I've got a real-world problem that I'm trying to solve with as few computations as possible. In this 2D problem, everything is parameterized on a unit square, $f(x,y); x, y \in [0, 1]$. Anecdotal evidence suggests that there is a smooth, potentially non-convex boundary in this range that splits the region into two domains $A$ and $B$ where $f_A \ll f_B$.

Given that it is very expensive to compute a single value of $f$ and that I've sampled $n$ points already, how can I choose the next point that will give me the most new information?

My rudimentary understanding of Gaussian Processes suggests that I can use them to sample from a posterior. From there I think I can find the region of most variance ... but I'm not at all sure if this is even possible.

  • $\begingroup$ The problem you want to solve is not clear. Do you want (1) an estimate of $f$ and the fact that there are regions $A$ and $B$ is additional (prior) information or are you (2) seeking a classification of points whether they belong to $A$ or $B$ or (3) would you like to test the hypothesis that regions $A$ and $B$ exist? $\endgroup$ – g g Jan 11 at 22:28

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.