rarchy of the probabilistic models corresponds directly to this hierarchical program structure. The ideas presented here have some commonalities to John Koza's ADFs and related tricks for putting reusable subroutines in GP trees, but there are also some very substantial differences, which we believe will m