A benchmark with facile adjustment of difficulty for many-objective genetic programming and its reference set
Makoto Ohki · 2020
In several works, Multi-Objective GP (MOGP) using Multi-Objective Evolutionary Algorithms (MOEAs) is effective on function estimation problems for the cutting process of steel, modeling of non-linear systems, and truss optimization. However, their targets are two or three objective GP problems. Only little research on GP problems with more than four objectives, or Many-Objective Genetic Programming (MaOGP), exists. This is not because real MaOGP problems are ere are rare, but probably because there are few MaOGP benchmarks. Therefore, this paper proposes a benchmark for MaOGP. The problem consists of an analytic function generated by GP and the well-known Many-Objective KnapSack Problem (MaOKSP). In this problem, the difficulty of the problem can be easily adjusted by changing the non-terminal node set, the number of knapsacks or the number of objectives, the number of items, and so on. Moreover, the IGD# indicator is proposed, in which this is a slightly improved version of the IGD+.