ѧ
JOURNAL OF SOFTWARE
1999ꡡ1011ڡVol.10No.111999



߽滮Ķǰ
ȱʤ㷨
ժҪһ߽滮ƶǰ.߽滮ڱͳ滮ȫĻ,пͨŨ໥õĻƺͶԱֻ.߽滮ȫԸ,ҾкǿӦ.ʵ֤߽滮ʱĸЧ.
ؼʡ滮,δ,,Ũ,Ա.
ͼšTP183
Design Multilayer Feed-forward Networks Based on Immune Evolutionary Programming
CAO Xian-bin LIU Ke-sheng WANG Xu-fa
(Department of Computer Science and Technology University of Science and Technology of China Hefei 230026)
Abstract The authors use an immune evolutionary programming to design multilayer feed-forward networks in this paper. The immune evolutionary programming retains the ability of stochastic global searching of traditional evolutionary programming, and draws into the interaction mechanism based on density and the diversity maintaining mechanism which exists in living beings' immune procedure. The immune evolutionary programming has better global convergence and very strong self-adaptive ability with enviornment. The experimental results prove the high efficiency of the immune evolutionary programming in designing neural networks.
Key wordsEvolutionary programming, premature convergence, immune, density, diversity maintain.
ǰоռʮҪĵλ,ҲĿǰӦΪ㷺ģ֮һ.10,ǰһֱܵձ,ģʽʶͷ෽˺ܴɹ1,2.
öǰĹҪʵṹ͹ӦȨֵĹ.ĿǰƷҪ.ǰһ̽ѵ3,4,Ȩֵͨǰͷ򴫲õ,ṹĻȡʽ̽ⷽ,Ȳܱⷴ򴫲㷨еĲ㣨ضṹȷ,ѵҾֲ),ҲؽʹѵĽھ̽;һǰ̽ʧ,ĿǰõмҲܼ.,ý㷨EA(evolutionary algorithms)5.㷨Ŵ㷨GA(genetic algorithm)ES(evolution strategy)ͽ滮EP(evolutionary programming).ΪһŻ,㷨˴ĵɺϢ˼,ʹŻȫּ,һڴͳŻ.㷨Ȼ಻.,ڳʼȺɵ,㷨ڽȺֲʱ׳δ,ֲ;Żжù̶Ľ,δŴȺĻ仯,ʹӦڻ.ý㷨Ҳͬһ.
һָĽĽ滮ƶǰ.㷨滮ͬ˼еŨȻƼԱֲ,ȱ˽滮ȫ,ں̶ܴǿӦ,δ.ʵʾ㷨ƶǰʱԽ,㷨ͬ.
1߽滮
滮ǽ㷨һ.Ŵ㷨ĲͬҪ,Ŵ㷨עǸΪŴ,֮ͨϢһ;滮ȺΪ仯,ֻñһ.ѡ滮ǻ㿼.һڽ滮ʵ뷽,ʡ˶뷽ӵʾͽ,ʺϱʾͲ;ǽ滮ֻñ,Ŵ㷨нڲʱƻѽṹһ.
ʱ,滮ǱȽϺõһ.1994,IEEE Transactions on Neural Network㷨ϵרǰ,漰ý滮Խ6,7.ڽ㷨ͬ޷ܺñָĶԼĽӦ,»ڽ㷨ƻԶԶҪ.
Ǵ߻Ƶõʾ.ԭΪ8:ϵͳַĿԭԶӦĿ,뿹ԭϺ,ͨһϵеķӦƻԭ,ҿ֮Ҳ໥̼ƷӦ,ַӦǸݿŨȽе.߹Եص.
.Ķ.߻Ʊֿ֤κʱ̶;
.໥ӦǻŨȽе.ŨԽ,Խܵ;ŨԽ,Խܵٽ.ʹӦ.
ݴ߻гļģ,ǽӦΪԭ,ĽӦΪ,ڽ滮ĻŨȻ,߽滮IEP(immune EP)㷨,.
߽滮㷨ͼͼ1ʾ.

