Here’s a link to a pre-print of our safe optimization paper that we have just submitted to be presented at this year’s Cognitive Science Conference:
Here’s what we found in a nutshell:
We assess how people try to optimize functions when they have to avoid outputs below a certain threshold at all costs. We find that participants are well-described by a Gaussian Process-based Safe Optimization routine, in which they seem to approach the problem step-wisely by first assessing whether or not a point is safe and then maximizing within this safe set. This result shows that participants try to resourcefully maintain homoeostasis (not sampling below the threshold instead of focusing on gains only) by applying short cuts (a tree-like, step-wise strategy) based on (close to-)optimal models (GP Safe Optimization).