A Generative Model-Guided Distributed DRL Framework for Scalable and Efficient SFC Orchestration in Future Network Architectures

Service Function Chain (SFC) provisioning plays a crucial role in 5 G and next-generation networks. It involves coordinating Virtual Network Functions (VNFs) in a predefined order to accommodate various SFC requests. Achieving optimal SFC provisioning necessitates advanced decision-making that can adapt to dynamic network conditions. While Artificial Intelligence (AI) modules and Deep Reinforcement Learning (DRL) algorithms have been extensively studied in the literature for this purpose, two critical factors must be considered: the algorithm's efficiency in large-scale networks and the model's ability to comprehensively capture environmental variations. Therefore, this paper introduces a novel Generative Model-Driven Distributed DRL framework for SFC provisioning, where the network is divided into multiple clusters, and each cluster is managed by a dedicated local agent equipped with a Generative-Assisted DRL module, enabling efficient handling of SFC provisioning within its respective region. Also, there is a general agent that can monitor and communicate with local agents to handle requests beyond their capacity. In this proposed approach, a distributed design reduces the workload of the centralized design, while a generative model, which is a Dreaming Variational Autoencoder, assists the DRL model in finding the proper data center for VNF placement by estimating the future state of the network. The proposed model is compared with a distributed DRL model to emphasize the impact of generative assistance under different network configurations. Results show that generative model-driven distributed DRL outperforms the distributed DRL in terms of SFC provisioning and improves SFCs' acceptance ratio from 4% to 11% under different network scale environments.