Combining Programs to Counter Code Disruption

In usual Genetic Programming (GP) schemes, only the best programs survive from one generation to the next. This implies that useful code, that might be hidden inside in-trons in low fitness individuals, is often lost. In this paper, we propose a new representation borrowing from Linear GP (LGP), called PhenoGP, where solutions are coded as or-dered lists of instruction blocks. The main goal of evolution is then to find the best ordering of the instruction blocks, with possible repetitions. When the fitness remains stalled, ignored instruction blocks, which have a low probability to be useful, are replaced. Experiments show that PhenoGP achieve competitive results against standard LGP