New evolutionary algorithm uses Pareto-optimal surrogates to beat SOTA SAEAs

Surrogate-assisted evolutionary algorithms (SAEAs) are a go-to approach when fitness evaluations are expensive, such as aerodynamic design or hyperparameter optimization. They replace costly simulations with cheap surrogate models, but the quality of the surrogate heavily influences search performance. Ensembles can improve prediction robustness, yet their effectiveness depends on mixing models with different degrees of smoothness—something most existing adaptive SAEAs ignore while focusing solely on prediction accuracy. In a new arXiv paper (2608.01777), researchers Kei Nishihara, Yaochu Jin, and Masaya Nakata tackle this gap with an adaptive ensemble SAEA that constructs its surrogates by solving a bi-objective optimization problem. Instead of just minimizing approximation error, the algorithm also minimizes model complexity, producing a Pareto frontier of radial basis function networks (RBFNs) with varied smoothness. This automatic structure search yields accurate surrogate models that are neither overfitted nor underfitted to the training data.

The authors also introduce a tailored infill criterion that lets surrogates with different smoothness levels contribute to prescreening candidate solutions. This avoids the common pitfall where one overly smooth model dominates selection. Experiments on a single-objective benchmark and real-world problem sets under an expensive optimization scenario show the proposed algorithm statistically outperforms state-of-the-art SAEAs. The source code has been released, allowing other researchers to reproduce results and apply the method to their own problems. The paper is accepted for IEEE Transactions on Cybernetics (Early Access, 2026), adding peer-reviewed validation. For practitioners, this means more reliable optimization in engineering and ML workflows where function evaluations are costly, with less manual tuning of surrogate models.