基于遗传算法的多任务学习  

Study on Multi-task Learning Based on Genetic Algorithm

作　　者：孟浩华[1] 李国正[1] 

机构地区：[1]上海大学计算机工程与科学学院,上海200072

出　　处：《计算机科学》2008年第10期186-187,203,共3页Computer Science

基　　金：国家自然科学基金资助(20503015)

摘　　要：机器学习中冗余特征会降低学习器的性能,而特征选择方法可以去掉一些冗余特征。然而,冗余特征也包含有用信息,因此可以利用多任务学习的概念,通过重复利用冗余特征提高预测精度。但是,如何确定哪些特征作为输入和输出仍然是一个待解决的问题。之前的工作是在多任务学习当中,运用遗传算法来确定哪些特征作为输入,哪些作为输出,取得了较好的效果,但是该算法不足之处是没有考虑到不相关特征。现将特征分为三部分:输入的特征、输出的特征和不相关特征,提出了对一个特征进行双位编码的遗传算法搜索策略。在基因芯片数据上的实验结果表明,提出的新算法e-GA-MTL比已有基于遗传算法的GA-MTL和其它启发式方法效果更好。Redundant features hurt the performance of learning methods. Feature selection methods were developed to remove some redundant features; however, the redundant features contain useful information, therefore, multi-task learning was developed to employ the removed redundant information to improve prediction accuracy. Adding which features to the target and/or the input during multi-task learning is still an open issue. The previous study on multi-task learning uses genetic algorithm to determine the features for the target and/or input, and which has been proved effective. In this paper,we classified the features into three parts： the input features,the output features and irrelevant features. We proposed a new search strategy of the genetic algorithm which encodes double bits for one feature. Experimental results on the microarray data sets show that the novel algorithm e-GA-MTL obtains better performance than the previous algorithm GA-MTL and the other heuristic algorithms.

关 键 词：遗传算法 多任务学习 双位编码 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] O242.23[自动化与计算机技术—控制科学与工程]