In the absence of Ockham's Razor, curve fitting is prone to overfitting. There is only one robust way of dealing with this problem:
Kolmogorov Complexity Even when you break your data up into a test set and a data set so that you have a way of testing to see if your fit is an overfit, once you have run your test set against your model any attempt to adjust your model is now subject to the criticism that you are now taking your test set into account as part of the data set you are fitting. In practice, this criticism needn't be a show-stopper but it is ultimately problematic. On Fri, Nov 9, 2012 at 2:21 PM, Jones Beene <[email protected]> wrote: > http://www.quantumheat.org/index.php/follow/109-fast-paced-progress > > Interesting technique seen in the second chart - curve matching using > http://zunzun.com/ > > Any comments on this? > > I suppose one rationale is that if the constantan curve shows no initial > anomaly - but it matches the formula up to a trigger point, and then > diverges -then that adds credence. Or else if they will be looking a > similar > shape with a different steepness. Not sure what this curve matching adds. > > On the last page new data is supposedly coming in - live ! but I cannot > get > the page to load > > http://www.quantumheat.org/data/calibrations/Master_Spreadsheet11-9.xls > > anyone have an updated url ? > > Jones >
