, ,
By now, drawing inference from predictions is considered as the most powerful statistical approach to generalization. While that is already known for over 30 years (Leo Breiman 2001), most practitioners and funders still ignore this basic scheme, and get hold up instead with lower performing mechanistic, parsimonious, linear and/or similar (‘homebrew’) models that do not live up to relevant scrutiny and performance metrics. The consistently under-predicting summer sea ice models for the Arctic make that very clear for years, as do the Population Viability Analysis (PVAs) by IUCN, medical disease models, permafrost models and many of the IPCC forecasts. This leaves us with disastrous impacts.
This presentation presents over 50 projects that have already used Machine Learning, namely ensemble algorithms like bagging and boosting, across a diverse set of applications worldwide to obtain best possible ‘classifiers’ (predictions) explicit in time and space. I will give a synthesis and an overview from these reviewed and published (polar) studies. Further, I will explain some of the roadblocks of the current scientific quagmire. Finally, I will present a good way forward on how to make predictions better and how to create a best-possible science-based institutional model culture for pre-cautionary decision-making and for reaching ‘true’ global sustainability while many problems (ecological, sociological, economical) currently appear to go ‘through the roof’.