The Effect of Function Noise on GP Efficiency

Genetic Programming (GP) has been applied to many problems and there are indications [1,2,3] that GP is potentially useful in evolving algorithms for problem solving. This paper investigates one problem with algorithmic evolution using GP — Function Noise. We show that the performance of GP could be severely degraded even in the presence of minor noise in the GP functions. We investigated two counternoise schemes, Multi-Sampling Function and Multi-Testcases. We show that the Multi-Sampling Function scheme can reduce the effect of noise in a predictable way while the Multi-Testcases scheme evolves radically different program structures to avoid the effect of noise. Essentially, the two schemes lead the GP to evolve into different “approaches” to solving the same problem.