Deep Reinforcement Learning Agent Based Analysis For Real-Life Evacuation Drill Applications

Real-life evacuation of large spaces (such as hotels, theaters, factory halls and so forth) in case of fires or other fast-spreading dangers, is to this day a major challenge for all the involved actors. This game-inspired simulation environment reconstructs the wide-range scenario, in the generation of statistics, and stencils in developing and optimizing procedures for evacuation during the live spread of the dangers in the said spaces. The game-inspired simulation environment’s main contribution is optimizing the control systems for evacuation management across the spaces by having an agent learn to make better choices in directing the people outside of the affected areas, and evacuating them without overcrowding certain areas and preventing blockages. During training, the agent had learned to control the arrows that point to the exits of the spaces in such a manner that more and more NPCs (Non-Player Character) have exited the spaces, gradually increasing the number of rescues and decreasing the number of losses. After training, in the agent play tests, the agent had positive results, with lower numbers of losses and higher number of rescues with every new episode.