Most signal prediction models are empirical and can be fine-tuned for particular environments. When used judiciously, prediction models will give a useful and well bounded degree of accuracy using the default model coefficients. Furthermore, where operational measurements are available, the prediction model coefficients can "tuned" to maximize accuracy.

All of Overture's built-in prediction models and any compatible prediction models can be manually or automatically tuned against measured data.

Prediction Comparison

Overture compares prediction accuracy on a per-sector basis, but can also aggregate comparison statistics to give a more complete view of prediction accuracy. A single model can be tuned for multiple sectors simultaneously, or all the models in a project can be tuned at once. This offers a significant saving in time and effort for the network engineer.

Measured data is expected to come as a Point data source, where each point contains signal strength information.

Network configuration data (sites, sectors, antenna models) is expected to be up-to-date and representative of the time the measured data was recorded. Note that the quality and accuracy of this data will have a significant impact on the ability of the prediction models to estimate signal propagation.

The following screen shot shows a single sector and data from a single drive test file, where the signal strength has been used to color the point map.

Once both these data sources are present in a project, Prediction Comparison components can be added. From the Start Page, select the Measurements page, which looks like this:

At the top of the page there is a New Comparison Wizard button. Press this to start adding per-sector comparison components.

The Signal Measurements source property should be set to point at the measured data stream (called Drive Data in this example). Once this is done, the signal column can be set:

Ensure that the base station that this drive test data was recorded for is selected in the Radio Selection section.

If the drive test file contains data for multiple base stations, the Server Resolution section must be completed. Server resolution is done by matching a column value from the drive test data (such as CellID) to the corresponding Radio flag value.

Once all the fields are complete, press the Add New Comparison button at the bottom of the form. Ticking the Add Delta Map checkbox will create a new map that is linked to this comparer and allows the user to visualize accuracy on a point-by-point basis.

The New Comparison Wizard remains active in case there are multiple sectors to add comparisons for. After this process is complete the user should return to the Measurements page, where the new comparer(s) will be listed:

The statistics for an individual comparer can be displayed by pressing the Compare button

Sector	Radio 1
Point Provider	Drive Data
Column	RSSI
Total # Binned Points	5726
# Binned Points on Signal Map	5726
Min Point Power	-105.90 dBm
Max Point Power	-44.38 dBm
Min Signal Map Power	-101.89 dBm
Max Signal Map Power	-55.22 dBm
Mean Error	1.98 dB
Standard Deviation	5.30 dB
RMS Error	5.66 dB

The key statistics here are Mean Error, which is the average difference between the measurements and the predictions, and Standard Deviation, which is a measure of the variation. In this example, the predictions are 1.98dB higher than the measurements on average, this is for the Okumura-Hata Presto model with default parameters.

Comparing Multiple Predictions

If a project contains multiple comparisons, pressing the Compare All button on the Measurements page will create an aggregated view of prediction performance.

The following is a partial table of the results from a five site cluster using the Generic Presto prediction model with default parameters.

Radio	Mean Error (dB)	Standard Deviation (dB)	RMS Error (dB)
All	-7.62	8.16	14.66
Site A 1	3.65	8.70	9.45
Site A 2	0.44	9.34	9.35
Site A 3	5.09	10.73	11.87
Site A 4	3.47	7.83	8.57
Site B 1	-2.83	7.33	7.86
Site B 2	-0.63	5.72	5.75
Site C 1	-1.29	8.13	8.23
Site C 2	5.76	7.82	9.72
Site C 3	5.69	9.61	11.17
Site D 2	5.89	11.59	13.00
Site D 3	2.13	7.51	7.81
Site E 2	-18.20	7.61	19.73
Site E 3	-20.25	7.52	21.61

Note that Site E has a significantly worse mean error than the other sites. This is because it is an umbrella site high on a mountain, and is a good candidate for it's own prediction model.

The Generic prediction model is based on the COST231 model, but exposes all the coefficients of the calculation. This makes it more flexible for tuning and allows for a closer fit to measured data.

Automatic Prediction Tuning

Whilst prediction parameters can be tuned manually, Overture provides an easy mechanism to automatically tune parameters and minimize prediction error.

Taking the previous multi-site project as an example, a second prediction model was added to the project and associated with Site E. Similarly a third model was added for Site A, which was significantly more urban than the other sites and likely to have different propagation characteristics. The Auto Tune Start button on the Measurements page was pressed to start the process of automatic tuning.

Note that both the Compare All and Auto Tune functions only include prediction comparisons that have their Active property set. This allows the user to control which models to compare and tune.

The automatic tuning was left to run for 1 hour and brought the average mean error to zero and the standard deviation to 6.72dB as shown below:

Radio	Mean Error (dB)	Standard Deviation (dB)	RMS Error (dB)
All	0.00	6.72	7.18
Site A 1	-1.50	7.62	7.77
Site A 2	-1.70	5.85	6.09
Site A 3	-4.21	10.57	11.38
Site A 4	1.90	5.66	5.97
Site B 1	-3.54	7.84	8.60
Site B 2	-10.00	5.74	11.53
Site C 1	-5.25	6.72	8.53
Site C 2	-2.14	7.31	7.61
Site C 3	-1.11	7.31	7.40
Site D 2	6.35	8.14	10.33
Site D 3	5.33	5.82	7.89
Site E 2	0.87	6.33	6.39
Site E 3	-1.16	6.39	6.50

Where signicant differences still exist (such as those for Site A 3 and Site B 2), the number of data points was very small. Overture weights the average error by the number of data points to avoid overfitting prediction model parameters to an unrepresentative sample size.