Learnable parameters for mean prior.

[lovecodng]

I am wondering if its possible to fit the following model using multitask GP, where all hyper-parameters should be learned by the model.

[gpleiss]

This should be possible since y1 and y2 are jointly normal.

Since y1 | y2 = N(const + a1 y2 - b1 ln(x) | k_1), we have that y1 = N(const + a1(const - b2 ln(x)) - b1 ln(x) | k1 + k2 + (const - b2 ln(x))^2).

You can implement a MultitaskMean function that parameterizes the mean in this way.