基于多种Transformer方法的服务功能链部署对比研究

With the development of network architecture, Service Function Chain (SFC) deployment has become crucial in modern communication networks. The Transformer model, with its self-attention mechanism and ability to model global information, can optimize resource allocation, scheduling strategies, and improve decision quality. It demonstrates stronger adaptability compared to traditional reinforcement learning methods in complex network states. Although the previously proposed OBSD algorithm performs well, it may face limitations in computational and storage resources in large-scale, complex networks. Moreover, its online fine-tuning mechanism may not make fast and accurate adjustments in dynamic environments. To further compare the performance of different Transformers in SFC deployment, this paper introduces Q-learning Decision Transformer (QDT) and Contrastive Decision Transformer (ConDT) models, proposing two Transformer-based deployment algorithms: QBSD and CBSD. Through simulation experiments, the performance of OBSD, QBSD, ConDT, and D3T models was evaluated in dynamic, static, and large-scale network environments. The experimental results show that in large-scale networks, OBSD outperforms the others, followed by QBSD, improving request acceptance rates by 7.1% and 3.7%, respectively. In static environments, OBSD performs the best, followed by CBSD, which improves request acceptance rates by 11.2% and 7.9% compared to D3T, and reduces end-to-end delay by 21.8% and 6.7%, respectively. These results provide theoretical support and practical guidance for the optimization of SFC deployment.