An Evolutionary Algorithm with Classifier Guided Constraint Evaluation Strategy for Computationally Expensive Optimization Problems

Abstract

Practical optimization problems often involve objective and constraint functions evaluated using computationally expensive numerical simulations e.g. computational fluid dynamics (CFD), finite element methods (FEM) etc. In order to deal with such problems, existing methods based on surrogates/approximations typically use cheaper and less accurate models of objectives and constraint functions during the search. Promising solutions identified using approximations or surrogates are only evaluated using computationally expensive analysis. In the event the constraints and objectives are evaluated using independent computationally expensive analysis (e.g. multi-disciplinary optimization), there exists an opportunity to only evaluate relevant constraints and/or objectives that are necessary to ascertain the utility of such solutions. In this paper, we introduce an efficient evolutionary algorithm for the solution of computationally expensive single objective constrained optimization problems. The algorithm is embedded with selective evaluation strategies guided by Support Vector Machine (SVM) models. Identification of promising individuals and relevant constraints corresponding to each individual is based on SVM classifiers, while partially evaluated solutions are ranked using SVM ranking models. The performance of the approach has been evaluated using a number of constrained optimization benchmarks and engineering design optimization problems with limited computational budget. The results have been compared with a number of established approaches based on full and partial evaluation strategies. Hopefully this study will prompt further efforts in the direction of selective evaluation, which so far had attracted little attention.