A Machine Learning Approach for Sensitivity Inference in Genetic Algorithms

摘要

Sensitivity in terms of genetic algorithms can be understood as the relative importance of a specific gene and its contribution factor to the fitness of the individual. We propose a novel approach to infer sensitivity for real-valued alleles based on machine learning. The presented algorithm utilizes neural networks and support vector machines on a genetic algorithm population. Experimental results of sensitivity inference on several fitness functions are provided. Our empirical results show that the presented algorithm is reliably able to detect sensitivity and will correctly rank the genes of an individual according to their relative importance. We demonstrate the usefulness of the obtained sensitivity information by a comparison of GA runs. Furthermore, the results in this paper are extended by a promising observation regarding sensitivity detection with correlation-based feature selection.