Efficient Calibration of Heterogeneous Agent-Based Pedestrian Model Using PSO With Deep Learning Surrogates

This paper presents an efficient framework for calibrating heterogeneous agent-based pedestrian models by combining particle swarm optimization (PSO) with deep learning-based surrogate modeling. Each agent was individually calibrated by predicting the next-step velocities using a neural network trained on real-world pedestrian trajectory data. By embedding the surrogate model into the PSO loop, the computational cost is significantly reduced while maintaining high accuracy. The proposed framework captures diverse agent behaviors by accounting for environmental influences such as walls, pillars, and crowd proximity, enabling dynamic and context-aware parameter adjustment. Comparative analysis revealed that although heterogeneous calibration is more time-intensive than homogeneous approaches, it consistently improves behavioral fidelity and model realism. Validation with a separate dataset confirms the effectiveness of the method in pedestrian dynamics applications such as flow analysis and the use of indoor spaces. The results demonstrate that agent-based modeling, when combined with deep learning surrogates and PSO, provides a scalable and accurate solution for simulating complex human movement patterns. This framework supports high-fidelity model calibration while offering practical runtime efficiency.