Towards Handling Uncertainty in Prognostic Scenarios: Advanced Learning from the Past

Zebrowski, P. ORCID: https://orcid.org/0000-0001-5283-8049, Jonas, M. ORCID: https://orcid.org/0000-0003-1269-4145, & Jarnicka, J. (2017). Towards Handling Uncertainty in Prognostic Scenarios: Advanced Learning from the Past. IIASA Working Paper. IIASA, Laxenburg, Austria: WP-17-015

Preview

Text WP-17-015.pdf - Published Version Available under License Creative Commons Attribution. Download (5MB) | Preview

Abstract

In this report we introduce the paradigm of learning from the past which is realized in a controlled prognostic context. It is a data-driven exploratory approach to assessing the limits to credibility of any expectations about the system’s future behavior which are based on a time series of a historical observations of the analyzed system. This horizon of the credible expectations is derived as the length of explainable outreach of the data, that is, the spatio-temporal extent which, in lieu of the knowledge contained in the historical observations, we are justified in believing contains the system’s future observations. Explainable outreach is of practical interest to stakeholders since it allows them to assess the credibility of scenarios produced by models of the analyzed system. It also indicates the scale of measures required to overcome the system’s inertia. In this report we propose a method of learning in a controlled prognostic context which is based on a polynomial regression technique. A polynomial regression model is used to understand the system’s dynamics, revealed by the sample of historical observations, while the explainable outreach is constructed around the extrapolated regression function. The proposed learning method was tested on various sets of synthetic data in order to identify its strengths and weaknesses, and formulate guidelines for its practical application. We also demonstrate how it can be used in context of earth system sciences by using it to derive the explainable outreach of historical anthropogenic CO2 emissions and atmospheric CO2 concentrations. We conclude that the most robust method of building the explainable outreach is based on linear regression. However, the explainable outreach of the analyzed datasets (representing credible expectations based on extrapolation of the linear trend) is rather short.

Actions (login required)

View Item