New computational intelligence algorithms and their applications toward automated intelligent decision making

Qian Xia, Purdue University

Abstract

Several new computational intelligence methods and their applications are investigated in this thesis. First, a multistage linear support vector machine (MLSVM) classification is developed. It iteratively bisects the dataset into an accepted and a rejected subset, and builds separate model on the accepted subset at each stage. The aggregation of individual models constructed delivers a mixture of submodels embedded in the dataset. A hybrid MLSVM by combining the MLSVM with other regression methods, such as logistic regression, is also investigated. Secondly, a framework referred to as interactive clustering and classification (ICC) is developed. A multiobjective desirability function is used in combining clustering and classification objectives. Optimizing the multiobjective desirability function involves solving the clustering and classification process simultaneously. The method is designed to identify sub-groups of samples with different model structures. Thirdly, the ICC is extended to regression by replacing the classification objective with a regression objective. The method is referred to as interactive clustering and regression (ICR). The ICC and the ICR have been tested rigorously on both synthetic and real world datasets. Design of Experiment (DOE) was used to generate a set of simulated datasets which represent a comprehensive set of scenarios. Lastly, a generalized partial least squares (GPLS) regression method is developed. It uses polynomial transformation on PLS components and model selection to find the best least squares (LS) model. The method is capable of handling small datasets with large number of features, which is crucial for fields like pharmaceutical and bioinformatics applications.

Degree

Ph.D.

Advisors

Ersoy, Purdue University.

Subject Area

Management|Electrical engineering|Artificial intelligence

COinS