10-13-06 - 1

I've done everything reasonably obvious with a linear predictor. There's a nasty trap in tweak factors, in that you are somewhat hard-coding a best fit to the data into your app with the tweak constants. That's fine as long as the data extrapolates well - eg. the fit on the "probe" is also close to the best fit for the "qualifying", but that will only be true if the data's not really strange and your clusters that you fit for are big enough. Any time you fit tweaks that only apply to a handful of elements you are in serious danger of fitting them to the training set too much and not really fitting the general case.

BTW I find the whole "Prize" aspect sort of annoying because it means no one is talking about their approaches publicly. Of course, even in fields where there is no prize, people are still super-secretive, because they always think they have some brilliant idea which is going to make them famous in the academic world, or get them patents and make them rich, blah blah blah.

Oh, I still could use some volunteers to run the heavy process if you have a fast machines that are idle.

No comments:

old rants