Transfer learning for dynamic RF environments

This paper presents transfer learning algorithms for adapting the radio-frequency (RF) propagation model to changing environments. RF variations are modeled using a Gaussian process (GP) whose hyperparameters capture how the propagation of communication signals varies spatially and temporally. These characteristics of the environment and radio hardware are transferred from one task to another by reusing the hyperparameters. The three transfer learning algorithms presented here have different tradeoffs between model efficiency and training cost amortizations. 12 sets of flight data from different days and using diverse emitter hardware are used to compare the performance of the algorithms. Results show that reusing the hyperparameters can give reasonable performance in terms of root mean square error over a set of validation measurements from the transferred tasks.