As a swarm intelligent optimization algorithm, particle swarm optimization(PSO) has been widely used in the practical engineering with its strongadvantage of global convergence ability, faster convergence speed and simpleimplementation, and it dose not require the objective function or constraintfunction continuous or differentiable. A variety of optimization problems inpractical engineering, however, tend to have a variety of constraint conditions.Therefore, how to process and evaluate the constraint conflicts is very importantfor PSO to solve the constrained optimization problems. However, with theincreasing of the complexity of the engineering problems, the time to evaluatethe design performance (that is, calculate the objective function within the scopeof constraints) increases. Because the particle swarm optimization algorithmneeds a lot of constraint and objective function evaluations before obtaining aglobal optimal solution, the application of particle swarm optimization algorithmin these problems will be impeded. Till now, researchers have studied ondifferent approaches for solving unconstrained optimization problems whoseobjective functions are time-consuming. While for constraint evaluationexpensive optimization problems, most researches transformed the constrainedoptimization problems to unconstrained ones using a penalty function, and thenfound optima using the approaches proposed before. However, how to set thepenalty factor is an optimization problem itself, and the value of penalty factorwill directly affect the accuracy of the algorithm.The constraint preserving method is a kind of constraint handling method forthe particle swarm optimization algorithm to handle the constraint problems.The constraint preserving method is easy to be implemented for its simplicity.The most important aspect is that it does not need to adjust penalty parameterslike penalty function methods. However, in order to judge whether a newposition is in the feasible region or not, a lot of constraint evaluations is needed.Therefore, the constraint preserving method is not suitable to be directly used insolving problems with time-consuming constraint functions value directly. Considering that any individual in the population only has two situations: satisfyor dissatisfy the constraints, so in this article we put forward using classificationwhich aims to reduce the constraint evaluations to assist PSO for solvingconstraint computation expensively problems. First, a classification wasproposed to be trained offline, and then it will be used to approximate whether anew position is in the feasible region or not. Only when it is approximated in thefeasible region, the constraints will be real evaluated. We call this algorithm theoffline_training classifier-assisted PSO algorithm. Although the offline_trainingclassifier-assisted PSO algorithm can reduce the number of the constraintevaluations in some extent, the quality of classification will influence theapproximation accuracy on feasibility of a new position, which consequentlyinfluences the constraint evaluations. In order to improve the quality of theclassification, we proposed to add a pair of feasible-infeasible positionsproduced in the evolution into the training database on the basis of theoffline_training classifier-assisted PSO algorithm, which we call theonline_training classifier-assisted PSO algorithm in this article.In this article, the support vector machine was used as our classifier toapproximate the computationally expensive constraints. And bothoffline_training and online_training classifier-assisted PSO algorithm have beenevaluated on13standard functions, and the results showed that the SVM-assisted particle swarm optimization algorithm can greatly reduce the number ofthe constraint evaluations with almost same results obtained by PSO withoutSVM assisting. In addition, the online_training classifier-assisted PSO algorithmobtained better approximate results than the offline_training one, whichcorrespondingly meant that the former can reduce more constraint evaluationsthan the later, and the optimization efficiency of the algorithms improved muchmore.