Integrating Knowledge-Driven RL into the 5G Scheduling

In the fifth-generation (5G) cellular networks, this paper offers a knowledge-assisted deep reinforcement learned (DRL) technique for developing wireless schedulers, specifically for time-sensitive traffic. Traditional scheduling policies may be optimised using deep probabilistic policy gradient (DDPG) since they are probabilistic mapping from route and buffer states for arranging actions. A simple adoption of DDPG, however, demonstrates delayed convergence and poor quality-of-service (QoS) effectiveness, making it unsuitable for actual, mobile 5G systems. We put forward a proposed DRL approach that uses theoretical communication models to create a Markov process of decision-making to solve these issues. We create a knowledge-assisted DDPG (K-DDPG) that takes into account expert knowledge about scheduler fashion, including QoS standards, target scheduling policies, and the significance of each example of training using the function's proximity errors and packet loss counts, in order to improve convergence speed and QoS for every user. In order to overcome the disconnect between offline sims and real-world situations, we also build an architecture for online instruction and inference where K-DDPG initialises the scheduler offline and optimises it online to respond to non-stationary real-world systems. Simulation findings show that our method outperforms conventional schedulers and considerably cuts down on DDPG's convergence time, resulting in a 30% to 50% reduction in packet losses. According to findings from experiments, our technique delivers higher initial QoS than randomly starting with offline beginning, and online tweaking converges quickly.