artificial intelligence
TRANSCRIPT
Soft Computing (Immune Networks)
In Artificial Intelligence Yasuhiko Dote
Muroran Institute of Technology Mizumoto 27-1, Muroran 050 8585,Japan
dote@,csse.muroran-it.ac.jp
ABSTRACT
iliis paper proposes a novel reactive distnbuted artificial ~ntcIiigeiice (dvnaniic)using immune networks and other soft Loiiiptitiiig inethods Fusth. extended sot? computing is defined
In .idding iiiuiiuiie networks and chaos theory including fractal and ivavelet to conventional sott computing which is the fusion or
coinbinatioii of tiizzv systenis.neural networks and genetic .~lgoritlinis and is suitable to cognitive distnbuted artificial
~ i i~c l l i~ence (static) Next, a novel fuuv neural net(genera1
parameter radial based function neural network) is developed in
order to use it for communication among agents in immune
iictnorhs The geiieral paraineter method is &ended to an adaptive structured genetic algorithm to obtain much faster convergence rate An unbiasedness criterion using distorter( a
radial based ftiiiction network in order to optimize parameters
resultiiy i n die reactive distributed artificial intelligence hnd of
(;MD!i) is applied to better generalization propertes. Then, t h s developed I'tvrv neural net is extended to a h g h performance
1. INTRODUCTION I<eactivit\ is a hehavior-bewd model of activitv,as opposed to
the svmbol inanipulation model u.wd in planning.This leads to the
iiotioii of cognitive c0st.i e.. the complexity of the over ,Irchitecture needed to achieve a task Cognitive agents support a coinplz\ architecture which inems that their cognitive cost is IiigIi.Copnitive agents have intenml representation of the world \ \ l i i ~ l i iiiiist he i n adequation with the morld itself The process of rslating the tiitenial representation and the world is considered as
'I ~oinple\- task On the other hand. reactive agents are simple. cash to uiiderstwd and do not support intemal representation of the world. Ilius. their cognitive cost is low, and tend to what is
~ a l l e d cognitive economy. the property of being able to perfom cvcn compls\r actions with simple architectures Because of their
complexit\. cognitive agents are otteii considered as self- ~iitlicicrit the\ can nork alone or nith a ten other agents.0n die c o i i t r m . reackive agcnts need companionshp 'I'hey can iiot work isolated and they usually achieve their tasks in groups. Reactive agents me companionship. They can not work isolated and thev
0-7803-4778-1 198 $10.00 0 1998 IEEE ,
usualh achieve theu tasks i n groups. llcactive a y i t s are sittiatcd
they do not take past e~eiits into account. a i d can iiot Ibrcsee rlic
ftiture. Their action is based on \\hat happens no\\. ho\\ the\ ~ C I I ~ ~
distmzuish situations ui Ilie aorld. on the ~ \ a v thev resognve world indexes and react accordingly llius. reacuve agents can not
plan ahead what they will do But, what can be considered as a
weakness is one of theu strengths because the! do not ha\^ to
revise their world model when perturbations chaiige the \\orld i n
an tme\pected u a ) Robustness and tatilt toleraiict arc tno 01 the
main properties of reactive agent swttiiis. j 2 group of r e ; ~ c t ~ w agents can coinplete tasks even when one o l them b r d s doun. The loss of one agent does iiot prohibit the coinpietion o l the whole task, because allocation of roles is achieved locall\ bv
perception of the enviroiunental needs. 'Thus, reactive ageiit
svstems are considered as ve: flexible and adaptive because[ I 1 In this paper ;I no\zcaI re;ictive distributrt l ;irtif'i(,i:il
int r llige nrr is proposvd us1 t i g Ish igrii ro's in i i i i i i ne
nr twork ;11~[~r(~~1t~Iil'l;in'li:(i : i nc l ot1ic.r +oft rwmputin?
approaches In section 11. soft coinpiit ing propoawl I)? L1r.L . i .Zadrh[-i] is r x t r n t l t d by :Ititling rhaos coinpiti ing
and iinmunr net,work theory. i\ novel fuzzy neural nrhvork with grneral p;lr;imrt.c-r statistics calculus taking advantages of both fiii.z\ ins arid neural iictnorhs i n section
111 In section IV this IS eaciided to a high perfonnoiice radial hasis tiiiiction iieural iitt\\urh using oii adaptive structure genetic
algorithinl5 I close to the seneral parciiiieter iiicthod arid o kind 01'
(iMnH[61. In section V these developed nct\\orks are applied to
optiiiiize Ishiguro's uimiune network reactive distributed artiticial intelligence.
x
11. EXTENDED SOFT COMPUTING Soft coiiiputing is proposed bv I)r I* A./.adch/-ll to coiistrtict
ne\\ generation Ai [ macliiiie .intelligeiicc quatient) and to solve
noiiliiiear and inatliematicallv iiimioJelld systems prohlenis
(tractability) especiallv for cognitive artilicial intelligence In this
section by adding chaos coiiiptituig arid iiiuiitiiie network thron.
