Genetic lineages were used to bias the selection method toward fit individuals from randomly chosen lineages. The results showed that genetic diversity was increased while producing smaller solutions, but with worse fitness on two of the three problem domains. A literature review of similar and related problem domains revealed that the results of using lineage selection clearly demonstrated the similarities between genetic programming and hill-climbing. When genetic programming benefits from a hill-climbing type search, increasing diversity worsened fitness by preventing a hill-climbing search from being carried out effectively. When genetic programming was likely to suffer from deception, the increased diversity of lineage selection improved fitness.
The idea that genetic programming may carry out a similar search to that of hill-climbing motivated an analysis of the sampling of structures and behaviours. In this case, a structure refers to a tree without any node contents. A behaviour was defined for each problem that reflected the fitness or complexity of solutions. These results demonstrated that genetic programming samples fewer unique structures of large sizes. Instead, when large structures were produced, the search effort was spent sampling different structures of an intermediate size. Problem specific results highlighted many areas where methods could be used to improve search performance.
The issue of code growth and bloat has received much attention in the genetic programming community. The mechanics of subtree crossover and the representation are thought to be responsible for code growth in genetic programming. However, there has been little effort to explain the varied rates of growth that occur in most experiments.