一种快速收敛的改进贝叶斯优化算法  

Improved Bayesian optimization algorithm with fast convergence

作　　者：王翔[1] 郑建国[1] 张超群[1] 刘荣辉[1] 

机构地区：[1]东华大学旭日工商管理学院,上海200092

出　　处：《华中科技大学学报（自然科学版）》2011年第6期66-70,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(70971020)

摘　　要：针对贝叶斯优化算法(BOA)中学习贝叶斯网络结构时间复杂度较高的问题,提出了一种可以快速收敛的基于K2的贝叶斯优化算法(K2-BOA).为了提升收敛速度,在学习贝叶斯网络结构的步骤中进行了2处改进:首先,随机生成n个变量的拓扑排序,加大了算法的随机性;其次,在排序的基础上利用K2算法学习贝叶斯网络结构,减少了整个算法的时间复杂度.针对3个标准Benchmark函数的仿真实验表明:采用K2-BOA算法和BOA算法解决简单分解函数问题时,寻找到最优值的适应度函数评价次数几乎相同,但是每次迭代K2-BOA算法运行速度提升明显;当解决比较复杂的6阶双极欺骗函数问题时,K2-BOA算法无论是运行时间还是适应度函数评价次数,都远小于BOA算法.K2-Bayesian optimization algorithm （BOA） with fast convergence was proposed to enhance the convergence rate figuring out the problem that the time complexity of learning Bayesian networks was high in the Bayesian optimization algorithm. There were two improvements in learning Bayesian network of the new algorithm, the topological sort of n variables was randomly generated for increasing the randomness of the algorithm, and on the basis of the sort K2 algorithm was used to learn Bayesian network structure to reduce the time complexity of the new algorithm. The simulation resuits for three benchmark functions show two conclusions. Firstly, when 3-deceptive function and trap-5 function are solved, the number of fitness function evaluation of K2-Bayesian optimization algo- rithm is almost the same as that of Bayesian optimization algorithm; however the running time of K2- Bayesian optimization algorithm is less than that of Bayesian optimization algorithm. Secondly, when 6-bipolar function is solved, the number of fitness function evaluation and the running time of K2- Bayesian optimization algorithm are much better than those of Bayesian optimization algorithm.

关 键 词：贝叶斯优化算法 快速收敛 分布式估计算法 K2算法 B算法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]