Fitness Landscape Optimization Makes Stochastic Symbolic Search by Genetic Programming Easier

Abstract

Searching for symbolic models plays an important role in a wide range of domains, such as neural architecture search and automatic program synthesis. Genetic programming is a promising stochastic method for searching effective symbolic models within an acceptable time. The genetic programming (GP) performance is closely related to the hardness of the fitness landscape (FL). A better FL with less local optima normally implies that it is easier to search for better solutions. In recent years, there have been many studies enhancing GP performance by forming better FLs. However, the better design of the FL highly relies on specific domain knowledge and consumes a lot of expert effort. This article proposes a FL optimization method to automatically design better FLs for GP search than the manually designed ones. We optimize the landscapes by optimizing the neighborhood structures of symbolic solutions. We verify the effectiveness of the proposed method in both supervised learning and combinatorial optimization problems. The results show that the proposed method significantly reduces the hardness of FLs. By simply searching against the automatically optimized FLs, a GP method can have a very competitive performance with state-of-the-art methods.