Deep reinforcement learning with a particle dynamics environment applied to emergency evacuation of a room with obstacles

Efficient emergency evacuation is crucial for survival. A very successful model simulating the social-force model. At heart of self-driven force that applied to an agent and directed towards exit. However, it not clear if application this results in optimal evacuation, especially complex environments with obstacles. In paper, we develop a deep reinforcement learning algorithm association social train agents find fastest path. During training, penalize every step room give zero reward at We adopt Dyna-Q approach, which incorporates both model-free Q-learning model-based method, update neural network used approximate action value functions. first show case without obstacles resulting points directly exit as To quantitatively validate our compare total time elapsed when escape one door employing result obtained using median intervals calculated two methods are significantly different. confirm proposed method obtains trajectories minimize travel by comparing generated geodesics-based adaptive pedestrian dynamics. Then, investigate obstacle produces similar convex. concave obstacles, sometimes can act traps governed purely prohibit complete approach clearly advantageous since derives policy object avoidance additional assumptions. also study multiple exits. able evacuate efficiently from nearest through shared trained single agent. Finally, test robustness environment exits Overall, model, based on handle modeling where difficult obtain intuitive rule fast evacuation.