In order to establish a prediction model for the taste of gelatin gummy， 90 agar-gelatin fondant samples with different ratios were subjected to texture profile analysis (TPA) and taste evaluation experiments. Then the comprehensive taste score of each sample was obtained by fuzzy mathematical method， using the texture parameters as input factors and the comprehensive taste score as output factors to build a neural network model. Through multi-model training comparison， the optimal topology is determined as 6×8×1， and the original model is optimized by using genetic algorithm to establish a GA-BP neural network prediction model. After the training of the experimental sample data， the results showed that the prediction model had a good fit， the correlation coefficient R of all the data was greater than 0.9， and there was no overfitting state. Meanwhile， the predicted taste errors for the three commercially available gelatin gummy samples were all within ±5% (maximum 4.11%)， which could meet the purpose of taste prediction. The model has good application and reference value， which can provide a new means to improve the quality of gelatinous fondant and can also give methodological reference for predicting the taste of other foods.