ͼ1 ߽滮㷨ͼ
IEPȽһԭ(Ӧض),Ȼһʼ(Ӧʼѡ,Ҳʼ);żÿһӦ;ͨԱֲάֿĶ;Կи˹;ͨŨȵȺ²һȺ;ֱֹ,㷨.,ŨCΪȺоӦȻӦȵĿtȺпNı,C=t/N.϶ǰƽһ˵IEP㷨.
2IEPĶǰ
NNʱ,(ڵֵ)(ڵֵ)֪,ڵתҲԤȷ,ȡSͺ,ڵڵֵΪ;,:
Rule1. ڵû;
Rule2. ڵû;
Rule3. ڵx,yֻһӴ.
ԱĿǵĶǰ,򻹱
Rule4. ǰڵ㵽ڵûӴ.
2.1
ÿӦһṹ,ǽڵĿȨֵлʵ,ÿһ봮ͼ2ʾ.

ͼ2 ʽ
볤ڵĿĲͬͬ.
2.2ӦȺ
迹PiӦΪei,ΪEi,ӦȺF(i)ֱӶΪeiEiĺ.ṩ3ֿѡķ:
F(i)=1/ei,
(1)
F(i)=e-,
(2)
F(i)=1/(Ei+const).
(3)
Ц,constΪĳ,constĿǱ㷨ΪĸΪ0.ڱǵĵǰ,

(4)

(5)
Tp,outYp,outֱΪpѵĵoutڵʵ.
2.3¿
IEP,뿹PiӦ¿PiҪ3׶.
(1) Ա
IEPĶԱֲֻڵ.ΪֿڵĶ,IEP涨
MinnumNum(k)Maxnum,
(6)
Num(k)ʾڵΪkĿĿ,Maxnum,MinnumֱֵСֵ,ɾ.ʵ,Maxnum=2(IN+ON),Minnum=1,IN,ONֱΪض.
(2) ˹
ֻÿеȨֵ.:
(a) PiΪӦṹ;
(b) ΰʽȨֵ

(7)
F(i)ΪPiӦ,aΪ-11֮һ,(0,1)Ϊ˹;
(3) ڵȨֵһ¿.
2.4ŨȵȺ
ڸÿһα˸Ӵ֮,ͳEPһqѡ񷽷ӸӴļѡɸ,һȺ.巽:(+)ѡq,ǰӦֵÿȽ,бaiĸĿmi,miΪaiĵ÷;ٰ÷ִӴСÿ,ѡǰɸһȺ.
IEPȺ²Խ߻пŨȵ໥,Ũӵq÷.Ũȹ,ڽδ,,IEPҪŨȹߵĿ,ͬʱ֤Ӧȸߵĸ屻ѡеĸʴ.巽,ڿiԭq÷P(i).scoresĻϼϻŨȺӦȵֵ,

(8)
Ц,Ϊ01Ŀɵ,Fitness(i)ΪiӦ,MaxFitnessΪӦ,CΪŨ.
ʽԿ,ڸŨȵt,ӦȽϸߵĿõĵ÷(ʽеĵ2);ŨȲ,ʽеĵ3֤ӦȵĿõĵ÷ֵҲ.
3ʵ 
ͨʵ֤IEP㷨Ч.
3.1ʵ1BPXOR
XORⱻ㷺Ӧһǰ.ʵ涨:ֵ=0,Maxnum=6,Minnum=1;ӦȺʽ(3),ʽconst=0.01;ȺģΪ50;ʽ(7)вa=0.1;ʽ(8)Ц=0.5,=0.5;ŨCΪ

ѡʱ,ÿΰ÷ִСӸߵѡǰ벿ָȺ.ڹ涨e0.001ʱ㷨ǰ,1ʾ㷨10εĽ.ͼ3ʾ˼ַѧϰ.Կ,ڼȫԺٶʱ,IEPЧΪ.

