Abstract:
A framework for regime dependent dynamic Bayesian networks for assessing climate risk
Terence O'Kane
Co-authors Dylan Harries, Mark Collier, Jana de Wiljes
CSIRO
Our current understanding of the future climate relies on general circulation models coupling the ocean, atmosphere, land, and sea ice domains. If these models can be shown to reasonably reproduce the observed climate modes of variability and the teleconnections between them, then inference of their associated structural causal models and the systematic changes in those networks in response to anthropogenic changes in the boundary conditions (radiative forcing) would enable a deep understanding of the potential future climate risks for a given emissions scenario. Here, we describe a framework to combine time-homogeneous dynamic Bayesian network models (DBNs) with regularized optimisation of a cost functional whose parameters are the regime dependent parents and link coefficients of direct acyclic graphs (DAGs) conditioned on a given regime. The DAGs are constructed from MAP estimates generated from timeseries of the major modes of climate variability to determine probabilistic graphical models. Reversible jump Markov Chain Monte Carlo is employed to provide a quantification of the uncertainty associated with the selection of a single network structure. The aim is to extend Bayesian inference via DBNs to nonstationary flow problems via embedding within an optimal regime learning process. Clearly this is a challenging problem, and we will discuss the advantages and limitations of the approach.
What is the Joint Oslo Seminar (JOS):
- Atmospheric and climate sciences have a stronghold in Oslo among the four institutions University of Oslo, the Meteorological Institute, CICERO and NILU.
- This joint seminar invites renowned international experts to contribute to an informal series of lectures, meant to create interaction with the Oslo atmospheric and climate science community on recent highlights and analysis in the field.
- All seminars will be held on Thursdays (12:15 pm -1 pm).