Robust Antenna Parameter Optimization in Cellular Networks via Adversarial Machine Learning

Cell shaping optimizes coverage and capacity in cellular networks by aligning cell coverage areas with the user distribution through antenna parameter tuning. As this distribution changes over time, ensuring the robustness of cell shaping involves (i) maintaining consistent performance across various user distributions and (ii) triggering retraining when the change in the user distribution is significant enough to erode performance. This work proposes a framework that achieves these objectives by leveraging adversarial machine learning techniques and introducing a dissimilarity metric based on antenna gain patterns and path loss. Our proposed framework maximizes network key performance metrics - such as data rate and outage probability - for given user distributions while also ensuring parameter robustness against worst-case scenarios. The effectiveness of the framework is demonstrated using the state-of-the-art 5G network simulator developed by AT&T Labs, revealing a 4.3 dB mean SINR gain and 22.7% sum-log-rate gain for the adversarial layout compared to antenna tuning based solely on the given layout.