A Modified Genetic Algorithm Used for Dimensionality Reduction in Record Classification

Abstract

This work proposes a modified Genetic Algorithm and compares its performance with the conventional Genetic Algorithms (GA) used for Dimensionality Reduction in record classification. A specialized elite voting crossover and mutation was introduced to the conventional GA and the population composition of every generation was compartmented into elite and non-elite individuals, and a proportion of offspring generated in each generation are derived from the elite individuals using the introduced voting crossover and mutation. The performance of the two algorithms was tested with 3 datasets from the UCI ML repository using different levels of elitism, crossover and mutation with the Extreme Learning Machine classifier. At higher rate of elitism, the results were highly in favor of the modified GA in both convergence time and classifier accuracy. While, at lower levels of elitism the two algorithms seen to be comparable in convergence time but the modified algorithm had better classifier accuracy. Furthermore, at higher rate of crossover, the modified algorithm tends to be slower in convergence than the conventional algorithm but better classifier accuracy. On the other hand, at higher mutation rate the modified algorithm tends to be faster in convergence than the conventional algorithm. In conclusion, except for the added computational cost due to the specialized voting crossover and mutation in the modified algorithm the results are in favor of the modified algorithm.