xApp Distillation: AI-based Conflict Mitigation in B5G O-RAN

The advancements of machine learning-based xApps in Open Radio Access Network (O-RAN) have created research and industrial opportunities. One of the major advantages of Machine Learning (ML)-based xApps over heuristic methods is the ability to learn the dynamics of the environment and predict upcoming instances. Typically, xApps are trained and fine-tuned for specific objectives. However, telecommunication companies often deploy multiple xApps in overlapping areas. Given the different design objectives of xApps, this deployment strategy can lead to conflicts. Current conflict mitigation schemes proposed by the O-RAN Alliance are rule-based, either ignoring some of the xApps or rolling back their actions. This leads to performing the same action for a different state of the network, resulting in suboptimal mitigation. To prevent this suboptimal mitigation, we propose the xApp distillation method. The proposed method distils historical network state, actions and outcome information from multiple xApps (either heuristic or ML-based xApps) and uses this knowledge to train a single model that has retained the capabilities of previous xApps. The simulation results show that xApp distillation has significantly more consistent performance than conventional conflict mitigation methods. Compared conflict mitigation schemes can cause up to 6 times more network outages than xApp distillation in some cases.