A Universal Framework for Automatically Generating Single-and Multi-Objective Evolutionary Algorithms

The automated design of evolutionary algorithms (EAs) is receiving more and more attention, which considerably reduces the labor-intensive process of algorithm design demanding expertise. While existing automated methods for selecting, tuning, and creating EAs mainly rely on existing operators connected in sequence, they can hardly surpass the performance thresholds of existing EAs across various problems. To break through the shackles of existing EAs, this work proposes a universal framework to automatically create EAs from scratch: First, a series of building blocks are designed without using existing operators, and thus offer the potential to exceed the performance thresholds of existing EAs. Second, these blocks are connected like the layers of deep neural networks, which can form non-sequential architectures to pursue good performance. Third, these blocks are parameterized and can be trained like neural networks, able to exhibit better performance on given problems. Using the proposed framework, non-sequential EAs characterized by dozens of parameters are trained on a few problems, which outperform 36 existing algorithms on more than 200 single-and multi-objective problem instances. Particularly, the non-sequential EAs without surrogate demonstrate superior convergence over surrogate-assisted EAs on expensive problems, and outperform gradient methods on unimodal problems.