Abstract for: Decoding Policy Interactions in System Dynamics: A Balanced Tree Clustering Approach

The challenges facing society today need quick and accurate responses. System dynamics (SD), which simulates time series to reveal the non-linear interactions among multiple factors, is a key tool for determining the combinations of policy actions that produce the most desirable outcomes. While SD is a powerful tool for exploring dynamic behaviors, understanding variable impacts and identifying critical thresholds across numerous scenarios remains challenging. We employ machine learning techniques to interpret the relationship between the policy parameters of the SD model and the time series. We propose a balanced tree clustering method that simultaneously considers the time series features and the policy parameters, clearly identifying the split points where these policy parameters influence the increases or decreases in the time series. We examined two models. The Bass model features a simple structure with few policy parameters, while the Earth4All model employs a complex structure with many policy parameters. We ran thousands of Monte Carlo simulations to extract the split points of these policy parameters. We show that the time series split clearly into two distinct groups—one with desirable trends, the other with undesirable trends—across a specific range of policy parameter values. This analysis enables us to identify the key policy parameters and their corresponding values. Grammar correction