New Survey Automates AI Model Design for Smarter Traffic Prediction

Traffic prediction is essential for intelligent transportation, but current deep learning models (graph convolutional networks, recurrent networks, Transformers) require labor-intensive manual architecture design and often fail to generalize across cities. Vu, Yang, and Pazzi’s new survey, accepted at UrbCom 2026, systematically reviews Neural Architecture Search (NAS) as a solution. NAS automates the search for optimal model architectures by exploring a design space of spatial and temporal operators. The authors categorize methods into three search strategies: gradient-based, evolutionary, and one-shot weight-sharing. For each, they analyze how the search space captures traffic-specific patterns and how search cost is balanced against architecture quality. The survey includes 8 pages and 3 tables, offering a structured comparison of recent approaches.

Beyond taxonomy, the paper identifies critical open challenges: computational scalability to large road networks, manual search space design that still requires expert input, poor cross-city generalization, handling dynamic graph structures (e.g., evolving road conditions), and the nascent question of applying NAS to spatial-temporal foundation models. By highlighting these gaps, the authors outline a roadmap for future research—moving toward fully automated, scalable, and transferable traffic prediction systems. For practitioners, this survey provides a clear entry point for adopting NAS in transportation AI, potentially reducing development time and improving model robustness across diverse urban environments.