SS
S.R.M. Salome
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Unfolding Futures in Feedback Loops
Integrating Loops That Matter into Scenario Discovery for System Dynamics Models
Decision making under deep uncertainty relies on simulation models to explore how complex systems may behave across a wide range of uncertain conditions. Scenario discovery is an important method within the field of deep uncertainty. It generates large ensembles of model runs, classifies these runs into decision relevant outcome classes, and then uses rule induction to identify the input space regions associated with those classes. In this way, scenario discovery helps analysts understand under which uncertain conditions problematic outcomes occur. However, the method usually focuses on the relation between model inputs and model outcomes, while the structure of the model remains outside the analysis.
This creates a tension for System Dynamics models. System Dynamics is a white-box modeling approach in which behavior is understood as the result of its model structure. Growth, decline, oscillation, and collapse are not treated as isolated output patterns, but as consequences of stocks, flows, delays, and feedback loops. When scenario discovery is applied to System Dynamics models without considering this structure, part of the potential of the model is lost.
The recently developed Loops That Matter method may help address this gap. Loops That Matter calculates loop scores that indicate how feedback loops contribute to the model behavior over time. This formalizes the quest to explain how model behavior is related to model structure. While feedback loops are commonly used to explain model behavior, such explanation is usually limited to a small number of exemplar runs and is rarely embedded in large uncertainty analyses.
This thesis therefore investigates how Loops That Matter can be integrated into scenario discovery to relate decision relevant model outcomes to the feedback structure of System Dynamics models.
This research develops four approaches. The first approach adds loop scores after a conventional scenario discovery workflow. The second approach uses a single loop score directly in the classification step. The third approach extends this logic by applying time series clustering to multiple loop scores. The fourth approach transforms dominant loop sequences into a network trie. These approaches are tested on two models, a small epidemic model and a more complex energy model. The epidemic model provides a straightforward relation between structure and behavior. The energy model provides more uncertain parameters, more feedback loops, and more diverse behavior.
The results show that Loops That Matter can meaningfully extend scenario discovery when the relation between feedback structure and model behavior is clear. In the epidemic model, loop scores helped explain the differences between mild, moderate, and severe outbreak classes. The dominant loop sequence strips showed how outbreak severity was related to the timing and duration of feedback loop dominance. In addition, clustering the R1 Infection loop score produced classes that could be related back to outbreak behavior and input space. This demonstrates that loop scores can be used both as an ex post interpretation of the model structure, and as a meaningful input to the classification step itself.
The results for the energy model were less successful. Multivariate clustering of two loop scores did not produce classes that were clearly distinct in the model outcomes or well separated in the input space. Similarly, the network trie did not provide a useful classification workflow for scenario discovery. However, the trie did provide an important insight. Namely, that the energy model generated many different dominant loop sequences across the ensemble. Similar behavioral patterns within a key model outcome were also associated with diverse loop dominancy sequences. This suggests that the model exhibits dynamic equifinality, where similar outcomes can arise through different shapes of the feedback loop structure.
The main conclusion is therefore twofold. Loops That Matter can be integrated into scenario discovery, but its usefulness depends on the clarity of the relation between input parameters, and model outcomes. A main contribution lies in the explanatory power of feedback loops after a scenario discovery workflow, as they can help with explaining why classes differ. Their use as a direct basis for classification is more uncertain, especially in larger models. The direct identification of dynamic equifinality is also novel. Previous work has only guessed at its existence, but this research exposes it, and discusses its impact on model based decision support. The thesis contributes a methodological bridge between scenario discovery and structural explanation in System Dynamics, while also showing that equifinality can limit the extent to which model structure can be used to explain decision relevant model behavior. ...
This creates a tension for System Dynamics models. System Dynamics is a white-box modeling approach in which behavior is understood as the result of its model structure. Growth, decline, oscillation, and collapse are not treated as isolated output patterns, but as consequences of stocks, flows, delays, and feedback loops. When scenario discovery is applied to System Dynamics models without considering this structure, part of the potential of the model is lost.
The recently developed Loops That Matter method may help address this gap. Loops That Matter calculates loop scores that indicate how feedback loops contribute to the model behavior over time. This formalizes the quest to explain how model behavior is related to model structure. While feedback loops are commonly used to explain model behavior, such explanation is usually limited to a small number of exemplar runs and is rarely embedded in large uncertainty analyses.
