Partial Evaluation Strategies for Expensive Evolutionary Constrained Optimization

Constrained optimization problems (COPs) are frequently encountered in real-world design applications. For some COPs, the evaluation of the objective(s) and/or constraint(s) may involve significant computational/temporal/financial cost. Such problems are referred to as expensive COPs (ECOPs). Surrogate modeling has been widely used in conjunction with optimization methods for such problems, wherein the search is partially driven by an approximate function instead of true expensive evaluations. However, for any true evaluation, nearly all existing methods compute all objective and constraint values together as one batch. Such full evaluation approaches may be inefficient for cases where the objective/constraint(s) can be evaluated independently of each other. In this article, we propose and study a constraint handling strategy for ECOPs using partial evaluations. The constraints are evaluated in a sequence determined based on their likelihood of being violated; and the evaluation is aborted if a constraint violation is encountered. Modified ranking strategies are introduced to effectively rank the solutions using the limited information thus obtained, while saving on significant function evaluations. The proposed algorithm is compared with a number of its variants to establish the utility of its key components systematically. Numerical experiments and benchmarking are conducted on a range of mathematical and engineering design problems to demonstrate the efficacy of the approach compared to state-of-the-art evolutionary optimization approaches.

Automated summary

This article proposes a constraint handling strategy for expensive constrained optimization problems using partial evaluations, where constraints are evaluated in a sequence based on their likelihood of being violated and the evaluation is aborted if a constraint violation is encountered. The proposed algorithm is compared with several variants to establish the utility of its key components, and numerical experiments and benchmarking are conducted on a range of mathematical and engineering design problems to demonstrate its efficacy compared to state-of-the-art evolutionary optimization approaches.