Abstract
The cooperative network, e.g. relay network and RIS network, is known to improve significantly the reliability in wireless communications. However, the upcoming Sixth Generation (6G) wireless networks requires not only high reliability, but also low-latency and high security. This challenge brings high-complexity in designing a proper optimization for future cooperative networks. As an emerging technology, deep learning plays an important role in solving complicated optimization problems in wireless communications. Compared with the conventional approaches, deep learning has remarkable power to deal with complicated optimization problems with low-complexity. Therefore, this thesis utilizes deep learning technologies to design the proper resource allocation such as relay selection, reconfigurable intelligent surfaces (RIS) coefficients optimization, power allocation and hybrid orthogonal multiple access (OMA)/non-orthogonal multiple access (NOMA) selection in cooperative networks.
First, to maximize the throughput with delay and secrecy constraint, deep reinforcement learning (DRL) is used to optimize the relay selection strategy in buffer-aided relay networks. To further improve the outage performance, a novel decision-assisted DRL method is proposed based on the a-priori information in the buffer-aided relay system. Compared with traditional DRL methods, the proposed novel algorithm reduces the exploration dimension and the impact of bad actions in the training. Moreover, the proposed algorithm can be used to solve more complex problems such as joint relay selection and power allocation, joint relay selection and OMA/NOMA selection in cooperative networks.
Second, this thesis investigates the performance of deep learning-based RIS optimization in RIS-assisted cooperative networks. To solve the low convergence efficiency problem in DRL-based RIS optimization methods, a novel deep cascade correlation learning (DCCL) method is proposed to optimize the RIS coefficients. Compared with DRL methods, the DCCL has higher sample efficiency to find a solution.
Third, both DRL and DCCL are considered in the optimization for hybrid relay and RIS networks. The DCCL can solve the optimization problem with higher sample efficiency in hybrid relay and RIS networks, compared to the DRL. On the other hand, the DRL-based method achieves superior secrecy performance with delay constraint in buffer-aided cooperative networks as it learns to maximize the long-term benefit.