ニューラルネットワークによる吟醸酒の官能評価値から総合評価値の推定

This paper discusses modeling of the total evaluation process of sake (Ginjo) using a neural network (NN) and multiple regression analysis (MRA). In each method, the total evaluation of 61 Ginjo samples was estimated from data sets of the sensory evaluations of 7 characteristics. The values of the performance index, J, based on errors between actual and estimated values were smaller in the NN model than in MRA. In particular, the NN model with 3 input variables selected from the 7 characteristics evaluated had a simple structure and the J value was sufficiently low. Samples with a large error were found to have the characteristic of "extraordinary flavor". Addition of "extraordinary flavor" as a new input variable in the NN decreased the error of the model. The results suggested that 4 variables (color, flavor base, aging and extraordinary flavor) are important in the total evaluation of Ginjo, and that a nonlinear modeling method such as a NN is effective in analyzing the sensory evaluation process.