Font Size: a A A

The Prediction Of Objective Function Difference Value In Expensive Multi-objective Optimization Problems

Posted on:2022-04-30Degree:MasterType:Thesis
Country:ChinaCandidate:R YuFull Text:PDF
GTID:2518306566951289Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Traditional evolutionary algorithms are highly dependent on the true evaluation of the objective function.And the expensive multi-objective optimization problem may bring huge evaluation cost for just one real evaluation,so it is necessary to use the surrogate model to assist the evolutionary algorithm to complete the problem.The evolutionary algorithm assisted by the surrogate model is a data-driven method that can alleviate the cost catastrophe to a certain extent.Although this method can effectively solve the black box problem,the modeling accuracy of the proxy model is also highly dependent on the prior information,and the it’s also difficulty to model such a high-accuracy proxy model.It brings additional modeling costs for solving problems.This paper mainly studies the objective function difference value prediction and the simplification strategy of the difference value sample set.The construction strategy of the difference value sample set and the method of the objective function difference value prediction are proposed.The Pareto dominated relationship between the candidate solutions is caculated via objective function difference values.In addition,this paper studies the prediction method of Pareto dominated relationship and characteristics of objective function difference value prediction,and a simplified strategy of objective function difference samples is proprosed,then a congenital advantage that difference value sample set constructed by known samples and unknown samples has high prediction accurcy is found.The prediction advantage establishes the basis for integrated application of the objective function difference prediction method in the following text.Since the nature of objective function difference value prediction is a series of scalar predictions of black box function difference values and it replaces the traditional Pareto dominance relationship prediction which is a two-level vector prediction,and reduces the learning difficulty of the surrogate model.However,in the promotion of the method,the high complexity of the objective function difference value sample makes this prediction method can not be used in some complex and excellent models.In order to reduce the additional computational cost caused by the high complexity of the order relationship samples,this paper proposes a K-means clustering subspace division strategy,and the simulation experiment proved that the method can effectively reduce the computational cost of the surrogate model under the premise of the smallest possible loss of accuracy.By analyzing the advantages and disadvantages of the method and the characteristics of the objective function difference valueship samples,a reduction algorithm for the training sample set is proposed.And through further experiments,the effectiveness of the sample set reduction algorithm is proved,then the objective function difference valueship prediction method is successfully applied to the Kriging model,and high prediction accuracy is achieved.In order to verify the feasibility of the objective function difference valueship prediction method,this paper selects the SPEAⅡ algorithm framework which is highly adaptable to the objective function difference valueship prediction.The objective function difference valueship prediction is integrated with the Kriging regression prediction model,and an improved OR-SPEAⅡ algorithm is obtained.Experiments have proved that the OR-SPEAⅡ algorithm can get an acceptable Pareto front in the ZDT and UF series of some two-objective test problems,and compared with other EMOEAs,it has also achieved certain advantages,which proves the feasibility of the objective function difference valueship prediction method.
Keywords/Search Tags:expensive multi-objective optimization, Pareto dominance prediction, subspace prediction, sample set reduction, Kriging model, SPEAⅡ
PDF Full Text Request
Related items