Heteroscedastic Multi-Task-GPR

[BerndSchuscha]

Hi,

I am looking similar to "FixedNoiseGaussianLikelihood", but for multiple dimensions.

What I have is N data points with 3 output variables.
I know the measurement error from these 3 outputs that they are correlated, but the correlation and error is varying with each measurement. For one measurement, this can be quite well approximated with a multi(tri-)variant Gaussian distribution. That means I can give for each of the N data points a 3x3 covariance matrix.

Could someone give me a point who this can be done?