Probabilistic developmental program evolution
Elmira Ghoulbeigi, Marcus Vinicius dos Santos · 2010
A Probabilistic Model Building Genetic Programming technique for automatic program synthesis is introduced. The approach, called Probabilistic Developmental Program Evolution (PDPE), draws on the Probabilistic Incremental Program Evolution (PIPE) learning algorithm, but employs the Developmental Genetic Programming representations of Gene Expression Programming (GEP). PDPE induces a population of programs, encoded as fixed-length GEP chromosomes, by iteratively refining and randomly sampling a probability distribution of program instructions stored in a vector called probability prototype chromosome (PPC). This refining, however, is accomplished solely by means of mutation of the PPC. We compared PDPE with PIPE and GEP on a function regression problem and the 6-bit parity problem. Our results show that PDPE outperforms PIPE in terms of solution quality and variance. It also outperforms GEP in terms of solution quality, but not in terms of variance.