Reinforcement learning for model transformations to support model-driven plug-and-play interoperability

Abstract

The digital transformation of industry is prompting companies to improve the interoperability of their systems, consolidating the Digital Thread at the core of their organizations. Interoperability ensures digital continuity between heterogeneous and isolated systems across the value chain and product life cycle, enabling a smooth, uninterrupted flow of data among stakeholders. However, the continual emergence of new technologies and increasing demands for flexibility and scalability in the value chain make establishing and maintaining interoperability in heterogeneous, dynamic environments a significant challenge. To address these requirements, the Model-Driven Interoperability (MDI) framework provides a concrete solution for achieving self-configuring (plug-and-unplug) and self-adjustable (update-and-play) mechanisms to enable plug-and-play interoperability between systems and support the digital thread. MDI employs two key mechanisms: projection mechanisms, which ensure syntactic mapping between systems, and model-to-model transformations, which guarantee semantic mapping between heterogeneous models. In this paper, we present an innovative reinforcement learning-based approach to automate the generation of model-to-model transformations, addressing the need for a self-configuring and self-adjusting interoperability mechanism. This method autonomously learns transformation rules that connect distinct metamodel concepts, significantly reducing the development time and human effort required to create transformation models. By addressing the challenges of dynamic semantic mapping between heterogeneous models, the proposed approach enables adaptive and on-demand interoperability. A case study on the digital thread in the space sector concludes the article and demonstrates the impact of the proposed solution in a complex context such as intelligent planning.