1382
ilii extended soft computing is detined for explaining, what thev
call. complex svstems(7). hunune networks are promising
approachos to construct reactive artiticial intelligence[21 and [ 3 ]as Illustnltcd 111 Fig I
Hiinian I~eir ig l i k e .AI Cognir iw
Fuzzy 1)istribur~d
~~-
Systern ;\I (Stnticl
Fig.1 Soft computing in AI
111. NOVEL FUZZY NEURAL NET
I irst. id consider the (iP approach to KHFN weights adjust~ng. ;\s soon iis IIRI,'N IS linear on its \\eights. the (iP method may bo impIementc.d in a straightfonxard manner The equation
dcscribiiig (if'-RBFN for a single output network is
U IirrI, 11.: fisrcl init ial va lues of' network we1ght.s: p :
si:;il;ir grnrr;il p a r a m e t e r t,o be adjusted with t h e
fol I ow 1 ng algori t hin
s t a r t i n g from the same initi;il conrhtions. In t h e first
case. :I coiivcsiit ioniil I2HF" wiis uswl w i t h in~livi~Iu;i l
adjusting of i t s \vciights. I n t h e s;c.c:ontl c:ise. (~ l ' - l i l$ l~N w a s simuliitrtl with leiirning iilgorit hili (2) 'l'h~, vfficirnc? of both mrthorls w e r r ~ o n i i ~ ~ r r d by using
t h e following m r a s u r e of convergrncr sprrd
I ' J
Figure2. The simplest GP-RRFN
0' 400 800 1
Figure 3 Im;trning algorithm c:onvt!rgc!nc:o:
t i ) conventional IZUCN: 11) [;I'-ltt3FN
1383
~linic.nsioniilit> IeiirninC sprrtl of (;P-RBFS hits iiicreasril reliitivelj- convent.iona1 RBFN. IitlFN to be used in adaptive fuzzy system (AFS) . in
comnion case. is assumed to be t.ririned by means of t h e iiiiniiiiuiii necessary nunibrr of rules (hidden unit
number) ilc.trrniin:ition a n d adjusting of t h e mean and
v;iri:incc~ vwtorh of' iiithvidu:il hidd[,n nodes a s well as
thrJir \\eight5 In th i s p:iper. t he simplest CP RBFN li;isc~tl :idapt ive fuzzy system for aut.oinatic fuzz!- rule niiinbrr tletemiination is proposed (Fig4). Only t h e nrtworli weights have been assumed t.o be adjust.ed by the' (:P algorithm while t he c rn t r e s a n d widt.hs of unit +nsit I V P zon6.s \yere ooiiipletel>- tleteriiiined wi th the n( , tworl i input Gign;il r:iiigr :inil u n i t
1
, CP KBFN BASED AFS - - ___
Figure4. G P RBFN adapt ive furLy system
nuiiihrr during riich t ra ining rpoch
.\ "s:iniplr" fuzz!- systrm has been present.ed by RBFN U i t h I he "unknown" number of hidden uni ts (i.e.. fuzzy
rules) Star t ing t'rom the single-unit-(;P-RBFN. the nr.twork learning has been performed by the scii1:ir grneriil piiriiiiieter iirljusting in the Learning l ' r ~ ~ c c ~ ~ l u r ~ ~ blorli The stratly st iitr general parameter
( ~ ~ ~ I T I : i t ion f < [ f l \ ; i n d viiriiince D { P ) have been
c,:ilcul:ittd hy GP Statist,ics Estimnt.or. The ;~pproxiniation quality cnter ion (1B) was evalutit.etl for ( h p current (:P KRFN st . ructurr . rind decision on rh;inging of' nrtwork s t ructure p:ir:iiiirter iicljusting
i i i t I I P 1,riiriiing Prow[iurr, L~lock. The stezicly s ta le
gr~iii~riil p:ir:iniet<Jr ~~spec t ; i t i on E[P}antl
viiriiince U(P: have been c;ilculatrd by GP Stat.istics
l%timwt.or. The approx" t . ion quality crit.enon (1:3)
\viis n ~ : i I i i ; i t c d fort he current (:P RBFN st.ructure. and
D(P: Q=-
Thereforr. the C:P KHFNAFS Jetc,rniines I h c , " t ru r "
fuzz! rulrj number b; incrt~iiir~nt;lli! rwrui i iny: I 1 1 ~
r ak i i l basis fuiiction uni ts ant1 cant inuous est in i i i t ion
of t.he approxlmtition quality through critrriii (4)
evwluat.ion for each fixed GP R B F S structure . The network t o be determined is the network with 1r:ist
v:ilue oi' i, anr! its unit n u m l ~ r r I C :issiinic,~l t o h.
