Զѧ
ACTA AUTOMATICA SINICA
1999ꡡ255ڡVol.25 No.5 1999



LVQܷĽ1)
־֣
ժҪȴϷ˹ѧϰʸ(GLVQ)GLVQF㷨ܣGLVQF㷨һ̶Ͽ˷GLVQ㷨ڵ.ȻԻʤ͵ѧϰкõܣıͣȴʮֲȶ.˲ԭֱӴӱ͵ѧϰʳѡȡѧϰʵ׼򣬲ָĽ㷨.ʹIRIS֤㷨ܣĽ㷨֮GLVQF㷨ԵȶԺЧ.

ؼʡӣģӣѧϰʣIRIS.
BEHAVIORAL ANALYSIS AND IMPROVING OF
GENERALIZED LVQ NEURAL NETWORK
ZHANG ZhihuaZHENG NanningWANG Tianshu
(Institute of AI and Robotics, Xi'an Jiaotong University, Xi'an710049)
AbstractIn this paper, the performance of GLVQ-F algorithm of GLVQ network is theoretically analyzed. The GLVQF algorithm, to some extent, has overcome the shortcomings that GLVQ algorithm possesses. But, there are some problems in GLVQF algorithm, for example, the algorithm has good performance on the winning prototype, and on other prototypes, its performance is very unstable. In this paper, the reasons of the problem are discussed. The rules of choosing the learning rates are proposed, and two modified algorithms are developed therefrom. Finally, the performance of the modified algorithms is verified with IRIS data, which shows the modified algorithms are more stable and effective than GLVQF algorithm.
Key wordsLoss factor, fuzzy degree factor, learning rate, IRIS data.
1
꣬Ϻṹϱֳ㷺ĻKohonenھ㷨1,2ݵоʮ.KohonenĹ2Ҫ棺һѧϰʸ(LVQ)㷨֯ӳ(SOFM).LVQԪδԼ;Ԫ֮Ϣ˷ѵҪ3.SOFMͬҲȱ4ݣһЩڵӦǣÿǻʤڵӦӰ.HuntsbergerHjjimarangsee5״ΰSOFMͬLVQϣΪͼ˷Եһµ˼·.Palȣ6˼һֹѧϰʸ(GLVQ).Gonzalezȣ8ݶGLVQܷȻһЩ.Karayiannisȣ9GLVQGLVQ-F㷨.GLVQ-F㷨һ̶ϽGLVQ㷨ڵ.ͬʱGLVQ-F㷨ڲͬģӾܲȶ.ĴѧϷGLVQ-F㷨ܣָĽ㷨.
2ѧϰʸ(GLVQ)GLVQ-F㷨
X=x1,x2,,xpRnΪϣXе(͵ĸ)Ϊc. GLVQѧϰǴŻһĿ꺯.Ŀ꺯Ϊxһ Lx
(1)
бͦ=v1,v2,,vcRnҪѰҵʸgrxԱ͵Ŀӣ䲻ͬĶͻᵼͬѧϰ㷨.GLVQ㷨Уʽ壺
.(2)
GLVQ㷨grΪʽ(2)ʵӦУʹ㷨˳ԣ7, 8.Ϊˣף8ģC-ֵеȹʽ1gr
(3)
mʣ01)һʾģ.ʱݶ½Ŀ꺯LxСֵGLVQF㷨xtεѧϰ
vj(t)=vj(t-1)+2(t).h(j,m).(x-v(t-1)),j=1,,c,(4)
h(j,m)Ϊѧϰ
h(j,m)=gj+F(j,m),j=1,2,,c,(5)
(6)
(7)
3GLVQ-F㷨ܷ
Ϊ˷vixĻʤͣvlxľСıͣ.ಿֶ.ڷGLVQF.
Ϊir1,-1ir-10,r=1,2,c,F(i,m)0,gi+F(i,m)gi.
֦lk1,-1lk-10,F(l,m)0,gl+F(l,m)gl.
ԿF(k,m)ǸŶk=1,2,c)ڻʤڵ㣬ӴѧϰԾСڵ㣬ѧϰ.һȽh(i,m),h(k,m)(k=1,2,cki)ĴС.ʽ(3),(4)򵥼㣬
(8)
1.1h(k,m)h(i,m),(k=1,2,,c;ki)2)F(k,m)F(i,m),(k=1,2,,c,ki).
ֱϿϣx-vk2ݡx-vj2ʱ㷨Ҳh(k,m)h(j,m).ȻźǸýۼʹھСıҲܳ.˵
.,c=3. x-v12=,x-v22=3,x-v32=4.
򵥼Ȼx-v22x-v32.ɴɼh(k,m)һ(ki). m2ʱ۳
2.m2x-vk2ݡx-vj20h(k,m)h(j,m).
֤.Ϊm2,.x-vk2ݡx-vj2