ͼ3 㷨ѧϰ
1IEP10εĽ

дeǷ
12000.000 607
22000.000 643
32000.000 687
42000.000 438
52000.000 141
62000.000 430
72000.000 167
82000.000 284
92000.000 002
102000.000 162

3.2ʵ2ũҵСԤ
ũҵΪҵ,ҹһֱܵر.ʵIEPԤʡũҵСϵͳ.һϵͳ10,ΪԤĲ.԰ʡ'93'953Ϊ,ӵݼ2. 
210ӵֵ

RR2RR2RR2S3RR4Rd3RRd3-4RRd4S3-4S3-4
9362.7062.752.36.36.7296.314181161.3
946.713.320.0532.71.371.7282.7166.3174.3
9514.7014.762.703.780.343.3168.3186.3

RRʾˮ,RdΪ,SΪʱ;±ʾʱ,S3-4Ϊ3Ѯ4Ѯʱ.
ڲIEP߽ʱ,涨Maxnum=22,Minnum=1,ʵ1һ.3г˱ĵԤ,ͬʱ˲ôͳרϵͳ(1)רϵͳ(2)ͳֵ(3)Լϵͳ(4)Ľ9,Կ,IEPƵϵͳȡԽϺõĽ.
3ַԤ

ʵʲ1234ĵĽ
'93456410-10.0%4907.5%409-10.5%420-7.6%426-6.6%
'94473432-8.6%40115.2%414-12.5%431-8.8%428-9.7%
'95506466-8.0%4697.4%465-8.1%473-6.5%484-4.3%

4
ʱ,ҪĹƳʵṹȨֵ.еһЩڲ,滮ȻԽϺ,ҲֲźͲӦȱ.Ĵ߻г֮ŨȵĹϵģ,һ߽滮㷨.˷ʱ,ȱ˽滮ǿȫ,ں̶ܴϿ˷Ĳ,ȫԺٶ涼ﵽ˽ϺõЧ. 
л оõȻѧ.ĿΪ69671029,ڴرʾл!
ĿоõȻѧͰʡ塱ص㹥Ŀʽ.
߼飺ȱ1969,ʿ,ʦ,ҪоΪ,,רϵͳ.
ʤ,1968,ʦ,ҪоΪ.
㷨,1948,,ʿʦ,ҪоΪ,˹.
ߵλйѧѧѧϵ Ϸ 230026
ͨѶϵ:ȱ,Ϸ230027,йѧѧѧϵ
ο
1Weymaere N, Martens J P. On the initialization and optimization of multilayer perceptions. IEEE Transactions on Neural Networks, 1994,5(1):738751
2Park Y. A comparison of neural net classifiers and linear tree classifiers: there simularities and differences. Pattern Recognition, 1994,27(11):14931503
3Ash T. Dynamic node creation in backpropagation networks. Connection Science, 1989,1(4):365375
4Azimi-Sadjadi M R et al. Recursive dynamic node creation in multilayer neural networks. IEEE Transactions on Neural Networks, 1993,4(2):242256
5Vittorio Maniezzo. Genetic evolution of the topology and weight distribution of neural networks. IEEE Transactions on Neural Networks, 1993,5(1):3953
6Fogel D B et al. Evolving neural networks. Biological Cybernetics, 1990,63(1):487493
7Angeline P J et al. An evolutionary algorithm that constructs recurrent neural networks. IEEE Transactions on Neural Networks, 1994,5(1):5465
8ع,ֲ.ҽѧѧ.Ϸ:йѧѧ,1995
(Wu Min-yu, Liu Gong-zhi. Medicine Immune Science. Hefei: Press of University of Science and Technology of China, 1995)
9ȱ.GANNܴϵͳо۲ʿѧλģ.Ϸ:йѧѧ,1996
(Cao Xian-bin. Research on intelligent processing system based on GA and NN Ph.D. Thesis. Hefei: University of Science and Technology of China, 1996)
1998-07-21յԭ,1998-12-01յ޸ĸ
