Machine learning based numerology assignment methods for 5G and beyond

5G and beyond (5GB) systems have flexible capabilities to meet different user communications requirements in parallel. The design of multi-numerology waveform is one of the key aspects of this flexibility. Different subcarrier spacing of a waveform have formed numerology structures, and an important capability has been gained to meet the communications requirements by changing the subcarrier spacing under the same coverage area for different users with 5G. However, while the communications needs of different users can be met simultaneously with multiple numerology structures, new problems have arisen due to interference between numerologies. Therefore, it is necessary to plan and manage multiple numerology structures appropriately. In this study, the Joint Sensing and Communications (JSAC) concept is applied with Machine Learning (ML) techniques to develop an intelligent numerology control mechanism for 5GB. Thanks to JSAC capabilities, an environment-aware system is designed and various Useful Sensing Information (USI) that affects the wireless channel characteristic is used. It is assumed that USI is achieved through 5GB smart city networks. At this point, synthetic data generation is carried out to form new datasets with USI-based features. Using the created datasets, three novel methods for numerology planning for 5GB are designed. In the first of these methods, the priority of the service type needed in a coverage area is determined. The second method makes a general decision in the same coverage area and determines the recommended subcarrier spacing parameter. Finally, in the third method, the ultimate numerology allocation that should be assigned to the user is realized. Different synthetic datasets and ML models are created for each of the novel methods developed in this thesis. Moreover, multiple ML algorithms, including ensemble learning methods, have been used for different ML models, and various performance results based on simulations are presented. Looking at the results, the use of ML methods for numerology planning in 5GB seems promising.