Path optimization of integrating crowd model and reinforcement learning

Exit choice and path planning are critical in emergency decision-making. Traditional research focuses on the shortest path, which is not sensitive to environmental factors such as the crowd congestion, obstacles distribution, air pollution, etc. To solve the path optimization problem, a behavior agent model is developed and integrated in the large-scale crowd simulation. The Q-Learning algorithm is applied to adjust the agent behavior. Considering the architectural space key exits and doors as network nodes, the paper presents combining dynamic crowd model and reinforcement learning strategy. The strategy with high training efficiency considering obstacles setup, crowd movement, and exits environment, the learning agent interacts dynamically with surrounding environment, and learns the shortest time path to exit. Simulation utilizes social force model for occupant movement, avoiding collisions with other occupants and obstacles. The path optimization is verified with the pedestrian library of Anylogic.