Ensemble-Assisted Multi-Objective Optimization Algorithm Combining Feature Perturbation and Allocation Strategy

Abstract

A surrogate model can use its approximate prediction to replace the real evaluation of an algorithm when applied to the multi-objective optimization problem, which greatly reduces the number of real fitness evaluations required by the multi-objective optimization algorithm for determining the optimal optimization algorithm.To improve the accuracy of the surrogate model under high-dimensional problems and reduce its computational overhead, this study proposes an ensemble-assisted multi-objective optimization algorithm based on feature perturbation and allocation strategy.First, the algorithm uses a Radial Basis Function Network(RBFN) and a Support Vector Regression(SVR) surrogate model as the basis model during the integration, which reduces the computational cost of the algorithm on high-dimensional problems.Second, the algorithm combines feature disturbance and memory-based influence factor allocation strategies to construct an integrated surrogate model and improve the accuracy of the integration.Finally, the algorithm uses the integrated prediction value and the weighted uncertainty information to assist in the management of the integrated surrogate model, balance the global search and local exploration, and enhance the optimization capability of the algorithm within the target space.The experimental results show that the distribution and convergence of the solution set obtained by this algorithm on the ZDT1~ZDT3 and ZDT6 test problems are better than those of a classical algorithm.In addition, when the number of dimensions of the decision variables increases, the ensemble surrogate model used by the algorithm reduces the number of fitness evaluations by approximately 90% in comparison with the Kriging surrogate model, and at the same time can obtain more accurate prediction results.