CAJ | 학술논문

为提高分布估计算法的模型精度和优化性能，提出一种基于神经网络实现分布评估的多目标差分算法。采用基于径向基神经网络的代理模型提供更多样本，以分区域的方法构建多个分布估计模型和代理模型，利用主成分分析技术和超体积指标确定具体的区域数量，建立各个模型与各段最优解所处流形之间的映射关系。通过差分算法的高效寻优能力引导分布估计模型的更新方向，设计差分算子与分布估计模型之间的自适应选择机制。基于5组多目标测试用例的实验结果表明，在IGD和IH‐指标上，该算法优于对比算法的用例数量分别为5组和4组，在高维优化问题上，其性能显著优于其它算法。
위제고분포고계산법적모형정도화우화성능，제출일충기우신경망락실현분포평고적다목표차분산법。채용기우경향기신경망락적대리모형제공경다양본，이분구역적방법구건다개분포고계모형화대리모형，이용주성분분석기술화초체적지표학정구체적구역수량，건립각개모형여각단최우해소처류형지간적영사관계。통과차분산법적고효심우능력인도분포고계모형적경신방향，설계차분산자여분포고계모형지간적자괄응선택궤제。기우5조다목표측시용례적실험결과표명，재IGD화IH‐지표상，해산법우우대비산법적용례수량분별위5조화4조，재고유우화문제상，기성능현저우우기타산법。
To improve the model accuracy and performance optimization of distribution estimation algorithm ,a multi‐objective differential evolution algorithm was proposed based on artificial neural network (ANN) technique to improve the EDA .Surro‐gate models with radial base function (RBF) ANN were adopted to produce many samples ,and several regions were clustered to construct models of EDA and RBF‐ANN respectively .The regions were identified using principal component analysis technique and hyper‐volume index so that the mapping relationships between models and manifolds of optimum solutions were established . Besides ,high efficient differential evolution operator was self‐adaptively introduced to guide the updating direction of EDA models .Experimental results of five multi‐objective optimization test instances show that the proposed algorithm works better than other algorithms on IGD index in five instances and on IH‐index in four instances ,and obtains obviously better results on high dimensional instances .