Quantifying Human Perception with Multi-Armed Bandits

Abstract

We study a variant of the continuous multi-armed bandits problem, where the objective is to estimate the sensitivity threshold for an unknown psychometric function Ψ. This setting models the conduct of a psychometric experiment, which aims at quantifying human perception. We show that this setting is akin to hierarchical multiarmed bandits and Black-box optimization of noisy functions, with both significant similarities and key differences. We introduce a new algorithm, DOS, for Dichotomous Optimistic Search, that efficiently solves this task, and show that DOS outperforms recent methods from both Psychophysics and Global Optimization for non Gaussian Psychometric functions in our experiments.1