Sensory evaluation data obtained from experts were analyzed by numerical methods. The aim of this study is to identify a model that can objectively estimate the sensory evaluation results based on the concentrations of components in sake. To this aim, a learning model in which Dempster-Shafer’s measure was learned by genetic algorithm (GA) was constructed. The learning process was performed by discovery of the assignments of basic probabilities according to the decrease in error between the observed and estimated data. When the model was compared with back propagation and multiple regression analysis by cross validation, the predictive faculty of the present model was as good as that of back propagation. The experiential rule by experts for time series data of sensory evaluation could be more sufficiently explained by the present model than by back propagation. The main advantage of this model was that its predictive faculty was compensated by Bayesian probabilities.