Anda di halaman 1dari 22

Natural Computing

)ecture 13: *article !+arm optimi!ation

Michael Herrmann mherrman@inf.ed.ac.uk phone: 0131 6 517177 nformatic! "orum 1.#$

N"%0&03' 5(11($010

,+arm intelligence

Collecti-e intelligence: . !uper/organi!m emerge! from the interaction of indi-idual! 0he !uper/organi!m ha! a1ilitie! that are not pre!ent in the indi-idual! 23i! more intelligent45 60he +hole i! more than the !um of it! part!7 Mechani!m!: Cooperation and competition !elf/organi!ation8 9 and communication :;ample!: ,ocial animal! 2incl. ant!58 !mart mo1!8 immune !<!tem8 neural net+ork!8 internet8 !+arm ro1otic!

=eni8 >.8 ?ang8 @.: ,+arm ntelligence in Cellular %o1otic ,<!tem!8 *roc. N.0A .d-. ?ork!hop on %o1ot! and =iological ,<!tem!8 0u!can<8 tal<8 $6B30(6 21&'&5

,+arm intelligence: .pplication area!

=iological and !ocial modeling Mo-ie effect! C<namic optimiDation

routing optimiDation !tructure optimiDation data mining8 data clu!tering

Arganic computing ,+arm ro1otic!

,+arm! in ro1otic! and 1iolog<


%o1otic!(.
Main intere!t in pattern !<nthe!i!
,elf/organiDation ,elf/reproduction ,elf/healing ,elf/configuration

=iolog<(,ociolog<
Main intere!t in pattern anal<!i!
%ecogniDing 1e!t pattern AptimiDing path Minimal condition! not 6+hat78 1ut 6+h<7

Con!truction

Modeling

Dumb parts, properly connected into a swarm, yield smart results. Ee-in Eell<

Comple; 1eha-iour from !imple rule!


%ule 1: ,eparation .-oid Colli!ion +ith neigh1oring agent!

%ule $: .lignment Match the -elocit< of neigh1oring agent!

%ule 3: Cohe!ion ,ta< near neigh1oring agent!

0o+ard! a computational principle

Evaluate <our pre!ent po!ition Compare it to <our pre-iou! 1e!t and neigh1orhood 1e!t Imitate !elf and other!

H<pothe!i!: 0here are t+o maFor !ource! of cognition8 namel<8 o+n e;perience and communication from other!.
)eon "e!tinger8 1&5#(1&&&8 ,ocial Communication and Cognition

*article ,+arm AptimiDation 2*,A5


Method! for finding an optimal !olution to an o1Fecti-e function Cirect !earch8 i.e. gradient free ,imple and Gua!i/identical unit! .!<nchronou!H decentraliDed control 3 ntermediate4 num1er of unit!: I 101/10JJ$3 %edundanc< lead! to relia1ilit< and adaptation *,A i! one of the computational algorithm! in the field of !+arm intelligence 2another one i! .CA5

@. Eenned<8 and %. :1erhart8 Particle swarm optimization8 in *roc. :::. nt. Conf. on Neural Net+ork!8 *i!cata+a<8 N@8 pp. 1&#$B1&#'8 1&&5.

*,A algorithm: nitialiDation

"itne!! function Num1er of particle! *article po!ition!

f :R R
m

n K $08 98 $00
m

xi R , i = 1, , n m vi R , i = 1, , n *article -elocitie! i x Current 1e!t of each particle


26!imple no!talgia75 >lo1al 1e!t 26group norm75 nitialiDe con!tant!

g
, 1 / 2

0he canonical *,A algorithm


"or each particle 1 i "or all mem1er! of the !+arm8 i.e. create random -ector! r1 , r2 +ith component! dra+n from U [0,1]

update -elocitie! update po!ition!

i xi ) + 2 r2 ( g xi ) vi vi + 1r1 ( x
component+i!e multiplication

xi xi + vi

update local 1e!t!

i xi if f ( xi ) < f ( x i ) x

update glo1al 1e!t

xi if f ( xi ) < f ( g ) g

minimiDation pro1lemL

Compari!on of >. and *,A

>enerall< !imilar:
1. $. 3. #.

%andom generation of an initial population Calculation of a fitne!! -alue for each indi-idual. %eproduction of the population 1a!ed on fitne!! -alue!. f reGuirement! are met8 then !top. Ather+i!e go 1ack to $. n >.: 1< genetic operator! n *,A: *article! update them!el-e! +ith the internal -elocit<. 0he< al!o ha-e memor<. Mutual in >.. ?hole population mo-e! a! a group to+ard! optimal area. Ane/+a< in *,A: ,ource of information i! onl< g=e!t 2or l=e!t5. .ll particle! tend to con-erge to the 1e!t !olution Guickl<. >.: di!crete *,: continuou! +++.!+armintelligence.org(tutorial!.php

Modification of indi-idual!

,haring of information

%epre!entation

*,A a! M=,
.! in >. the 6model7 i! actuall< a population 2+hich can 1e repre!ented 1< a pro1a1ili!tic model5 >enerate ne+ !ample! from the indi-idual particle! of the pre-iou! iteration 1< random modification! M!e memor< of glo1al8 neigh1orhood or per!onal 1e!t for learning

nitialiDation
N nitialiDe the particle po!ition! and their -elocitie! O K lo+erPlimit Q 2upperPlimit / lo+erPlimit5 R rand2nPparticle!8 mPdimen!ion!5 a!!ert O.!hape KK 2nPparticle!8 mPdimen!ion!5 S K Dero!2O.!hape5 N nitialiDe the glo1al and local fitne!! to the +or!t po!!i1le fitne!!Pg1e!t K inf fitne!!Pl1e!t K fitne!!Pg1e!t R one!2nPparticle!5 +K0.1 N omega range 0.01 9 0.7 a1Ka$K$ N alpha range 0 9 #8 1oth eGual nK$5 N range $0 9 $00 ma; -elocit< N no larger than: range of ; per !tep or 10/$0T of thi! range

