Regulating jaywalking behaviour in adaptive traffic signal control using a novel deep reinforcement learning approach

This paper presents a deep reinforcement learning based adaptive traffic signal control framework that explicitly models jaywalking at urban intersections. We integrate a behaviourally grounded Jaywalking Decision model, which endogenises red light violations through waiting time and dynamic gap acceptance, with a Branching Double Deep Q-Network and a comprehensive hybrid action space that controls both phase selection and subphase timing. A multiobjective reward balances delay and jaywalking related safety risk, enabling the controller to respond to non-compliance as it emerges. The framework is evaluated in a multimodal microsimulation of a real intersection in Melbourne across four naturalistic demand scenarios and against both actuated control and established reinforcement learning baselines. Results show that the proposed approach reduces observed pedestrian delay and jaywalking relative to actuated control while achieving a more balanced safety and efficiency profile than single objective or alternative learning architectures. The analysis highlights context-dependent trade-offs that are relevant for policy, since the controller can adapt timing to mitigate non-compliance without assuming full pedestrian obedience. The contributions are threefold: a realistic jaywalking model linked to observable states, a high-granularity action space for multimodal control, and an integrated learning framework that jointly manages delay and safety risk. The proposed framework not only facilitates a more equitable traffic operation system but also offers the first scalable approach to managing risky behaviours in urban traffic environments.