Characterizing Crowd Preferences on Stadium Facilities through Dynamic Inverse Reinforcement Learning

Understanding people’s preferences for facilities in sports stadiums is important to optimize the facility allocation and predict the crowd distribution for public safety purposes. Crowd preferences refer to people’s priority in seeking food, game swags, restrooms, finding seats, and exploring the venue over time. Existing approaches include pre-event surveys and interviews, which are costly and prone to biases when the response rate is low. In recent years, the development of advanced sensing systems enables automatic tracking of crowd distributions in public spaces, enabling the characterization of human preferences based on sensor data. However, the main challenge of human preference characterization is that their behaviors are not static, but rather dynamic and context-dependent, leading to high uncertainties over time and space. To overcome this challenge, we present a novel method to infer the dynamic preferences of the crowd on facilities such as food stands, restrooms, and game swag stations using dynamic inverse reinforcement learning (DIRL). Our method leverages the facility layout in sports stadiums and learns a time- and context-dependent reward function of the crowd by observing their distribution around various facilities over time. We then infer the crowd’s preference for facilities based on the weight of each type of facility in the reward function. To evaluate our approach, we conduct 6 real-world experiments for NCAA Pac-12 sports games at Stanford Maples Pavilion. We demonstrate how our method captures the changing crowd preferences on facilities over time and estimates the crowd distribution trend at 5 entry doors with 70% accuracy, allowing dynamic human preference characterization through sensor data.