Integrate Transportation Planning Models with Machine Learning Algorithms: A Computational Graph Framework in a Data-Rich Environment

With the advent of new mobility services and technologies, the complexity of understanding the mobility patterns has been gradually intensified. The availability of large datasets, in conjunction with the transportation revolution, has been increased and incurs high computing costs. These two critical challenges require us to methodologically handle complex transportation problems with numerical performance: fast, high-precision solutions, and reliable structure under different impact factors. That is, it is imperative to introduce a new type of modeling strategy, advancing the conventional transportation planning models. In order to do this, we leverage the backbone of the underlying algorithm behind machine learning (ML): computational graph (CG) and automatic differentiation (AD). CG is a directed acyclic graph (DAG) where each vertex represents a mathematical operation, and each edge represents data transfer. AD is an efficient algorithm to analytically compute gradients of necessary functionality. Embedding the two key algorithms into the planning models, specifically parametric-based econometric models and network optimization models, we theoretically and practically develop different types of modeling structures and reformulate mathematical formulations on basis of the graph-oriented representation. Three closely related analytical and computational frameworks are presented in this dissertation, based on a common modeling methodology of CG abstraction. First, a two-stage interpretable machine learning framework developed by a linear regression model, coupled with a neural network layered by long short-term memory (LSTM) shows the capability of capturing statistical characteristics with enhanced predictability in the context of day-to-day streaming datasets. Second, AD-based computation in estimating for discrete choice models proves more efficiency of handling complex modeling structure than the standard optimization solver relying on numerical gradients, outperforming the standard methods, Biogeme and Apollo. Lastly, CG allows modelers to take advantage of a special problem structure for the feedback loops, a new class of problem reformulation developed through Lagrangian relaxation (LR), which makes CG based model well suited for reaching a high degree of the integrated demand-supply consistency. Overall, the deep integration of the practically important planning models with the underlying computationally efficient ML algorithms can enhance behavioral understanding of interactions in real-world urban systems, and the proposed differentiable mathematical structures will enable transportation decision-makers to accurately evaluate different demand-side and supply-side scenarios with a higher degree of convergency and optimality in more complex transportation systems.