Eigen RadarAI
Archive
Analysis

KDDI plans network settings that follow hourly congestion

KDDI plans to start rolling out time-adaptive AI across Japan in November after a commercial-network trial. The system adjusts resources assigned to phones as congestion changes during the day. It builds on an earlier distributed-AI deployment, using shared learning and compressed network data to prepare different settings for busy and quieter hours.

Artificial Intelligence··Midday
A cellular antenna and connected radio enclosure overlook commuter paths converging at a busy Japanese transit station.

November rollout follows a commercial-network trial

Japanese mobile operator KDDI and its research organization, KDDI Research, plan to start introducing time-adaptive artificial intelligence nationwide in November. They tested the approach on their commercial network between June and August, choosing 100 congested areas from across Japan for the trial. The sites supplied operating conditions from the network already serving customers. The timetable calls for full operation during fiscal 2026. The new system changes base-station settings to match conditions at different hours rather than keeping the same allocation through the day.[1], [2]

The work builds on distributed AI first introduced at some KDDI base stations in February. That earlier deployment has since reached the nationwide network. KDDI describes improvements in communication quality and a reduction in the work needed to adjust coverage. The latest trial tackles a further operating problem: a location’s congestion can change substantially within one day, requiring adjustments as demand moves between quieter and busier periods.[1]

Base stations share what they learn

KDDI uses distributed reinforcement learning, a method in which several AI components learn from experience and cooperate. Components serving individual base stations calculate and apply their settings. The learning component collects knowledge from those experiences and shares it back across the system. Changing hourly conditions produce substantial amounts of network data, while the parameters offer many possible combinations for the system to consider.[1]

The trial adds a way to compress that input before learning. A model prepared using nationwide base-station data keeps the features needed for adjustment while removing redundant information. Reducing the number of data dimensions is intended to make learning about different hourly conditions more efficient. This preparation supplies the learning process with a smaller representation of the network conditions it needs to handle.[1]

Busy and quiet hours receive different allocations

The setting adjusted in the trial was the signalling capacity assigned to each smartphone. During congestion, the AI allocated what each device needed, providing more opportunities for data transmission and reception. During quieter hours, it left a larger resource margin to keep connections stable. KDDI says relieving shortages of control-signalling resources allowed data-signalling resources to be assigned to phones. These adjustments concern the network’s management of communications as demand changes.[1]

References

  1. News sourceKDDIKDDI plans rollout of AI that adjusts base stations to hourly congestion↩1↩2↩3↩4↩5
  2. News sourceAI WatchKDDI tests AI that changes network settings during busy hours↩
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →