Archive assisted fully informed evolutionary algorithm for expensive many-objective optimization

In many real-world engineering and scientific optimization scenarios, practitioners often face expensive many-objective optimization problems where evaluating candidate solutions incurs prohibitive computational costs. The inherent scarcity of truly calculated data often leads to the construction of models with high uncertainty using limited datasets. This uncertainty can adversely affect the Surrogate-assisted Many-Objective Evolutionary Algorithms (SAMaOEAs). To address this issue and enhance performance, this paper introduces an Archive-assisted Fully Informed Evolutionary Algorithm (AFIEA). In AFIEA, two kinds of models are constructed from archive data to simultaneously predict objective values and uncertainty trends (whether the predictions are overestimated or underestimated). With this foundation, both the optimizer and infill criterion processes are fully guided by the predicted objective values and uncertainty trends. In the optimization phase, a novel Uncertainty Trend Classification (UTC)-based Upper Confidence Bound is employed as the acquisition function. During the infill criterion phase, UTC is used to preprocess the population, enhancing the selection probability of under-estimated solutions, while an archive-based metric selects more precise solutions, guided by the archive in terms of convergence and diversity. The performance of AFIEA is compared with six state-of-the-art SAMaOEAs on artificial benchmark problems and one real-world expensive optimization problem within a limited budget. In the benchmark tests, AFIEA outperforms the six advanced SAMaOEAs across most of the test functions, demonstrating that the proposed mechanism offers strong generality and enhanced search performance. Additionally, in the optimization of electromagnetic devices, AFIEA achieves superior population quality in a shorter time with a limited number of simulations.