Authors: László Sipos, Viktor Losó, Ákos Nyitrai, Zoltán Kókai, Attila Gere
According to our knowledge, there are only a few publications in available literature sources on the sensory characteristics and consumer preferences of sweet corn varieties. In our research, practical application of artificial neural networks (ANNs) is presented. In our study, 41 frozen sweet corn varieties were evaluated by a panel of expert sensory panelists (14 persons), by the method of profile analysis (MSZ ISO 11035:2001; ISO 13299:2003), on an unstructured scale of 0 to 100, then, in large-scale tests, 6 of the 41 varieties were evaluated by consumers (167 people) according to preference, on a structured scale of 1 to 9. Artificial neural networks require large amounts of data, therefore, on the expert and consumer data for the 6 varieties, 1,000 Monte Carlo simulations were run. 80% of the resulting dataset was used to train the created neural networks, and 20% was utilized to test them. The best prediction was given by the 4-node multi-layer feedforward neural network (MLFN), the smallest residues were obtained in this case during the training and the test, which were also validated by predictions on random numbers and cross-checking. Preference values of the other 35 corn varieties were predicted by this model. The most preferred variety was ‘Shinerock’ (8.46), while the least preferred ones, according to the predictions, were ‘Madonna’ and ‘Rustler’, with and average preference value of 2.7 (on a scale of 1 to 9).
During the establishment of the artificial neural network model, product characteristics that are the main drivers of consumer acceptance were successfully identified: sweet taste, global taste intensity and juiciness. In general, it can be stated that prediction of the preference of different varieties is made possible by the validated product-specific artificial neural network presented.
2. Introduction and literature overview
For the development of artificial neural networks, revealing the analogy, the structure and functioning of the human nervous system was of key importance. Neural network programs were originally developed as a model for the nervous system, where signals coming from other neurons are collected by the inputs, summation is carried out by the processing unit (neuron), and then, depending on the result, the signal is transmitted by the outputs , , , , . A breakthrough in the research of artificial neural networks was achieved by the work of Hopfield , and Rumelhart et al. , in which non-linear mapping was achieved by the dynamic modeling of neural network programs, as well as feedback between the outputs and the inputs.