Presentation Information
[B-6-51]Training Data Generation for Communication Quality Estimation Based on Device Information
〇Akihiro Kimura1, Tomoki Tajimi1, Toshihiko Nakano1, Satoshi Nishiyama1, Ken Takahashi1 (1. NTT, Inc.)
Keywords:
Training Data Generation,Communication Quality Estimation,Machine Learning
In network design and operation, the quality attainable at a given place and time directly informs capacity planning and degradation diagnosis. We estimate, via machine learning from device information, the maximum quality (the "latent quality") a terminal with sufficient demand could reach upon starting communication. Active measurement reliably elicits it but needs on-site work and offers limited coverage; log observation scales broadly but depends on demand, undershooting it when demand is low. We propose deriving, from ordinary operational logs, the relation between background load and attainable throughput. Intervals of sharp throughput rise are treated as naturally occurring load from rising demand; the device state just before the rise is the input and the post-rise throughput a candidate label, avoiding input contamination. Whether the value is near the limit is judged from resource utilization, buffer occupancy, saturation, and radio quality; if not, the limit is extrapolated from incremental ratios, with confidence-based weights. In simulation, the method cut RMSE from 20.6 to 7.6 Mbps (63%) and nearly removed the systematic underestimation of naive log learning.