Latent Learning-Based Intelligent Resource Allocation for Dynamic Spectrum-Sharing Networks

Resource allocation is paramount to improve spectral efficiency in spectrum-sharing networks. However, numerous existing resource allocation schemes, especially those based on deep reinforcement learning techniques, overlook the impact of time-variant channel quality caused by high dynamics of wireless environment and heterogeneous action space due to discrete actions and continuous parameters, which may significantly degrade the desired system performance. To tackle these issues, in this paper, two intelligent resource allocation schemes that can jointly optimize channel allocation and transmit power in a dynamic spectrum-sharing network are proposed. In particular, an intelligent framework, enhanced by channel prediction, is first proposed to capitalize fully on the latent evolutionary characteristics of time-varying channels, facilitating efficient resource allocation design. Subsequently, a hybrid action representation-based intelligent framework is proposed to learn the latent dependence between channel allocation and transmit power for each secondary user. Simulation results demonstrate that our proposed schemes achieve superior performance compared with several benchmark schemes, highlighting that the sum rate can be improved by exploiting latent channel characteristics and latent hybrid actions dependence.