Impact of surrogate model accuracy on performance and model management strategy in surrogate-assisted evolutionary algorithms

Surrogate-assisted evolutionary algorithms (SAEAs) are widely used to solve expensive optimization problems where evaluating candidate solutions is computationally intensive. To reduce this cost, SAEAs employ surrogate models—machine learning models that approximate expensive evaluation functions. While previous studies have investigated how surrogate model prediction accuracy affects SAEA performance, they are limited in two key ways: (1) lacking a comprehensive analysis of commonly used model management strategies, and (2) comparing these strategies under inconsistent surrogate accuracy settings. To address these limitations, we construct a pseudo-surrogate model with adjustable prediction accuracy, enabling fair comparisons across different strategies. We evaluate three representative strategies — pre-selection (PS), individual-based (IB), and generation-based (GB) — using a common pseudo-surrogate model on six benchmark problems in 10 and 30 dimensions from the CEC2015 competition. Results show that although higher surrogate accuracy generally enhances search performance, the impact varies by strategy. PS exhibits steady improvement with increasing accuracy, while IB and GB maintain robust performance beyond a certain threshold. Notably, SAEAs with accuracy above 0.6 consistently outperform the baseline without surrogates. In strategy comparisons, GB performs best across a wide accuracy range, IB excels at lower accuracies, and PS at higher ones. These findings support developing guidelines for selecting model management strategies based on surrogate accuracy.