Mixed traffic control and coordination: optimization and learning-based approaches

Abstract

Traffic crashes in the United States resulted in over 44,000 fatalities in 2024, with human error contributing to approximately 94% of these incidents. Automated driving technologies offer a promising solution to reduce human error and improve road safety. Beyond safety, connected and automated vehicles (CAVs) are expected to alleviate congestion, lower energy consumption, and enhance traffic efficiency. However, realizing these benefits at scale requires effective control and coordination strategies in mixed-traffic environments where CAVs share the road with human-driven vehicles (HDVs). This dissertation addresses the control and coordination problem for CAVs in mixed-traffic interaction-driven scenarios, such as intersections and merges. Three key challenges need to be addressed: (i) real-time optimization and control for interacting multi-agent systems, (ii) understanding human behavioral models and integrating those prediction models into control, and (iii) developing human-compatible control designs that are safe, robust, efficient, and able to adapt. To address these challenges, we leverage the intersection of control, optimization, and machine learning to propose three contributions: (i) a learning-to-adapt framework for model predictive control under varying human driving styles; (ii) a stochastic time-optimal trajectory planning approach incorporating data-driven car-following models; and (iii) a distributed, learning-aided mixed-integer quadratic programming framework for the joint coordination of traffic lights and CAVs. ☐ The first contribution of this dissertation is the development of a learning-to-adapt control framework that enables model predictive control (MPC) to account for varying human driving behaviors. It integrates a game-theoretic MPC formulation for CAV-HDV interactions, online inverse learning of human driving intent, and an optimal weight adaptation strategy. Since computing the optimal adaptation is computationally expensive and involves black-box evaluations, in a sim-to-real manner, the framework employs Bayesian optimization and contextual Bayesian optimization to efficiently learn how to adapt MPC weights based on the inferred human objectives. ☐ The second contribution of this dissertation introduces a stochastic time-optimal trajectory planning framework for coordinating multiple CAVs in mixed-traffic merging scenarios. The approach leverages a time-optimal control problem formulation for CAVs and incorporates a data-driven Newell's car-following model with Bayesian linear regression to predict HDV behavior, together with uncertainty qualification. Safety constraints are enforced probabilistically for robust yet efficient trajectory planning. To account for prediction errors, a replanning mechanism is developed based on monitoring the accuracy of HDV trajectory prediction. ☐ The third contribution focuses on the coordination of CAVs and traffic lights at complex intersections in mixed traffic. We propose a joint optimization framework allowing simultaneous green signals on conflicting lanes when no HDV-related lateral conflicts exist, enabling more efficient use of CAV coordination. We formulate this as a multi-agent mixed-integer quadratic program and develop a distributed solution using a variant of the maximum block improvement algorithm with penalization of local constraints. Moreover, we proposed a graph machine learning approach that learns the mapping from problem parameters to optimal binaries offline, enabling fast online control by solving convex quadratic programs. ☐ Collectively, this dissertation contributes toward safe, interaction-aware, robust, and efficient trajectory planning methods for CAVs in mixed traffic. Moreover, by leveraging the coordination of CAVs both with and without traffic signals as CAV penetration rates increase, overall traffic performance can be remarkably improved. Beyond mixed-traffic connected and automated driving, the proposed frameworks, particularly learning-based control, learning to adapt, or distributed and learning-aided mixed-integer optimization, can be extended to broader cyber-physical human systems.