8278

7676

كارشناسي ارشد

آمار اقتصادي و اجنماعي

اصفهان: دانشگاه صنعتي اصفهان، دانشكده رياضي

1392

سيزده، 137ص: مصور، جدول، نمودار

ص.ع:به فارسي و انگليسي

سروش عليمرادي

محمد رضا احمد زاده

25/09/1392

علي رجالي، علي زينل همداني

رياضي

ID7676

به فارسي و انگلسيي: قابل رويت در نسخه ديجيتال

Logistic regression modeling by means of evolutionary radial basis function neural networks Arezu mojiri khuzani Master of science ABSTRACT Classification is a subject which has many practical applications in various areas Objects or indiviuals classification can be made by observing their attributes or features One statistical approach for classification is logistic regression model Logistic regression by assuming a linear combination of the covariates or attributes is a modeling for class membership posterior probabilities It is possible for some practical cases that nonlinear combinations of covariates also include in the model In this thesis we focus on this issue by considering a type of nonlinear transformations of covariates named radial basis functions in logistic regression model But the main problem of this model is finding maximum likelihood estimation of its parameters Existence of nonlinear transformations of covariates in logistic model causes many local extremums in likelihood function Therefore finding of global maximum for this function using usual maximum likelihood methods is almost impossible The proposed method in this thesis for overcoming to this difficulty is a hybrid method which solves the problem of maximum likelihood estimation for these logistic regression models by the use of combining the idea of neural networks evolutionary algorithms and maximum likelihood methods This method includes three steps In first step nonlinear part of model is shown by using a radial basis function neural network then parameters in this neural network and its structure is determined optimality by using an evolutionary programing algorithm In second step best radial basis functions determined in pervious step as new covariates is added to space of the initial covariates The resulted model in this new space is linear with respect to all covariates Therefore in final step the remaining coefficients in model are estimate by using two distinct algorithms In first algorithm ridge estimators of the coefficients are calculated and therefore is avoided the overfitting problem in model The second algorithm incrementally constructs the model by automatic covariate selection Each of these algorithms lead to different models After estimating the parameters by hybrid method we compare the resulting logistic model to other classification methods using several data sets In first data set effective factors in heart disease is analyzed Second and third data sets are associated with a large factory producing steels In these data sets effective factors which causes defects in steel sheets is analyzed Experimental results show that the proposed method is superior to other conventional methods MSC J Keywords logistic regression radial basis function neural networks evolutionary algorithms PDF created with pdfFactory trial version www pdffactory com

سروش عليمرادي

محمد رضا احمد زاده

علي رجالي، علي زينل همداني