Vitaly Kupisk schrieb: > > Thanks for all the replies. I am passing along your comments to my > client but my job is just to find an easy way to implement orthogonal > L1 regression in C#. > I transcribed toms algorithm 478 (L1 optimization of a linear system > of equations) into C#, but it doesn't do orthogonal distance, so I am > following up with rotating the axes so that the hyperplane is > orthogonal to the "y" axis and trying again until the changes in the > parameters are small. Problem is, things get unstable with more than > a few parameters, so I am using ad hoc methods to stop the params from > becoming huge. That's hardly satisfying. > > Any suggestions? > > So, nobody tried to read the Bargiela/Hartley paper? > > Vitaly
Hi Vitaly, I'm not experienced with the term "orthogonal regression". But it sounds to me like a relative to the principal component analysis. What makes me wonder is your remark on "parameter becoming huge". Maybe my components-analysis-tools contain an instrument to do this type of regression. In a quick overview of an article just found with google, it seems to me, that only the assumption of item-specific errors is the difference to common mult regression. If it's just that, then I could provide a simple example, how to do that. Gottfried Helms . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
