Optimization and Combination of Classifiers Using Genetic Programming

Created by W.Langdon from gp-bibliography.bib Revision:1.2031

@PhdThesis{Majid:thesis,
  author =       "Abdul Majid",
  title =        "Optimization and Combination of Classifiers Using
                 Genetic Programming",
  school =       "Ghulam Ishaq Khan Institute of Engineering Sciences \&
                 Technology",
  year =         "2006",
  address =      "Topi, Swabi, NWFP, Pakistan",
  month =        may,
  keywords =     "genetic algorithms, genetic programming",
  URL =          "http://bpt.hec.gov.pk/2511/",
  URL =          "http://prr.hec.gov.pk/Thesis/349S.pdf",
  size =         "155 pages",
  abstract =     "The success of pattern classification system depends
                 on the improvement of its classification stage. The
                 work of thesis has investigated the potential of
                 Genetic Programming (GP) search space to optimise the
                 performance of various classification models. In this
                 thesis, two GP approaches are proposed. In the first
                 approach, GP is used to optimize the performance of
                 individual classifiers. The performance of linear
                 classifiers and nearest neighbour classifiers is
                 improved during GP evolution to develop a high
                 performance numeric classifier. In second approach,
                 component classifiers are trained on the input data and
                 their predictions are extracted. GP search space is
                 then used to combine the predictions of component
                 classifiers to develop an optimal composite classifier
                 (OCC). This composite classifier extracts useful
                 information from its component classifiers during
                 evolution process. In this way, the decision space of
                 composite classifier is more informative and
                 discriminant. Effectiveness of GP combination technique
                 is investigated for four different types of
                 classification models including linear classifiers,
                 support vector machines (SVMs) classifiers, statistical
                 classifiers and instance based nearest neighbour
                 classifiers.

                 The successfulness of such composite classifiers is
                 demonstrated by performing various experiments, while
                 using Receiver Operating Characteristics (ROC) curve as
                 the performance measure. It is evident from the
                 experimental results that OCC outperforms its component
                 classifiers. It attains high margin of improvement at
                 small feature sets. Further, it is concluded that
                 classification models developed by heterogeneous
                 combination of classifiers have more promising results
                 than their homogeneous combination.

                 GP optimisation technique automatically caters the
                 selection of suitable component classifiers and model
                 selection. Two main objectives are achieved, while
                 using GP optimisation. First, objective achieved is the
                 development of more optimal classification models. The
                 second one is the enhancement in the GP search strategy
                 itself.",
  notes =        "

                 Item Type: Thesis (PhD)

                 ID Code: 2511 Deposited By: Ch Abdulla fayyaz Chattha
                 Last Modified: 28 Jul 2009 21:16",
}

Genetic Programming entries for Abdul Majid