lower online computational latency compared to the traditional routing algorithm. Additionally, with an increasing network size, the advantages of DQN-based approach will become more significant.