A Machine Learning based Framework for KPI Maximization in Emerging Networks using Mobility Parameters
This work addresses network management challenges for telecom operators in 5G networks, though it is incremental as it combines existing methods for a specific optimization task.
The authors tackled the problem of manually tuning thousands of configuration parameters in emerging 5G networks by proposing a machine learning and genetic algorithm framework to optimize mobility parameters, achieving a three orders of magnitude faster convergence time compared to brute force.
Current LTE network is faced with a plethora of Configuration and Optimization Parameters (COPs), both hard and soft, that are adjusted manually to manage the network and provide better Quality of Experience (QoE). With 5G in view, the number of these COPs are expected to reach 2000 per site, making their manual tuning for finding the optimal combination of these parameters, an impossible fleet. Alongside these thousands of COPs is the anticipated network densification in emerging networks which exacerbates the burden of the network operators in managing and optimizing the network. Hence, we propose a machine learning-based framework combined with a heuristic technique to discover the optimal combination of two pertinent COPs used in mobility, Cell Individual Offset (CIO) and Handover Margin (HOM), that maximizes a specific Key Performance Indicator (KPI) such as mean Signal to Interference and Noise Ratio (SINR) of all the connected users. The first part of the framework leverages the power of machine learning to predict the KPI of interest given several different combinations of CIO and HOM. The resulting predictions are then fed into Genetic Algorithm (GA) which searches for the best combination of the two mentioned parameters that yield the maximum mean SINR for all users. Performance of the framework is also evaluated using several machine learning techniques, with CatBoost algorithm yielding the best prediction performance. Meanwhile, GA is able to reveal the optimal parameter setting combination more efficiently and with three orders of magnitude faster convergence time in comparison to brute force approach.