Abstract
The important process of choosing between algorithms and their many module choices is difficult, even for experts. Automated machine learning allows users at all skill levels to perform this process. It is currently performed using aggregated total error, which does not indicate whether a stochastic algorithm or module is stable enough to consistently perform better than other candidates. It also does not provide an understanding of how the modules contribute to total error. This paper explores the decomposition of error for the refinement of genetic programming. Automated algorithm refinement is examined through choosing a pool of candidate modules and swapping pairs of modules to reduce the largest component of decomposed error. It is shown that a pool of candidates that are not examined for diversity in targeting different components of error can provide inconsistent module preferences. Manual algorithm refinement is also examined by choosing refinements based on their well-understood behaviour in reducing a particular error component. The results show that an effective process should exploit both the advantages of targeted improvements identified using a manual process and the simplicity of an automated process by choosing a hierarchy of the most important modules for reducing error components.