In universal kriging the general model is $$y(x)=h(x)^{t}\boldsymbol\beta + f(x)+\epsilon(x)$$ where $h(x)^{t}$ is a regression function such as $\left [1,x,x^{2} \right ]$ and $\boldsymbol\beta$ are the set of regression coefficients and $$f_{i}(x)\sim GP(0,K(x,x^{'}))$$ and $\epsilon(x)$ is normal with zero mean and $\sigma^{2}$ variance

I am wondering whether $h(x)^{t}\boldsymbol\beta$ is assumed deterministic and thus to find $\boldsymbol\beta$ we need to optimize the the multivariate gaussian likelihood function to find the hyperparameters in $K(x,x^{'})$ and $\boldsymbol\beta$ where the joint distribution of $y(x)$ is $$y(x)\sim GP(h(x)^{t}\boldsymbol\beta,K(x,x^{'})+\sigma^{2}I)$$ and thus the predictive distribution is $$y^{*}|\mathbf{y} \sim N(K_{*}K^{-1}(\mathbf{y}-h(\textbf{x})^{t}\hat{\boldsymbol\beta}),K_{**}-K_{*}K^{-1}K_{*}^{T})$$

However, In the famous Gaussain process for machine learning book Ch2 page 28, it seems that $\boldsymbol\beta$ is a random variable and the likelihood function and predictive distribution are different. I need help understanding the distinction between the method I described above (if valid) and the approach in the book. The main related text is attached as a pic $$$$

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.