TL;DR
- AFP assigns frequencies and neighbours against an interference objective; ACP tunes physical and electrical parameters against a coverage and quality objective.
- Both are only as good as the constraints you define — unconstrained optimisation produces plans no engineer would accept.
- Run them on a validated model, never on an uncalibrated one, or the optimiser will faithfully optimise the wrong prediction.
What does automatic frequency planning actually optimise?
AFP assigns frequencies and neighbour relations across a cluster to minimise interference while satisfying separation constraints. The inputs are the site database, the antenna configuration, the available carrier list and the constraint set — minimum reuse distance, forbidden adjacencies, BCCH restrictions and so on.
The output is a frequency and neighbour plan plus an interference estimate per cell. The engineering judgement lives in the constraint set: if you allow aggressive reuse, AFP will produce a plan with tight reuse and correspondingly higher interference, and it will be right to do so given what you told it.
| Task | What automation does | What still needs a human |
|---|---|---|
| Frequency assignment | Minimises interference under separation constraints | Setting the constraints; judging acceptable interference |
| Neighbour planning | Builds neighbour lists from predicted overlap | Deciding handover policy and priorities |
| Parameter tuning | Searches azimuth, tilt and power combinations | Defining the objective; reviewing outliers |
| Cluster consistency | Applies one constraint set across all sites | Deciding what the standard should be |
What does automatic cell planning change?
ACP adjusts physical and electrical parameters — azimuth, tilt, antenna height where changeable, transmit power — to meet coverage and quality objectives. It is a search over a large parameter space where each candidate configuration requires a prediction, which is why it is computationally heavy and why model speed matters.
The classic failure is running ACP with a coverage-only objective. It will happily produce a plan that maximises coverage probability while creating pilot pollution and handover problems, because nothing in the objective told it those mattered. Multi-objective constraints are not a refinement; they are the whole point.
When is automation better than manual planning?
Scale is the first answer. A cluster of a few hundred sites has more parameter combinations than a team can evaluate by hand, and the interactions between neighbouring cells are exactly what humans are bad at holding in their heads.
The second answer is consistency. Manual planning produces plans that reflect whoever did the work; automated planning produces plans that reflect the constraint set. For a multi-country rollout where the same standard has to be applied across markets, that consistency is worth more than the marginal creativity of a manual plan.
What makes teams distrust AFP and ACP?
Usually one of three things: an uncalibrated model, an objective function that did not match the real acceptance criteria, or a constraint set copied from a different network. In each case the optimiser does exactly what it was asked and the result is rejected by engineers who can see it is wrong.
The fix is procedural rather than technical. Validate the model first, define the objective against the same KPIs used in acceptance, and review the first output cell by cell before scaling it. Teams that do this tend to keep using automation; teams that skip it go back to manual planning and stay there.