Presentation Information
[B-8-13]Delay-Consistent Cross-Traffic Generation via Stepwise Feature Representation Learning
◎Ryohei Yamada1, Hideaki Kimura1, Takashi Nakanishi1, Tatsuya Shimada1 (1. NTT)
Keywords:
Network Digital Twin,Machine Learning,Surrogate Network Generation
Network digital twins (NDTs) have attracted attention as a promising technology for supporting network performance estimation and design verification. However, in unidentified networks that include segments where network configuration information cannot be obtained, conventional NDTs cannot be directly applied because input information such as topology, routing, and traffic volume is missing. This paper focuses on a scenario where the topology and routing are known, but the traffic volume on communication paths other than the E2E communication path used by a user, namely cross-traffic, is unknown. We propose a method for generating cross-traffic conditioned on the delay and transmitted traffic observed on the user’s E2E communication path. The proposed method first generates path feature representations using a conditional variational autoencoder (CVAE), and then generates link feature representations and cross-traffic in a stepwise manner using graph neural networks (GNNs). Experimental results show that, by complementing part of the missing NDT input with the generated cross-traffic, the delay estimated by the NDT closely matches the conditioning delay, confirming that the proposed method can generate cross-traffic consistent with the observed information.