Multi-agent immune recognition of water mine model

Journal of Marine Science and Application ,Vol.4 ,No. 2 ,June 2005Multi-agent immune recognition of water mine modelLIU Hai-bo , GU Guo-chang , SHEN Jing ,and FU YanSchool of Computer Science and Technology , Harbin Engineering University , Harbin 150001 , ChinaAbstract :It is necessary for mine countermeasure systems to recognise the model of a water mine before destroy ing because the destro-ying measures to be taken must be determined according to mine model. In this paper , an immune neural network( INN ) along withwater mine model recognition system based on multi-agent system is proposed. A modified clonal selection algorithm for constructingsuch an INN is presented based on clonal selection principle. The INN is a two-layer Boolean network whose number of outputs is a-daptable according to the task and the afinity threshold. Adjusting the afinity threshold can easily control different recognition preci-sion , and the afinity threshold also can control the capability of noise tolerance.Keywords imulti-agent system ; immune neural network ; clonal selection ; pattern recognition ; water mine modelCLC number TP18Document code AArticle ID :1671 - 9433( 2005 )02 -0044 - 06statistical snake ) model is presented in Ref.[ 5 ] which1 INTRODUCTIONextracts the highlight and shadow of the object. TheMCM ( complete mine countermeasures ) systems ,detection model used spatial a priori knowledge on theare usually composed of a detection process and a clas-size and geometry of object signatures in side-scansification process such as the systems by Dobeck etwithin the framework of an MRF model to provide ac-al.[1-2] , Ciany et al.[3-4] , and Reed et al.[5]. Allcurate detection results even when large amounts ofthree of these systems operate using the detection/ clas-clutter were present. This technique demonstrated howsification framework although they operate using verythe inclusion of a priori information could again providedifferent models. Dobeck implements a matched filtermore accurate results.in Ref.[ 1 ] to MLOs ( detect mine-like objects ) afterMost of the previous researches on MCM focusedwhich both a -nearest neighbor neural network classifiermain attentions on the detection and classification ofand a discriminatory filter classifier are used to classifywater mines. The proposed techniques can classify thethe objects as mine or not-mine. The detection processdetected MLOs as mine or not-mine , but cannot offeris relatively simple and is primarily for identifying re-the information about model. However ,it is very nec-gions that definitely do not contain MLOs. The classifi-essary for MCM systems to recognise the model of a wa-cation process then uses up to 45 features for everyter mine for accurate destroying.possible MLO to determine which are real MLOs andIn recent years ,the AIS ( artificial immune sys-which are false alarms. The system in Ref.[ 3 ] utilizestem ) is widely used for pattern recognition. The im-an adaptive threshold technique for the detection aftermune system is responsible for the production and ma-which geometric features are extracted , allowing eachintenance of an antibody repertoire capable of properlyMLO to be classified as mine or not-mine. They imrecognizing any antigen population ( foreign substancesprove their approaches in Refs.[ 2 ] and[ 4 ] respec-or i中国煤化Ttion principle Osuggeststively. Reed presents unsupervised models for both thea meYHC N MH Gasic features of an im-detection and the shadow extraction phases of an auto-mune response to an antigenic stimulus. In this paper ,mated classification system. A novel CSS ( co-operatingan INN( immune neural network ) is constructed basedon clonal selection principle and the INN is used in a .Received d海方数据06 -24.water mine model recognition system based on MASLIU Hai-bo et al Multi-agent immune recognition of watermine model. 45.( multi-agent system )patterns we want to recognise are usually called learn-2 ARCHITECTURE OF MULTI-AGENT WATERing samples ( or training patterns ) and the units com-posing the network are called neurons. In the shape-MINE MODEL RECOGNITION SYSTEMspace domainJ , the input patterns are the antigensWe propose an MAS-based water mine model rec-and the neurons are simply called cells. We assumeognition system as shown in Fig. 1. In such a systemthat the antibody repertoire contains a single individualthere are three sensor agents ,i.e. , VSA( visual sensorat the beginning of the learning process , so that theagent ), DSA( depth sensor agent ) , and SSA( sonarrepertoire will have to be constructed while submittedsensor agent ) , and one immune recognition agent , i.to the Ag population. For