Intervention Strategy Discovery in Complex Systems through Causal Modeling and Sensitivity AnalysisZhou, Gaoxiang (2023) Intervention Strategy Discovery in Complex Systems through Causal Modeling and Sensitivity Analysis. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractIn many scientific disciplines, there exists a notable lack of comprehensive analysis concerning the dynamics of complex systems. Many researchers have attempted to model such complex systems, test hypotheses, and propose strategies for intervention. Existing methodologies for modeling these systems can generally be classified into event-based, agent-based, and element-based. This work focuses specifically on the element-based approach, wherein elements are updated over time based on a diverse range of rules, including regulatory functions and dependencies. Boolean Network, Probabilistic Boolean Network, and Dynamic Bayesian Network are three commonly employed element-based techniques. However, formal translation and conversion mechanisms among these approaches are missing. In this dissertation, we automated the conversion process between various element-based modeling approaches, elucidating disparities in the information required by these approaches and identifying their respective use cases. Secondly, we developed a framework that facilitates exploration of models created by element-based approach, with a specific focus on sensitivity analysis. Lastly, we devised a strategy for identifying intervention targets through influence transmission. This dissertation describes the development of a sensitivity assessment that quantifies the susceptibility of a system and its elements to internal or external changes. Unlike many static measures of node influence such as in/out-degrees, centralities, or dynamic properties like determinative power, we have defined a measure for immediate influence between directly connected elements, as well as remote influences between indirectly connected elements. Additionally, our work entails the creation of a framework capable of parsing any element-based model and systematically evaluating causal influences between elements. Two sensitivity computation methods, namely function-based and hybrid-based, have been proposed and implemented, ensuring flexibility and balance between efficiency and accuracy. Our approach to assessing sensitivity extends to various applications like signaling networks in biomedicine and causal networks in social science. For instance, we have utilized sensitivity as benchmarks for edge weights assignment and extracted influential pathways. The validity of these pathways has been verified theoretically using model checking method and practically through user-embedded experiments in a human-machine interface. Through applications, our framework bridges the goal to synthesize network paths, culminating in a practical intervention strategy that guides the target toward desired states. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|