On Wednesday, 31. January 2007 22:54, Matthias Troyer wrote:
Which accumulators would then have a non-trivial "combine" method?
There are two types of accumulators that I can think of immediately:
Thanks for your answer, Matthias; I already suspected that you had some accumulators in mind which I don't know, thanks for giving examples. (I will have to look them up somewhere.)
1. approximate median and quantile estimators such as the p^2 methods cannot be combined but are very important for some applications
2. accumulators for correlates samples, e.g. autocorrelation estimators, or jackknife-bins cannot easily be combined. Instead of just merging the results one has to store the results from the individual accumulator_sets separately and combine them only in the final evaluations. Please keep in mind that correlated samples are very common.
For 1. there is simply no way, and that is just it - one might either have to make this cause a compile-time error or drop the median and quantile estimators when combining.
I'd say that a compile-time error would be perfectly OK. It's just that having a method for combining accumulators would open your library up to many more applications. (E.g., in my case I would be interested in using it for image analysis, where I am using sort of accumulators for measuring properties of image regions, which are eventually merged.)
For 2. the problem is that combined accumulator sets need a different data structure than individual ones.
I would be very happy if you found a way to integrate a simple way for combining soon, and postpone a solution of that latter problem. Right now, I don't think it would be impractical to require the user of your library to specify in advance that additional data structures shall be maintained in case he/she wants to combine accumulators later. Ciao, / / /--/ / / ANS