基于多主体系统的多分类直推学习  

Multi-agent-system-based multi-class transductive learning

作　　者：潘俊[1] 孔繁胜[1] 王瑞琴[1] 

机构地区：[1]浙江大学人工智能研究所,浙江杭州310027

出　　处：《计算机集成制造系统》2009年第8期1656-1663,共8页Computer Integrated Manufacturing Systems

摘　　要：针对少量样本已标记和大量样本未标记的多分类问题,提出了一种新颖的基于多主体系统的直推学习方法。该方法将以Agent表示的样本点随机映射到输出空间构成初始空间格局,空间格局随时间演化的过程是一个自组织的马尔可夫过程,它将在有限时间内达到平稳分布,从而求得最佳的标记分布。根据该方法,给出了两个多主体系统直推学习算法,并讨论了算法的收敛性和复杂度。最后在两个数据集上进行了仿真测试,表明了算法的有效性与实用性。Aiming at the problem of multielass classification in which both a few labeled data and lots of unlabeled data were given a novel approach called Multi-Agent-System-Based Multi-Class Transductive Learning was presented. All the data objects were carried by agents and then mapped to the output space, and the spatial configuration of the agents formed a self-organizing Markov stochastic process. The Markov process finally converged to a stationary probability distribution, in which an optimal label distribution was obtained. Based on the proposed approach, two MMTA （Multi-Agent-System-Based Multi-Class Transductive Algorithms） algorithms were put forward and their convergence as well as time complexities were discussed. The simulations were provided to demonstrate the effectiveness and practicability of MMTA algorithms.

关 键 词：直推式学习 多主体系统 自组织 多分类 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]