Dynamic radio resource management for radio access networks by using machine learning techniques

In the next generation networks era, telecommunication area will serve any device that would leverage from being connected to environment such as household appliances, massive industrial machines, remote surgery robot, automated vehicles, etc.. Hence, the spectrum that has been used up until now, gets more crowded and not suficient to serve all these requests. In LTE which is current communication technology uses the 700 MHz - 2.7 GHz range of the spectrum where in 5G networks, 450 MHz - 6 GHz domain and 24.25 GHz - 52.6 GHz range frequencies is planned to use. In this thesis, radio admission and resource allocation process in the downlink channel of a next generation network is discussed. With the vertical industries' needs from the 5G networks, supporting all mobile service users who have different Quality of Service (QoS) requirements becomes the main challenge. To manage and satisfy the heterogeneous requirements, splitting the network into slices which have different properties (e.g., bandwidth requirements, delay tolerance, user density, etc.) can be a solution over a common physical infrastructure. Network slicing concept allows to manage and schedule the requests under the constraint of limited resources in an optimal way. The expected behavior from the admission mechanism to learn to serve the user requests which have low latency requirement with high priority. While the network is admitting these high priority users and allocate them required resources, it must provide the maximum spectrum efficiency. The problem is set on a single-cell scenario. Deep Q-learning algorithm is used as a solver to tackle this highly complex radio resource management problem instead of traditional convolutional optimization algorithms.