Optical transport networks are evolving from closed, vertically integrated systems to open and disaggregated architectures able to support diverse services while adapting to rapidly increasing and variable traffic demands. This evolution is enabled by programmable hardware, standardized data models, and cloud-native control platforms. The thesis proposes a microservice-based architecture that combines open and disaggregated optical systems with GitOps and Network as Code practices to achieve automation, continuous integration, and lifecycle management of optical network control functions.
On this foundation, the research investigates how Machine Learning (ML) and Reinforcement Learning (RL) can address major control and management challenges. The work focuses on routing and spectrum assignment, traffic and state prediction, quality of transmission estimation, and adaptive transponder configuration. To enhance these tasks, a Compositional Machine Learning (CML) framework is introduced, decomposing complex operations into modular ML models that can be stitched into optical control pipelines. This approach improves scalability, interpretability, and reuse of ML-driven functions. MLOps methods—including automated retraining, drift detection, and continuous monitoring—are integrated into the framework to ensure robust and adaptive model performance under dynamic network conditions. Simulations and experiments demonstrate significant gains in spectrum utilization, service blocking probability, and operational automation compared with existing approaches.
By bridging cloud-native architectures with ML-driven control, this thesis advances the integration of automation and intelligence in optical networks. It also outlines perspectives such as cross-layer IP/optical coordination, self-healing closed loops, and AI-native resource control, paving the way towards autonomous and sustainable optical transport infrastructures.
