DWMODOA: a Dreamwalk-inspired multi-objective evolutionary algorithm with archive-assisted partition guidance

Non-convex, disconnected, and degenerate Pareto fronts, together with inconsistent objective scales, can bias selection and reduce diversity in multi-objective optimization. This paper proposes DWMODOA, a Dreamwalk-inspired multi-objective evolutionary algorithm with archive-assisted partition guidance for HPC-assisted optimization workflows. DWMODOA has two main components. First, an archive-assisted objective-space partition reference strategy converts non-dominated archive solutions into partition-specific references for multi-directional guidance. Second, a Dreamwalk two-branch mutation operator combines phase-guided sparse exploration and reference-guided differential aggregation to balance exploration and exploitation. Robust normalized crowding distance truncation supports environmental selection and reduces scale-induced bias. Because candidate evaluations in population-based evolutionary algorithms are mutually independent, DWMODOA is suitable for parallel and distributed execution when objective evaluations are computationally expensive. Experiments on 22 ZDT, WFG, and IMOP benchmarks show competitive convergence and diversity, especially on irregular Pareto fronts.