the sake of simplicity , thee. ,IRA( immune recognition agent ). The sensor a-cells will be uniquely represented by their receptorsgents detect the mine-like objects in the underwater en-( A,={Ab, ,Ap,.. Ao. }). The antibody repertoire willvironment and offer the feature information for IRA，be modeled as a Boolean competitive network with l in-and IRA offers the recognised water mine model forputs , called INN. We developed the MCSA for con-DMA( destroying mines agent ) that is not part of thestructing an INN.recognition system.To plain depict the MCSA , we define some varia-VSAble and data structures by Visual Basic expression as .follows :DSAIRAType Antigenx As String' antigen patternDMASSA0 As Integer' index of the highestafinity antibodyFig. 1 The Architecture of MAS-based water mineEnd Typemodel recognition systemType AntibodyDifferent kinds of sensor agent can offer specialw As String' network weight vectoraspect of feature data about the detected mine-like ob-c As Integer' concentration leveljects. The VSA offers about shape and color informa-m As Boolean' affinity maturationtion , the DSA offers about depth information , and theEnd Type .SSA offers about state information. But all the informa-tion is unifiedly represented as binary bit- strings in dif-Dim Ag( ) As Antigen' antigen arrayferent length , and the bit-strings from each sensor a-Dim Al( ) As Antibody'antibody arraygent are sent to IRA and are combined to one bit-stringDim e As Integer' afnity thresholdto be recognised by INN that is the core component ofFor conveniently , the index of a specified antigenIRA.or antibody will be written as a subscript , and theIMMUNE NEURAL NETWORKmember variables will be written as superscript ,e. g.3. 1 Modified clonal selection algorithm ( MCSA )Ag( i).x and Abl(j). w will be written as Ag and a" re-for INN constructionspectively.Primarily ,it will be assumed the existence of an中国煤化工bothx and 10 are bit-antigen population( Ag = {Ag ,A.g... Ag。}) to bestrin.MYHCNM H G', is regarded as an l di-recognised by the antibody repertoire ,e. g. the com-mensional vector , then Aj; represents the weights con-bined feature bit-strings of water mines from the sensornecting the l inputs to a single output unit j of network.agents , and the antigen can be represented as l-lengthThe afinity A; between Agand A,yis given by the Ham-bit-string..万方熬据a neural network perspective, theming distance of Ag; and A": according to Eq.( 1 ). .，46*Journal of Marine Science and Application ,Vol. 4 ,No.2 ,June 2005A,= IA-A"I.( 1)i=arg，minA..(6)For each antigen i , label 0 indexes the highest af-According to Eq.( 7 ) ,clone a new cell hi withfinity antibody. Label D can be calculated according tothe weights of its antibody are the exact complement ofEq.(2 ).Ag.A" =j = arg maxA,.(2)Alx =A"(7)For each antibody j , concentration level c of anti-Add Ask to INN.gens for the antibody can be calculated according to .Step 4. UpdateEq.( 3 ).For each antigen i , update A"; according to Eq.Ai =〉i(Ag=j,1 ρ),(3) (2).For each antibody j ,update A and A" accord-where i( A"j= j 1 0 )means that if A": = j holds thening to Eq.(3 )and Eq.(4).return 1 else 0.If all antibodies are mature then delete the anti-The member variable m of antibody is defined asbodies with Aζ=0 and stop ,else go to step 2.an affinity maturation indicator and if An; is true thenWhen MCSA stops ,an INN is constructed , eachstop cloning antibody Ay. The afinity threshold e isoutput of which represents one or several antigen pat-defined as a parameter to control the match precision ofterns ,i. e. , water mine models. The number of pat-antigen and antibody. For antibody A" ,if for arbitraryterns that an output represents is determined by the af-i∈{ilA"i= j},l-A;>e holds , then set A"; is true.finity threshold e. When e =0 ,one output correspondsThis can be given as Eq.( 4 )to one pattern. The bigger the value of e is , the moreA"; = A (l-A;> e).(4)patterns an output represents along with the lower accu-Lastly , the variable e is defined as an afnityracy.threshold. For A。;and Aij ,ifl-A; >e then we can say3.2 Recognition process of water mine modelthat Agi and Ayare matched.The process of recognition is very simple. GivenNow , the MCSA can be shown as follows :an antigenAgi ,i. e. ,an l-bit string in length represen-Step 1. Initializationting the attributes of a water mine received from theInitiate an INN with n antigens( Ag={Agl ,Apsensor agents , as input , the INN computes the anti-. Ag})and 1 antibody(An ).body being activated according to Eq.( 8 ), then theFor each integeri=[ 1 ,n],setA"i= 1.water mine model can be recognised.Set each bit ofA"I to 1.j = argmaxA; ,where 0 = {jl l-A; > e}Set Ai= n.