A neuro-diversified benchmark generator for black box optimization

No Free Lunch Theorem presents a dilemma in the evaluation of emerging evolutionary algorithms in terms of handling various real world problems and their unknown internal structures, since the performances of these algorithms are related to the corresponding benchmarks. Although white and black box schemes have made impressive progress in overcoming this dilemma, such as clear property definition and basis function composition, the evaluation of algorithms on sophisticated suites remains insufficient on account of the limited quantity and diversity of such benchmarks, which can induce bias in a narrow problem domain. Therefore, this study proposes a novel framework for randomly generating diversified benchmark functions to comprehensively evaluate evolutionary algorithms in a black box scenario. The proposed approach adopts a recurrent neural network with various activation functions to produce test problems with important characteristics such as ruggedness and multi-funnels. In addition, the proposed framework can generate virtually limitless chaotic benchmarks by using random weights. The experimental results demonstrate a distinct difference among the performance of the tested optimizers on the proposed problems and the well-known BBOB and CEC problems, which implies the necessity of the proposed benchmarks when facilitating a more comprehensive evaluation.