Optimization on Metamodeling Supported Iterative Decomposition

Abstract

The recently developed metamodelbased decomposition strategy relies on quantifying the variable correlations of blackbox functions so that highdimensional problems are decomposed to smaller subproblems, before performing optimization. Such a twostep method may miss the global optimum due to its rigidity or requires extra expensive sample points for ensuring adequate decomposition. This work develops a strategy to iteratively decompose highdimensional problems within the optimization process. The sample points used during the optimization are reused to build a metamodel called principal component analysishigh dimensional model representation (PCAHDMR) for quantifying the intensities of variable correlations by sensitivity analysis. At every iteration, the predicted intensities of the correlations are updated based on all the evaluated points, and a new decomposition scheme is suggested by omitting the weak correlations. Optimization is performed on the iteratively updated subproblems from decomposition. The proposed strategy is applied for optimization of different benchmarks and engineering problems, and results are compared to direct optimization of the undecomposed problems using trust region mode pursuing sampling method (TRMPS), genetic algorithm (GA), cooperative coevolutionary algorithm with correlationbased adaptive variable partitioning (CCEAAVP), and divide rectangles (DIRECT). The results show that except for the category of undecomposable problems with all or many strong (i.e., important) correlations, the proposed strategy effectively improves the accuracy of the optimization results. The advantages of the new strategy in comparison with the previous methods are also discussed.