PRETTY: A parallel transgenerational learning-assisted evolutionary algorithm for computationally expensive multi-objective optimization

Many real-world optimization problems, called computationally expensive optimization problems, often require a time-consuming fitness evaluation through computer simulation or neural network training. A persistent challenge is to obtain reasonable solutions to these problems quickly. In this study, we design a parallel transgenerational learning-assisted evolutionary algorithm, called PRETTY, to accelerate the search process using three novel mechanisms. First, we fuse the transgenerational learning mechanism with metaheuristic algorithms to reduce unnecessary fitness evaluations. Then, a Hypervolume-Cauchy (HV-Cauchy) algorithm is proposed to balance population convergence and diversity adaptively. Finally, a parallel-computing framework is proposed to fuse the transgenerational learning mechanism and the HV-Cauchy algorithm into a multipopulation evolutionary algorithm. To verify the performance of the proposed algorithm, empirical experiments are conducted on nine benchmarks and four real-world problems, including feature selection, critical node detection, flocking, and neural network training. In terms of execution time, the experimental results demonstrate that PRETTY has significant advantages over several expensive multi-objective algorithms. Furthermore, PRETTY can maintain population diversity and fast convergence while accelerating the computation. In particular, in tackling real-world problems that require expensive fitness calculations, PRETTY can reduce the execution time by more than five times while ensuring high-quality solutions.