Deep reinforcement learning based approaches for capacity sharing in radio access network slicing

Abstract

Network slicing has become a fundamental capability for 5G networks to support the expected high variety of service requirements over a common physical network infrastructure. Each network slice can be customized for a specific application, making that the radio resources have to be accordingly managed by the Radio Access Network (RAN) part of the slice. In this thesis, three different Deep Reinforcement Learning (DRL) based approaches are presented to optimize the resource allocation among slices. A RAN slicing simulator scenario is developed, where the DRL mechanisms build knowledge about the network and learn how to optimize the capacity allocation for each tenant at every moment of time. The performance of each approach is studied based on simulation results, and before the comparison between the algorithms, the set of hyperparameters of each approach is tuned to optimize the learning process.