An Effective Many-Objective Extension of Genetic Programming and Its Benchmark

This paper describes many-objective extension (MaO extension) of GP problem. The extension reduces a bloat phenomena what often happens in GPs. MaO extension of GP consists of addition of objective functions, a partial sampling (PS) operator, a structural distance, elimination of identical trees and subset size scheduling. In this paper, MaO extension of GP are applied to NSGA-II and SPEA2. Because few researchers have considered effective benchmark problems for many-objective genetic programming (MaOGP), MaOGP is less developed than the many-objective evolutionary algorithm (MaOEA). This paper proposes a new benchmark problem for MaOGP. The benchmark problem is based on the many-objective knapsack problem (MaOKSP). In the benchmark, a function given by GP returns a binary vector for MaOKSP. A research on such benchmark problems contributes to the future development of MaOGPs.