Analysis of an individual‐based influenza epidemic model using random forest metamodels and adaptive sequential sampling

Abstract

This study proposes a three‐step procedure for the analysis of input–response relationships of dynamic models, which enables the analyst to develop a better understanding about the dynamics of the system. The main building block of the procedure is a random forest metamodel capturing the input–output relationships. We utilize an active learning approach as the second step to improve the accuracy of the metamodel. In the last step, we develop a novel way to present the information captured by the metamodel as a set of intelligible IF–THEN rules. For illustration, we use the FluTE model, which is an individual‐based influenza epidemic model. We observe that the number of daily applicable vaccines determines the success of an intervention strategy the most. Another critical observation is that when the daily available vaccines are constrained, nonpharmaceutical strategies should be incorporated to reduce the extent of the outbreak.

Pattern‐oriented analysis of system dynamics models via random forests

“…Subsequently, they used decision trees to perform interpretation. Edali and Yücel (2020) applied an adaptive sampling method combined with a rule-extraction technique to support policy analysis based on an agentbased influenza epidemic model. Lamperti et al (2018) used XGBoost metamodels to calibrate agent-based models.…”

“…Although such an approach may be feasible for small‐scale SD models, it can prove to be time consuming when the number of parameters increases. Moreover, manual exploration is susceptible to human bias, resulting in a more focused exploration and causing a significant portion of the parameter space to remain unexplored (Edali and Yücel, 2018, 2020; Lee et al ., 2015; Saleh et al ., 2010). Herein, we propose an approach to reduce the time required in identifying the combinations of input parameters responsible for generating SD model behaviors (either continuous or categorical) by using machine‐learning models and extracting intelligible rules.…”

“…In early studies, linear regression techniques were widely used to construct metamodels (Blanning, 1975; Friedman and Friedman, 1985; Kleijnen, 1975, 1979). Recent advancements in the machine‐learning domain have enabled analysts to use different techniques, such as neural networks (Can and Heavey, 2012; Sabuncuoglu and Touhami, 2002; Sharifnia et al ., 2021), random forests (RFs) (Edali and Yücel, 2019, 2020; Stolfi and Castiglione, 2021), Gaussian processes (Betancourt et al ., 2020; Dosi et al ., 2018), support vector machines (Edali and Yücel, 2018; Ten Broeke et al ., 2021), and radial basis functions (Jakobsson et al ., 2010; Mullur and Messac, 2006). In addition, comparative analyses of some of these techniques based on their accuracy, interpretability, robustness, and efficiency have been presented in some studies.…”

“…Metamodels have been used to learn input–output relationships in the agent‐based modeling domain (Edali and Yücel, 2018, 2019, 2020). In this respect, Edali and Yücel (2018) generated support vector regression (SVR) metamodels of agent‐based simulation models to learn input–output relationships.…”

“…Owing to the low interpretability of SVR models, Edali and Yücel (2018) utilized decision trees to distill the information captured by complex SVR metamodels. In another study, Edali and Yücel (2020) showed a process by which IF–THEN rules can be extracted from an RF metamodel to guide policymakers in evaluating intervention strategies, such as quarantine, school closures, and vaccination via an agent‐based influenza‐epidemic model.…”