Efficient Region-of-Interest Detection with Guidance

Abstract: Motivated by the challenges of detecting extremely rare failures in sophisticated circuit design specifications, we study the problem of efficiently identifying multiple disjoint regions of interest (ROIs) — regions of a continuous, high-dimensional parameter space where a complex, expensive-to-evaluate target function falls below or above a pre-specified threshold. We reframe this problem as a Bayesian sequential decision-making problem and, more broadly, as an instance of strategic agent behavior in continuous spaces: an agent that must allocate a hard evaluation budget…