Pattern‐oriented analysis of system dynamics models via random forests

Abstract

System dynamics (SD) modeling studies aim to reveal the causes of problematic dynamic behaviors and eliminate them through policy design and analysis. The analyst conducts sensitivity/scenario analyses and what‐if experiments to reveal the input–output relationships during modeling. However, during these analyses and investigations, the identification of input‐parameter spaces that cause the generation of different SD model behavior patterns is time consuming and susceptible to human bias. Therefore, we propose a metamodel‐based procedure for SD models that considers the necessity for unbiased and automated analysis and insight generation. The approach uses the random forest algorithm for metamodel generation and extracts interpretable IF–THEN rules from the metamodel, thereby identifying input subspaces that generate different qualitative or numerical SD model outputs. We illustrate the proposed approach using two well‐established SD models. These case studies reveal how the model analyst can utilize the proposed method to capture input–output relationships. © 2022 System Dynamics Society.

What is (quantitative) system dynamics modeling? Defining characteristics and the opportunities they create

“…Classical statistical methods are certainly still useful for model development, but machine‐learning methods can facilitate more generalized discovery of predictive patterns from data (Bzdok et al ., 2018). Machine‐learning techniques have recently been incorporated into system dynamics model parameterization, calibration (Chen et al ., 2011), and analysis (Edali, 2022; Ozik et al ., 2016; Pruyt and Islam, 2015), although significant room for innovation exists. The system dynamics field should continue to incorporate modern data science techniques into all aspects of the system dynamics modeling process.…”