c l i t * " s : l n i l i l t ~ " r,qu;il t o t he f'uzzy r d c ~ nuiii lwr C I ~
i'uzz> .ystc"
Let consider t he proposed procrdure i n c1et.d for r h r siiiiplest case of the (:P RBFN AFS Lvith sciil;rr input
signal
i n p u t slgniil I I (E ' :u : =0) iintl lino\vn nuni1ii.r of
(:aussian uni ts r] (for t he first stage. y = I ) t h r
sensitive zone center coor&n;ites :ire calculiitcd by relationship (5).
whrrr. I is i i current unit number For y = I ;incl I = I
for rsiiiiiplr. onr ['tin recrivr (': = 0
3 ) The initial (basic) sensitive zone w i d t h rqu;il i'or all netu.orli uni ts I:, c;ilcul;itt.tl as ((5)
1384
IC; p r t o r n i e d biised on input-out,put sample d;itii In this section. the 1Jnbiasediiess Criterion tisiiig Distorter I I K I ) ) ;icwrtlingly to the ;iIgorithm ( 2 ) . Simiiltmeousl>- t h r approach( 8 I is used. which has been sho\\n provldlng iiiiproved
features i n coiiipare to conveiitioiial methods. such as ~ k a i k e
Infomiation Cntenon ( A I C ) [ 9 ] and its modification for neural networks Network Infonnation Criterion (MC) [ I O ] , f i n l n i u m
Descnption Length (MDL)[ I I].
gc'nrr;il p;ir:iiii~ter iJspwt ;it ion E { P ) and viiriiince
D[,& :ire es t imated with some conventional method.
for rs-ample. by t h e movlng average calculation. Let consider the IJCD method application to the GP RBFN AFS The overall svstein block diagram IS shown ui Fig. 6.5
Both of them are (iP RBFN with a lemiing procedurs llie same signals are ted uito the network inputs The diiYerelici: I \ 111
the u a y of the teaclung signal usage While the reaching signill is
fed mto uppa loop without any changes, the lower iietuork is
trained by "distorted", i.e. nonliiirarly traiisfonned, sample d ~ t a
The output of the lower network is also changed hv the
transfoniier of the same transfer function as fir teachins sgiitl
The critenon ol' the iietuork structure optimality is derivedI61.
nhich IS otthe tonnc 7) :1 0 6
%ax (*,; (:2 - 0-. C: umax I .( 'D = 5 (U ) - I.-? (7 ) ]
/ = I
( 7 ) IJiguref,. Definition of GP RBFN basic parameters
where ' ' j-th set (vector) ofthe network input data. 17 overall
c.v;iIii,ii N I : i i i ( I iiiemorizrtl. variables of the both networks. Tlie structure of the netnork n i th
the least value of the cntenon 7 1 is assumed to be a soliition ot the problem
- , - - - - ':I ,- ' 5 ... \ : , ,_ :,! \ : ... '
8 ) The strucciirr of GP RHFN is modified by one inore '. - . . . [:iiussiaii un i t recruiting: y = q + l . The st.eps 1) - 6)
: I r i a rvl)i.at 6.i I .Y , .VI
The, r rs i i l t of the algorif hni 1 ) - 8) imp1ement;ition is :I
111 I'uiivtion uni ts i n c:P IiBFS $1 h i ( . h provic1c.s I he best :ipproxiniating accur:icy In the
car ti is of' fuzxy system theory i t iiieiiiis t h e fuzzy rule ii ii m1)c.r clrtc~rminiition problriii solution. [8]
1V 1 IJ(il.1 PERFOIWANCE RI3k.N l'he prohiein of the reliahiliiy n1' the denved model is one of thc
iiiost iiiipottaiit ones. ansing duruig the identitication task solving
Hic model over-titting prevention IS a crucial point tor inam
y l c t i c a l iinplrmeiitations 11s i t WJS discussed ui the preceding sections, there are several approaches to cope with this ditficultv
Fig.6 Determinution of number of units by dibtorter The proposed general paaiiieter method in scctioii Ill I,
again illustrated i n Fig.7.. This idear is extended to aii adaptive structure genetic algonthm[j]. Geiiotvpe has an adaptive
structure . The string representation is constructed by two l a y s
One is nanied locus l a y . the other .operon l+er as slio!!ii i i i
