Reinforcement Learning based Adaptive Access Class Barring for RAN Slicing in 5G Networks

Abstract

Machine-to-Machine (M2M) communication is one of the major drivers of 5G networks as M2M traffic might soon surpass Human-to-Human (H2H) traffic. Network slicing is a promising technique for supporting M2M traffic on 5G networks as there is a need to concurrently support varying Quality-of-Service (QoS) requirements of M2M devices. A major bottleneck for M2M traffic is the Random Access Channel (RACH) procedure, which has to be performed for all devices, which results in the same latency for all service types. Due to the event-driven simultaneous access behavior of M2M devices, this procedure can cause severe congestion. Legacy congestion control schemes such as Access Class Barring (ACB) are not adequate to handle the overload in bursty traffic scenarios, which can happen frequently in M2M communications. There is also no clear guideline to adjust ACB parameters dynamically in such situations. Here we propose a multi-rate ACB algorithm using Reinforcement Learning (RL) to tune the barring rates and barring times of different service classes. Our priority-based algorithm not only reduces the congestion but also slices the RACH among different service types. Comprehensive simulation results show that our proposed algorithm maximizes the RACH utilization. In the meantime, based on each service priority, it reduces the delays and increases the access probability even when the connection requests exceed the RACH capacity.