Pareto Active Learning with Gaussian Processes and Adaptive Discretization

Andi Nika, Kerem Bozgan, Sepehr Elahi, Cagin Ararat, Cem Tekin

Abstract: We consider the problem of optimizing a vector-valued objective function $\bs{f}$ sampled from a Gaussian Process (GP) whose index set is a well-behaved, compact metric space $(\XX,d)$ of designs. We assume that $\bs{f}$ is not known beforehand and that evaluating $\bs{f}$ at design $x$ results in a noisy observation of $\bs{f}(x)$. Since identifying the Pareto optimal designs via exhaustive search is infeasible when the cardinality of $\XX$ is large, we propose an algorithm, called Adaptive $\bs{\epsilon}$-PAL, that exploits the smoothness of the GP-sampled function and the structure of $(\XX,d)$ to learn fast. In essence, Adaptive $\bs{\epsilon}$-PAL employs a tree-based adaptive discretization technique to identify an $\bs{\epsilon}$-accurate Pareto set of designs in as few evaluations as possible. We provide both information-type and metric dimension-type bounds on the sample complexity of $\bs{\epsilon}$-accurate Pareto set identification. We also experimentally show that our algorithm outperforms other Pareto set identification methods on several benchmark datasets.