A Deep Reinforcement Learning Approach to Queue Management and Revenue Maximization in Multi-Tier 5G Wireless Networks

It is envisioned that 5G systems will increasingly leverage on the network slicing concept to meet the demand of diverse services, each tailored for specific user requirements.  In this context, slice admission algorithms that admit slices to the system, that optimize a given objective while ensuring the efficient allocation of resources, are required.  Reinforcement learning has been used successfully to implement optimal slice admission policies.  But as the 5G wireless network becomes more extensive and intricate, the state and action spaces become large.  The efficiency and convergence of reinforcement learning slice admission algorithms is negatively impacted in such a scenario.  To improve on this, deep reinforcement learning, a combination of reinforcement learning and deep learning, has been adopted.  In this paper, a Deep Q-Learning slice admission algorithm is designed; to this end a utility, was developed.  Results show that using the utility as a maximization objective enabled the designed algorithm to (i) optimize the infrastructure provider’s revenue while (ii) providing queue management, in terms of queue length and queue delay