This thesis therefore investigates how Loops That Matter can be integrated into scenario discovery to relate decision relevant model outcomes to the feedback structure of System Dynamics models.
This research develops four approaches. The first approach adds loop scores after a conventional scenario discovery workflow. The second approach uses a single loop score directly in the classification step. The third approach extends this logic by applying time series clustering to multiple loop scores. The fourth approach transforms dominant loop sequences into a network trie. These approaches are tested on two models, a small epidemic model and a more complex energy model. The epidemic model provides a straightforward relation between structure and behavior. The energy model provides more uncertain parameters, more feedback loops, and more diverse behavior.
The results show that Loops That Matter can meaningfully extend scenario discovery when the relation between feedback structure and model behavior is clear. In the epidemic model, loop scores helped explain the differences between mild, moderate, and severe outbreak classes. The dominant loop sequence strips showed how outbreak severity was related to the timing and duration of feedback loop dominance. In addition, clustering the R1 Infection loop score produced classes that could be related back to outbreak behavior and input space. This demonstrates that loop scores can be used both as an ex post interpretation of the model structure, and as a meaningful input to the classification step itself.
The results for the energy model were less successful. Multivariate clustering of two loop scores did not produce classes that were clearly distinct in the model outcomes or well separated in the input space. Similarly, the network trie did not provide a useful classification workflow for scenario discovery. However, the trie did provide an important insight. Namely, that the energy model generated many different dominant loop sequences across the ensemble. Similar behavioral patterns within a key model outcome were also associated with diverse loop dominancy sequences. This suggests that the model exhibits dynamic equifinality, where similar outcomes can arise through different shapes of the feedback loop structure.
The main conclusion is therefore twofold. Loops That Matter can be integrated into scenario discovery, but its usefulness depends on the clarity of the relation between input parameters, and model outcomes. A main contribution lies in the explanatory power of feedback loops after a scenario discovery workflow, as they can help with explaining why classes differ. Their use as a direct basis for classification is more uncertain, especially in larger models. The direct identification of dynamic equifinality is also novel. Previous work has only guessed at its existence, but this research exposes it, and discusses its impact on model based decision support. The thesis contributes a methodological bridge between scenario discovery and structural explanation in System Dynamics, while also showing that equifinality can limit the extent to which model structure can be used to explain decision relevant model behavior. ...
Decision making under deep uncertainty relies on simulation models to explore how complex systems may behave across a wide range of uncertain conditions. Scenario discovery is an important method within the field of deep uncertainty. It generates large ensembles of model runs, classifies these runs into decision relevant outcome classes, and then uses rule induction to identify the input space regions associated with those classes. In this way, scenario discovery helps analysts understand under which uncertain conditions problematic outcomes occur. However, the method usually focuses on the relation between model inputs and model outcomes, while the structure of the model remains outside the analysis.
This creates a tension for System Dynamics models. System Dynamics is a white-box modeling approach in which behavior is understood as the result of its model structure. Growth, decline, oscillation, and collapse are not treated as isolated output patterns, but as consequences of stocks, flows, delays, and feedback loops. When scenario discovery is applied to System Dynamics models without considering this structure, part of the potential of the model is lost.
The recently developed Loops That Matter method may help address this gap. Loops That Matter calculates loop scores that indicate how feedback loops contribute to the model behavior over time. This formalizes the quest to explain how model behavior is related to model structure. While feedback loops are commonly used to explain model behavior, such explanation is usually limited to a small number of exemplar runs and is rarely embedded in large uncertainty analyses.
This thesis therefore investigates how Loops That Matter can be integrated into scenario discovery to relate decision relevant model outcomes to the feedback structure of System Dynamics models.
This research develops four approaches. The first approach adds loop scores after a conventional scenario discovery workflow. The second approach uses a single loop score directly in the classification step. The third approach extends this logic by applying time series clustering to multiple loop scores. The fourth approach transforms dominant loop sequences into a network trie. These approaches are tested on two models, a small epidemic model and a more complex energy model. The epidemic model provides a straightforward relation between structure and behavior. The energy model provides more uncertain parameters, more feedback loops, and more diverse behavior.
The results show that Loops That Matter can meaningfully extend scenario discovery when the relation between feedback structure and model behavior is clear. In the epidemic model, loop scores helped explain the differences between mild, moderate, and severe outbreak classes. The dominant loop sequence strips showed how outbreak severity was related to the timing and duration of feedback loop dominance. In addition, clustering the R1 Infection loop score produced classes that could be related back to outbreak behavior and input space. This demonstrates that loop scores can be used both as an ex post interpretation of the model structure, and as a meaningful input to the classification step itself.
