M.Sc. Ferdinand Mütsch, Prof. Dr. J. Marius Zöllner

Abgeschlossen

The effective generation of heterogeneous spatio-temporal graphs (HSTGs) is a crucial  step in bridging graph neural networks and traffic scenario generation, particularly addressing the scarcity of critical real-world scenarios that are vital for the safety validation  of autonomous vehicles. This thesis introduces SHGDiff, a novel generative model designed to tackle the unique challenges of HSTGs, including their heterogeneity, scalability,  and temporal dynamics. By leveraging advancements in deep generative models such as  variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion  models, SHGDiff provides a scalable and flexible approach to generating large, spatiotemporal graphs. The results highlight that SHGDiff is the first model in its field to fully  harness the expressive power of HSTG representations. Its performance on molecular  datasets and large networks is comparable to existing state-of-the-art models, while its  application to traffic scenarios demonstrates its functionality and potential. Furthermore,  SHGDiff marks a significant step forward in the foundational research of traffic scenario  generation, setting a new baseline for future work by enabling the creation of traffic scenarios with a level of expressivity and complexity previously unattainable. These findings  underscore the potential of deep generative models to advance simulation frameworks for  complex, relational systems, paving the way for future research into heterogeneous graph  representations.