On the Transferability of Tilt-Dependent Radio Map Prediction Models
09 March 2018
The fifth generation of wireless networks (5G) is expected to address many of the challenges present in current architectures by improving the quality of service, reducing latency, adding more capacity users and devices being connected to the network, and reducing power consumption, among others. In this complex scenario, Machine Learning and, in our case, Transfer Learning are promising tools; reducing the need for collecting huge volumes of labeled data, which could be expensive for operators and end users. In this paper, we apply Machine Learning and introduce the idea of Transfer Learning for Tilt Radio Map prediction by developing two different approaches. One based on spatial correlation and the other on the traditional Path loss model, using real data recorded in a 5G testing scenario. We evaluate the suitability of both methods to predict the Radio Map based on the domain similarity between the source and target domains. Finally, we conclude that given enough source labeled data, we can predict the Radio Map for new PCIs and antenna tilts under the assumptions that domains are relatively similar. We propose solutions to make the domains similar and show further research directions.
