Towards Automated RAN Configuration Tuning in Cellular Networks with Causal Learning

Adaptive RAN configuration tuning is key to effective and efficient operation of cellular networks. This task is however very challenging due to the large number of diverse and inter-dependent RAN configuration parameters. Consequently, current practice is still dominated by manual tuning with data-driven approaches limited to serving as an aid. These existing approaches also lack inter-pretability and are limited in their generalizability and transfer learning capability. Motivated by the above, we introduce a causal learning driven approach as a first step towards automated RAN configuration tuning. Specifically, we propose a tailored causal model to isolate the effect of configurations on RAN performance. This model not only is interpretable but also allows predicting RAN performance outcomes under unseen configurations as well as transferring causal learning from data-rich to data-sparse regions. Evaluations of this framework with real-world RAN configuration and performance data from over one million cells across six months demonstrate its effectiveness. Our results show that the proposed causal learning based approach can provide reliable estimation of RAN configuration effects on network performance, outperforming prior methods by 6x.