Expensive Multiobjective Optimization With Adaptive Contribution Sampling and Enhanced Search

In surrogate-assisted expensive multiobjective evolutionary optimization, selecting individuals for expensive function evaluations has received widespread attention to improve the model. However, most algorithms search for optimal solutions based on estimated function values obtained by the regression model, followed by choosing the infill individual to update the model following the estimated population distribution, which may mislead the location of optimal solutions. Therefore, this article proposes an extended estimation and adaptive contribution sampling approach to find optimal solutions to expensive optimization problems. In evolution, the estimation value of each individual is composed of the prediction of each objective, together with the estimated uncertainty obtained by the GP model to enhance the exploration of the optimum. Subsequently, the adaptive contribution-based sampling strategy is applied to select the promising individual for expensive function evaluations to guide the approximate model search. In this contribution mechanism, the sparseness change of each sample is used to guide the selection of the infill solution with a performance indicator or the estimated confidence level. Experiments on two multiobjective test suites and the practical problem demonstrate the effectiveness of the extended estimation and adaptive contribution sampling approach compared to seven recently prevailing algorithms.