Precision aggregated local models

Abstract Large‐scale Gaussian process (GP) regression is infeasible for large training data due to cubic scaling of flops and quadratic storage involved in working with covariance matrices. Remedies recent literature focus on divide‐and‐conquer, example, partitioning into subproblems inducing functional (and thus computational) independence. Such approximations can be speedy, accurate, sometimes even more flexible than ordinary GPs. However, a big downside loss continuity at partition boundaries. Modern methods like local approximate GPs (LAGPs) imply effectively infinite are both good bad this regard. Model averaging, an alternative maintain absolute but often over‐smooths, diminishing accuracy. Here we propose putting LAGP‐like experts‐like framework, blending partition‐based speed model‐averaging continuity, as flagship example what call precision aggregated models (PALM). Using LAGPs, each selecting from total pairs, our scheme most , linear . Extensive empirical illustration shows how PALM least accurate LAGP, much faster, furnishes continuous predictions. Finally, sequential updating that greedily refines predictor up computational budget.