On July 16, 2003 11:50 pm, Zlatin Balevsky wrote: > Possible improvements over using a fixed number of reference points and > straight lines between them would be :
Once ng is fully working, this is worth playing with. We need to optimize it for good estimations with a minimum of data. Maybe we can get the 'shape' of the initial curve from a global estimator and change its co-efficients to make the node look better and adjust as we route. Think we want to consider more than 10 points. Ed > I. Non-linear estimators: instead of straight lines, user some sort of > interpolation to create the graph : > a. If the number of points is fixed around 10, we could use newton > interpolation or chebyshev polynomials. > b. For larger number of more dynamic reference points we could use cubic > or quadric splines - these have the benefit that when one reference > point moves only its immediate neighbors need to be re-calculated which > spares us some cpu time while providing even better interpolation than > the polynomials. > c. If after some experimental data we notice that the reference points > tend to form periods, we could even use FFT for the very best of the > interpolating world ;). > > these are only the interpolations that I could remember from a Numerical > Methods course I took last year; when the textbook finds itself I'll > probably come up with more information. Each of the above methods beats > using straight lines by far though. > > > AND/OR > > II Dynamic number of reference points based on specific events: > a. adding new reference point for every new sample (yeah right) > b. adding new reference point only if the estimate turns out to be 9x% > accurate and removing one if the estimate is less than x% accurate > c. adding a new ref point if the sample is x distance from the > specialization of the other reference points (to encourage multiple > specialization). and removing one if it they start clustering too much > d. the opposite of c > > _______________________________________________ > devl mailing list > [EMAIL PROTECTED] > http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl _______________________________________________ devl mailing list [EMAIL PROTECTED] http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
