TL;DR
- ACP searches azimuth, tilt, height and power against a composite objective; it cannot optimise anything the objective does not mention.
- A coverage-only objective produces plans that look excellent on a prediction map and fail on accessibility and handover KPIs.
- Below roughly fifty sites, manual tuning with a good prediction map is usually faster than configuring and reviewing an ACP run.
What does automatic cell planning actually change?
ACP is a constrained search over the physical and electrical parameters of a cell: azimuth, mechanical tilt, electrical tilt, antenna height where the tower allows it, transmit power, and in some implementations antenna type selection from a permitted set. Each candidate configuration requires a fresh prediction across the cluster, which is why ACP runtime is dominated by propagation model speed rather than by the optimiser itself.
The output is a parameter set per cell plus a predicted performance delta. What the optimiser cannot do is decide what "better" means. That definition — the objective function and its constraints — is the entire engineering contribution, and it is usually written in a few lines by whoever configures the run.
| Parameter | Typical search range | Real-world limit that shrinks it |
|---|---|---|
| Electrical tilt | 0° to 10° in 1° steps | Remote electrical tilt units may not be installed on older sectors |
| Mechanical tilt | 0° to 8° | Tower-sharing contracts and rigging cost; often frozen for the contract term |
| Azimuth | ±30° from design | Structural mounts and planning permission on rooftop sites |
| Antenna height | ±5 m | Only where a free position exists on the tower; rarely available in practice |
| Transmit power | Full range per port | Regulatory EIRP limits and neighbour-cell fairness |
How do you write an objective that survives acceptance?
Acceptance criteria are written in KPIs, so the objective should be written in KPIs. A composite objective typically combines a coverage term (percentage of pixels or bins above a received-power threshold), a quality term (SINR or RSRQ distribution), an overlap term (fraction of area served by more than three cells above a threshold), and a penalty term for cells that fall outside permitted parameter ranges.
The weights matter less than the presence of every term. An optimiser given only a coverage term will raise power and reduce tilt until coverage is maximal, producing a plan with heavy overlap, poor SINR in the centre of the cluster, and a handover count that the field team will reject. Adding an overlap penalty changes the result more than any weight tuning.
When is ACP worth the compute?
Below roughly fifty sites, a competent RF engineer with a good prediction map and a tilt table will usually reach an equivalent or better plan faster than the time it takes to configure, run and review ACP. The optimiser's advantage is combinatorial: it holds the interaction between every neighbouring cell pair simultaneously, which a human cannot do beyond a handful of sites.
Above a few hundred sites the advantage is decisive, and the constraint becomes review capacity rather than compute. That is why mature teams run ACP on clusters of 100 to 300 sites, review the outliers, and then extend the accepted parameter standard to adjacent clusters rather than optimising the whole network in one run.
What goes wrong most often?
Four failure modes account for nearly every abandoned ACP deployment. The first is an uncalibrated propagation model: the optimiser faithfully optimises a prediction that is 8 dB optimistic, and the field result is uniformly worse than predicted. The second is an objective that omits overlap or interference. The third is a constraint set copied from a different network with different site density and clutter. The fourth is reviewing only aggregate KPIs instead of cell-by-cell output, which lets a handful of badly degraded cells hide inside a good average.
How do you review an ACP result before rollout?
Sort cells by predicted degradation, not by predicted improvement. The cells that got worse are where the objective and reality disagree, and they are the only places where a human judgement changes the outcome. Check whether each degraded cell had a constraint that forced the change, and whether that constraint reflects the actual contract or a default value nobody edited.
Then check the parameter delta distribution. If ACP moved two hundred cells by more than three degrees of tilt, the input site database probably contains a systematic error — wrong antenna height, wrong antenna model, or a clutter class assigned to the whole city. A large uniform correction is almost always a data problem expressed as a parameter change.