受噪聲影響的復擬隨機樣本的STL關鍵定理
杜二玲1,李俊華2
(1.中國地質大學長城學院 基礎課教學部,河北 保定071000;2.河北大學 數學與信息科學學院,河北 保定071002)
摘要:引入了擬概率空間上復擬隨機樣本受噪聲影響的復經驗風險泛函、復期望風險泛函、復經驗風險最小化原則以及嚴格一致性的定義,提出并證明了擬概率空間上復擬隨機樣本受噪聲影響的學習理論關鍵定理,為系統建立擬概率空間上基于噪聲影響的復擬隨機樣本的統計學習理論奠定了基礎.
關鍵詞:復擬隨機樣本;噪聲;復經驗風險最小化原則;關鍵定理
DOI:10.3969/j.issn.1000-1565.2015.05.001
中圖分類號:O29;TP181文獻標志碼:A
收稿日期:2014-11-30
基金項目:河北省教育廳科研項目(QN20131055);河北省高等學??茖W技術研究項目(Z2013038)
Key theorem of statistical learning theory with complex
quasi-random samples corrupted by noise
DU Erling1, LI Junhua2
(1. Basic Teaching Department, China University of Geosciences Great Wall College, Baoding 071000,
China;2. College of Mathematics and Information Science, Hebei University, Baoding 071002, China)
Abstract:Some new concepts, such as complex empirical risk functional, complex expected risk functional, complex empirical risk minimization principle, and strict consistency built on quasi-probability space and based on complex quasi-random samples corrupted by noise are introduced. The key theorem of learning theory is given and proved on quasi-probability space and based on complex quasi-random samples corrupted by noise. The investigations will help lay essential theoretical foundations for the systematic and comprehensive development of the complex quasi-random samples corrupted by noise.
Key words: complex quasi-random samples; noise; complex empirical risk minimization principle; key theorem
MSC 2010: 28B99
第一作者:杜二玲(1975-),女,河北安國人,中國地質大學長城學院講師,主要從事不確定統計學習理論.
E-mail:duerling@126.com
統計學習理論(statistical learning theory, SLT)是Vapnik等[1-2]在20世紀60年代末提出,于90年代中期發展較成熟, 被學術界公認為較好地處理小樣本的學習理論. SLT是建立在概率空間上且所研究的樣本是實隨機樣本. 而概率的可加性條件非常強, 現實中還存在大量的非實隨機樣本. 為此, 一些學者已經開始從事非概率空間上和復隨機樣本的統計學習理論的研究, 得到了一些重要的成果[3-10].其次, SLT所研究的樣本總是事先假定不受外界干擾.這種假定在實際應用中往往得不到滿足. 噪聲是影響樣本的因素之一,也是人們考慮比較多的一種, 有學者開始了樣本受到噪聲影響的統計學習理論的研究[11-13].基于上述考慮, 本文在擬概率空間上引入了復擬隨機樣本受噪聲影響的一些基本定義, 討論了復擬隨機樣本受噪聲影響的學習理論的關鍵定理, 從而擴大了支持向量機等應用性研究領域的理論基礎, 拓展了統計學習理論的應用范圍.
1基本概念


定義1設Q′(z,α)=Q(z,α)+ξ是考慮噪聲之后的損失函數,ξ1,ξ2,…,ξl與ξ是獨立同分布的, 定義擬概率空間上復擬隨機樣本受等均值噪聲影響的復期望風險泛函為
R′(α)=E[Q′(z,α)]=E[Q(z,α)+ξ]=E(Q(z,α))+p=R(α)+p.
擬概率空間上復擬隨機樣本受等均值噪聲影響的復經驗風險泛函為


定義2假設復期望風險泛函的最小值在Q′(z,α0)上取得,復經驗風險泛函的最小值在Q′(z,αl)取得.用Q′(z,αl)逼近Q′(z,α0)的值. 這種在擬概率空間上解決最小化復期望風險泛函問題的方法被稱為復經驗風險最小化原則(CERM原則).
定義3對于擬概率空間上的復可測函數集Q′(z,α),α∈Λ和擬概率μ,如果對于該函數集的任何非空子集Λ(c)={α:‖R′(α)‖≥c},c∈(-∞,∞)和任意ε>0,收斂性
(1)
成立,則稱復經驗風險最小化原則對于擬概率空間上的復可測函數集Q′(z,α),α∈Λ和擬概率μ是嚴格(非平凡)一致的.
定義4對于擬概率空間上的復可測函數集Q′(z,α),α∈Λ和擬概率μ,如果對于任意ε>0,

(2)
則稱式(2)為在擬概率空間中的給定復可測函數集上復經驗風險泛函到復期望風險泛函的一致單邊收斂性.
2主要結論

1)對于給定的擬概率μ,復經驗風險最小化方法對擬概率空間上的復可測函數集Q′(z,α),α∈Λ嚴格一致成立.

