Classi®cation of aged wine distillates using fuzzy and neural network systems C.G. Raptis, C.I. Siettos, C.T. Kiranoudis * , G.V. Bafas Department of Chemical Engineering, National Technical University, GR 15780 Athens, Greece Received 22 November 1999; accepted 16 May 2000 Abstract The classi®cation of aged wine distillates is a non-linear, multi-criteria decision-making problem characterized by overwhelming complexity, non-linearity and lack of objective information regarding the desired ®nal product qualitative characteristics. The most ecient solution for the evaluation of aged wine distillates estimations with emphasis on the properties of the aroma and the taste, when an appropriate mathematical model cannot be found, is to develop adequate and reliable expert systems based on fuzzy logic and neural networks. A fuzzy classi®er and a neural network are proposed for the classi®cation of wine distillates for each of two distinct features of the products namely the aroma and the taste. The fuzzy classi®er is based on the fuzzy k-nn algorithm while the neural system is a feedforward sigmoidal multilayer network which is trained using the back-propagation method. The results show that both fuzzy and neural classi®cation systems performed remarkably well in the evaluation of the aroma and the taste of the products. Ó 2000 Elsevier Science Ltd. All rights reserved. 1. Introduction Food processing is most often characterized by severe complexity, non-linearity and lack of objective infor- mation regarding the qualitative ®nal product charac- teristics. The increasing and strong need for total quality management in food industries has rendered the con- struction of ¯exible and robust automotive decision- making systems for product evaluation (Davidson & Smith, 1995; Zhang & Chen 1997; Geeraerd, Herre- mans, Cenens, & Van Impe, 1998; Linko, 1998). During the last 20 years, fuzzy logic based systems and arti®cial neural networks have emerged as two very promising and powerful approaches for handling and solving real decision-making problems. The major advantage of fuzzy logic based and neural network systems over traditional techniques, is their eciency in handling complex and non-linear problems due to their inherent non-linear character, their capability of adaptation and integration of expert knowledge. They can be used either in addition to other approaches or as self-reliant meth- odologies providing thereby a plethora of alternative structures and schemes to work out. This provides ef- fective and ¯exible means of capturing the complex, imprecise and non-linear nature of practical decision- making and classi®cation problems. In contrast to other approaches that are mostly quantitative approaches, fuzzy logic addresses the problem of data classi®cation in a rather uni®ed qualitative and quantitative manner. Classi®cation is the procedure for the assignment of input data into some speci®ed classes, based on the in- formation acquired by training. The more and accurate a priori input information is known about the problem, the more precise the classi®cation algorithm can be constructed to re¯ect the actual situation. Most often, the derivation of an exact mathematical model relating the process to product characteristics is a very expensive or even impossible task due to the in- herent ambiguity that many of the characteristics may possess. Arti®cial neural networks utilize a strictly nu- merical, black-box structure for input±output data classi®cation while fuzzy logic may incorporate human knowledge and expertise in the form of a rule base. Utilising the concept of fuzzy sets, products with dif- ferent degrees of possession of certain qualitative properties are allowed to have degrees of membership in more than one class. The advantage of the fuzzy ap- proach in comparison to other classi®cation black-box techniques is that it provides, through class membership, a more representative picture of the system, providing thereby more information. Journal of Food Engineering 46 (2000) 267±275 www.elsevier.com/locate/jfoodeng * Corresponding author. 0260-8774/00/$ - see front matter Ó 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 2 6 0 - 8 7 7 4 ( 0 0 ) 0 0 0 8 7 - X