Thesis

Modelling parameter interdependence in system dynamics : a data-driven Bayesian network approach to assessing uncertainties in simulations

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T17628
Person Identifier (Local)
  • 201958018
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • System Dynamics (SD) modelling supports decision-making by simulating and projecting keyperformance indicators (KPIs) of a system. To assess uncertainty in these KPIs, modellerstypically vary model parameters and run multiple simulations. This process illustrates potentialscenarios and generates KPI distributions that provide quantified measures of uncertaintythrough statistical analysis.However, parameters often do not vary independently. Studies across sectors have reportedcorrelations and dependencies among parameters (Krefeld-Schwalb et al., 2022; Li and Vu,2013). When varying parameters in SD models, whether these dependencies are accounted forcan shape the combinations of parameter values, influence the distributions of projected KPIsand the derived insights. This issue has not been thoroughly addressed in the SD literature.To highlight the importance of parameter dependence, we present a copula-based experiment,modelling dependencies between SD model parameters in several different ways andcomparing the resulting KPI distributions. The experiment demonstrates that both the strengthof correlations and the structure of dependencies can affect KPI uncertainty. These findingsmotivate the adoption of more flexible approaches to adequately model such dependencies.The main contribution of this thesis is a method that models dependencies among SD modelparameters using Bayesian Networks (BNs) to improve KPI uncertainty assessment. BNsprovide a flexible framework and algorithms for uncovering complex conditional relationshipsfrom data and integrating them with expert knowledge. We apply the approach to an epidemicSD model, where dependencies among epidemiological parameters are estimated from a crosscountry COVID-19 dataset using a BN and validated against domain knowledge. The learnedBN produces input for analysing KPIs of the epidemic SD model and yields uncertaintyenvelopes that differ from those generated by independent or single-copula priors.This study offers SD practitioners a practical way to incorporate empirically grounded multiparameter dependencies into their models, enhancing the defensibility of uncertaintyassessments while keeping additional data-collection effort manageable. The proposed methodcontributes to both the SD and mixed-methods research literatures.
Advisor / supervisor
  • Howick, Susan
Resource Type
DOI

Relations

Items