RETRACTED: QoE-Driven Resource Allocation for SUs with Heterogeneous Traffic using Deep Reinforcement Learning (Retracted Article)

Quality-of-experience (QoE) driven resource allocation play a huge role in the deployment of cognitive radio networks (CRNs) towards the 5G era. Cognitive approaches need to exhibit a certain degree of intelligence in order to draw optimal decisions when dealing with heterogeneous traffic. However, with current methods such as reinforcement learning (RL), efficient exploration of the state space still remains a major challenge. The epsilon (epsilon)-greedy exploration strategy suffers from poor convergence since it does not carry out deep exploration. This consequently results in more convergence cycles or exponentially large data requirements. The docitive paradigm, which is an extension of the cognitive paradigm, performs well using RL but it becomes unfeasible when the network becomes distributed with a large number of nodes. Thus, it becomes a huge challenge to perform user experience evaluation of multimedia services on-the-fly since this requires the collection and correlation of a mixture of variables on network conditions, the service, as well as the user. As our first step towards addressing this problem, we perform resource allocation using the mean opinion score (MOS) for heterogeneous traffic. We then demonstrate that deep Q-networks (DQNs) can combine deep exploration with deep neural networks (DNNs) for exponentially faster learning and convergence. Our results indicate that docitive approaches have a 55% better convergence rate compared to individual learning at 5 SIN and 70% better at 42 SUs. Also, perfect docition performs 6.67% better than the other approaches when the number of SUs reaches 32.