New AI optimizes traffic lights for buses without retraining
A single policy tunes transit priority vs traffic delay at runtime...
A team from Philip-Roman Adam and Stefanie Schmidtner introduced a novel reinforcement learning approach for transit signal priority (TSP) that enables dynamic runtime tuning without retraining. Their method, detailed in their arXiv paper (29th IEEE ITSC 2026), implements a preference-conditioned policy π(a|s,w) that selects optimal signal phases while respecting minimum/maximum green time constraints.
The system leverages IntersectionZoo as a foundation and enhances scenario generation with bus-prevalence augmentation and timetable-based bus insertion to address sparse transit events during training. Experiments demonstrated that a single learned policy could span a smooth trade-off frontier across different runtime preferences, outperforming fixed-time control, rule-based TSP overlays, and fixed-weight PPO specialists while maintaining constraint feasibility. However, tail-delay diagnostics revealed that non-bus externalities increase substantially under high bus-priority weights, highlighting the need for careful parameter tuning.
- First runtime-tunable TSP controller using preference-conditioned RL (π(a|s,w)) that balances bus priority against traffic delay without retraining
- Implemented on IntersectionZoo with bus-prevalence augmentation and timetable-based bus insertion to improve training scenarios
- Single policy spans smooth trade-off frontier and outperforms fixed-time/rule-based baselines, though high bus priority increases non-bus delays
Why It Matters
City planners can now dynamically optimize traffic signals for buses vs general traffic without costly model retraining or infrastructure changes