Surrogate-Assisted Global Optimization Using Structured Low-Discrepancy Sampling and Neural Network

Surrogate-assisted optimization is an effective strategy for solving global optimization problems involving expensive and highly multimodal objective functions, but its performance strongly depends on the structure of the experimental design and the reliability of the surrogate model. This paper proposes a structured surrogate-assisted optimization framework that integrates hybrid low-discrepancy sampling, informed neural network surrogate modeling, and adaptive evolutionary optimization under limited evaluation budgets. The initial training dataset is generated using a hybrid Latin Hypercube-Sobol sampling strategy, ensuring both stratified marginal distributions and low global discrepancy. To improve scalability, an informed surrogate modeling approach is adopted, combining analytical trend decomposition with trigonometric feature encoding to reduce the effective complexity of the learning task. The surrogate is embedded within a genetic algorithm and progressively refined through adaptive enrichment based on a small number of exact evaluations. The framework is evaluated on the multimodal Rastrigin benchmark in two- and five-dimensional settings. Results demonstrate high surrogate accuracy, stable optimization behavior, and effective mitigation of surrogate-induced artifacts, highlighting the benefits of structured sampling and informed surrogate modeling for efficient global optimization.