Model Calibration for Agent-Based Simulation Using a Pattern Clustering Network

Agent-based simulation has become essential for simulating complex evolving systems, such as social systems and complex manufacturing systems. However, due to the uncertainty in agent behaviors, model calibration poses a significant challenge. Traditional methods like try-and-error and regression are inefficient, requiring numerous simulation runs under different parameter settings. To address this, we propose an online pattern clustering network-based calibration method, termed PCN-Calibration. This method establishes a pattern clustering network using reference data from the target system and compares simulation results under random parameter settings with this network. Ten weighting rules are introduced to estimate the best parameters based on the distance between the simulation results and the clustering network. Experimental results demonstrate that the proposed method can find feasible parameter settings within seconds. Furthermore, the experimental discussions provide guidelines for selecting suitable weighting rules for calibrating different models.