Uplink Prescheduling Improvement in 5G Networks Using Machine Learning

In a 5G network, the scheduling algorithm at the base station (gNB) decides when and how much data connected User Equipments (UEs) are allowed to transmit. Normally, a UE signals that it has data to transmit and the gNB grants the user permission to send data. Prescheduling is an alternative approach where the gNB speculatively sends grants to UEs that have not requested resources, aiming to reduce user latency. One existing prescheduling algorithm at Ericsson uses a timer to track recent active UEs and applies the Round Robin algorithm to issue speculative grants if there are resources left after prioritizing UEs known to have data. However, this method sometimes sends grants to UEs unlikely to have data, leading to transmissions with no data, increased radio interference, and higher UE battery consumption. A Machine Learning (ML)-based prescheduling algorithm was created to issue speculative grants only if the probability of data presence is above a certain threshold, aiming to reduce latency and lower transmissions containing no data, thus reducing radio interference and UE battery consumption. Since the probability of data in the UEs buffer depends on the user type of originating data, the thesis also explored creating a classification model from UE reports to categorize user types. Two predictive models were developed: one specific to user types and another spanning all user types, both trained with data from an Ericsson 5G network simulator of UE reports at the gNB and at UEs. The time series data was transformed to a tabular format, and the XGBoost ML algorithm was used for training. The model that spans all user types was integrated into the simulator's scheduler, and end-to-end simulations evaluated the ML-based prescheduling algorithm for latency and transmissions sent with no data, comparing it to the timer-based approach. Results show that the user type classifier predicted user types with an accuracy and F1 score of 0.93. The user type specific model predicted data in the buffer with an average accuracy of 0.88 and recall of 0.87, while the model trained across all user types achieved 0.85 accuracy and recall. Results from simulations show that the ML-based prescheduling algorithm could not reduce latency compared to the timer-based. However, results show that the percentage of transmissions sent that contain no data was reduced for the ML-based prescheduling algorithm whilst still achieving lower latency compared to when no prescheduling was used. Specifically for webpages, the ML-based algorithm with a 0.4 threshold achieved a mean object delay from request of 335-370 ms, close to the timer-based method's 320-355 ms, and improved upon the no-prescheduling delay of 447-580 ms. It also reduced transmissions with no data to 35-25%, compared to 56-45% for the timer-based method.