ORIGINAL ARTICLE Prediction of feed abrasive value by artiﬁcial neural networks and multiple linear regression M. A. Norouzian • S. Asadpour Received: 6 June 2010 / Accepted: 11 March 2011 / Published online: 6 April 2011 Ó Springer-Verlag London Limited 2011 Abstract In order to evaluate rapid testing methods based on the relationship between feed abrasive value (FAV) and physicochemical properties (particle size, bulk density, dry matter (DM), soluble dry matter, water-holding capacity (WHC), ash, crud protein, neutral detergent ﬁber (NDF), physically effective NDF and non-ﬁbrous carbohydrates (NFC)), 100 empirical dataset were used. Relationships were investigated using multiple linear regression (MLR) and artiﬁcial neural networks (ANNs). The mean relative error was signiﬁcantly (P \ 0.01) lower for ANN than MLR model. Globally, the non-linear ANN model approach is shown to provide a better prediction of FAV than linear multiple regression. Keywords Feed abrasive value  Multiple linear regression  Artiﬁcial neural network 1 Introduction The gastrointestinal tract of the newborn ruminant under- goes appreciable changes between birth and the time when the rumen becomes functional. The lining of the rumen is composed of tissue (stratum corneum) that keratinizes on the surface [1]. Use of concentrate feed in starter diets results in increasing incidence of ruminal parakeratosis and a thickening of the stratum corneum in calves [2] and lambs [3]. Parakeratosis creates a physical barrier, restricting absorptive surface area and volatile fatty acid absorption, reducing epithelial blood ﬂow and rumen motility, and causing papillae degeneration and sloughing in extreme cases [4]. Initial evidence of parakeratosis is papillae clumping and branching, followed by papillae degeneration and sloughing [5, 6]. Concentrate diets with small particle size and low abrasive value [7] increased volatile fatty acid production, decreased rumen buffering capacity, and subsequently decreased rumen pH [5] are factors commonly associated with occurrences of para- keratosis [8]. Feed abrasive value (FAV) is deﬁned as a feed efﬁcacy in physically removing keratin and/or dead epithelial cells from the rumen epithelium [9]. Greenwood et al. [9] designed a new method to measure FAV and conducted an experiment to determine whether FAV and rumen development were related. In order to determine FAV, a mixer hook was evenly coated with parafﬁn and used to mix moistened feedstuffs at different particle sizes, including ﬁne, intermediate, or coarse. This laboratory technique can provide high precision, but they can be expensive in terms of time and resources and are therefore not attractive to many laboratories or even practical for farming applications. In this study, artiﬁcial neural networks (ANNs) are employed to investigate the relationship between feed abrasive value and physicochemical properties. ANNs are new analytical tools that are based on the models of neu- rological structures and processing function in the brain. The main advantage in using ANNs for prediction is that a priori assumptions about the relations between indepen- dent and dependent variables are not necessary. However, those relations learned by an ANNs are hidden in its neural architecture and cannot be expressed in traditional mathe- matical terms. The comparative advantage of ANNs over M. A. Norouzian (&) Department of Animal Sciences, College of Aboureihan, University of Tehran, P.O. Box: 11365-4117, Tehran, Iran e-mail: manorouzian@ymail.com S. Asadpour Department of Chemistry, Faculty of Sciences, Ferdowsi University of Mashhad, Mashhad, Iran 123 Neural Comput & Applic (2012) 21:905–909 DOI 10.1007/s00521-011-0579-5