Accueil > Actualités > Séminaires > Séminaire de Frank selten


Titre : Improving predictions by combining imperfect models into a supermodel
Nom du conférencier : Frank Selten
Son affiliation : Royal Netherlands Meteorological Institute
Laboratoire organisateur : LMD
Date et heure : 22-01-2015 14h30
Lieu : ENS - Paris - 29 Rue d'Ulm - salle 235C
Résumé :

In climate and weather prediction multi-model ensembles have proven their benefits over single model predictions as they err in different ways. Forecasts from different models are combined into a single probability forecast through various statistical techniques. In a new approach we aim to take the multi-model approach a step further by allowing the models to interact during the simulations.

This approach is illustrated with simple chaotic systems in a perfect model setting. The interaction between the models is achieved by adding additional terms to the tendency equations that damp the differences between the model states. A new dynamical system is created, a supermodel (SUMO), with as additional parameters the strength of the connection terms. For strong enough connections the models within the supermodel synchronise on a single solution that, after training the connection coefficients on past, observed data, produces improved forecasts.

Recent applications of the SUMO approach using more complex climate models will be presented and unexpected dynamical surprises for weak connections will be revealed.