摘要
GLA具有较强的抗噪声能力,但是其收敛的稳定性和学习速度是一对矛盾。通常为保证收敛的稳定性,需要选取足够小的步长,但过小的步长会导致训练时间过长。结合自适应步长的原理,提出改进型的算法TDBDGLA。实验结果表明,与采用同种强化方案的GLA相比,TDBDGLA取得更低的误分率,并且对于给出的衡量稳定性和学习速度的指标,TDBDGLA比GLA提高了11%以上。
GLA has strong noise-tolerant ability, while its stability of convergence is in contradiction with learning speed. In general, to guarantee the stability of convergence, it requires a step size that is small enough. However, if the step size is too small, it will result in very long training time. The principle of self-adaptive step size is incorporated to propose the improved TDBDGLA algorithm. The experimental results show that compared with GLA using the same reinforcement scheme, TDBDGLA achieves lower classification error rate. In the given metric for measuring stability and learning speed, TDBDGLA improves by more than 11% compared with GLA.
出处
《信息安全与通信保密》
2014年第3期76-79,共4页
Information Security and Communications Privacy
基金
国家自然科学基金资助项目(批准号:61271316
61071152)
国家973计划重大基础研究资助项目(编号:2010CB731403
2010CB731406
2013CB329605)
国家十二五科技支撑计划(编号:2012BAH38 B04)
上海市信息安全综合管理技术研究重点实验室基金
关键词
广义学习自动机
抗噪声
线性分类器
自适应
步长
generalized learning automata
noise-tolerant
linear classifier
self-adaptive
step size
