| M. Trigka, E. Dritsas, Ph. Mylonas |
| Learning Surrogate Models for Wireless Channel Descriptors from Low-Dimensional Spatial Features |
| 14th EETN Conference on Artificial Intelligence (SETN 2026), September 9-11, 2026, Chania, Greece |
|
ABSTRACT
|
| Accurate characterization of millimeter-wave (mmWave) wireless channels in fifth-generation (5G) and beyond systems typically requires high-dimensional channel state information (CSI) and a significant measurement overhead. In this study, we propose a data-driven surrogate modeling approach that predicts low-dimensional, physically interpretable channel descriptors from spatial and propagation features without relying on explicit CSI. Using the Deep-MIMO dataset, we estimated the channel energy and root mean square (RMS) delay spread as a supervised learning task, evaluating multiple regression models under baseline and angular-enhanced feature representations. The results show that although spatial features alone suffice for accurate energy prediction, angular descriptors significantly improve delay spread estimation, highlighting the importance of geometric propagation information. In particular, Random Forest (RF) effectively captures the nonlinear relationship between the spatial context and channel behavior. Generalization analysis across different spatial segments and base station (BS) configurations reveals that performance depends on distribution similarity, indicating both the potential and inherent limitations of low-dimensional surrogate models for wireless propagation.
|
| 09 September, 2026 |
| M. Trigka, E. Dritsas, Ph. Mylonas, "Learning Surrogate Models for Wireless Channel Descriptors from Low-Dimensional Spatial Features", 14th EETN Conference on Artificial Intelligence (SETN 2026), September 9-11, 2026, Chania, Greece |
[
BibTex] [
Print] [
 Back] |
