ABSTRACT Disaster situations impose extreme pressure on emergency medical services, where ambulance routing decisions directly affect patient survival outcomes. We consider the post‐triage ambulance routing problem, where patient locations and survival times are known following initial assessment. Unlike classical vehicle routing focused on operational efficiency, this problem introduces nested temporal constraints where each patient has an individual hospital arrival deadline. We adopt a static formulation—appropriate for the initial deployment phase—to isolate the core algorithmic challenges of survival‐aware multi‐agent coordination before extension to dynamic scenarios. We address this problem through multi‐agent deep reinforcement learning with attention mechanisms, modelling the task as a semi‐Markov decision process where action durations reflect travel times and fleet coordination emerges through learned attention weights. Our actor‐critic approach employs centralised training with decentralised execution, enabling scalable deployment while maintaining coordination benefits. Unlike constraint programming frameworks that struggle with forward‐looking temporal dependencies in multi‐patient routes, the attention‐based policy learns these complex constraint interactions implicitly through reward‐driven training. Experimental evaluation demonstrates that learned policies achieve substantially lower costs while maintaining superior survival compliance compared to OR‐Tools baseline, with the critical distinction that the learned model treats constraint violations as patient mortality rather than optimisation penalties. The results indicate that deep reinforcement learning can effectively capture the trade‐offs between operational efficiency and medical priorities in emergency response scenarios.
Ahmadi et al. (Thu,) studied this question.