Location-To-Channel Mapping Using Model-Based Learning

Abstract

Accurate channel estimate is crucial for modern communication systems to ensure dependable and efficient delivery
of information. A neural network may be used to convert the user's geographic coordinates into the coefficients of
the communication channel, as this information is highly dependent on the user's position. These latter, however, are
changing location-dependently at a rate on the order of wavelengths. Because classical neural networks have a
spectrum bias that makes them better at learning low-frequency functions, learning such a map is quite challenging.
To get over this issue, this research introduces a cheap model-based network that divides the target mapping
function into low-frequency and high-frequency components. The result is a hypernetwork design in which the
neural network learns sparse coefficients with low frequencies from a lexicon of components with high frequencies.
On realistic synthetic data, the suggested neural network outperforms conventional methods, according to the
simulation findings. Index Terms—Information Modeling, Channel Estimation, Spectral Bias, and Implicit Neural
Representations