Andrew Gelman has me thinking about how you'd want to model-check climate models. It seems as if the predictive machine-learning of withholding parts of the data set might be inappropriate in his framework, both because you're modelling a process and not simply a functional relationship (so withholding a random subset of data points from the domain doesn't prevent over-fitting--and the data you have aren't gathered independently from the input domain in the first place) and because Gelman advocates an iterative approach, which seems reasonable. So, how would you check them? Gelman talks some about comparing the generated model to the original data inputs and seeing whether it "looks" like the original data, particularly in ways that you weren't fitting for. That seems like a start but I'm not sure how to do it, or if it's sufficient.
Anyway, in this context I read about checking climate models by comparing their regional predictions for climate change, which naturally are much richer than the single "global temperature change" metric. Seems like a reasonable approach. The models are compared against each other and against simple statistical models which incorporate no knowledge of physics or climatology.
-Max
--
Hahahahaaaa!!! That is ME laughing at YOU, cruel world.
-Jordan Rixon
I could not love thee, dear, so much,
Loved I not Honour more.
No comments:
Post a Comment