Learning and Control for Adaptive Transportation Systems

Transportation is a permeating factor of modern society, effectively enabling the accomplishment of everyday tasks such as going to work, school, or simple personal errands.
The efficiency of the transportation network highly influences the quality of life of individuals, and it is widely recognized that investments in transport infrastructure can generate large developmental payoffs throughout society.

Currently, transportation within cities is intrinsically system-first, whereby users have to adapt to a fixed transportation system (e.g., fixed bus routes and schedules).
In contrast, we believe in the vision of a user-first system, whereby supply and demand for transportation are seamlessly co-adapted into an Adaptive Transportation System (ATS).
To pursue this vision, transportation systems need to acquire two fundamental capabilities: 1) being able to understand and predict current mobility patterns, and 2) converting this knowledge into effective decisions that can satisfy the need for mobility of a broad and diverse audience.
At the same time, the high complexity of the transportation system, together with the vast availability of mobility data (e.g., consider the daily stream of data generated by our smartphones), make this a perfect scenario for applications of machine learning and data-driven strategies broadly.

In this context, we aim to develop learning algorithms that are capable of addressing the challenges emerging within adaptive transportation systems, with a focus on mobility-on-demand systems.
In the first part of the thesis (i.e., “Perception and Prediction'') we develop tools for predicting the evolution of mobility demand.
Specifically, we focus on developing learning algorithms capable of (i) handling the multi-modality of mobility demand through structured uncertainty quantification, and (ii) exploiting properties of the prediction problem by injecting sensible prior assumptions, or inductive biases, within the learning process.

On one hand, we cast the learning problem under the lens of Bayesian statistics while leveraging tools from modern literature on deep generative models (e.g., deep latent variable models and normalizing flows).
On the other hand, we improve the search for solutions of learning algorithms by embedding assumptions about the data-generating process within learning.
In the second part of the thesis (i.e., “Optimization and Control'') we develop autonomous decision-making strategies for an efficient allocation of supply.
In particular, we first investigate the effectiveness of traditional predict-then-optimize approaches, and highlight a number of shortcomings in literature by providing insights into the interplay between forecasts, model assumptions, and decisions.
Equipped with this intuition, we then develop a fully end-to-end strategy for learning how to control Autonomous Mobility-on-Demand (AMoD) systems.
This strategy combines methods from reinforcement learning and control theory with the relational representative power of graph neural networks to yield effective, scalable, and generalizable control policies.

We conclude the thesis with a discussion on short, intermediate, and long-term next steps in extending the ideas developed herein towards the ultimate goal of adaptive transportation systems.