بررسي صفات رشد گوسفند لري با استفاده از مدل‌هاي غير خطي و شبكه عصبي مصنوعي بهينه شده با الگوريتم ژنتيك

Study of Lori growth traits using nonlinear models and artificial neural network optimized by genetic algorithm

زمينه مطالعاتي: در اين پژوهش از اطلاعات تعداد 7054 راس گوسفند نژاد لري براي برازش منحني رشد اين نژاد استفاده شد. هدف: صفات رشد مورد بررسي شامل وزن تولد، از شيرگيري، شش ماهگي و نه ماهگي بود كه با استفاده از سه مدل غير خطي شامل گمپرتز، برودي و لجستيك و همچنين شبكه عصبي مصنوعي (ANN) برازش شد. روش كار: تيپ تولد، جنسيت، سال تولد، سن مادر و فصل تولد به همراه وزن تولد، شيرگيري و شش ماهگي به عنوان عوامل ورودي به ANN معرفي شدند و براي وزن نه ماهگي پيش بيني انجام شد. براي اين منظور يك شبكه Feed-forward بهينه شده با الگوريتم ژنتيك مورد استفاده قرار گرفت. مقايسه مدل­هاي غيرخطي بر اساس ضريب تبيين (R2)، ميانگين مربعات خطا (MSE)، تعداد تكرار و معيار آكائيك (AIC) انجام شد و بر اين اساس مدل برودي به عنوان مدل مناسب براي برازش صفات رشد انتخاب شد. پارامترهاي A، B و K بر اساس مدل برودي براي دو جنس ماده و نر برآورد شدند. نتايج: همبستگي بين پارامترهاي A و K منفي گزارش شد. اثر عوامل محيطي بر روي پارامترهاي منحني رشد معني دار بود (۰/۰۱>P). بر اساس بررسي­هاي انجام شده ANN با R2 برابر با 36/84 و 49/85 درصد قادر به پيش بيني وزن نه ماهگي براي جنس ماده و نر بود. همچنين با تعداد 10 و 9 نورون در لايه مياني براي جنس ماده و نر، در MSE همگرايي ايجاد شد. نتيجه­گيري نهايي: بر اساس ميزان R2 گزارش شده، مدل­هاي برودي، لجستيك، گمپرتز و ANN به ترتيب مناسب­ترين مدل­ها براي برازش صفات رشد در گوسفند لري بودند.

Introduction: Machine learning methods such as artificial neural network (ANN) are already widely used in agriculture because these methods are fast, powerful and flexible tools for classification and forecasting requirements. In the field of animal science, these methods are used for the detection of mastitis, estrous and removal reasons of animals (Shahinfar et al. 2012). ANN is a machine learning method that simulates brain function. The most important advantage of ANN is related to its ability to accept large volumes of data and find interesting and complex relationships between these data. Feed - forward neural network is a type of neural network training methods that is a training monitored. The network contains neurons that are composed of several layers. The first layer of input data, the last laye r of the data is output, and between these two layers are hidden layers. In this way the genetic algorithm is programming technique that uses a process of genetic evolution as a problem solving model ( Ahmed and Simonovic 2005). Non - linear regression models are developed form of classical models. This models are includes fixed and random effects that used to describe the growth of their data ( Bahreini Behzadi et al. 2014). Often growth traits of livestock described by non - linear growth models such as Gompert z, Logistic , Richards , Weibull, Brody and von Bertalanffy (Aman Ullah et al. 2013) . Material and m ethods : The data for this research was related to number of 7054 Lori sheep and including birth weight, weaning weight, weight six and nine months of age that were collected by the Agricultural Organization of the Lorestan province between the years 2001 to 2010 y ears. This data was related to nomadic herds in the Khorramabad city. At first, this data was edited using Excel 2010 and Fox Pro 3 ( Hentzen 1995) software. To check the normality of the data, the software SAS (Institute SAS 2004) univariate procedures was used. Also for evaluation of growth traits, Gompertz, Brody and Logistics models were used. These models were performed by non - linear procedure (PROC NLIN) and the Gauss - Newton Iterative methods using SAS 9.2 (2003) software and then the growth paramete rs were calculated. Different models were validation and compared with each other based on the coefficient of determination (R 2 ), mean squared error (MSE), the number of iteration and Akaike information criterion (AIC). In the ANN environmental effects such as sex of lamb, type of birth, birth season, birth year, mother's age and birth weight, weaning weight and weight at six months of age were introduced as input to the neural network and ultimately weight at nine months of age was predicted. When neura l network structure was formed, Mean Square Error (MSE) was used to evaluate and determine the optimal number of neurons in the middle layer. Results and discussion : Compare models based on the coefficient of determination shows that the models are not muc h different from each other and coefficient of determination range was varied between 96.79 to 98.84 percent. The highest R 2 for male and female was related to the Brody model. High R 2 and low iteration for all 3 models show that these models are suitable to describe the growth curve of Lori sheep. Year of birth, birth type, lamb sex, mother age and birth season had significant effect on the A, B and K parameters (P<0.01). In this study, in all 3 models the growth rate of females (0.017 ± 0.0003 to 0.026 ± 0.0004) was higher than the growth rate in males (between 0.016 ± 0.006 to 0.022 ± 0.0005). The difference in growth rates in males and females in Lori sheep is down and f emales have higher growth rates and, consequently, lower maturity weight. The Matlab software was used to implement neural network code and calling the information and their implementation. the neural network with 8 inputs including birth weight, weaning and six months and fixe effects was formed. Conclusion : Based on R 2 in this study sugge st that the Brody model is the best model for fit the growth traits and also three models of Brody, Logistic and Gompertz have the higher performance for forecasting and analysis of growth traits in Lori sheep than ANN. The birth and weaning weights and th e other growth traits in Lori lambs is impressed by the change in weather conditions, followed b y changes in natural conditions.