IFig 8 For this reprcseiitatioii .live ne\\ genetic o~)er~i~i~ii is iirs
detined in order to scll~orgaiii/t: the siring itriicture and dsvclo1)
adaptive genctic change 111 the evolutioiial pro approach bnngs attractive optiiiiiimoii results fbr probizins
including (iA-dilticultv.Suice genetic algorithm and chaos
1385
Loinputnips are heuristic approaches, they have capabilities of a creative thinking ivav or evolution
H i these techniques the Iuzzv neural net in section III turns Into <I high pcrlbnnancr radial basis fuiictlon neural network
Fig.7 General parameter method
S t r i n g
fashion.Namelv.onlv one antibodv is allowved lo activate and act
its corresponding its action to the ivorld 11' its coiiceiitratioii
surpasses the prespecitied the threshhold As sho\vii i n Fig10 . ilic
concentration of the aiitibodv is influenced b! the stimulatioii iuid
suppression from other antibodies . the stiiiiulation froin antigeii. and the dissipation Factor t i c. natural death ). The concentration 01
I-th antibody .which is denoted by a, . is calculated b! ( X ) ( I and
0 are the rate of interaction ainong antigens and antibodird.
+. ..... .... +
~ a l u e list t i x e d lenzth .~
. . . . . . . . _ - ~ _ . . _ ~ ~ - ~ ............ General Parameter Locur libel V V V
.~ ......
................. ....................
N!:eight layer (fixed nominal value) - ....... -. __ . .- -- / I ; , ... Ili,, ......... li:, II, ... It,;,"
_- r
-.: * .. ~
.-- -.; ._. . ~
............ -~ . ? -- ~ ~-~ *. i ~ _ _ _ :-
Inputs : blutually
1
Inputs : Mutually ' Correlated - - _ _ - I Correlated
... - .... I - .....
Fig.8 Adaptive string structure o f genetic algorithm
V. SOFT COMPUTLNG IN REACTIVE DISTRIBUTED ARTIFICIAL
INTELLIGENCE \ Is1 l l G [ J R O 3 REACTIVE IIISTRIBIJTED ARTIFICIAL
IN'TEI.Ll(;ENCE WITH MMJNE NETWORKS[Z] and [i] i 'he detected current situation and competence modules as .\iitigciis and Antibod~es,respzctiveI~ lo inake a iinonoido(antihody) select a suitable antibodv against ilw wrreiit antigen, it IS highlv important I i o ~ the antibodies
arc described .Moreover.it is noticed that the unmunogical
dntration inecliamsm select an antibody in bottom up manner by ~ommuiiicating aiiioiig the antibodies. To rwlize the above
rcquireineiits. the descnptioii the description of antibodies are
defined as follons The identitv of a specific antibody is generally ilcleniiinzd h? the stncture of its paratope and idiotope F i g 5 dcplcts thc represetitation of antibodies As shown iii this tigure.a
pair of precondition action t o paratope .the nuinher of l l~wllo\\rd antibodies and thc degrce ot' disallowance to idiotope
,irc respectively assigned In addition, the structure of paratope is J I \ ided into four portions: objects, direction,distance, and action.
For adequate selection of antibodies . one state variable called
concentration is assigned to each antibody. The selection of
;Ilitibodics I S simply carried out in a wiimer-take a lb
N N N tlA,(tvdt=( (L ( XI11 i l ( 1 ) XI11 ) n i Llll .<I. ( 1 )
J - I 1 1 k 1
I X IN:, - 0 111: ~ k. ii: ( t)
il. ( t - I ) -1.. (l.rxp(O. 5 - A . ( t ) ) )
(8)
k = I
\\liere N IS die number of antibodies. a i d nil denotc~ inatclinis
ratio hrtneen antibod! I and antigen .m), that denotes dcgrce 01
disallo\\ance of antibod\ I for antibod! I 'The first and sccond
tenns of nght hand side denote the stiiiiulatioii and supprzssioti
from other antibodies, respectively The thrd tenii represents lhr
stimulation from antigen, and the forth tenn thtl natural death . . ~ _ _ _ ~ - .
