Self-Adaptive Systems (SASs) can modify itself according to the change of working environment. As a typical SAS, the dynamic bus system monitors the willingness and orientation of the passengers, aiming to provide the dynamic path decision scheme for buses in an open environment. In a large number of passengers scene, the complex crowd distribution due to the increased number of passengers makes the decision challenging, in that the increase of space dimension in the complex environment leads to exponential data growth and requires a large number of computational resources. To enable the dynamic bus system to operate properly in a large number of passengers scenario, we propose a deep Q-network (DQN) based path optimization scheme for dynamic bus systems. By using deep neural networks to fit Q values end-to-end instead of the table storage approach in Q-Learning, in order to avoid the problem of dimensional explosion in large number of passenger scenarios. We explore the impact of different crowd distributions on path decisions to represent the impact of crowds on the adaptive system. We also compare the decision capability of DQN, Q-learning and Pareto methods for dynamic bus paths and the carrying capacity for large number of passenger scenarios to demonstrate the feasibility and effectiveness of the scheme in this paper.