Development of neuro-fuzzifiers for qualitative analyses of milk yield

In fuzzy logic, the calibration of membership functions is often cumbersome and the use of univariate models, such as triangular or trapezoidal functions, is not always ideal. In this study, artificial neural networks were used as multivariate models for the fuzzification of milk yield, with the aim of providing a more convenient approach than traditional techniques. The objective was to develop a neuro-fuzzifier that would mimic an expertʹs process of assigning dairy production records to fuzzy milk-yield sets and assessing their corresponding degrees of membership. Data consisted of 313 dairy production records fuzzified by an expert, according to herd-average 305-day milk yield, cow lactation number, days in lactation, standard milk yield on test day, test day milk-yield and deviation of test day production from a standard value. Five fuzzy sets for milk yield were predetermined as being very low, low, medium, high and very high. Two neuro-classifiers were first trained to identify the sets to which a record belonged and then, four specialized networks were devised to predict the corresponding degrees of membership. In order to evaluate this approach, network classification and predictions were compared with those obtained from a previously developed univariate model and results showed that it was generally better in both the classification of milk yields into fuzzy sets, and the determination of corresponding degrees of membership. These results suggest that the approach of neuro-fuzzification may be superior to traditional methods for applications that require multivariate fuzzifiers.