聚类多目标演化算法及其应用研究

发布时间：2018-11-08 19:30

【摘要】：无论生产或生活过程中,人们总能遇到大量的复杂多目标优化问题,此类问题一般具有多个自变量,多个等式或不等式约束条件以及多个非线性的目标量等。利用传统的方法,如加权法,约束法等不能很好地解决此类问题,而多目标演化算法可以不受问题规则特性的限制,具有多种优点,获得了显著的成果。多目标演化算法主要由新解产生和环境选择两部分组成,目前大部分关注与研究内容聚集在环境选择方面,对新解产生算子钻研与学习非常少。故本文将目前所盛行的机器学习中一种典型的方法——聚类技术与多目标演化算法合理融合,充分考虑问题的规则特性,为算法研究高效的,经改良的新解产生方式,使算法具有更佳的求解性能。首先,本文针对多目标分布估计算法对问题的规则特性考虑不够,对群体演化过程中得到的异常解的处置方法欠佳,群体中解的多样性容易丢失,巨大的计算开销用于构建最优概率模型等不足,研究了一种基于聚类技术改进的多目标分布估计算法(CEDA)。CEDA在每一次循环迭代中利用凝聚层次聚类算法对种群数据进行分析,得出群体解分布结构信息,基于此结构信息,为所有解均建立一个多元高斯模型,依据此模型选择适当的样本,获得新个体。为了降低建模计算开销,邻近个体共享相同的协方差矩阵建立高斯模型。基于标准测试题对比结果显示CEDA可以解决十分复杂的问题。然后,本文针对多目标粒子群算法在求解过程时,虽然具有很高的收敛速度,但是容易丢失种群多样性的不足,研究了基于聚类技术改进的MOPSO(CPSO)。CPSO在每一次迭代循环产生新解过程中,运用聚类算法对所有个体聚类分析,每一个个体的配对个体分别以确定的几率从全局或局部种群挑选,另外为了更好的维持种群解的多样性与算法的收敛速度之间的平衡,自适应的调整新解产生方式为粒子群算法或多样性保持好的复合差分进化算法。基于标准测试题对比实验表明CPSO同样能够解决复杂的问题。最后,本文将新研究的两种基于聚类的多目标演化算法应用于返回式卫星舱布局优化与某轻型飞机的齿轮减速器优化设计问题中,求证了新算法在解决实际工程应用中表现。
[Abstract]:Whether in production or life, people can always encounter a large number of complex multi-objective optimization problems, such problems generally have multiple independent variables, multiple equality or inequality constraints, as well as a number of nonlinear objective quantities and so on. The traditional methods, such as weighted method and constraint method, can not solve this kind of problem well, but the multi-objective evolutionary algorithm can not be restricted by the characteristic of the rule of the problem, so it has many advantages, and has obtained remarkable results. The multi-objective evolutionary algorithm is mainly composed of two parts: new solution generation and environment selection. At present, most of the research contents focus on the environment selection, and the research on the new solution generation operator is very little. In this paper, a typical method of machine learning, clustering technique and multi-objective evolutionary algorithm, is combined reasonably, and the rule characteristic of the problem is fully taken into account in this paper, which is an efficient and improved new solution generation method. The algorithm has better solution performance. First of all, the algorithm of multi-objective distribution estimation is not enough to consider the rule of the problem, and the method to deal with the abnormal solutions in the process of population evolution is poor, and the diversity of solutions in the population is easy to be lost. The huge computational overhead is used to build the optimal probability model and so on. In this paper, an improved multi-objective distribution estimation algorithm based on clustering technique, (CEDA). CEDA, is proposed to analyze the population data in each cycle iteration, and to obtain the distribution structure information of the population solution, based on the structure information. A multivariate Gao Si model is established for all solutions, according to which suitable samples are selected and new individuals are obtained. In order to reduce the overhead of modeling, neighboring individuals share the same covariance matrix to build Gao Si model. The comparison results based on standard test questions show that CEDA can solve very complex problems. Then, in order to solve the problem of multi-objective particle swarm optimization, although it has a high convergence rate, it is easy to lose the deficiency of population diversity. In this paper, we study the application of clustering algorithm to the clustering analysis of all individuals in the process of generating new solutions in each iteration cycle of MOPSO (CPSO). CPSO based on the improved clustering technology. In order to maintain the balance between the diversity of the population solution and the convergence rate of the algorithm, each individual is selected from the global or local population with a definite probability. Adaptive new solutions are generated by particle swarm optimization (PSO) or composite differential evolution (DEA) with good diversity. The contrast experiment based on standard test shows that CPSO can solve complex problems as well. Finally, in this paper, two new multi-objective evolutionary algorithms based on clustering are applied to the optimization of recoverable satellite cabin layout and the optimal design of gear reducer of a light aircraft, and the performance of the new algorithm in solving practical engineering applications is verified.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【参考文献】