о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999Vol.36No.6P.675-680



һ־пʽѧϰ㷨
Ǭ־Զ
ժҪһ־пʽѧϰ㷨IHMCAP.㷨ڸ۵ķѧϰѧϰϣͨFTART磬ʵֲͬ˼άεĿɹؽ˷ѧϰѧϰ֮ľ.صѧϰƲʹֻһѧϰɶʾѧϰʹ㷨нϺõĿӶӦʵʱѧϰ.
ؼʣģͣ ѧϰ 磬 
ţTP18
AN INCREMENTAL HYBRID LEARNING ALGORITHM WITH
NOISE RESISTANCE ABILITY
CHEN Zhao-QianZHOU Zhi-HuaJIANG YuanCHEN Shi-Fu
(State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210093)
AbstractIn the paper here, an incremental hybrid learning algorithm IHMCAP is proposed, which has the ability of noise resistance. This algorithm successfully combines probability based symbolic learning with neural learning. By adopting the FTART neural network proposed before, the IHMCAP not only proportions the learning accuracy between symbolic and neural parts, but also lays the two different thought levels aboard. The unique incremental learning mechanism employed enables the IHMCAP performs only one round learning to cover new training patterns. Moreover, it also depresses the noise sensibility of the learning system, which makes IHMCAP fit for tasks that require real-time online learning.
Key wordshybrid model, incremental learning, neural network, noise disposal
1
ѧϰϵĻѧϰѳΪѧϰоȵ㣬ⷽҪɹУEITHER1ݵϵͳKBANN2һ㷨.ЩûгַӷѧϰĴֻEBLexplanation-based learningELempirical lear
ningĽϣڹ÷Χкܴ.
IHMCAP㷨HMCAP㷨3ݻһʽͶȡ㷨ֵԱʾʽڸ۵ķѧϰѧϰϣܴڶʾйɳԻͶжʾĸԴԣЧʽϸߣݴԽϺ.㷨ַFTART45ŵ㣬ɹڷѧϰѧϰ֮ﵽ˾.ΪҪǣ㷨ѧϰʹнϺõĿ.ʹIHMCAP㷨Ӧ㷨Դġḻʵʱѧϰж㷨÷Χ.㷨Windows98Visual C++ 5.0 ʵ.
2IHMCAP㷨
IHMCAP㷨ֵԵʾʾʽܴڶӼйɳԶжʾĸѧϰԣǿЧʸ.㷨ȸɢԶʾз࣬Ȼ޷ɢԾȷֵʾдЩʾһѧϰ.
ڷѧϰѧϰϵĻѧϰжһߵľ⣬ԭĹҪѵ.һ棬ѧϰֵЧʺܸߣṩѵͻ٣ľȾͺѵõ֤.һ棬ѧϰֵЧʽϵͣṩѵͿԱȽ϶࣬ľȾߣϵͳ徫ȴȽϵ.ѡڱַѧϰȵͬʱѧϰľȣʹܹһʵٽﵽ.
IHMCAP㷨ɹؽ⣬FTART45ݣֻһѧϰʱҲʵֶռЧ֣ٶȼ죬ҲֲС⣬ǳʺڻѧϰʹ.ΪҪǣ봫ͳǰ͵BP㷨ͬڸѵõFTARTµģʽʱеṹֻӦԪʵȨɸģʽ.FTART㷨ΪIHMCAP㷨ѧϰṩ֧.
3ѧϰ
3.1ѧϰ
ΪѧϰҪʵжƥ䣬ɵĻжа㣬ﲻʹID46ݣID5R7Ѱ״̬Ȼļѧϰԣ״̬޷㷨ʧ.Ǹݶжһֶصѧϰƣ8ݣѧϰ⣬㷨ݵж.
ΪIHMCAPʹöжʵжƥĹ䲻ͬID4ID5Rƥƥ䵽Ҷ֮ǰֹͣ۵һҶ㣬漸֮һ
(1) 㣬ʵķҶķͬ
(2) 㣬ʵķҶķ಻ͬпڼֵɢԣ
(3) 㣬ʵķҶķ಻ͬûпڼֵɢԣ
(4) 㣬ʵγɵģʽѱԭ縲ǣ
(5) 㣬ʵγɵģʽûбԭ縲.
Ե(1)(4),ǲҪκιԵ(2)ҪԸҶнһ֣Ե(3)Ҫһµ㣻һƾFTARTѧϰܣԭ˽ṹ.
⣬ID4ID5R㷨ӹжʼͽѧϰ͵¶ѧϰжṹңƥ.IHMCAP㷨÷ֶѧϰԣṩһѵѧϰһʼṹ֮ٽѧϰӶΪƵؽ.
3.2
ʱΪѧϰ㷨֪ὫݵȷϢӶԭĽṹд.ж㷨ǻƥģɱƥĿԱȷ֦ö࣬ϵĴɵӰҲȷ֦ϵĴö.
ǰIHMCAP㷨ѧϰʱָģڸݲӼγһʼж֮IHMCAPѧϰʱĽṹ仯ȫҶ.Ϊ˸ӵӣ㷨Ҫ䶯ԭжṹҪеӶѧһعжIHMCAPֻ΢.ӵаʱIHMCAPͿԽݵӰжķ֦ϣ㷨󣬴Ӷ㷨жȣ9.
4IHMCAP㷨
㷨1.
Step1. ǷѵõĻж
ǣGOTO Step7
GOTO Step2
Step2. ʼ㣻
Step3. ǰʾǷͬһ
ǣ򡡷أ
򡡡GOTO Step4
Step4. Ƿпڼֵɢԣ
ǣGOTO Step5
򡡡㣬FTART㷨ѧϰ
أ
Step5. ǰʾɢȡֵǷͬ
ǣ㣬FTART㷨ѧϰ
أ
򡡡GOTO Step6
Step6. ڻֵɢɵǰ
ңGOTO Step3
Step7. ǰǷ㣿
ǣGOTO Step8
򡡡GOTO Step9
Step8. ģʽǷѱԭ縲ǣ
ǣ򡡷أ
򡡡FTART㷨ѧϰ
أ
Step9. ǰǷҶ?
ǣGOTO Step3
򡡡GOTO Step10
Step10. ʾжƥ䣻
GOTO Step7
5ʵ
鱨Ϊָ豸бزٵĲ֣ƵѵǴȷԣĿĿٶȡ˶ԼŵȲȷ.ԭ򣬾ɰǧβԵ鱨ԲɱػһЩ.μʱϹʹָУᴩ鱨ڵĸ׶.йҵܹ˾716оṩݣIHMCAP㷨һ鱨ϵͳ10ݣȡ˺ܺõЧ.ϵͳõı֪ʶ1ͱ2ʾ
1鱨ͼ京

