Mlops as a Service for AI-Native 6G Networks

This paper concerns the lifecycle and runtime management of AI components of AI-native networks, in which the AI components are deeply integrated with network functions. The paper presents an overview of the standardised, IT-based, and research approaches to ML pipeline management. To address the identified gaps, we introduce the automated MLOps as a Service concept in which an ML model deployed in the network is continuously monitored for data and concept drifts, as well as network performance degradation, and on that basis, specific ML pipeline operations are triggered to improve the model's operation. The proposed approach allows for selecting local (inplace) or remote training by a network operator or external ML service providers. The benefits of selecting different training options are demonstrated in the AI-based network analytics scenario.