Two Ways of Discovering the Size and Shape of a Computer Program to Solve a Problem

The requirement that the user of a problemsolving paradigm prespecify the size and shape of the ultimate solution to a problem has been a bane of automated machine learning from the earliest times. This paper compares two techniques for automatically discovering the architecture of a multi-part computer program while concurrently solving the problem during a run of genetic programming. In the first technique, called evolutionary selection, the initial random population is architecturally diverse and there is a competitive selection among the various architectures during the run. In the second technique, called evolution of architecture, six new architecture-altering operations provide a way to evolve the architecture of a multi-part program in the sense of actually changing the architecture of the program dynamically during the run. The new architecture-altering operations are motivated by the naturally occurring operation of gene duplication, as described in Susumu Ohno's provocative book Evolution by Means of Gene Duplication, as well as the naturally occurring operation of gene deletion.