Multi-objective traffic signal control considering pedestrian crossings via constraints-decoupled reinforcement learning

This study proposes a multi-objective adaptive signal controller for pedestrian crossings using a constraints‑decoupled reinforcement learning framework. Unlike direct optimization of signal timings based on predefined phases, the approach decouples phases into vehicle and pedestrian movements via an option‑action structure. A linear scalarization mechanism enables collaborative optimization of pedestrian and vehicle efficiency under different weight preferences. Additionally, a value function approximation method based on phase Q‑value decomposition is developed. It learns movement‑level Q‑values and aggregates them into phase‑level estimates using a conflict matrix and combination rules, reducing repetitive learning and improving generalization. Simulation results show that the proposed controller converges faster than phase‑based adaptive controllers, reduces delays for both pedestrians and vehicles under varying demands, and maintains stability under high demand. Through multi‑objective optimization, it achieves an adjustable balance between pedestrian and vehicle flows, offering an effective solution for adaptive traffic signal control involving multiple traffic participants.