This work studies robust learning in dynamic wireless environments. Modern wireless data networks are complex and modeling their behavior accurately is difficult. As a result, machine learning and artificial intelligence have been making significant inroads into wireless networks. State-of-the-art machine learning algorithms (e.g., neural networks) assume a stationary learning environment and they generally require large training sets. However, modern radio access networks (RANs) are dynamic, and by the time a large training set is collected the environment may have changed so much as to render the learning useless. Therefore, in dynamic networks learning frameworks must work with small training sets. Assuming that each training sample is informative, the lack of a large training set results in uncertainty about the underlying phenomenon/function to be learned. In light of these facts, we study "hybrid" learning approaches in which the above-mentioned uncertainty is combated by the inclusion of model based prior knowledge in the proposed learning frameworks. In Chapter 2 we study cell-load approximation in RANs using a small sample set and a robust learning framework armed with model based prior knowledge. To this end, we study the nonlinear load-coupling model and prove some salient properties of cell-load as a function of user rates. We show how this prior knowledge can be used to decrease the uncertainty resulting from a small training set. In Chapter 3 we study robust multiuser detection in dynamic wireless networks in which users transmit sporadically. Though it is known that the optimal multiuser detector is nonlinear, learning this detector using conventional methods requires a large number of training samples. Additionally, all nonlinear detectors are sensitive to small changes in the environment. In modern wireless applications, such as machine-type communications, users transmit sporadically and as a result performance of nonlinear detectors may deteriorate. To address this issue, we propose a novel online learning framework that combines the expressive power of a nonlinear filter with the robustness of a linear filter. The proposed "sum filter" is designed in a reproducing kernel Hilbert space (RKHS) constructed by taking the direct sum of an RKHS associated with a linear kernel and an RKHS associated with a nonlinear kernel. We derive the nonlinear kernel from the multiuser detection model by exploiting the connection between the optimal nonlinear filter and certain RKHSs. Working in RKHSs and, in general, Hilbert spaces allows for low-complexity projection based algorithms which are well-known for their robustness to noise and their numerical stability. In Chapter 4 we use the celebrated projection onto convex sets (POCS) technique to learn probability density functions (pdfs) in a Hilbert space. Here again, we combine a small training sample set with prior knowledge based on general properties of pdfs. We then show how to apply our learning framework to distributed multiuser detection in a cloud RAN network.
Robust learning in wireless networks.
Wirksamkeit von Modellen und Vorkenntnisse beim Lernen aus kleinen Datensätzen
efficacy of models and prior knowledge in learning from small sample sets
Robustes Lernen in drahtlosen Netzwerken
2021
Sonstige
Elektronische Ressource
Englisch
DDC: | 629 |
Europäisches Patentamt | 2021
|Wissensbasierte Plausibilisierung von Datensätzen
Tema Archiv | 2005
|Erheben von fahrzeugbasierten, ortsbezogenen Datensätzen
Europäisches Patentamt | 2020
|ERHEBEN VON FAHRZEUGBASIERTEN, ORTSBEZOGENEN DATENSÄTZEN
Europäisches Patentamt | 2021
|Europäisches Patentamt | 2021
|