Delay-Aware VNF Migration: A Lagrangian-Constrained Soft Actor-Critic Method

Virtual Network Function (VNF) migration is a critical mechanism in Mobile Edge Computing (MEC) for accommodating user mobility while maintaining low-latency services. However, existing reinforcement learning approaches often struggle with the trade-off between minimizing end-to-end latency and migration-induced service downtime. Furthermore, their effectiveness is frequently limited by insufficient policy exploration. To address these challenges, we propose a Soft Actor-Critic (SAC)-based framework, augmented with a Lagrangian constraint mechanism. Our approach leverages the maximum entropy objective of SAC to foster robust policy exploration, while employing a Lagrangian method to constrain service downtime to a predefined threshold. This combined approach yields an adaptive and delay-aware VNF migration policy that minimizes latency while strictly adhering to service continuity requirements. Numerical results demonstrate that our method significantly outperforms conventional approaches, achieving lower latency under stringent downtime constraints and thereby enhancing the overall service quality in dynamic MEC environments.