(8)Set A" = False.There is no output can be activated if 0 is anSet e to an appropriate value.empty set , which means that A gi cannot be recognisedStep 2. Selectioni.e. ,the water mine is not any known model or mayDetermine the candidate An; to be cloned ac-be it is not a water mine at all. The affinity threshold ecording to Eq.(5).can be set to a value different from that in constructingj = arg maxAi;，(5 )algon中国煤化工le of e is ,the greater .Ifj is null then stop else continue step 3 tonoise:MYHC N M H Gappears.clone A;4EXPRERIMENTS AND ANALYSISStep 3. CloneSelect the worst matching Agi of An according toThe main purpose of simulation is to test the per-Eq.( 6).formance of INN. For constructing an INN , a waterLIU Hai-bo et al Multi-agent immune reognition of watermine model. 47.mine data set with models and binary attributes wassponds to the 21 models ) , and e =0 , which meansprepared as shown in Table 1. A value of 1 in this ta-that the MCSA will create a network with cells highlyble corresponds to the presence of an attribute , whilst 0specific for each input pattern ( antigen ). Thus ，itcorresponds to the lack of this attribute. In the datawould be expected that the final network was composedset ,the relationship between the different symbols mayof 21 cells , but the resultant network contained only 18not be directly detectable from their encodings , thuscells( see Fig.2 ). This is because the models mappednot presuming any metric relations even when the sym-into the same node (" WSM110 , MK65 , MK67" andbols represent similar items.NM103 , WSM210" ) are described by the same at-We applied MCSA to the data set with settinng l =tribute strings( see Table 1 ).20 ( corresponds to the 20 attributes ) ,n =21 ( corre-Table 1 Water mine data set with their models and binary attributesShapehemisphere)010000000cylindertruncated-coneoblateclutter .)0.010100000000000000Colorgreenblackgrayred)0orangeyellowbrownStatefloating0000000mooredbottomnavigationadsorption0000..0.0000000000010000Depthdeep000000000111，11110，1111shallow111111111000000010000HasantennaeIf there is certain correlativity existing between the e to a中国煤化工ore useful when INN isattributes , we can get more abstract classification of the applieYHc N M H Gi here ,this caracerisantigens ,i.e. ,the output of INN will be less ,by setting tic is unnecessary , we don not discuss it in detail.，48.Journal of Marine Science and Application ,Vol.4 No. 2 June 2005PD-3YAApparently ,the value of e in recognition should be e-qual to or bigger than that in construction. So , to attain .ws 110.K65,K67a satisfied CR , the value of e must satisfy the in-K60U/Iequations in( 9 ).N103. WS210PDM-2BDUMI0Ife,≥e。，(9 )PDM-1BPLARM-IAe,≥b.DUMI02Nm102where e, is the construction affinity threshold ,e, is theVS-RM-30MSHPDM-IMrecognition affinity threshold , and b。is the number of89A2noise bits. It is an acceptable strategy to set e, to biggerFig.2 The INN constructed by MCSA , convergingvalue when the noise level is uncertain.to 18 outputs after 18 iterationsIn most cases ，the sensor data maybe incomplete100--and/or inaccurate ,e. g. , being affected by the bright-3050.ness of environment , the color of a black mine may be10-perceived as gray by the VSA , or being interfered b20)other objects , the shape of a mine cannot be detected ,0such that some bits of the attribute strings of water mines8 10maybe mutated ,i.e. ,0 to 1 or1 to0. INN can easilyNoise bits002deal with these cases by adjusting the value of the afini-Fig. 3 Percentage of correet recognition for the INNty threshold e when recognises ，which can be demon-with relation to number of noise bits and afinitystrated by the following experiments.threshold e.The input patterms used to test INN noise tolerancewere arificially generated. Random noise was inserted 5 CONCLUSIONSinto the learning samples in Table 1 by simply revertingWe propose an INN along with a water mine modela bit0 into a 1 ,or vice-versa. We tested eleven differ-recognition system based on MAS in this paper. Theent noise levels from 0% to 50% , corresponding to thesystem architecture , the MCSA for constructing theshift of0 to 10 bits , respectively. The INN( see Fig. 2 )INN ,the recognition process of water mine model , andwas tested for different values of the afinity threshold( ethe simulation experiments are discussed in detail.