Cognitive radio networks (CRNs) or self-organizing mobile cellular networks are a promising technology for 5G that manages the spectrum frequency domain more efficiently. At the heart of CRNs is the cognitive engine (CE), which is responsible for decision making on the optimal configuration settings for the CRN in real time if possible. In this paper a novel paradigm for decision making in the CE will be presented called hierarchical random neural networks (HRNNs). The proposed HRNN model decomposes a large complex neural network into a network of loosely interconnected localized subnets, which allow the simplified understanding of network behaviour and also allows the addition of more nodes for long-term memory(LTM). The model can also accurately capture the dynamic nature of the system. Simulation results of the proposed HRNN structure has shown improvements in learning efficiency (based on required execution time for convergent result) in the range of 33% to 35% with reduced computations.