A mobile operator wanted more capacity out of the 4G+ network it already had. The levers were physical - which way the antennas point, how wide they spread - and the evidence was millions of signal measurements reported by ordinary handsets.
Tilt one antenna and you have changed its neighbours
Adding capacity by building sites is slow and expensive, and often the capacity is already there and misaimed. The obstacle is that tuning a network is not a local decision: tilt an antenna to serve one area better and you have changed what its neighbours receive, usually as interference. Two constraints pull against each other. Coverage wants reach, and capacity wants cells that do not talk over each other. Optimising either one alone produces a worse network.
Working from what the handsets already report
Every subscriber device continuously reports how well it hears the network - signal power and signal quality, RSRP and RSRQ. Those reports arrive in the millions, they come from where users actually are rather than where a planning model assumes they are, and they are already being collected. That is a rare combination: a ground-truth measurement of the thing you are trying to improve, at the density of the real user population, at no acquisition cost. The work was making it usable at that volume.
What the models produce
Not a score. A list of specific antenna reconfigurations on specific base stations, each with the modelled consequence attached: what it does to interference, what it does to capacity, what it does to the neighbours. SINR modelling and antenna patterns are what let a proposed change be evaluated before anybody climbs a mast, and the reason that matters is that the alternative - try it and watch - is a change to a live network carrying real traffic.
The part that decides whether any of it lands
A recommendation that cannot be executed is a slide. The output had to arrive inside the operator's own network management systems, in the form those systems accept, or it would have been a study rather than an optimisation. This is routinely the least interesting part of a project like this and the one that determines whether it changes anything.
The two builds that skip the physics
- Drive-test the network
- A vehicle with measurement equipment produces clean, trustworthy readings of the roads it drove, at the hours it drove them. Users are not on the roads at those hours — they are indoors, on upper floors, in the places a car cannot go. Optimising against drive-test data improves the network where nobody was standing.
- Train a model to output the configuration directly
- End to end, measurements in and tilt settings out, is the shape this looks like from a distance. The training signal does not exist: nothing in the data says what would have happened at a different tilt, because that network was never run. And a recommendation with no reason attached does not get a crew sent to a mast.
The same shape as the first days of a build, on a bigger dataset
Somebody had a large amount of production data, a suspicion that something was recoverable in it, and no arithmetic either way. The work was turning that into named changes with numbers attached. That is the same shape as the first days of a build, on a longer timescale and a bigger dataset.
What it proves
We can dig through millions of real measurements from a production system and come back with concrete, quantifiable levers to tune.
Category
Consulting
Built with
Machine Learning · Network Optimization · SINR Modeling · Antenna Patterns · RSRQ Analysis · Telecommunications
Need something similar?
The cheapest way in is two weeks. The first days work out which task would pay for itself in your processes; the rest builds that agent on your own data and measures it.