Comprehensive learning gene expression programming for automatic implicit equation discovery

Abstract

Implicit equation is loose in form, which makes it more powerful than explicit equation for data regression. The mainstream method for automatic implicit equation discovery is based on calculating derivatives. However, this derivative-based mechanism requires high time consumption and it is difficult to solve problems with sparse data. To solve these deficiencies, this paper proposes a new mechanism named Comprehensive Learning Fitness Evaluation Mechanism (CL-FEM). The mechanism learns knowledge from disturbed information collected from several previously generated stochastic datasets, to check the validity of the equation model. Only the valid equations can be candidates of selection, which is a process to pick out the equation with the smallest output. We integrate the proposed mechanism with the simplified Self-Learning Gene Expression Programming (SL-GEP) and propose the Comprehensive Learning Gene Expression Programming (CL-GEP). The experiment results have demonstrated that CL-GEP can offer very promising performance.