Accurately characterising the performance of black-box RF demodulation pipelines normally
requires dense sweeps over many parameters, incurring high computational cost[1]. Gaussian
process (GP) surrogate models provide smooth interpolants and predictive uncertainty, enabling
efficient exploration[1]. This paper investigates how many targeted runs are needed to reconstruct a smooth performance field and whether the GP uncertainty estimates are well calibrated.
A synthetic two-dimensional performance function serves as the ground truth. We sample N
points uniformly at random, fit a GP to the observations and evaluate the root mean-squared error (RMSE) and mean predictive uncertainty on a dense grid. Results show that modest sample
sizes (N in the tens) yield low RMSE and that predictive uncertainty decays at a similar rate.
Calibration curves illustrate that the GP standard deviation provides reasonably accurate confidence intervals[2]. These findings support using few-shot GP characterisation for deployment
planning in RF systems.