A dependent sampling approach to Scenario Discovery

Master thesis (2020)

Authors

I.S. van Droffelaar Technology, Policy and Management

Contributors

J.H. Kwakkel Policy Analysis - TPM (mentor)
N.Y. Aydin System Engineering - TPM (mentor)
Martijn Warnier Multi Actor Systems (mentor)
Faculty
Technology, Policy and Management, Technology, Policy and Management
Copyright: © 2020 I.S. van Droffelaar
More Info
expand_more

Published Date

26-08-2020

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Technology, Policy and Management

Abstract

The use of scenario planning has a long history in decision-making and public policy (Bryant and Lempert, 2010). Traditional scenario planning, as, for example, used by the Shell Scenarios group, provides tools to communicate and characterize uncertainty,allowing decision-makers to anticipate the future and create more robust strategies (Bradfield et al., 2005).

However, the classical qualitative approach, where scenario narratives are developed, has some severe limitations. While this approach provides results which are readily communicated to decision-makers, it often overlooks truly unexpected (but plausible) scenarios (Kwakkel and Cunningham, 2016). Besides, they are not readily implemented for problems where the structure is also disputed (Bryant and Lempert, 2010).

Scenario discovery, a quantitative, brute-force approach to scenario development,developed by Bryant et al. (2010) aims to address these limitations. This approach has been successfully applied to numerous grand challenges, among others climate change (Kwadijk et al., 2010), sustainable water management (Haasnoot et al., 2011), and global natural resource management (Kwakkel et al., 2013).

The original approach to quantitative scenario discovery relies on three consecutive steps: sampling, labeling and searching for subspaces leading to the regions of interest.The sample is typically collected using Latin Hypercube Sampling (Kwakkel, 2017). For uniform, independent sampling, the chance that a random sample falls within the region of interest decreases quadratically with increasing size of the bounds (Vrugt, 2016). Therefore, if the prior distribution cannot be estimated, or the problem at hand demands a wide range of possible values, and the problem has a high number of dimensions, these sampling techniques require an unreasonably high number of samples to adequately delineate the region(s) of interest.

Therefore, this thesis proposes using dependent sampling, which concentrates the sampling on the regions of interest, rather than attempting to represent the entire input space. After convergence, the resulting sample approximates the true posterior distribution. By performing a Kolmogorov-Smirnov test for uniformity, the uncertain factors leading to the behavior of interest can be derived. The relevant ranges of parameter values can be recognized and communicated to decision-makers.