Main loop 2ne;t page5

for k K 1 .. 0Piteration!: N loop until con-ergence fitne!!PO K e-aluatePfitne!!2O5 N e-aluate fitne!! of each particle for K 1 .. nPparticle!: N update local 1e!t! if fitne!!POU V J fitne!!Pl1e!tU V: fitne!!Pl1e!tU V K fitne!!POU V for @ K 1 .. mPdimen!ion!: OPl1e!tU VU@V K OU VU@VH end @H end lH minPfitne!!Pinde; K argmin2fitne!!PO5 N update glo1al 1e!t minPfitne!! K fitne!!POUminPfitne!!Pinde;V if minPfitne!! J fitne!!Pg1e!t: fitne!!Pg1e!t K minPfitne!!H OPg1e!t K OUminPfitne!!Pinde;8:V for K 1 .. nPparticle!: N update -elocitie! and po!ition! for @ K 0 .. mPdimen!ion!: %1 K uniformPrandomPnum1er25 %$ K uniformPrandomPnum1er25 SU VU@V K 2+RSU VU@V Q a1R%1R2OPl1e!tU VU@V / OU VU@V5 Q a$R%$R2OPg1e!tU@V / OU VU@V55 OU VU@V K OU VU@V Q SU VU@V end 8 end @8 end kH

llu!trati-e e;ample

Marco .. Monte! de Aca *,A ntroduction

Ho+ doe! it +orkW


:;plorator< 1eha-iour: ,earch a 1road region of !pace :;ploitati-e 1eha-iour: )ocall< oriented !earch to approach a 2po!!i1l< local5 optimum *arameter! to 1e cho!en to properl< 1alance 1et+een e;ploration and e;ploitation8 i.e. to a-oid premature con-ergence to a local optimum <et !till en!ure a good rate of con-ergence to the optimum.

Con-ergence
:;ploration: ,+arm collap!e! 2or rather di-erge!8 o!cillate!8 or i! critical5 :;ploitation: >lo1al 1e!t approache! glo1al optimum 2or rather8 for a collap!e of the !+arm8 a local optimum5 Mathematical attempt! 2t<picall< o-er!implified5: Con-ergence to glo1al optimum for a 1/particle !+arm after infinite time 2". -. d. =ergh8 $0015
!ee *,A at en.+ikipedia.org

%epul!i-e *,A algorithm


"or each particle 1 X i X n create random -ector! r 8 r 8 r +ith component! dra+n 1 $ 3 from MU081V

update -elocitie! Y 1e!t of random neigh1or!8 Z$J0 D random -elocit<

component+i!e multiplication

update po!ition! etc. *ropertie!: !ometime! !lo+er8 more ro1u!t and efficient

Con!triction factor in canonical *,A


ntroduced 1< Clerc 21&&&5 ,imple!t form:

Ma< replace interia Meant to impro-e con-ergence 1< an enforced deca< 2more a1out thi! later5

0opolog<: %e!tricted competition(coordination


0opolog< determine! +ith +hom to compare and thu! ho+ !olution! !pread through the population 0raditional one!: g1e!t8 l1e!t >lo1al -er!ion i! fa!ter 1ut might con-erge to local optimum for !ome pro1lem!. )ocal -er!ion i! a !ome+hat !lo+er 1ut not ea!< to 1e trapped into local optimum. Com1ination: M!e glo1al -er!ion to get rough e!timate. 0hen u!e local -er!ion to refine the !earch. "or !ome topologie! analogou! to i!land! in >.

nno-ati-e topologie!

,pecified 1<: Mean degree8 clu!tering8 heterogeneit< etc.

Compari!on of >. and *,A

>enerall< !imilar:
1. $. 3. #.

%andom generation of an initial population Caclulate of a fitne!! -alue for each indi-idual. %eproduction of the population 1a!ed on fitne!! -alue!. f reGuirement! are met8 then !top. Ather+i!e go 1ack to $. n >.: 1< genetic operator! n *,A: *article! update them!el-e! +ith the internal -elocit<. 0he< al!o ha-e memor<. Mutual n >.. ?hole population mo-e! a! a group to+ard! optimal area. Ane/+a< in *,A: ,ource of information i! onl< g=e!t 2or l=e!t5. .ll particle! tend to con-erge to the 1e!t !olution Guickl<. >.: di!crete *,: continuou! +++.!+armintelligence.org(tutorial!.php

Modification of indi-idual!

,haring of information

%epre!entation

)iterature on !+arm!

:ric =ona1eau8 Marco Corigo8 >u< 0heraulaD: ,+arm ntelligence: "rom Natural to .rtificial ,<!tem! 2,anta "e n!titute ,tudie! on the ,cience! of Comple;it<5 AM* M,. 21&&&5 @. Eenned<8 and %. :1erhart8 Particle swarm optimization8 in *roc. of the ::: nt. Conf. on Neural Net+ork!8 *i!cata+a<8 N@8 pp. 1&#$B1&#'8 1&&5. [ ,hi8 %C :1erhart 21&&&5 *arameter !election in particle !+arm optimiDation. ,pringer. :1erhart [. ,hi 2$0015 *,A: Ce-elopment!8 application! re!!ource!. :::. +++.engr.iupui.edu(Ie1erhart(+e1(*,A1ook.html 0utorial!: +++.particle!+arm.info( =i1liograph<: icd+e1.cc.purdue.edu(Ihu;(*,A.!html

Anda mungkin juga menyukai