Ted Dunning a écrit : > Hmmm... > > I may have misunderstood our conversation (and therefore probably sounded a > bit incoherent), but I was talking about weights on observations which may > or may not be what you are talking about here.
This is exactly what I am talking about. > > Also, if you *are* talking about the same thing and the vector is a vector > of residuals such that there is one element per observation, then there is a > different impedance mis-match in that I would typically expect the user to > never see this vector of residuals, but rather to pass in a weight vector > beside the observations. I have tried to generalize the concept behind the existing estimation package WeightedMeasurements. This class was designed to represent one observation. The three parts associated with an observation in an optimization context are weight, observed value, expected theoretical value. The residual is observed - expected. In the new design, I have separated the weight from the residual but left the residual as one entity. This was expected to be provided by a user defined MultiObjectiveFunction. If I understand you correctly, the MultiObjectiveFunction should rather compute only the theoretical part and the converter should gather a weights vector, an observations vector, the objective function and perform both the subtraction and the multiplication ? This way, the residual would not be seen and only the pure observation vectors should be provided by user. Is this right ? Luc > > On Thu, Feb 26, 2009 at 11:34 AM, Luc Maisonobe <[email protected]>wrote: > >> The first tool is LestSquaresConverter which can convert a >> MultiObjectiveFunction representing residuals into a simpler >> ObjectiveFunction. I have taken into account both Gilles and Ted >> remarks, the conversion can use no weights at all, vector weights or a >> complete matrix representing both weights and correlations. >> > > > --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
