A generalized many‐objective optimization approach for scenario discovery

Abstract

Scenario discovery is a model‐based approach for scenario development, aimed at finding one or more subspaces within the uncertainty space associated with a model that is decision‐relevant. These identified subspaces can subsequently be translated into narratives or shared in other ways a broader participatory process. Finding such as subspace involves solving a three‐objective optimization problem. A subspace should cover many of the decision relevant model runs, while containing as few as possible nondecision relevant model runs, and being easy to interpret. Existing techniques for scenario discovery, however, focus only on finding a subspace that minimizes the number of nondecision relevant model runs. Adopting a single objective optimization approach for a many‐objective optimization problem implies that the full trade‐off space is not identified. In this paper, we introduce a many‐objective optimization approach for scenario discovery. We compare this with an improved usage of Patient Rule Induction Method (PRIM) for identifying the multidimensional trade‐offs amongst coverage, density, and interpretability. We find that the many‐objective optimization approach produces results which slightly dominate those of the improved version of PRIM on all three objectives. Qualitatively, however, both approaches identify essentially the same subspaces. The prime benefits of the many‐objective optimization approach are its potential in bringing additional scenario relevant concerns such as consistency or diversity into the scenario discovery framework, as well as its ability to avoid overfitting. Potentially more important, it also paves the way for future work on using more sophisticated metaheuristic optimization approaches for scenario discovery.

Defining Robustness, Vulnerabilities, and Consequential Scenarios for Diverse Stakeholder Interests in Institutionally Complex River Basins

“…Scenario discovery is most typically performed using the Patient Rule Induction Method (PRIM) (Friedman & Fisher, 1999) or Classification and Regression Trees (CART) (Breiman, 1984). One of the main criticisms of these methods is that they rely on orthogonal subspaces to classify the parameter space into success or failure, making their application to nonlinear systems inadequate (Kwakkel, 2019). Logistic regression is a common technique in binary classification problems that overcomes this limitation by using a nonlinear function to describe the probability that a SOW belongs to the scenarios that lead to failure (Lamontagne et al, 2019; Trindade et al, 2019).…”

Uncertainty Analysis in Multi‐Sector Systems: Considerations for Risk Analysis, Projection, and Planning for Complex Systems

“…Scenario discovery is a computer‐assisted approach to scenario development that identifies regions of the uncertainty space that are tied to outcomes of interest (Bryant & Lempert, 2010; Kwakkel, 2019). These methods begin by sampling possible values of uncertain factors, which are then simulated using one or more system models to generate a large ensemble of potential future system conditions.…”

Identifying Actionable Compromises: Navigating Multi‐City Robustness Conflicts to Discover Cooperative Safe Operating Spaces for Regional Water Supply Portfolios

“…Beyond the discovery of robust planning alternatives, exploratory modeling approaches also enable decision makers to discover consequential future scenarios that cause vulnerabilities in candidate alternatives (Lempert et al, 2006). Typically, the process of scenario discovery utilizes data mining, machine learning, and/or multi-objective optimization to delineate critical thresholds in the uncertainty space that cause proposed alternatives to fail to meet performance criteria defined by stakeholders (Bryant & Lempert, 2010;Groves & Lempert, 2007;Kwakkel, 2019). In the MORDM framework, scenario discovery is a core component of an iterative problem formulation process, which allows stakeholders to learn from exploratory modeling results to improve their understanding of their preferences and potential actions (Kasprzyk et al, 10.1029/2019WR025462 2013; .…”