Learning and Control for Intelligent Transportation Systems with Human Driver Interaction
Intelligent transportation systems are developing rapidly, yet fully autonomous operation has not been achieved in most real-world traffic environments. At this stage, human drivers remain in the loop, so intelligent systems must make decisions that are not only efficient but also compatible with human driver preferences, while connected and automated vehicles (CAVs) must operate in traffic alongside human-driven vehicles (HDVs). To address these challenges, this thesis develops two frameworks for transportation applications. First, an adherence-aware reinforcement learning approach is proposed for a lane-changing recommendation system in a semi-autonomous driving environment, where human drivers may not always follow recommendations. The method formulates the recommendation problem as a Markov decision process and incorporates driver adherence into Q-learning to learn recommendation policies that account for partial human compliance. Second, a functional learning framework is developed for team-optimal traffic coordination in intersection scenarios involving CAVs and HDVs. The framework is studied in both an offline setting, through a policy iteration method based on an implicit discrete Fréchet update, and an online receding-horizon setting, through model estimation and rolling policy updates under unknown dynamics introduced by HDVs. Both proposed frameworks are validated through realistic simulation environments, including CARLA and SUMO. Collectively, these approaches contribute to intelligent transportation systems that better account for human behavior and support safe and efficient traffic coordination in mixed traffic environments.