Adaptive Knowledge Fusion via Dynamic Stacking for Surrogate-Assisted Transfer Optimization

Abstract

The integration of transfer learning with surrogate-assisted optimization represents a highly promising paradigm for complex engineering problems. However, it is fundamentally constrained by insufficient cross-domain adaptation and the absence of robust knowledge-fusion techniques. To address these limitations, this study proposes an adaptive knowledge-fusion method via dynamic stacking for surrogate-assisted transfer optimization. Domain adaptation is achieved through polynomial chaos expansion/augmented radial basis function surrogate modeling of source tasks, coupled with domain transformation coefficients derived from target task data. A cosine-similarity metric computed directly from surrogate predictions is introduced to quantify intertask relationships. This similarity measure is then integrated into an iterative updating framework, in which transfer-weight coefficients are dynamically computed to enable the adaptive construction of stacked surrogate models. Target task optimization is further refined via a balanced sampling criterion. The proposed method is validated through benchmark functions, a two-dimensional airfoil optimization case, and a multidisciplinary solid rocket motor design problem. Compared with existing transfer optimization approaches, the proposed method substantially improves optimization efficiency while achieving superior knowledge transfer capability and robustness.