定理第1部分得證.
下面證明充分性. 假設一致單邊收斂性式(2)成立.




(3)



因此B?(B1∪B2).由大數定理知μ(B1)→0, 由契比雪夫不等式知μ(B2)→0,所以
μ(B)<μ(B1∪B2)→0,
由式(3)得到μ(A1)→0.
(4)


(5)
根據式(4),(5)得μ(A)=μ(A1∪A2)→0.定理得證.
參考文獻:
[1]VAPNIK V N. Statistical learning theory[M]. New York: A Wiley-Interscience Publication, 1998.
[2]VAPNIK V N. An overview of statistical learning theory[J]. IEEE Transactions on Networks, 1999, 10(5):.
[3]SUN Xiaojing, WANG Chao, HA Minghu, et al. The key theorem of learning theory based on hybrid variable [Z]. Proceedings of the International Conference on Machine Learning and Cybernetics, Guilin, 2011.
[4]YAN Shujing, HA Minghu, ZHANG Xiankun, et al. The key theorem of learning theory on uncertainty space[J]. Sixth International Symposium Neural Networks, 2009(6): 27-29.
[5]HA Minghu, BAI Yunchao,WANG Peng. The key theorem and the bounds on the rate of uniform convergence of statistical learning theory on a credibility space[J]. Advances in Fuzzy Sets and Systems, 2006, 1(2): 143-172.
[6]張植明, 田景峰, 哈明虎. 基于復擬隨機樣本的統計學習理論的理論基礎[J]. 計算機工程與應用, 2008, 44(9): 82-86.
ZHANG Zhiming, TIAN Jingfeng, HA Minghu. Theoretical foundations of statistical learning theory of complex quasi-random samples[J]. Computer Engineering and Application, 2008, 44(9): 82-86.
[7]HA Minghu, PEDRYCZ W, ZHANG Zhiming, et al. The theoretical foundations of statistical learning theory of complex random samples[J]. Far East Journal of Applied Mathematics, 2009, 34(3): 315-336.
[8]HA Minghu, LI Yan, LI Jia, et al. The key theorem and the bounds on the rate of uniform convergence of learning theory on Sugeno measure space[J]. Science in China, Series F Information Sciences, 2006, 49(3): 372-385.
[9]周彩麗, 哈明虎, 鮑俊艷, 等. 基于模糊數的模糊學習理論的關鍵定理[J]. 河北大學學報:自然科學版, 2008, 28(5): 449-451.
ZHOU Caili, HA Minghu, BAO Junyan, et al. Key theorem of learning theory based on fuzzy number[J]. Journal of Hebei University:Natural Science Edition, 2008, 28(5): 449-451.
[10]哈明虎, 馮志芳, 宋士吉, 等. 擬概率空間上學習理論的關鍵定理和學習過程一致收斂速度的界[J]. 計算機學報, 2008, 31(3): 476-485.
HA Minghu, FENG Zhifang, SONG Shiji, et al. The key theorem and the bounds on the rate of uniform convergence of learning theory on quasi-probability spaces[J]. Chinese Journal of Computers, 2008, 31(3): 476-485.
[11]李俊華, 高林慶, 李海軍. 噪聲影響的泛空間上的學習理論關鍵定理[J]. 計算機工程與應用, 2012, 48(27): 49-52.
LI Junhua, GAO Linqing, LI Haijun. Key theorem of learning theory with samples corrupted by noise on pan-space[J]. Computer Engineering and Application, 2012, 48(27): 49-52.
[12]哈明虎, 李俊華, 白云超,等. 樣本受零均值噪聲影響下的學習理論的若干理論研究[J]. 自然科學進展, 2006(12): 1668-1673.
HA Minghu, LI Junhua, BAI Yunchao, et al. Several influence theoretical research of learning theory with samples corrupted by zero-mean noise[J]. Progress in natural science, 2006(12): 1668-1673.
[13]李俊華, 李海軍. 受噪聲影響的復hybrid樣本的學習理論關鍵定理[J]. 河北大學學報:自然科學版, 2014, 34(1): 14-18.
LI Junhua, LI Haijun. Key theorem of learning theory with complex hybrid samples corrupted by noise[J]. Journal of Hebei University:Natural Science Edition, 2014, 34(1): 14-18.
(責任編輯:王蘭英)