The results for the energy model were less successful. Multivariate clustering of two loop scores did not produce classes that were clearly distinct in the model outcomes or well separated in the input space. Similarly, the network trie did not provide a useful classification workflow for scenario discovery. However, the trie did provide an important insight. Namely, that the energy model generated many different dominant loop sequences across the ensemble. Similar behavioral patterns within a key model outcome were also associated with diverse loop dominancy sequences. This suggests that the model exhibits dynamic equifinality, where similar outcomes can arise through different shapes of the feedback loop structure.
The main conclusion is therefore twofold. Loops That Matter can be integrated into scenario discovery, but its usefulness depends on the clarity of the relation between input parameters, and model outcomes. A main contribution lies in the explanatory power of feedback loops after a scenario discovery workflow, as they can help with explaining why classes differ. Their use as a direct basis for classification is more uncertain, especially in larger models. The direct identification of dynamic equifinality is also novel. Previous work has only guessed at its existence, but this research exposes it, and discusses its impact on model based decision support. The thesis contributes a methodological bridge between scenario discovery and structural explanation in System Dynamics, while also showing that equifinality can limit the extent to which model structure can be used to explain decision relevant model behavior.
This creates a tension for System Dynamics models. System Dynamics is a white-box modeling approach in which behavior is understood as the result of its model structure. Growth, decline, oscillation, and collapse are not treated as isolated output patterns, but as consequences of stocks, flows, delays, and feedback loops. When scenario discovery is applied to System Dynamics models without considering this structure, part of the potential of the model is lost.
The recently developed Loops That Matter method may help address this gap. Loops That Matter calculates loop scores that indicate how feedback loops contribute to the model behavior over time. This formalizes the quest to explain how model behavior is related to model structure. While feedback loops are commonly used to explain model behavior, such explanation is usually limited to a small number of exemplar runs and is rarely embedded in large uncertainty analyses.
This thesis therefore investigates how Loops That Matter can be integrated into scenario discovery to relate decision relevant model outcomes to the feedback structure of System Dynamics models.
This research develops four approaches. The first approach adds loop scores after a conventional scenario discovery workflow. The second approach uses a single loop score directly in the classification step. The third approach extends this logic by applying time series clustering to multiple loop scores. The fourth approach transforms dominant loop sequences into a network trie. These approaches are tested on two models, a small epidemic model and a more complex energy model. The epidemic model provides a straightforward relation between structure and behavior. The energy model provides more uncertain parameters, more feedback loops, and more diverse behavior.
The results show that Loops That Matter can meaningfully extend scenario discovery when the relation between feedback structure and model behavior is clear. In the epidemic model, loop scores helped explain the differences between mild, moderate, and severe outbreak classes. The dominant loop sequence strips showed how outbreak severity was related to the timing and duration of feedback loop dominance. In addition, clustering the R1 Infection loop score produced classes that could be related back to outbreak behavior and input space. This demonstrates that loop scores can be used both as an ex post interpretation of the model structure, and as a meaningful input to the classification step itself.
The results for the energy model were less successful. Multivariate clustering of two loop scores did not produce classes that were clearly distinct in the model outcomes or well separated in the input space. Similarly, the network trie did not provide a useful classification workflow for scenario discovery. However, the trie did provide an important insight. Namely, that the energy model generated many different dominant loop sequences across the ensemble. Similar behavioral patterns within a key model outcome were also associated with diverse loop dominancy sequences. This suggests that the model exhibits dynamic equifinality, where similar outcomes can arise through different shapes of the feedback loop structure.
The main conclusion is therefore twofold. Loops That Matter can be integrated into scenario discovery, but its usefulness depends on the clarity of the relation between input parameters, and model outcomes. A main contribution lies in the explanatory power of feedback loops after a scenario discovery workflow, as they can help with explaining why classes differ. Their use as a direct basis for classification is more uncertain, especially in larger models. The direct identification of dynamic equifinality is also novel. Previous work has only guessed at its existence, but this research exposes it, and discusses its impact on model based decision support. The thesis contributes a methodological bridge between scenario discovery and structural explanation in System Dynamics, while also showing that equifinality can limit the extent to which model structure can be used to explain decision relevant model behavior.