A novel surrogate-assisted evolutionary algorithm with an uncertainty grouping based infill criterion

Abstract

For tackling expensive optimization problems (EOPs), surrogate-assisted evolutionary algorithms (SAEAs) will run the evolutionary search and then select some promising solutions to be evaluated as predicted by the surrogate models. Different model management criteria for the surrogate models, such as improvement probability, expected improvement, and lower confidence bound, have shown their effectiveness when solving EOPs. In this paper, a novel SAEA with an uncertainty grouping based infill criterion, called SAEA-UGC, is proposed, in which the uncertainty is treated as an indicator to select solutions for training the models. The selected solutions are adopted to train an ensemble surrogate model and a radial basis function model, which run the global search and local search respectively. After obtaining the predicted value and uncertainty for all solutions in global search, they are evenly grouped according to uncertainty and a best predicted solution from each group is selected to form a new population. The global search and local search are cooperative to find the optimal value for the target EOP, i.e., global search or local search will continually run if an improved value for the target EOP can be found in each iteration; otherwise, the switch between global search and local search will happen. The performance of SAEA-UGC is validated when tacking 20 widely used test problems with various properties. The experimental results confirm the superiority of SAEA-UGC over four representative SAEAs in solving a majority of test EOPs.