You Calculate and I Provision: A DRL-Assisted Service Framework to Realize Distributed and Tenant-Driven Virtual Network Slicing

Deep Reinforcement Learning-Based Policy for Baseband Function Placement and Routing of RAN in 5G and Beyond

“…The results showed the proposed policy outperformed benchmark heuristics in terms of the profit of infrastructure providers. The authors in [28] performance. An online multi-tenant secret-key assignment policy based on DRL was proposed in [29].…”

Admission Control for 5G Network Slicing based on (Deep) Reinforcement Learning

“…Sun et al [24] propose an RA algorithm for vehicular networks based on DRL to provide QoS inefficiently in which the DRL-agent adjusts the [3] RL (Q-Learning) Provider profit [6] Queuing theory Utility rate, admission rate, request waiting time [7] Big Data Analytics Provider profit [8] Heuristic algorithm QoE, resource utilization [9] Heuristic algorithm System resource utilization [10] Heuristic algorithm Provider revenue [11] Knapsack problem Monetization ratio [12] RL (Q-Learning) Provider revenue [13] DRL (DQN) Provider revenue [15] Heuristic algorithm Acceptance ratio and Execution time [16] Heuristic algorithm Acceptance ratio, provider revenue [17] Queuing Theory Running time [18] Complex Network Theory Resource efficiency, acceptance ratio, execution time [19] Integer Programming CPU Utilization [20] Alternating Direction Method of Multipliers Latency [21] Artificial Neural Networks Latency [22] DRL (DQN) Spectrum efficiency and QoE [23] DRL (DQN) Resource Utilization [24] DRL (DQN) Resource Utilization and Satisfaction ratio [25] DRL (DDPG) Provider revenue SARA/DSARA Q-learning and DQN Provider profit, resource utilization, acceptance ratio allocated resources for a slice as a function of the demand. Zhang et al [25] perform slicing in data center networks by using a DRL-agent. The DRL-agent maximizes the revenue from provisioned slices while avoiding overutilized resources.…”

“…The DRL-agent maximizes the revenue from provisioned slices while avoiding overutilized resources. The aforecited papers [15]- [25] focus on mapping NSLs. They map the arriving NSLRs but without controlling admission.…”

“…To the best of our knowledge, no other paper has proposed a solution based on RL and DRL that jointly performs AC and RA, differentiates core and edge nodes, and considers the typical standardized 5G use cases (Table I). Zhang et al [25] perform slicing in data center networks by using a DRL-agent. The DRL-agent maximizes the revenue from provisioned slices while avoiding overutilized resources.…”

“…These papers propose making admission decisions considering individual NSLRs, which can lead to sub-optimal decisions since more profitable requests arriving in the short term can be rejected due to the unavailability of recently allocated resources [14] Moreover, most of these proposals neither consider the QoS requirements of different service types (use cases) nor the allocation of resources in the 5G core network. In addition, various other solutions based on heuristics [15], [16], Queuing Theory [17], Complex Network Theory [18], Linear Programming [19], Alternating Direction Method of Multipliers [20], Artificial Neural Networks [21], and DRL [22]- [25] have considered only RA and neglected AC, although this prevents the achievement of NSP goals. On the other hand, jointly performing AC and RA in an intelligent way can considerably improve the achievement of target objectives.…”