Low Risk Antenna Configurations for Mobile Communication Systems: A Safe Reinforcement Learning Method

Reinforcement learning offers an effective framework for antenna angle setting since it enables the autonomous and adaptive tuning of antenna parameters based on continuous interaction with the environment. However, due to the inherent trial-and-error nature of reinforcement learning, the agent may execute unacceptable decisions that reduce the network coverage and result in unbearable network performance degradation. In this paper, we propose a safe reinforcement learning (SRL) method to ensure that the policies made by agents can always maintain network valid coverage ratio above a threshold. The optimization task is formulated as a finite-horizon constrained Markov decision process and a confidence ball is introduced to limit the search scope within a safe range. Numerical results show that our proposed method provides an efficient and low risk scheme for antenna configuration in practical scenarios.