Clustering and learning Gaussian distribution for continuous optimization
Clustering and learning Gaussian distribution for continuous optimization
复制标题
DOI：
10.1109/tsmcc.2004.841914
复制
发表时间：
2005-05
期刊：
影响因子：
--
通讯作者：
Qiang Lu;X. Yao
中科院分区：
文献类型：
--
作者：
Qiang Lu;X. Yao
由于引入了分布算法（EDA）的估计，因此已经开发了连续域中的不同方法。最初，在构建概率模型时，通常会广泛使用单个高斯分布，在处理多模式函数时，通常会误导搜索。后来，一些研究人员构建了通过使用聚类技术利用混合概率分布的EDA。但是，他们的算法在应用聚类之前都需要先验知识，这在现实生活中是不合理的。在本文中，提出了两个用于连续优化的新EDA，两者都将聚类技术纳入估算过程中，以打破单个高斯分布假设。基于BGE度量和聚类以及对高斯分布算法的估计的高斯网络算法的新算法，群集和估计，不仅显示出极大的优势，在使用几个本地Optima优化多模式功能方面，还可以在聚类之前克服限制要求先验知识的限制。通过使用非常可靠的聚类技术，竞争对手受到竞争的竞争学习。这是EDA首次能够自动检测全局最佳的数量。已经实施了一组实验来评估新算法的性能。除了根据没有免费的午餐理论，除了对某些多模式功能的改进外，还显示了它们的弱方面。
Since the Estimation of Distribution Algorithm (EDA) was introduced, different approaches in continuous domains have been developed. Initially, the single Gaussian distribution was broadly used when building the probabilistic models, which would normally mislead the search when dealing with multimodal functions. Some researchers later constructed EDAs that take advantage of mixture probability distributions by using clustering techniques. But their algorithms all need prior knowledge before applying clustering, which is unreasonable in real life. In this paper, two new EDAs for continuous optimization are proposed, both of which incorporate clustering techniques into estimation process to break the single Gaussian distribution assumption. The new algorithms, Clustering and Estimation of Gaussian Network Algorithm based on BGe metric and Clustering and Estimation of Gaussian Distribution Algorithm, not only show great advantage in optimizing multimodal functions with a few local optima, but also overcome the restriction of demanding prior knowledge before clustering by using a very reliable clustering technique, Rival Penalized Competitive Learning. This is the first time that EDAs have the ability to detect the number of global optima automatically. A set of experiments have been implemented to evaluate the performance of new algorithms. Besides the improvement over some multimodal functions, according to the No Free Lunch theory, their weak side is also showed.