Fast Adaptive Meta-Learning Framework for Resource Allocation in Optical Networks

Abstract

With new applications arising in 6G, digital twins and generative artificial intelligence, optical networks are facing challenges from diverse and constantly changing environments. However, existing global optimization schemes for resource allocation such as meta heuristics and machine learning based methods are designed for specific environments, and thus struggle to maintain knowledge when encountering different environments. Consequently, new resource allocation scheme that can fast adapt across various environments (resource types, requests, and network topologies) becomes crucial. In this paper, we propose a fast adaptive meta-learning framework for resource allocation (FAML) to achieve high resource efficiency and fast adaption across environments. FAML is a systematic solution that reformulates the resource allocation problem by explicitly considering environment changing, and leverages the meta-learning approach to learn the meta policy that can adapt across environments. Besides, FAML adopts graph attention networks (GAT) to represent information of various networks, and deep reinforcement learning (DRL) to learn an optimized policy. We conduct simulations under various environmental distributions for evaluation, showing that FAML performs significantly better than previous methods with respect to both optimization performance and adaptation efficiency. In particular, FAML has 80%/57% reduction on average in training episodes compared to the DRL/TL, respectively.