Exploration of Taste Generation Parameters using Preferential Bayesian Optimization

By using taste sensors and taste substance mixing devices, taste can be described as a mixing ratio of taste substances (taste generation parameters) and can be generated by controlling these parameters. This allows individuals to explore the parameter space to obtain a target taste or a taste that matches their own preferences. However, this exploration requires the user to understand each taste substance and to be able to describe their preferences in numbers or words. As a result, the search range tends to be biased toward mixtures the user is already familiar with, or the search becomes difficult because the user cannot effectively express the desired taste numerically or linguistically. In this paper, we propose a method for exploring taste generation parameters using a pairwise comparison UI. Users only need to taste two presented tastes and answer "which one and to what extent it is preferable." This enables exploration even without knowledge of taste substances or the ability to express preferences in numbers or words. By using preferential Bayesian optimization, we have made it possible to converge toward the target parameters within approximately 15 to 20 comparisons in a five-dimensional space. In addition, we prototyped and utilized a high-speed cutlery-type taste-mixing device to enable rapid iterations. The effectiveness of the proposed method was verified through several exploration case studies.