͡͡
C1̫C6δս
C2λC7δյ
C3ʼ̫C8вģ
C4ʼģC9Ŀģ
C5ģC10ۺ

2Ժ弰ȡֵΧ

š塡ȡֵΧ
A1ʱV10<=2%V112%<<=4% 
V124%<<=6%V136%<
A2źV20ޣV21
A3ܷۺV30ܣV31
A4λ״̬V40޵㼣V41е㼣δͣ
V42ܱͣV43
A5ۺ״̬V50V51
A6ս״̬V60ڷǽս״̬V61ڽս״̬
A7в仯V70вޱ仯V71в仯С
V72в仯ϴ
A8Ŀ仯V80ޱ仯V81仯СV82仯ϴ
A9ڽܵĵ㼣0255
A10ʼ0127
A111ϵɵ㼣0255
A122ϵɵ㼣0255
A133ϵɵ㼣0255
A144ϵɵ㼣0255
A155ϵɵ㼣0255
A16ʼ0127
A170127
A18ǰĻǰ08192
A19016
A201ŵ㼣Ĵ㼣0255
A212ŵ㼣Ĵ㼣0255
A223ŵ㼣Ĵ㼣0255


йҵܹ˾716оṩ352ʾԣЩݶȷûбȾ.IHMCAP÷ֶѧϰԣȴȡ172ʾѵѵõͼ1ʾĻжڴ˻ٽѧϰ.Ϊ˽бȽϣͬʱHMCAP㷨Ӽѧϰõжʽͼ1ͬõBP㷨.


ͼ1ʼĻж
ǶIHMCAP㷨HMCAP㷨ѧϰ˱Ƚϣǰߵѧϰ½ԶСں.Ϊģݣṩһѵ{ۺϣ4%<<=6%Уܣе㼣δͣڽս״̬вޱ仯仯С8115752117944825580474105722}.IHMCAPʽѧϰõжͼ2ʾ.HMCAP㷨ڲ߱ѧϰֻܽӼԭӼѵõжͼ3ʾ.ȻIHMCAPжṹ䣬ڳʼжײһҶһµ㣬NN10HMCAPµ㣬ʹжҲ˽ϴı仯.


