Less is More: Improved Path Loss Prediction Using Simple Interpolation Models

Accurately predicting path loss in wireless channels is essential for effective network planning, particularly in dynamic spectrum scenarios where users must minimize interference to other devices. We evaluate multiple path loss prediction methods on three large city-scale datasets, testing out-of-distribution predictions to assess the robustness of each technique. Surprisingly, interpolation methods outperform computationally intensive physics-based techniques and complex machine learning (ML) models. In our evaluation, a radial basis function interpolator featuring a linear kernel emerges as the most accurate method due to utilizing local information while producing a smooth interpolation surface for OOD regions. Even a simple linear interpolation model is more accurate on average than a sophisticated physics+ML method, demonstrating the efficacy of simpler approaches in practical wireless channel prediction.