The impact of external dependency in genetic programming primitives

Abstract

The power of genetic programming arises from its ability to identify and promote appropriate subprograms of the “true” solution via its fitness based selection and inheritance mechanism (“survival of the fittest”) and then combine them via blind variation in terms of subtree crossover. Both control and data dependencies among primitives impact the behavioural consistency of subprograms in genetic programming solutions which in turn taxes the efficiency of selection. We present the results of modelling dependency through a parameterized problem in which a subprogram exhibits internal and external dependency levels that change as the subprogram is successively incorporated into larger subsolutions. We find that the key difference between non-existent and “full” external dependency when a solution is composed of subsolutions with exponentially scaled fitness contributions is a longer time to solution identification and a lower likelihood of success as shown by increased difficulty in identifying and promoting correct subprograms.