As Chief Data Officer for Rakuten Mobile, I’ve worked with my team to build and deploy AI models that have been deployed and fine-tuned to optimize and monetize our network.
Our Remote Electrical Tilt (RET) model has been live in Japan for two years, successfully maximizing coverage and capacity by increasing spectral efficiency. We have achieved 25% spectral efficiency improvement and raised handover rates from 97% to 99%.
Our success is due in part to the granularity of real-time data processed. We are capturing data points at millisecond levels across all the nodes in the network, configuring antenna elements as needed.
Every day is different. New base stations are continuously added and subscriber demand fluctuates. A Sunday in downtown Tokyo has drastically different traffic requirements than a Monday. A bank holiday is different from a normal working day.
Even while our subscribers average more than 23 GB of usage per month, our AI model helps us maintain competitive data rates using just one frequency band while we compete with operators using multiple. Still, our subscribers enjoy a more stable and reliable user experience.
From once per week to every day with AI
RET is an important function of any mobile network. Typically, adjustments are made every seven days and this was generally best practice in our network before implementing AI-powered management.
The lack of detail in the data driving these weekly change decisions was evident. We didn’t always understand what was happening in the network and offline processing did not allow us to respond effectively to the daily challenges being presented.
It is an entirely new world with AI modeling now making daily RET adjustments possible at Rakuten Mobile. We can be much more responsive to real-time data, allowing the network to adapt faster to changes like unexpected traffic patterns and subscriber density.
Let’s break down our implementation process:
- Real-time data collection. Our model continuously collects and analyzes performance metrics, user data and interference stats from each network cell with millisecond granularity.
- Dynamic antenna adjustment. Following data analysis, the model dynamically adjusts the antenna tilt with a focus on optimizing network coverage and minimizing interference.
- Learn and improve. Upon introduction, we immediately saw a 17% improvement on spectral efficiency. The reinforcement deep learning algorithm we deployed incorporates network feedback for rapid reconfiguration that leads to ongoing performance improvements. We have since doubled our spectral efficiency as a result of these improvements.
The model is playing an outsized role in our network management as it constantly increases or decreases areas of coverage. This is especially important given the difficulty imposed by the terrain in certain areas of Japan or densely built-up areas, which can affect signal distribution.
As an example, to address challenges posed by Tokyo’s high-rise buildings, RET can avoid signal overshooting or undershooting. We are able to optimize coverage with minimal user interference at any given data point.
We can better manage interference between cells by adjusting the antenna tilt to reduce overlapping signals, which tend to degrade performance. By targeting uplink and downlink interference, we’re able to keep performance high.
To reiterate, spectral efficiency has improved by 25%. The result is improved capacity as more data can be transmitted via the same spectral resource. A 2% improvement in handover rates means a vastly better customer experience. Optimized infrastructure and reduced physical adjustments means a reduced reliance on expensive equipment upgrades and manual labor all while improving cost efficiencies.
Proven in Japan, with global applicability
Our patented RET AI model is continuously being improved and can be applied to any operator’s network. Specifically, settings for individual cells can be customized based on unique requirements and subscriber behavior. Increasing spectral efficiency directly improves the operational efficiency of the most valuable and usually scarce asset a mobile operator has: licensed space for radio waves.
Beyond the use cases highlighted here, our RET AI model can be tuned for specific applications, such as disaster recovery. For example, if a cell goes down, the model can reconfigure neighboring cells immediately to provide coverage without increasing interference. This delivers a much-needed network resiliency boost for service continuity in emergency scenarios.
In greenfield situations, special roaming algorithms can be employed to more strategically deploy new cells, focusing on ROI and reduced construction costs while maximizing coverage. Millions of dollars in roaming fees can be saved via strategic expansion based on advanced data analysis.
In Japan, we’ve shown how improved strategic planning can accelerate site rollout, hitting population coverage targets years ahead of schedule.
When it comes to site planning, the algorithm informs cell deployment decisions with a level of granularity down to a 50 square meter area. This is leading to unprecedented precision in planning. Any operator can benefit from improvements like this.
One way to predict how the RET AI model will perform in your network is to use a digital twin to validate the impact of changes and adjustments in a controlled virtual environment. With more sophisticated testing capabilities comes confidence to surgically deploy AI one model at a time.
.svg)
