Adaptive multi-surrogate and module-based optimization algorithm for high-dimensional and computationally expensive problems

Surrogate-assisted evolutionary algorithms (SAEAs) have been widely used for solving computationally expensive optimization problems. However, the performance of SAEAs deteriorates drastically for high-dimensional problems since the accuracy of surrogates may decrease to an insupportably low level in high dimensions which profoundly affect subsequent evolution optimization. To overcome this issue, this paper proposes an adaptive multi-surrogate and module-based optimization algorithm named AMSMO. Unlike existing algorithms that adopt sole surrogate models, AMSMO develops a real-time autoselection of kernels approach for building the most promising RBF. AMSMO makes use of five modules to promote the optimization quality, including: a. global search module; b. local search module; c. semi-local search module; d. convergence module; e. restart module. Specifically, we provide a novel distance- and objective-value-based criterion for global search. The semi-local search module applies a dynamic search criterion that makes the algorithm explore at the early search stage and exploit at the final stage. A cluster-based restart criterion is proposed to address the stagnation of optimization. Several benchmark functions with dimensions varying from 30 to 100 and computationally intensive practical simulation–optimization problems are adopted to validate the proposed algorithm. The empirical experiments demonstrate that the proposed algorithm outperforms other state-of-the-art algorithms in most cases.