Work

Learning to Empower Surfaces: Machine Intelligence for Reconfigurable Paradigms

Year: 2023
Type: dissertation
Abstract: The goals of the sixth generation (6G) of wireless communication systems are expected to be transformative and revolutionary encompassing applications like data-driven, instantaneous, ultra-massive, and ubiquitous wireless connectivity, as well as connected intelligence. As a new paradigm in 6G, reconfigurable intelligent surfaces (RISs) are able to proactively modify the wireless radio environment to support these new applications and services by changing the electric and magnetic properties of the surface. The feature of RISs is to serve blocked subscribers by varying the channel, which also consumes less energy compared to some communication-assisting technologies like amplify-and-forward (AF) and decode-and-forward (DF) relays. For RIS-assisted networks, machine learning (ML)-based techniques are quite useful to solve important signal processing issues, such as phase-shift optimisation and resource allocation, etc. due to their ability to learn underlying trends between the operating and efficiency parameters for large search space. This thesis provides a systematic analysis of ML-empowered this promising technology, from the basic principles of RISs to its devotion to resource allocation of RIS-aided networks, to apply in the multi-user navigation system and URLLC discussion in Metaverse scenarios based on resource allocation, as well as its emerging advancement on achieving full-space coverage. Firstly, the thesis examines the resource allocation of RIS-assisted wireless networks, which forms the foundation for optimising the performance of RIS-aided networks. By designing algorithms and techniques, the resources can be efficiently allocated, such as reflection coefficients and power levels, for the RIS elements. In this Chapter, a new protocol is proposed, in which the machine leaning-based algorithms are invoked to treat the resource allocation for dynamic scenarios. The users' positions are predicted by the long-term short memory (LSTM) algorithm, user clustering is achieved via a K-means-based Gaussian mixture model (K-GMM) and the sum of communication rate is optimised utilising the deep Q-network (DQN) algorithm. Secondly, once the resources are allocated, intelligent trajectory design becomes important. By considering factors such as task objectives and resource constraints, trajectory planning algorithms can be developed that enable the users to navigate the environment effectively while utilising the empowered surfaces to their fullest potential. In this Chapter, an integrated ML scheme for RIS-aided multi-user navigation systems is proposed, which combines LSTM and autoregressive integrated moving average (ARIMA) models and duelling double deep Q-network (D$^3$QN) algorithm. The LSTM-ARIMA is able to overcome the problem of gradient vanishment of non-stationary and non-linear sequences of channel modelling. For jointly determining the phase shift matrix and users’ trajectories, D$^3$QN is invoked for solving the problem of action value overestimation. Thirdly, the goal is to explore how the empowered surfaces and the RIS-aided multi-user system can facilitate reliable and low-latency communication in the metaverse. Techniques can be investigated to achieve real-time interactions, low-latency data exchange, and high reliability for immersive experiences in the metaverse environment. RISs are capable of expediting the Metaverse service deployment between providers and users. By creating a reflected link between the base station (BS) and the user to build a line-of-sight, communication can be established between the BS and the blocked users. In addition to localising the Metaverse users, the RISs can assist to enhance the performance of key performance indicators in URLLC. A meta-learning-based position-dependent multi-objective soft actor and critic (MO-SAC) algorithm is proposed to approach the Pareto front between the total service cost and transmission latency in Metaverse and to further optimise its latency-dependent reliability. Fourthly and last, to break through the limitation that conventional RISs have maximal 180$^{\\circ}$ coverage, a new concept named simultaneously transmitting and reflecting RISs (STAR-RISs) becomes appealing. In contrast to conventional RISs, STAR-RISs are able to transmit and reflect the incident signal simultaneously, which contributes to full-space coverage. Considering the spectrum efficiency, it is an ultra-interesting question how STAR-RISs perform in terms of coverage and capacity. To solve this problem, a multi-objective proximal policy optimisation (MO-PPO) algorithm is proposed to handle long-term effects, which provides a set of optimal solutions to approach a Pareto front to obtain the optimal balance between the capacity and coverage. Important insights are concluded from the generated simulated results.
Author: Gao, X
Language: English
Open Access status: green
OpenAlex ID: W6992344894
Cites: 107
Cited by:
Related to: 10