基于机器学习的地方鸡产蛋曲线拟合探索

Exploration of Egg Production Curve Fitting of Local Chickens Based on Machine Learning

[目的]本研究探索以机器学习方法对2个地方鸡品系周产蛋率建模,并将其与非线性回归方法比较,旨在提高养鸡生产中产蛋曲线的拟合精度。[方法]产蛋数据采集自地方鸡杂交组合试验群,自22周龄开始统计产蛋数,至50周龄截止。试验鸡于全封闭鸡舍单笼饲养,产蛋期人工补光16 h。试验鸡分为两组,每组150只鸡。第Ⅰ组是黄羽肉鸡合成系,第Ⅱ组是兼用型地方鸡种。以IBM SPSS Statistics 21.0软件中的非线性回归方法拟合产蛋曲线,所用模型包括Logistic模型、McNally模型、杨宁模型以及Grossman模型。以MATLAB R2014a构建机器学习模型,神经网络选用多层感知器,用300次迭代的拟牛顿法训练数据。以贝叶斯最小二乘支持向量机构建产蛋模型,针对正则项系数和核函数参数进行优化。[结果]依据MSE、R 2、AIC评判标准,Grossman模型在4种非线性回归模型中拟合度最好,McNally模型表现最差。McNally模型预测的高峰产蛋率偏离真实值;Logistic模型、杨宁模型以及Grossman模型高峰产蛋率统计值与真实值基本相符。两组试验鸡的模型参数不同,Ⅱ组持续产蛋能力优于Ⅰ组。基于MSE、R 2以及图形评估结果,神经网络优于传统的非线性方程拟合,而支持向量机略好于神经网络。优化后神经网络参数为2个隐藏层,每个隐藏层包含5个神经元。第Ⅰ组支持向量机的正则项系数为30.97,核参数为0.0701;第Ⅱ组支持向量机的正则项系数为566.53,核参数为0.1754.[结论]机器学习方法可用于产蛋模型构建,相比于传统单变量回归方法,机器学习方法可加入更多变量,提供更准确的预测。

[Objective]In order to improve fitting accuracy on egg production curve in chicken,the machine learning method was explored to model the weekly egg production rate for 2 indigenous breeds,and compared with the nonlinear regression method.[Method]Data were collected from local chickens populations recorded from 22 to 50 weeks.The hens were housed in a single cage in an enclosed house,kept in artificial light for 16 hours during the laying period.The chickens were divided into two groups,each with 150 chickens.GroupⅠwas a synthetic line of Meat-type Yellow chickens,and groupⅡwas a local dual-purpose breed.The non-linear regression method in IBM SPSS Statistics 21.0 software was used to fit the egg production curve,including the Logistic,McNally,Yang,and Grossman models.The machine learning model was constructed with MATLAB R2014a,using a multilayer perceptron.The neural network was trained with the quasi-Newton method for 300 iterations.The egg production curve was fitted with least squares support vector machine,and the regularization and kernel function parameters were optimized with Bayesian inference.[Result]According to the evaluation criteria of MSE,R 2,and AIC,the Grossman model had the best degree of fitting among the four non-linear regression models,and the McNally model performed the worst.The peak egg production rate predicted by the McNally model deviated from the real value,the peak egg production rates obtained from the Logistic,Yang,and Grossman models were consistent with the observed value.There were differences between the parameter estimates of the curves fitted for the two groups.The persistence of the lay for groupⅡwas better than that of groupⅠ.Based on the MSE,R 2 and graphical evaluation,the neural network fitted better than the traditional nonlinear regression,and the support vector machine was slightly better than the neural network.The optimized parameters for the neural network were 2 hidden layers,containing 5 neurons in each layer.For support vector machine,the regularization parameter of groupⅠwas 30.97,and the kernel parameter was 0.0701;The regularization and kernel parameter of groupⅡwas 566.53 and 0.1754,respectively.[Conclusion]Machine learning method could be used to fit the egg production curve in these populations.Compared to those classic univariate regression,machine learning methods could harbor more variates and provide more accurate predictions.