Automated Transformation Design via Online Variable-Grouping Optimization for Large-Scale Multi-Objective Problems
Year: 2026
Type: article
Abstract: The curse of dimensionality makes it difficult to coordinate many decision variables while preserving objective trade-offs in large-scale multi-objective optimization. Problem transformation reduces the active search dimension, but its variable grouping can also exclude useful directions. Variable relationships and random partitions do not directly measure the search progress a transformation can produce from the current population. We propose the Automated Transformation Design-based Multi-objective Evolutionary Algorithm (ATDMOEA), which treats grouping as an optimization task rather than a preprocessing rule. A meta-population evolves candidate transformation topologies over successive generations. Each candidate is tested through a budget-limited search from shared anchors, and hypervolume (HV) improvement guides further topology search. This design process repeats as the main population changes. The selected topology guides reduced-dimensional exploration before full-space refinement. On nine large-scale multi-objective optimization benchmarks (LSMOP1–LSMOP9), ATDMOEA obtains the unique lowest mean inverted generational distance (IGD) in 20 of 27 two-objective and 19 of 27 three-objective cases. Its IGD on the two-objective, 1000-variable LSMOP1 case is 78% lower than that of the next best comparator. It also obtains the highest mean HV on four of five voltage-transformer ratio-error estimation instances among methods with valid results.
Source: Algorithms
DOI: 10.3390/a19111014
Institutions: Hunan College of Information, Xiangtan University
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,930
OpenAlex ID: W7235914666
