LLM-Assisted Initial Population Generation for Multi-Objective Joint Service Placement and Traffic Routing in 6G Networks

Abstract

In forthcoming 6G networks beyond 5G, services with increasingly stringent performance requirements are expected to emerge, requiring dynamic and flexible service placement and traffic routing. Furthermore, the integration of non-terrestrial networks for global coverage introduces continuously changing network topologies and conditions, making network optimization significantly more challenging. To address these challenges, AI-native networking has attracted considerable attention by enabling autonomous network optimization. However, AI-based approaches generally require costly training processes, while most network optimization problems remain NP-hard, making it difficult to obtain optimal solutions within polynomial time. Genetic algorithms (GAs), one of the most widely adopted metaheuristic approaches, have been extensively used to efficiently search large solution spaces by evolving an initial population according to fitness values. Since the convergence speed and optimization performance of GAs depend on the quality and diversity of the initial population, generating effective initial solutions is crucial. However, conventional methods typically rely on random initialization, which often limits the exploration of the solution space. To overcome this limitation, this paper proposes an LLM-assisted initial population generation method for jointly optimizing service placement and traffic routing under dynamic network conditions. The proposed method leverages in-context learning to generate informative and distinct initial solutions, which are subsequently evolved by a GA to improve the starting point and convergence behavior of the evolutionary search.