Toward Handling Uncertainty in Prognostic Scenarios: Advanced Learning from the Past (Prognostic Uncertainty)

The objective of this 1-year, call-winning proposal is to advance our insights in retrospective learning in order to better handle (reduce?) uncertainty in prognostic scenarios; in particular, to provide a reference (standard), which is readily understandable and which prognostic modelers can use in order to inform experts as well as non-experts about the “predictive” power, and its limitation, of their models. To our knowledge, such a reference is not applied at all.

ASA’s three questions motivating the proposed research

1.     Given the one reality – the past – a system has experienced, can we develop an understanding of uncertainty, which proves reliable in the future, i.e., under prognostic conditions, irrespective of how the system unfolds dynamically?

2.     How far into the future will our understanding of uncertainty prove reliable?

3.     What are the practical consequences: Can we apply our improved understanding of uncertainty in ESS (Earth system sciences) models that project future change and even contribute to handling (reducing?) uncertainty in these models?

The research niche

The research proposed by ASA addresses the call’s theme Predictability and Handling of Change. It centers on uncertainty and learning, not on error and perfect projections.

Prognostic uncertainty is at the heart of many disciplines, Earth systems sciences being a prominent one. ASA’s research will focus on global greenhouse gas [GHG] emissions and concentrations, and mean global temperature, not precluding the option of branching off into sub-global scales.

Uncertainty and learning are studied under diagnostic conditions by means of available, observation-based data or estimates; and, thereafter, are tested under “controlled prognostic” conditions by means of part of the available data, which are held back and are used exclusively for checking our improved understanding of uncertainty and learning. We call this experimental procedure “retrospective learning”, which we will explore by applying an advanced methodology (granular computing) and a (or more) standard methodologies to facilitate comparison. Retrospective learning does not aspire knowing the perfect forecast. Instead, retrospective learning accepts an increase in prognostic uncertainty with time in return for our ignorance of the perfect forecast. It is this optimal balance that is sought between our ignorance of the perfect perfect forecast and the increase in prognostic uncertainty.


Print this page

Last edited: 21 April 2015

CONTACT DETAILS

Matthias Jonas

Guest Senior Research Scholar Advancing Systems Analysis Program

Timeframe

June 2015 - May 2016

International Institute for Applied Systems Analysis (IIASA)
Schlossplatz 1, A-2361 Laxenburg, Austria
Phone: (+43 2236) 807 0 Fax:(+43 2236) 71 313