Record Details

Development of Statistical Models using Nonlinear Support Vector Machines

KrishiKosh

View Archive Info
 
 
Field Value
 
Title Development of Statistical Models using Nonlinear Support Vector Machines
M.Sc.
 
Creator Mohan Kumar, T. L.
 
Contributor Prajneshu)
 
Subject sets, biological phenomena, forecasting, hybrids, spacing, yields, marketing, productivity, crops, solutes
 
Description Statistical modelling plays a very important role in comprehending underlying relationships among crucial variables in an agricultural system. In this thesis, several statistical models are developed using Nonparametric Nonlinear Support Vector Machines (SVM) methodology. SVM is a relatively new Generalized portrait algorithm proposed for solving problems in classification, function estimation and density estimation. Basic concepts of linear and nonlinear SVM and their formulation for binary classification problems are discussed. Methodology for extending binary classification problem to multiclass classification problem is considered. Further, optimal hyper-parameters of this model are estimated using Particle Swarm Optimization (PSO) algorithm. As an illustration, this methodology is applied for classification of three varieties of banana based on their morphological characters. Result shows that the above methodology has performed the best vis-à-vis other competing methodologies for the data under consideration. Extension of SVM methodology to regression problem called as Support Vector Regression (SVR) is thoroughly studied. PSO technique is employed to estimate its optimal hyper-parameters. Superiority of this methodology is demonstrated over Artificial neural network and Multiple linear regression for some maize crop yield data. Various hybrid models are developed to tackle complex time-series data by combining Linear Seasonal Autoregressive Integrated Moving Average (SARIMA) and Nonlinear SVR models for time-series forecasting. Optimal hyper-parameters of these models are estimated using PSO technique. Subsequently, as an illustration, these models are applied to all-India monthly marine products exports time-series data. Superiority of hybrid models over individual Linear SARIMA and Nonlinear SVR models is demonstrated for the data under consideration. Further, Least squares version of SVM, known as Least Squares Support Vector Machine (LS-SVM) methodology is thoroughly investigated. Optimal hyper-parameters of this model are estimated by employing PSO technique. As an illustration, the methodology is successfully illustrated for modelling and forecasting all-India monthly rainfall time-series data. Linear Kalman Filter (KF) and Nonlinear LS-SVM methodologies are also combined to develop various hybrid
models for forecasting complex time-series data. PSO is employed for estimating optimal hyper-parameters of hybrid models. All-India monthly rainfall time-series data is considered to illustrate the superiority of developed hybrid models over individual Linear KF and Nonlinear LS-SVM methodologies. In order to achieve the above tasks, relevant computer programs are written in R and MATLAB software. Various software packages Viz. R, MATLAB, SAS and STATISTICA are used to carry out data analysis.
 
Date 2016-03-15T20:00:14Z
2016-03-15T20:00:14Z
2013
 
Type Thesis
 
Identifier http://krishikosh.egranth.ac.in/handle/1/65166
 
Language en_US
 
Format application/pdf
 
Publisher IARI, INDIAN AGRICULTURAL STATISTICS RESEARCH INSTITUTE INDIAN AGRICULTURAL RESEARCH INSTITUTE NEW DELHI -