Large-scale transportation governance: A tensorization-parallelization co-empowered framework

Transportation systems are complex mega-constructs exhibiting involved interactions between human mobility activities and transportation infrastructure supplies. We need models to characterize how such systems operate and decision-making methods to optimize them toward sustainable development goals. While theory and methods are constantly being innovated, scientific endeavors are far from enough to break through the fundamental computational bottlenecks in large-scale transportation governance, which refers to transportation planning, design, control, and management with extensive spatiotemporal coverage. In the era of artificial intelligence (AI) and smart cities, a unique opportunity—model tensorization and parallel computing co-empowerment—has emerged.

The need for computational breakthroughs in large-scale transportation governance

Transportation governance generally encompasses two interconnected steps. We must first understand how transportation systems operate.1 Computational models that are digital twins of the physical world are the primary means of achieving this goal.

Transportation network models typically fall into three scales, namely macroscopic, mesoscopic, and microscopic models.2 The choice of which scale to use depends on the problem and its modeling needs; macroscopic models reproduce static network flow distribution to facilitate planning and strategic design, mesoscopic models capture spatiotemporal flow evolution to support dynamic management and control, and microscopic models characterize agent behaviors and interactions to meet fine-grained design and individualization needs. The great challenge, however, is that all models suffer from computational bottlenecks in large-scale networks with millions of trips.

What connect these models for transportation governance are decision-making methods, often in the form of optimization algorithms. With AI, models are naturally losing their analytical tractability and facing challenges in deriving explicit input-output relationships required by mathematical programming.3 Simulation-based optimization (SBO)—in particular, metamodeling that constructs surrogate models to approximate “black-box” input-output relationships—provides a unique perspective on building intelligent decision-making methods.4 Unfortunately, given expensive-to-evaluate models, the iterative sampling-guided resolution of SBO under serial computing frameworks is computationally prohibitive.

Intelligent transportation governance and smart city construction cannot be realized without addressing such computational bottlenecks.5 Stacking computing hardware simply fails because models and methods are not computationally compatible with high-performance computing power, such as graphics processing units (GPUs). We advocate for software-hardware synergy in large-scale transportation governance concretized through a model tensorization and parallel computing co-empowered framework, where tensorization reformulates models into tensor structures and parallel computing enables them to run simultaneously across hardware.