=[0 ,10]), and a CR ( correct recognition) is as-The INN is a two-layer Boolean network whosesumed when the network maps the corrupted patternnumber of outputs is adaptable according to the task and( pattern with noise ) into the same node as the originalthe afinity threshold , which is significant for solvingpattern( pattern without noise ). Fig. 3 depicts the ex-machine-learning problems , like knowledge acquisitionperiment results that are the average taken over tenand clasification. Adjusting the affinity threshold ,runs. Fig. 3 shows that the CR is affected both by thewhich endows INN with satisfied noise tolerance capabil-number of noise bits and the affinity threshold e. Keep-ity , can easily control different recognition precision.ing a value of e constant , the CR decreases approxima-Moreover , The INN is straightforward to implement intive linearly with the noise level increasing. If we keep ahardw中国煤化工concerning about theconstant noise level , the CR increases nonlinearly , i.quanti:MYHC N M H Gts and network parame-e. . ,the CR almost keep constant when the value of e is .ers.bigger than the number of noise bits. In these experi-The MCSA is a kind of unsupervised learning algo-ments , we did not take into account the effect of adjus-rithm. How to adjust the affinity threshold to attain theting the afimwytmeshold e when the INN is constructed.optimal numbers of classification is an open challenge49uu Hai-bo et al suli-agen innne rcogition of watemine modelHis research interests focus on autono-needs more in-deph study. In future work ,we sall ury Engnering Unrvenitito investigate an adaptive strategy to adjust the afinitmous underwater vehicle ( AUV ) and moulti-agent systemthreshold. Further more , we shall analyze how the cor-( MAS). In rcent years ,he has published more than 30 academ-ic papers.rect recogition rate is aeted by the noise in more de-GU Guo-chang was borm in 1946. He gradu-ated in computer from Harbin Miltary Institu-REFERENCES :te in 1967. He is currenly a professor and[ I ]DOBECK GJ ,HYLAND JC ,SMEDLEY L Automated de-doctoral supervisor in School of Computer Sci-tection/ easfcationon of sea mines in sonar imagery[J]. Na-ence and Technology , Harbin Engineering U-val Research Reviews ,1997 A9(3):9-20niversity , a commitee menber of the Itell[2 ]ARIDGIDES T, FERNANDEZ M. DOBECK GJ Imprvedgent Rotics Soiety of Chinese Assc iaion for Arifcial Itellprocessing string fusion approach investigation for automatedgence( CAAI ),a comite member of Heilongjiang Istiute ofsea mine lasfationin in sallwv waler[ A] Proceedings ofCommunicaion ( HIC ), a member of Heilongiang Clleges andDetetion and Remediation Technologies for Mines and Mine-Universities Science and Technology Advisory Comnitee ,an edi-like Targes IX[ C], Orlando ,USA :2004 ,315 -326.torial board member of Jourmal of Harbin Engineering University ，[3 ]CIANG c M , HUANG J. Computer aided deletion/ compurand serves as the vice - pesident of Heilongiang Computer Soce-er aided lasficaionin and dala fusion agrithms for auloma-ity. His research interets mainly include AI theory and roboticsted detection and lasifcationin of underwater mine[ A] Pro-and he has published over 80 papers in the areas.ceedings of MTS/IEEE Oceans Conference and Exhibition[ C] Providence , USA 2000 277 -284.SHEN Jing was born in 1969. She is current-[ 4 ]CIANY C ,ZURAWSKI w. Improvements in computer aidedly a Ph. D candidate for computer applicationdetetionv computer aided lascation(o CAD/CAC)of bo-technology in Harbin Enineeing University.tom mines through post analysis of a diverse set of very shal-She received bachelor and master' s degree inlow waer( vsw )environmental test data[ A]. Poccedingsgcomputer pplication in 1990 and 1996 re-of Detion and Remediation Technologies for Mines andspetively from Northeast China Intitute of EMinelike Targets IX[ C] Orlando , USA 2004 327 -335.letric Power Engineering. His research interests focus on rein-[ 5 ]REEDS , PETLOT Y , BELLJ. An automatic apprach to forcement leaming and arificial immune system.the detection and extraction of mine features in sidescan sonar[J] IEEE Joumal of Oeeanic Enineering ,2003 28( 1):FU Yan was bom in 1978. She is currenly a90 - 105Ph. D candidate for computer application tech-[6 ]BURNET r M. The Clonal Section Theory of Acquired Im-nology in Harbin Engineering University. Shmmity [ M] Cabridge : Cambridge Univesity Press ，received bachelor' s degree in computer and1959.its application in 2001 and master' s degree in[7 SECEL L ,PEREISON A s. Computations in shape space :Acomputer appication technology in 2004 re-new apprach to immune network theory[ A ] Theoretialspetively from Harbin Engineerng University. Her research inImmunology ( Part2)[ C] Redwood City :ddson-Wesley，terests focus on AUV and AI.1988. 321 -343.LIU Hai bo was borm in 1976. He is current-中国煤化工technology in Harbin Engineering UniversityMHCNMHGand a student member of IEEE/CS. He re-ceived bachelor' s degree in computer and itsapicaionin 1998 and master' 8 degree incmpuer明穷数据chnoly in 200 rspetively fom Hatin