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