Learning to Tackle Task Variations in Control - A Transportation Context

Abstract

Real-world control tasks are messy and often exhibit task variations. Practical solutions to these problems must exhibit generalization across task variations. For example, in the task of controlling traffic signals, control strategies must adapt to different intersection topologies (the variations), each with distinct dynamics. In this thesis, we consider the challenge of coping with task variations in the context of transportation problems, specifically in roadway interventions where many such variations are both common and imperative to handle. We develop machine learning techniques to address three key challenges: 1) quantify the impact of task variations in control, 2) model them to align with the real world, and 3) optimize in the presence of them. To this end, we begin with a large-scale case study of cooperative eco-driving and illustrate how explicitly modeling task variations can surface otherwise overlooked insights. Building on this, we argue for the necessity of formally incorporating task variations into problem specifications, emphasizing that task underspecification due to loosely defined task variations can severely impair decision-making. We then introduce a contextual reinforcement learning algorithm capable of leveraging the structure of task variations to generalize effectively in cooperative eco-driving with autonomous vehicles. We also present IntersectionZoo, a benchmark designed to promote the development of learning algorithms that generalize by exploiting task variation structures, thus standardizing progress in the field. Last, we explore task variation modeling through a generative modeling lens, using human driver behavior modeling as a case study. Overall, this thesis lays the groundwork for robust control methods by leveraging machine learning to tackle task variations, specifically in roadway intervention designs.