Runtime-Tunable AI Optimizes Transit Signal Priority

Transit signal priority (TSP) systems aim to reduce bus delays but must also manage the impact on general traffic and prevent excessive waits for other vehicles. Current reinforcement learning (RL) approaches for TSP often use fixed reward functions, limiting their adaptability when operational priorities, such as time-of-day or disruption conditions, change. This research presents a novel preference-conditioned TSP controller, denoted as $\pi(a \mid s,w)$. This controller can select the optimal signal phase while adhering to minimum/maximum green times and transition feasibility constraints. Crucially, it can be dynamically tuned at runtime using a preference parameter 'w' to adjust the trade-off between prioritizing buses and minimizing overall traffic delay, all without requiring a complete retraining of the model. Implemented within the IntersectionZoo framework, the system was extended with bus-prevalence augmentation and timetable-based bus insertion for robust training. Experiments comparing it against fixed-time control, rule-based TSP, and fixed-weight PPO specialists showed that the single learned conditioned policy effectively spans a smooth empirical trade-off frontier. It consistently outperformed baselines, maintained constraint feasibility, and demonstrated that non-bus externalities remained limited under moderate bus-priority settings.