Current students


ZAMBIANCO MARCOCycle: XXXIV

Section: Telecommunications
Advisor: VERTICALE GIACOMO
Tutor: CESANA MATTEO

Major Research topic:
Deep Reinforcement Learning for Inter-numerology Interference Minimization in 5G RAN Slicing

Abstract:
The research activity focuses on the design of spectrum allocation policies for 5G network slicing on the radio interface. The analysis differentiates from the work in this field in two key aspects that we believe are worth studying to improve the effectiveness RAN spectrum slicing policies in practical scenarios. In most of state-of-the-art work, the design of radio resource allocation algorithms for 5G network slicing is treated separately from other technologies that characterizes 5G systems. In addition, slice isolation, that expresses the concept of making the performance of each network slice independent one from the other, has received little attention from the research community, despite being of critical importance for radio interface. As a matter of fact, the allocated spectrum to each service is affected by the channel propagation condition and external interference that hinder the service provisioning quality. 

Following these observations, we filled the gap in the research analysis by designing spectrum allocation policies for mixed-numerology network slices. Differently from conventional OFDM scheme, mixed-numerology schemes allows to tune the subcarrier spacing on the physical layer in order to tailor the transmission performance with respect to the network service. However, such flexible physical layer design causes the loss of the orthogonality between subcarriers of different services, thus leading to inter-numerology interference (INI). For this reason, an INI-aware spectrum allocation policy are required to enhance each network slice performance by mitigating such unwanted interference effect. 

To overcome the computational complexity of the considered resource allocation problem, we leverage deep reinforcement learning theory to design agents capable of autonomously learning an effective spectrum slicing policy according to the wireless channel status, inter-numerology interference and service requirements of each network slice. In particular, we also investigate the practical deployment of such agents by designing suitable schemes to increases the agent learning efficiency and scalability.