The generalisation ability of a selection architecture for genetic programming

Abstract

As an alternative to various existing approaches to incorporating modular decomposition and reuse in genetic programming (GP), we have proposed a new method for hierarchical evolution. Based on a division of the problem's test case inputs into subsets, it employs a program structure that we refer to as a selection architecture. Although the performance of GP systems based on this architecture has been shown to be superior to that of conventional systems, the nature of evolved programs is radically different, leading to speculation as to how well such programs may generalise to deal with previously unseen inputs. We have therefore performed additional experimentation to evaluate the approach's generalisation ability, and have found that it seems to stand up well against standard GP in this regard. © 2008 Springer-Verlag Berlin Heidelberg.