ͼ2IHMCAPõĻж(ֲ)


ͼ3HMCAPɵĻж(ֲ)
ͼ2ͼ3ͼ1ͬĲԡʾʣ180ʾɲԼ4жͼ1ʼжֱIHMCAP㷨HMCAP㷨ѧϰвԣͼ1ʾжIHMCAPľΪ96.11%HMCAPľΪ86.67%.ȻIHMCAPѧϰЧHMCAPõö࣬ҪΪIHMCAPFTART磬ڷѧϰѧϰ֮ﵽ˾.ͼ2жΪ92.22%ͼ3Ϊ73.89%.Ȼ IHMCAP㷨нϺõĿѧϰȵ½ԶСHMCAP.
⣬ǻIHMCAP㷨ID46ݣID5R7ݵ㷨ѧϰ˱ȽϣȽϽԣIHMCAPжԶ㷨IHMCAPĿǿ.ͬʱǻ֣˿ԽϺõش⣬IHMCAP㷨ѧϰʽѧϰõľÿȡһӾ¹õľȷǳӽʱȴٵö࣬ҲֱIHMCAPһٶȿ졢ȸߵʽѧϰ㷨.
6
IHMCAP㷨ڻͻѧϰģ͵ֵ롢ȡѧϰϵľԼѧϰȷĳ.㷨вͬж㷨ѧϰƣܺõؽ˽Ļ㷨ѧϰ⣬ʹ㷨Ӧж㷨Դġḻʵʱѧϰ.ѾIHMCAP㷨Ϊйҵܹ˾716оһʵʱϵͳнϺõʵüֵ.
ĿõȻѧ(Ŀ6957500869875006)뽭ʡȻѧ(ĿBK97029).
עͣͼ2ͼ3ͼ1ͬĲԡʾ
߼飺ǬŮ19402ڣҪоΪѧϰ硢רϵͳ.
־У197311ʿоҪоΪѧϰ硢ھ.
ԶŮ19765˶ʿоҪоΪѧϰרϵͳ.
У193810ڣʿʦҪоΪ֪ʶ̡ѧϰͼֲʽ˹.
ߵλǬϾѧ¼صʵҡϾ210093
־Ͼѧ¼صʵҡϾ210093
ԶϾѧ¼صʵҡϾ210093
Ͼѧ¼صʵҡϾ210093
οף
1Mooney R, Ourston D. A multistrategy approach to the theory refinement. In: Proc of the Int'l Workshop on Machine Learning. Harper's Ferry, WV, 1991. 115130
2Towell G G, Shavlik J W, Noordewier M O. Refinement of approximately correct domain theories by knowledge-based neural networks. In: Proc of the Eighth National Conf on Artificial Intelligence. Boston: AAAI Press, 1990. 861866
3ݳǬ, . һֻͶȡ㷨HMCAPӦ. ѧ,1996,19(10): 753761
(Chen Zhaoqian, Liu Hong et al. A hybrid algorithm for multi-concept acquisition and its application. Chinese Journal of Computers (in Chinese), 1996, 19(10): 753761)
4ݳǬ, ֵ. һµӦг㷨FTART. ѧ,1996,7(8): 458465
(Chen Zhaoqian, Zhou Rong et al. A new adaptive resonance theory algorithm. Journal of Software (in Chinese), 1996, 7(8): 458465)
5ݳǬ, . FTART㷨оĽ. ѧ,1997,8(4): 259265
(Chen Zhaoqian, Li Hongbing  et al . Analysis and improvement of FTART algorithm. Journal of Software (in Chinese), 1997, 8(4): 259265)
6Schlimmer J C, Fisher D. A Case study of incremental concept induction. In: Proc of the Fifth National Conf on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann, 1986. 496501
7Utgoff P E. Incremental induction of decision trees. Machine Learning, 1989,65(4): 161186
8ݳǬ, ־. ѧϰģHLMеѧϰ㷨. ѧ, 1997, 8(11): 875880
(Chen Zhaoqian, Zhou Zhihua  et al . The incremental learning algorithm in hybrid learning model HLM. Journal of Software (in Chinese), 1997, 8(11): 875880)
9־, Ǭ. IHMCAP㷨ݵĴ. 廪ѧѧ(Ȼѧ), 1998, 38(S2): 118122
(Zhou Zhihua, Chen Zhaoqian  et al . The noisy data disposal by IHMCAP algorithm. Journal of Tsinghua University (Natural Science Version)(in Chinese), 1998, 38(S2): 118122)
10ݺμ, ־. IHMCAP㷨һģ. 1998йԶѧļ.ϺйԶѧԶרҵίԱ,1998. 969974
(He Jiazhou, Zhou Zhihua  et al . A fault diagnosis model based on IHMCAP algorithm. In: Proc of the CIAC'98 (in Chinese). Shanghai: Intelligent Automation Committee of Chinese Society of Automation, 1998. 969974)
ոڣ1998-03-30
޸ڣ1999-02-10
