Data-driven Transfer Learning Methods for Wireless Networks

Abstract

Radio maps provide information about spatial signal strength and network coverage in a designated geographical area. The estimation of accurate radio maps is necessary to improve the performance of many applications of future wireless networks. For instance, localization, network planning, and resource allocation, to name a few. To obtain accurate radio maps, the exact knowledge of transmitter(Tx) and receiver(Rx) locations can be used. This is known as the location-based method. However, in practice, wireless networks incur a high degree of multipath. As a result, it is difficult to obtain accurate locations of Rxs. Alternatively, time of arrival (ToA) features of radio signals, which are easier to obtain, can be used. This is known as the location-free method. One of the ways to incorporate both methods is the mixture of experts (MoE). Due to changes in the propagation characteristics of wireless networks, a radio map model designed under a particular wireless environment (source environment) can not be directly used in a new wireless environment (target environment). Moreover, designing a new radio map model for each new wireless environment requires a huge amount of measurement samples and may need substantial computational resources and data acquisition costs. To address these issues, in this dissertation, we propose a series of transfer learning (TL) schemes using each of the aforementioned methods to estimate radio maps in new wireless environments where there is a scarcity of measurement samples. To this end, we first train a radio map model in a source wireless environment and then transfer it to another similar but still different target wireless environment. It is then fine-tuned using a small amount of samples of the target wireless environment to reduce the data acquisition cost. For such a scheme, the similarity between two wireless environments controls the effectiveness of the TL operation. Therefore, to quantify the similarity, we investigate different classical similarity measures including the widely used Wasserstein distance. Numerically, we show that these classical measures do not perform well in the context of TL for radio map estimation. To overcome the limitations of these classical measures, we design a data-driven similarity measure (DDS), which is able to capture all the variations of wireless environments and can learn the wireless propagation characteristics directly from the data. Additionally, our DDS can predict the amount of training data needed to estimate radio maps in new target wireless environments when performing the TL operation. Experiments show that our proposed TL schemes perform efficiently with high model accuracy and save a substantial amount of sensor measurement data. Different models are designed for each of the cases of location-based, location-free, and MoE-based radio map estimation. Numerical experiments showcase the performance of each case, respectively. Finally, we investigate the application of TL between two different optimization problems of joint resource allocation (channel assignment and power allocation) in underlay D2D communication. The resource allocation model trained on the dataset obtained from the perfect channel state information (CSI) scenario is transferred to the imperfect CSI scenario and then fine-tuned. The experiment shows that TL improves the performance of the imperfect CSI scenario with less amount of training data.