Idiotour - 7 z E E D -~
. . . ~ ~~
Food Bark Middle H w k w u d Obsmclr I v t l FW KlEhi EnrrgY and c , r . i.cn ni>d et,' . - _ ~
Fig.9 Represent;rtion of antibodies
1386
hi order to optimize this reactive distributed artificial intelligence.
h e deve1opr:d ftiziv neural net is applied to communication
aiiioiig agents( antigens and antibodies ) The developed radial
hasis function neural net is used to optimize parameters in (8) and lbr a inetadyaniics whch produces and removes antigens and ailtibodies to make reactive tables.[f]
VL. CONCLUSION 1111s paper proposes extaidtxl sott computing to construct 10%
cos^ reactive distrihuted artificial intelligence resutmg in excellent decision iiiahng. Table I shows the comparison of the proposed system vvith fuzzy svstems on decision making.
Tirblel Comparison of immune network- based with fuzzy reiisoning approach
Iiiiiiiuiic iietnork-bawd T'wn reasoning
t3ottoiii-up decentralized Top-dow~ centralized IIsplicit uiteraction Implicit interaction 1)viiamir: static
REFERENCES
_1 lcrhcr."Reactive I)istnhwed Arti ticial
Intt.lli~eiice.Principles and Applications".Chapter
I I .I:oundations of Distnbuted Artificial
hitelligence,cdited bv GM.P.O'harc and N. R.Jemngs,John Wilev&Sons Innc.,New York, 1 9 9 6 . ~ ~ 2 8 7 - 3 14.
A.Ishiguro.T.Kondo.Y.Watanabe and Y.lJchka\va."ki Iniiiiunogical Approach to Behavior Control of .~lutoiioimious Mobile Ilobots-Coiistructioii Immune Netuorks Through Leanimg--'.Proceedmgs of the
hitexnational Workshop on Solt Computing in
Industry( IWSCI'96),Muroran,Jap~i.April17-28.l996,pp 253-267.
A 1shiguro.Y Watanahe.'l:Kondo and Y I Ichil;a\\a."Constrctioii of a Decentralized
Consensus-Maklng Netaork Based on the Inunune
S! stem-Application lo Action Arbitration for an Autonornous Mobile Robot-",The SlCE
Trans. .Vol.33,No.h. I097,pp 524-5.32(inJapanese). I. A %adeh?The Role of Soti Computing mid Fuzzv
Logic iii the Conception.Design. Development of 1111211 igeiit Svstems".€'roc Of the
IWSC I'Oh.Mtiroraii.Japaii.ApnI27-ZX. I990 PP I .3b-
I .37. (I'leiian: Speaker)
IC Ohkura and K.11eda..'Srlf-Orgaiii/;ing of Stnng Structure arid Adaptive < imetic Swrch".Proceedings of the IWSC 1'06 pp 172- I77
t r TaLeuclu and T Mpos1ii.H Ishihashi and H.Tanaka,"A
Heunstic Model Selection Cnterion I king Distorter and
Its Application to Detenmiumatioii of the Nuinher oI
Hidden IJIUIS in RBFN', .louiial o t rhr: .lap Soc 01'
Syt.Contr. and Inf.,Vol Il,N0.2,l99X.pp6 1-70 Y.Dote,"Sott Coniputmg( Immune Networks) 111
Artificial Intelligence". Web.site:http-//bik.csse Muroraim.Japan. I997 D FhE;hntetov.Y.Dote and M S ShaiMi."S\striii
Identilicetion bv the (iciieral l'urumeier N e t d Netuorks nith Fuzzy self-or~anizaiion"f'rep. o t the I I " '
IFAC SVmP on Svsrelll IdentiIication,Kitak~shu,Japaii,Vol.2, I 997.~~829-8.34 H.Al;aike."A New Look at the Statistical Model
Ideiiti!ication".IEEE Tran. On AC.Vol 19.I974.pp71b
72 3
M.Murata.S Yosluka\va uid S.Aiiian."Nt.r\~orL
Infonnation Cntenoii-l)eieniuiuimg die Nuinher ol'
Hidden IJiUts for Anilicial Neural Nelnork
Model".IEEE Tran. on Neural Net,Vol.j,No.j, I994,pp865-872. J kssanen,"A IJniversal Prior tor Integers and
Estimation bv M " u m Descriptloii I .engtIiC. Annals 01'
Statistics.Vol I I.No 12.l9X3.pp4l(~-i.~l
1387