Safe RL for Core Network autoscaling

Core Network autoscaling is a critical scalability challenge, where traditional methods and existing advanced Reinforcement Learning (RL) approaches often overlook safety considerations leading to exploration of unsafe actions during training. This paper explores Safe Reinforcement Learning (Safe RL) as a solution to tackle these safety challenges. This paper formulates autoscaling as a Safe RL problem to integrate safety constraints. Then autoscaling methodologies are developed by optimizing Safe RL algorithms through delicate function design and hyperparameter tuning. Finally, Safe RL models are evaluated on a Gym-based simulation environment and the Magma Open-Source Core Network platform. Results confirm that Safe RL models have a safer behavior during exploration and can help real-world applicability of RL.