A nonlinear multiple regression model of taste sensor data for components in sake

A nonlinear function that expresses the relationship between taste sensor data and components in sake was approximated using a polynomial of Legendre functions. First, the number of components in sake was reduced using principal component analysis. Second, the number of Legendre functions of the polynomial and their degrees were selected using a genetic algorithm. Third, the coefficients of the polynomial were calculated using multiple regression analysis. The approximation error was estimated using cross‐validation, and the number of Legendre functions and their degrees were optimized so as to maximize the generalization of the polynomial. As a result, sufficiently small approximation errors were obtained, and the explicit relationship between taste sensor data and components in sake was clarified using the polynomial. Furthermore, it was possible not only to confirm the taste sensor response but also to improve manufacturing processes of sake using the estimates of the variations in the taste sensor data.