Predicting the unseen: improving robustness in koopman surrogate models for crowd dynamics at a bottleneck

Abstract

Crowds moving through bottlenecks form a dynamical system, with its density fluctuating in time and space. The system dynamics can be learned and predicted using the Koopman operator framework. But how reliable are predictions for previously unseen crowd sizes? How significant is the impact of stochastic observations? In this work, we show that enriching the state space with head counts and using diffusion maps as part of our learning pipeline facilitates the robustness of Koopman-based surrogate models.