Kisuk Kweon, David M. Gutierrez-Estevez, Joan S. Pujol Roig, and S. Jeong
The evolution experienced by cellular networks towards systems of extremely high-complexity is accelerating the need for technologies enabling network automation. Furthermore, the current trend to embed support for data analytics in the network architecture as well as the rise of Artificial Intelligence (AI) are jointly providing a solid framework to tackle the extraordinarily complex task of achieving full end-to-end network automation. In this paper, we introduce a novel technique to automate the determination of an adaptive inactivity timer value associated with 5G Packet Data Unit (PDU) sessions that may belong to multiple services within the same User Equipment (UE). The proposed approach leverages the standardized Network Data Analytics Function (NWDAF) within the 5G Core (5GC) to train and execute a Deep Reinforcement Learning (DRL) algorithm yielding a solution to the adaptive inactivity timer problem. Performance evaluation shows our proposed technique striking the best balance in the tradeoff between resource waste and number of reconnections when compared to legacy techniques, reducing both UE power consumption and network resource usage.