A key issue in Gaussian Process modeling is to decide on the locations where measurements are going to be taken. A good set of observations will provide a better model. Current state of the art selects such a set so as to minimize the posterior variance of the Gaussian Process by exploiting submodularity. We propose a Gradient Descent procedure to iteratively improve an initial set of observations so as to minimize the posterior variance directly. The performance of the technique is analyzed under different conditions by varying the number of measurement points, the dimensionality of the domain and the hyperparameters of the Gaussian Process. Results show the applicability of the technique and the clear improvements that can be obtain under different settings.
展开▼
