Spherical Bounding Classifier using CGP Generated Transforms

Abstract

A Cartesian Genetic Programming (CGP) approach was applied to generate data projection equations (transforms) to project original N-dimensional features space into new 3-dimensional coordinate space followed by a spherical bounding for classification. The solution finding mechanism differs slightly from conventional CGP based machine learning in four ways. Firstly, inputs were weighted to introduce more flexibility for solution finding. Secondly, a dual mutation sequence was introduced to encourage faster convergence. Thirdly, chromosome vector included a 3-dimensional coordinate point and spherical bounding mechanism for classification. Fourthly, the probability of selecting either of the input types (feature input, node sequence, or constant) are made approximately equal to ensure that the input types have equal probability to be in the active equation nodes. The best classification rate on test datasets achieved using 10 fold cross validation was 98.57% (Wisconsin Breast Cancer dataset), 87.78% (Heart disease dataset) and 80.5% (PIMA Indian diabetes dataset).