Fig. 4: Human-in-the-loop of Dual-GP exploration.
From: Active oversight and quality control in standard Bayesian optimization for autonomous experiments
a Amplitude of the synthetic peak model data, random distortion is added in the raw data in the space x ∈ [0.3, 0.6], leading to resultant scalarizers diverging from the true function. b Examples of a good raw spectrum and distorted spectra. c Single GP exploration result after 50 iterations, d Dual-GP exploration result after 50 iterations, here human experts assess raw spectra every 9 iterations (this can be flexible) and assign a quality score to the raw spectra based on the assessment; low-quality score indicates the raw spectrum is not compatible with the predefined scalarizer function for various reasons. e–g Show that Dual-GP approximated the true function parameter better than the single GP.
