A fine-grained view of GP locality with binary decision diagrams as ant phenotypes

Abstract. The property that neighbouring genotypes tend to map to neighbouring phenotypes, i.e. locality, is an important criterion in the study of problem difficulty. Locality is problematic in tree-based genetic programming (GP), since typically there is no explicit phenotype. Here, we define multiple phenotypes for the artificial ant problem, and use them to describe a novel fine-grained view of GP locality. This allows us to identify the mapping from an ant’s behavioural phenotype to its concrete path as being inherently non-local, and show that therefore alternative genetic encodings and operators cannot make the problem easy. We relate this to the results of evolutionary runs