Machine learning-enhanced dynamic path decisions for emergency stewards in emergency evacuations

In emergency situations such as indoor fires, emergency stewards can significantly influence the evacuation behavior of trapped individuals, and the rationality of their own path decisions is crucial to the overall evacuation effectiveness. This paper introduces an improved social force model to reflect the impact of stewards on the behavior of those in distress, and designs a decision-making framework based on simulation models and Deep Reinforcement Learning (DRL) technology to optimize the path decisions of stewards in dynamic scenarios. The simulation model is used to simulate various scenarios to obtain sufficient sample data; the role of DRL is to interact with the environment and dynamically guide individuals towards optimal paths using learned effective strategies. Within this framework, the decision training for emergency stewards employs a Modified Priority Experience Deep Q-Network (MPE-DQN), avoiding areas with high personnel density to optimize evacuation path decisions. The safety metric during the evacuation process is measured by personnel density per unit area, and evacuation time is chosen as the efficiency metric. Simulation experiments conducted in AnyLogic show that compared to the standard DQN algorithm, our framework, using the MPE-DQN algorithm, increased safety by 58.77 % and improved efficiency by 14.2 %.