Computational aspects of soft computing models to predict sorption isotherms in Nutrimix (weaning food)

Abstract

Soft computing models have been proposed to model adsorption and desorption isotherms for weaning food called fortified Nutrimix, at four different temperatures and over a specific water activity range. The two soft computing paradigms: connectionist models and Adaptive Neuro Fuzzy
Inference System (ANFIS) hybrid models have been investigated. Also, several conventional empirical
sorption models have been explored for fitting the sorption data. The dataset comprised 192 data points. The Error Back-Propagation (EBP) learning algorithm with Bayesian Regularisation and Levenberg-Marquardt optimisation techniques as well as various combinations of connectionist network parameters were employed. The ANFIS model used was based upon the Sugeno-type Fuzzy Inference System (FIS) consisting of 2 to 10 Gaussian membership functions for the input variable. The FIS was generated using the grid partitioning method; which was optimised with the EBP training algorithm. The neuro-fuzzy hybrid model was the best with prediction accuracy of 99.91% as compared to that of the simple connectionist model with accuracy as 97.51% and that of the best classical empirical sorption model, i:e:, Guggenheim-Anderson-de Boer (GAB) model having accuracy as 94.52%. Evidently, soft computing paradigm especially the hybrid ANFIS models outperformed the conventional sorption models for predicting moisture sorption isotherms in Nutrimix powder.