Research On Ordinal Regression Prediction Method Of Candidate Solution In Evolutionary Algorithm

Surrogate-assisted evolutionary algorithm is the most commonly used method to solve expensive optimization problems.Most surrogate-assisted evolutionary algorithms reduce the number of real evaluation by establishing a surrogate that approximates the fitness function and replacing the real evaluation with the predicted results of the surrogate,so as to overcome the computational cost disaster.However,the establishment of high-precision surrogate of objective function often depends on the prior knowledge of the original black box problem and model type,and the lack of such knowledge will lead to a significant reduction in the prediction performance of the model.In order to obtain a good generalization performance of the surrogate and reduce the accuracy requirements of the algorithm on the surrogate,the ordinal regression prediction method and its application in evolutionary algorithms are studied in this thesis.An ordinal evaluation method for the merits of candidate solutions is proposed and the realization method of ordinal regression prediction is studied.Compared with the conventional regression(function approximation surrogate)prediction method,the ordinal regression prediction method has strong robustness in single objective and multi-objective optimization problems.The training data provided by Pareto rank is not enough to fully reflect the useful information contained in the data,and it is difficult to predict the unknown candidate solutions effectively.By mining the relative position information of candidate solution set in the object space,a Pareto ordinal method is proposed to evaluate the performance of candidate solutions,which increases the diversity of layer.The comparison experiments with the non-dominated rank(Pareto rank)show that Pareto ordinal regression prediction improves the accuracy of model prediction.In view of the uncertainty of prediction results,in order to use the prediction results to assist the evolutionary algorithm more reliably,a method to analyze the uncertainty of model prediction is obtained by using DCGS(discounted cumulative gain of surrogates),and the rationality and effectiveness of the method were verified by the simulation data.For single objective optimization problem,ordinal regression evolutionary strategy(ORES),which integrates scalar ordinal regression prediction with evolutionary strategy,is implemented.Experimental results on classical single-objective test functions show that ORES algorithm has obvious advantages over function approximation surrogate-assisted evolution strategy in convergence performance.For multi-objective optimization problem,an ordinal regression genetic algorithm(ORGA)integrating Pareto ordinal regression prediction method and NSGA-II was implemented.Simulation experiments on ZDT and UF test functions show that compared with the main surrogate-assisted evolutionary algorithm,the non-dominated solution set obtained by ORGA under limited resources has better convergence and diversity.