|
Abstract:
|
Gaussian Process (GP) models have long been used as a flexible method for nonparametric regression. Despite their accuracy, they are infeasible for larger data sets due to the scaling of the computational burden and the storage requirements. Typical methods to combat intractability have focused on splitting the larger GP into smaller problems by partitioning the domain space, or partitioning the data themselves into discrete sets. These methods get around the functional dependence of the individual GP models by assigning no weight to most of the models during prediction, or selecting global points to have low dependence, however each solution has its drawbacks. Partition models maintain accuracy, but lose continuity at the boundaries. Typical averaged models, on the other hand, maintain absolute continuity in both the mean and variance surface while over-smoothing the function as a whole. Using Local Approximate Gaussian Processes as a method to build local experts, the Precision Aggregated Local Models approach bridges the gap between these two approaches to create a locally accurate global model that maintains absolute continuity.
|
