Dependency-aware evolutionary search using population representation learning for large-scale many-objective optimisation

Many-objective optimisation problems characterised by large decision spaces and complex objective interactions remain challenging for conventional evolutionary algorithms. Existing approaches primarily rely on predefined variation operators and dominance-based selection mechanisms, which often struggle to exploit latent structural relationships among candidate solutions. This study introduces a representation-guided evolutionary optimisation framework that learns population-level dependencies during the search process and incorporates the acquired knowledge into evolutionary variation and selection. The proposed framework employs a population encoding mechanism based on self-attention to capture global interactions among candidate solutions and generate latent representations describing the evolving search landscape. These representations are subsequently utilised to guide adaptive crossover, mutation, and diversity preservation strategies, enabling a more informed exploration–exploitation balance. Unlike conventional learning-assisted evolutionary algorithms that use machine learning models as external predictors, the proposed method integrates representation learning directly within the optimisation cycle. Experimental evaluation was conducted on selected configurations from the DTLZ, WFG, MaF, and CEC 2020 MMOP benchmark families, involving up to 20 objectives and 1000 decision variables. On the evaluated configurations, MOEA-T achieved improved mean convergence and diversity indicators compared with the considered Pareto-based, decomposition-based, reference-vector-based, and learning-assisted algorithms. These findings provide evidence of the potential effectiveness of representation-guided evolutionary search under the tested high-dimensional and many-objective conditions. However, they do not imply uniform superiority across every problem instance in the complete benchmark families. Additional ablation and scalability studies confirm the effectiveness of dependency-guided search and population representation learning in addressing the challenges associated with high-dimensional many-objective optimisation. The findings suggest that integrating population-level structural learning into evolutionary search provides a promising direction for next-generation large-scale optimisation frameworks.