Temporal Feature-Enhanced Deep Reinforcement Learning for RAN Slicing With User Mobility

Radio access network (RAN) slicing is driving the growth of 5G verticals. A large number of mobile user equipment (UE) increases the dynamicity of service requests. However, in different 5G scenarios, the mobility of UEs can easily lead to disordered and fluctuating traffic, thus yielding latency and rate. It brings challenges to effective inter-slice resource allocation. In this letter, we construct a temporal feature-enhanced deep reinforcement learning (DRL) framework for resource allocation in RAN slicing, considering the multi-transformer-encoder (MTE) and an improved advantage actor-critic (A2C), namely MTE-A2C. Here, MTE can extract temporal features of multi-scenario service traffic via powerful capture of data relationships. Such a feature can provide a strong time-related environment state for the latter A2C resource allocation. Specifically, to improve the efficiency of system training, we make the policy and value networks share the same loss function for gradient updates. Finally, the superiority and universality of MTE-A2C are verified by comparing experiments under different mobile scenarios.