HAS-EA: a fast parallel surrogate-assisted evolutionary algorithm

Many real-world phenomena are modeled as expensive optimization problems (EOPs) that are not readily solvable without extensive computational cost. Surrogate-assisted mechanisms and parallel computing techniques are effective approaches to improving the search performance of evolutionary algorithms for these EOPs. However, the search efficiency of existing methods are limited by a combination of synchronization barriers and a failure to use heterogeneous computing resources fully. Therefore, we propose an efficient heterogeneous asynchronous parallel surrogate-assisted evolutionary algorithm (HAS-EA). The proposed HAS-EA incorporates an improved asynchronous parallel evolutionary algorithm module on the CPU, a surrogate module on the GPU, and an improved asynchronous recommendation module on the CPU. By performing these operations in parallel on heterogeneous computing resources, the search performance can be accelerated. Test results of our proposed method with several benchmark problems and a real-world model calibration problem demonstrate that HAS-EA offers better performance than other recently published methods in solving EOPs.