Deep reinforcement learning-based panic crowd evacuation simulation

Crowd evacuation simulation can provide guidance for public emergencies or mass casualties which is easy to cause crowding and stampede. Deep reinforcement learning is an effective method for the path planning of crowd evacuation simulation which can reduce the dependence on data and has strong generalization. However, current methods of deep reinforcement learning ignore the consideration of panic emotion which could result in weak authenticity since the panic emotion has a significant impact on crowd evacuation. To address this problem, we propose a novel deep reinforcement learning for the panic crowd evacuation. First, a quadrant based crowd network is constructed to guide the movement of agents according to the change of panic degree in different quadrants. Second, a method of quantifying panic degree is proposed. Through the mean field equation, the dynamic evolution of agent states is realized, and the panic degree in different regions is quantified. Finally, a panic crowd evacuation model based on multi-agent deep deterministic policy gradient (MADDPG) was established. The computated panic degree is introduced into the reward function to recommend a path for multiagent to avoid panic. Experimental results show the effectiveness of our method.