h(k,m)-h(j,m)0.Ƕ֤.
һǶGLVQ-F㷨̶krh(k,m)=gk+F(k,m)mĺз.ȸһ.
17.1)2)
3)
h(k,m),gk(k=1,2,c)mĵ
(9)
(10)
ʽ1ף1ݣ׵õ漸
3.giΪm(1,)ĵݼ
4.glΪm(1,)ĵ
5.h(i,m)Ϊm(1,)ĵݼh(i,m)=1,h(i,m)=
5h(i,m)ƶ3ĽźǱĲ֤h(l,m)ƶ4Ľ۳.ߵĿϣҵh(k,m)
1),h(i,m)Ϊmĵݼh(i,m)=,h(i,m)=1
2)ki,h(k,m)Ϊmĵh(k,m)=h(k,m)=0.
ͿԲף9еķѧϰӣmɴ𽥼Сʼʱб͵ѧϰȣֻлʤͼ.ֱϿǺ.
4GLVQ-FĸĽ㷨
ߵĿǹh(k,m)(k=1,2,c).ף8h(k,m)ĹǴgkֵ.ʵѧϰʱֻҪh(k,m),ֱӴh(k,m)֣gk.
h(k,m)gkһŶF(k,m)ڻʤ伤̶ȣھСıƳ̶ȣıͣF(k,m)ŶʱܲЧϽڵģ.ȥF(k,m)ֱ
h(k,m)=gk,(11)
ڻʤͺ;Сı˵ǵļƳ̶ȣͣȶЧ㷨ԶҪȶЩ.ɶ3Ͷ4֪h(k,m)1)2)ֻǶԾСıͳ.ǰѻʽ(11)㷨ƸĽ㷨1.
һָĽƸĽ㷨2.
(12)
ʽȿɵáh(k,m)gk(k=1,2,,c).
Ϊ
ԡ
Ԫ
ʽ(13)ɵ0h(k,m)gkgi=h(i,m)=1(k=1,2,cki).ٸʽ(12)Ȼh(k,m)=h(k,m)=0,(k=1,2,cki).


(k=1,2,,c;ki),
ԡh(k,m)(k=1,2,,c;ki)m(0,)ϵĵ..
5ֵʵ鼰
IRISݣ9֤Ľ㷨.IRISݾ(޼ල)(мල)ģʽʶ𷽷ܣܹ150ݣÿ4.IRISݷΪ3࣬ÿ50.мලƱ׼ĴΪ05޼ලƱ׼ĸ15.GLVQ-F㷨Ľ㷨12ֱIRISݽо.Ϊ˸õرȽ㷨ܣֱȡͬmֵʵ飬2.1˱͵ĳʼֵʵ鶼Ǹʼֵе.ΪһȽ׼1-NP()㷨ֱӶо࣬Ҳ1.
1ֵʵб͵ĳֵ㷨ʵ

