ACGP is a new method to explore regularity

GP inherently is a technique for, and relies upon, exploration and exploitation of regularity. With ADFs, regularity translates into modularity. Regularity is explored both implicitly (evolution via crossover) and explicitly (as in ADFs). ACGP (Adaptable Constrained GP) offers a new way for GP to explore regularity of a different kind. It will have an impact on modularity as well. I will here briefly describe what kind of regularity ACGP explores, and how it differs from ADFs. As to how this is done, I will be presenting this during my regular presentation (ACGP was developed for tree-based chromosomes, so this discussion is on trees). GP solves problems by implicitly combining Building Blocks via evolution (same as GA, just different kind of BBs). BBs don’t have to be complete subtrees, they are simply tree fragments, with possible subtrees represented by a single “don’t care” node. When it comes to explicit exploration of regularity, such as leading to modularity by ADFs, the regular structures are complete subtrees, that is every leaf must be a terminal instance (or an ADF, which in fact relaxes the terminal requirement I have to admit I need to catch up on your recent work). Whether ADFs are complete subtrees or not, they change the alphabet (by adding ADFs to the set of terminals), but not the probability distribution of crossover and mutation. In other words, they explore regularity among BBs, or rather subtree instances, but not regularity in the probability distribution. The latter is what ACGP does.