24 November 2021, Shenzhen, China – ZTE Corporation (0763.HK / 000063.SZ), a major international provider of telecommunications, enterprise and consumer technology solutions for the mobile Internet, today announced that it ranked in the third position of the Graph Neural Networking Challenge 2021 for the data model it proposed in the ITU AI/ML (Artificial Intelligence/Machine Learning) in 5G Challenge.
With the data model, ZTE got the MAPE (Mean Absolute Percentage Error) of 1.85 in a real large-scale network, which is significantly better than the test result (MAPE>300) without the algorithm optimization. The Graph Neural Networking Challenge is one of the series of AI/ML in 5G Challenge and is held by the Barcelona Neural Networking Center and Universitat Politecnica de Catalunya BarcelonaTech (BNN-UPC).
Graph Neural Networking Challenge aims to construct a large-scale digital twin with Graph Neural Networks (GNN) and simulate network performance by using GNN algorithm. The traditional network digital twin technology cannot achieve high simulation accuracy with low computing cost. While GNN is the only machine learning technology to keep a balance between them. Besides, GNN can provide higher simulation accuracy with low computing resource requirements and facilitate the deployment of laboratory network training results in real networks.
The difficulty in applying GNN in a digital twin network is that the performance indicators of a small-scale network used for machine learning training are largely different from those of a real large-scale network. Therefore, the training model obtained in the small network cannot be applied in real networks.
To solve this challenge, ZTE has changed the predictive variable in the model from the latency to the port utilization, based on the principle that the port utilization ranges in different scales of networks are close to each other. Then, ZTE has calculated the latency according to the relationship between the port utilization and the latency, to solve the problem of different ranges and highly improve the prediction accuracy of small-scale network training models.
To date, ZTE has integrated this technology into the research and planning of network self-healing and quick service provisioning to improve their reliability. Based on the principle of this algorithm, ZTE has developed an automatic network optimization algorithm, so that the intelligent management and control system can recommend the most desirable network configuration solution based on the current network status and user requirements. Moving forward, ZTE will continue using its AI/ML technologies to provide more intelligent, efficient and convenient communication networks for all customers.