ʼĻ
5.006,3.428, 1.462, 0.2465.006,3.428,1.462,0.2465000
5.936,2.770, 4.260, 1.3265.936,2.770, 4.260, 1.32604647.3%
6.588,2.974, 5.552, 2.0266.588,2.974, 5.552, 2.0260743

ڱ2ÿһϲĻ(Confusion) ²ʾĴ.ӱ2ԿGLVQ-F㷨m2ʱܴmֵ1ʱ㷨ڲܽо.Ľ㷨1GLVQF㷨ЧһߣӦͬ.ԭģʽ(5),(11)ѧϰh(j,m)(j=1,2,,c)mȡֵͬܲȶӶƻ㷨ľЧ.Ľ㷨2ȫ㷨m=5Լm=3ʱĴΪ1115ʱGLVQF㷨͸Ľ㷨1Ĵܸ.mֵ1ʱm=1.01㷨Ѿ޷о࣬Ľ㷨2ȻЧضо࣬㷨ľʮȶʮЧ.
2ѧϰ㷨ʵȽ

m5.03.02.01.51.21.11.01
GLVQF㷨005050005000500050005000硡
0050491024620464043704010ޡ
83210740301139024801490149ۡ
93.3%64.0%10.0%4.0%5.3%7.3%
Ľ
㷨1500050005000500050005000硡
2921054501463046404550446ޡ
11138010400545024801490149ۡ
27.3%10.0%6.0%4.0%4.0%4.7%
Ľ
㷨25000500050005000500050005000
3443344304640446043704370437
0545054502480149014901490149
7.3%7.3%4.0%4.7%5.3%5.3%5.3%

ʵĽǰĵ۷Ǻϵ.һ֤˻ѡȡѧϰh(k,m)(k=1,2,c)ѧϰ㷨Ч.
6ᡡ
GLVQ㷨ǴŻһĿ꺯ģ㷨ӱΪ˷ѧϰʸ㷨ڵṩһµ˼·.KarayiannisͨģcֵȺGLVQ㷨һµѧϰ㷨GLVQF.ĵĹΪGLVQF㷨ṩһѧۻ.ѧϰʸ㷨ۣԼЩƳЧľ㷨㷨Ӧͼн߽һоĿ.
1) ȻѧصĿ(No.69735010).
߼飺־ʦͨѧʿѧλ.Ҫģʽʶͼģ߼ϵͳ񾭼ܵо.
֣ʿʿʦ.ҪģʽʶͼӾӦоڹҪѧڿϷһƪ.
ߵλͨѧ˹о710049
ο
1Bezdek J. Pattern Recognition with Fuzzy Objective Function Algorithms. New York: Pelnum, 1981
2Kohonen T. Selforganization maps. Berlin: SpringerVerlag, 1995
3Chung F L, Lee T. Fuzzy competitive learning. Neural Networks, 1994, 7(3): 539-551
4Tsao E C K, Bezdek J C , Pal N R. Fuzzy Kohonen clustering networks. Pattern Recognition, 1994, 27(5): 757-764
5Huntsberger T, Ajjimarangsee P. Parallel selforganizing feature maps for unsupervised pattern recognition. Int. J. General Systems, 1989, 16(2): 357-372
6Pal N R, Bezdek J C, Tsao E C K. Generalized clustering networks and Kohonen's selforganizing scheme. IEEE Trans. on Neural Networks, 1993, 4(4): 549-558
7Gonzalez A Z, Grana M, Anjar A D. An analysis of the GLVQ algorithm. IEEE Trans. on Neural Networks, 1995, 6(4): 1012-1016
8Karayiannis N B, Bezdek J C, Pal N R, Hatharay R J. Repair to GLVQ: a new family of competitive learning schemes. IEEE Trans. on Neural Networks, 1996, 7(5): 1062-1071
9Bezdek J C, Pal N R. Two soft relatives of learning vector quantization. Neural Networks, 1995, 8(5): 729-743
ոڣ1998-07-03
޸ڣ1998-12-10


