Optimising traffic signal timings for safety and efficiency using graph-based reinforcement learning

Reinforcement learning-based adaptive traffic signal controllers have traditionally prioritised efficiency, with only recent studies beginning to incorporate safety considerations into signal design. However, most safety-oriented approaches rely on simulated trajectories that oversimplify driver behaviour, leading to potentially unreliable safety estimates. In contrast, studies based on real-world traffic conflicts often depend on rare critical events and fail to capture the safety conditions across the broader range of traffic interactions. Moreover, existing safety-oriented controllers typically employ fixed state representations tied to specific intersection layouts, limiting the transferability of learned signal policies across different geometric configurations. To address these limitations, this study proposes a graph-based reinforcement learning framework to optimise signal timings for both safety and efficiency. The safety objective minimises time spent in conflict, measured using the proportion of stopping distance, while the efficiency objective reduces queue length. A multi-objective deep reinforcement learning algorithm was used to optimise both objectives simultaneously. The intersection was represented as a graph of lanes and movements, which was processed through a graph neural network to capture underlying traffic dynamics. Video data were collected over two weekdays from two four-legged signalised intersections in Queensland, Australia. Road user trajectories were extracted using artificial intelligence-based video analytics and subsequently used to calibrate a microscopic traffic simulation environment via genetic algorithms. The trained signal control policy was then evaluated within these calibrated environments and compared with the existing field signal plan, as well as safety-oriented and efficiency-oriented reinforcement learning variants. Relative to the field signal plan, the proposed safety-efficiency framework achieved reductions of 5.1 % in average waiting time, 7.9 % in queue length, and 17.3 % in time spent in conflict. When transferred to a new intersection with a different layout, the learned policy continued to outperform the field signal plan across both safety and efficiency metrics. These findings demonstrate that traffic signals can be redesigned to improve safety without compromising efficiency. Overall, the proposed graph-based reinforcement learning framework offers a promising direction for intelligent signal control systems.