`
`Fig. 11. 3 A n exampl e o f matchin g a s
`parameter optimization , (a ) Initia l
`parameter se t (displaye d a t lef t a s three
`dimensional surfac e (se e Fig . 9.8 ) (b )
`Fitting process : iterativel y adjus t a base d
`on M (se e text) , (c ) Fina l paramete r se t
`yields thi s threedimensiona l surface .
`{See color inserts.)
`
`Ch. 11 Matching
`
`356
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 367
`
`
`
`continuum, an d tha t labele d arc s coul d b e replace d b y node s t o yiel d a directe d
`graph wit h labele d nodes .
`Depending o n th e attribute s o f th e relationa l structur e an d o f th e correspon
`dence desired , th e definitio n o f a matc h ma y b e mor e o r les s elegant . I t i s alway s
`possible t o translat e powerfu l representation s suc h a s labele d graph s o r «ar y rela
`tions int o computationa l representation s whic h ar e amenabl e t o forma l treatmen t
`(such a s undirecte d graphs) . However , whe n grap h algorithm s ar e t o b e imple
`mented wit h compute r dat a structures , th e freedo m an d powe r o f programmin g
`languages ofte n tempt s th e implemente r awa y fro m pur e grap h theory . H e ca n re
`place elegan t (bu t occasionall y restrictiv e an d impractical ) graphtheoreti c con
`cepts an d operation s wit h arbitraril y comple x dat a structure s an d algorithms .
`One exampl e i s th e "grap h isomorphism " problem , a ver y pur e versio n o f
`relational structur e matching . I n it , al l grap h node s an d arc s ar e unlabeled , an d
`graphs matc h i f ther e i s a 1: 1 an d ont o correspondenc e betwee n th e arc s an d node s
`of th e tw o graphs . Th e lac k o f expressiv e powe r i n thes e graph s an d th e require
`ment tha t a matc h b e "perfect " limit s th e usefulnes s o f thi s pur e mode l o f match
`ing i n th e contex t o f nois y inpu t an d imprecis e referenc e structures . I n practice ,
`graph node s ma y hav e propertie s wit h continuou s range s o f values , an d a n arbi
`trarily comple x algorith m determine s whethe r node s o r arc s match . Th e algorith m
`may eve n acces s informatio n outsid e th e graph s themselves , a s lon g a s i t return s
`the answe r "match " o r "n o match. " Generalizin g th e graphtheoreti c notion s i n
`this wa y ca n obscur e issue s o f thei r efficiency , power , an d properties ; on e mus t
`steer a cours e betwee n th e "elegan t an d unusable " an d th e "genera l an d uncon
`trollable." Thi s sectio n introduce s som e "pure " graphtheoreti c algorithm s tha t
`form th e basi s fo r technique s i n Section s 11. 3 an d 11.4 .
`
`11.2.1 Th e Algorithm s
`
`The followin g ar e severa l definition s o f matchin g betwee n graph s [Harar y 1969 ;
`Berge 1976] .
`• Graph isomorphism. Give n tw o graph s (V\, E\) an d (V 2, E 2), find a 1: 1 an d
`onto mappin g (a n isomorphism ) / betwee n V\ an d V 2 suc h tha t fo r
`V\, v 2 € V\, V 2, f(v\) = v 2 an d fo r eac h edg e o f E\ connectin g an y pai r o f
`nodes v i an d v' i € V\, ther e i s a n edg e o f E 2 connectin g f(v\) an d
`f(y\).
`• Subgraph isomorphism. Fin d isomorphism s betwee n a grap h (V\ i E\) an d sub
`graphs o f anothe r grap h ( V2i E 2). Thi s i s computationall y harde r tha n isomor
`phism becaus e on e doe s no t kno w i n advanc e whic h subset s o f V 2 ar e involve d
`in isomorphisms .
`"Double"subgraph isomorphisms. Fin d al l isomorphism s betwee n subgraphs o f
`a grap h (V\ i E\) an d subgraphs o f anothe r grap h (V 2i E 2). Thi s sound s harde r
`than th e subgrap h isomorphis m problem , bu t i s equivalent .
`• A matc h ma y no t confor m t o stric t rule s o f correspondenc e betwee n arc s an d
`nodes (som e node s an d arc s ma y b e "unimportant") . Suc h a matchin g cri
`terion ma y wel l b e implemente d a s a "computational " (impure ) versio n o f on e
`of th e pur e grap h isomorphisms .
`
`•
`
`Sec. 11.2 GraphTheoretic Algorithms
`
`357
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 368
`
`
`
`Figure 11. 4 show s example s o f thes e kind s o f matches .
`One algorith m fo r findin g grap h isomorphis m [Cornei l an d Gotlie b 1970 ] i s
`based o n th e ide a o f separatel y puttin g eac h grap h int o a canonica l form , fro m
`which isomorphis m ma y easil y b e determined . Fo r directe d graph s (i.e. , nonsym
`metric relations ) a backtrac k searc h algorith m [Berztis s 1973 ] work s o n bot h
`graphs a t once .
`Two solution s t o th e subgrap h isomorphis m proble m appea r i n [Ullma n
`1976]: Th e first i s a simpl e enumerativ e searc h o f th e tre e o f possibl e matche s
`between nodes . Th e secon d i s mor e interesting ; i n i t a proces s o f "parallel
`iterative" refinemen t i s applie d a t eac h stag e o f th e search . Thi s proces s i s a wa y o f
`rejecting nod e pair s fro m th e isomorphis m an d o f propagatin g th e effect s o f suc h
`rejections; on e rejecte d matc h ca n lea d t o mor e matche s bein g rejected . Whe n th e
`iteration converge s (i.e. , whe n n o mor e matche s ca n b e rejecte d a t th e curren t
`stage), anothe r ste p i n th e tre e searc h i s performe d (on e mor e matchin g pai r i s hy
`pothesized). Thi s mixin g o f paralleliterativ e processe s wit h tre e searc h i s usefu l i n
`a variet y o f application s (Sectio n 11.4.4 , Chapte r 12) .
`"Double" subgrap h isomorphis m i s easil y reduce d t o subgrap h isomorphis m
`via anothe r wellknow n grap h problem , th e "cliqu e problem. " A clique of siz e ./Vi s
`a totall y connecte d subgrap h o f siz e TV (eac h nod e i s connecte d t o ever y othe r nod e
`in th e cliqu e b y a n arc) . Findin g isomorphism s betwee n subgraph s o f a grap h A
`and subgraph s o f a grap h B i s accomplishe d b y formin g a n association graph G fro m
`the graph s A an d B an d finding clique s i n G (fo r details , se e Sectio n 11.3.3) . Cliqu e
`
`(b)
`
`(c)
`
`(d)
`
`(e)
`
`Isomorphism s an d matches . Th e grap h (a ) ha s a n isomorphis m wit h
`Fig. 11. 4
`(b), variou s subgrap h isomorphism s wit h (c) , an d severa l "double " subgrap h iso
`morphisms wit h (d) . Severa l partia l matche s wit h (e ) (an d als o (b) , (c) , an d (d)) ,
`depending o n whic h missin g o r extr a node s ar e ignored .
`
`358
`
`Ch. 11 Matching
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 369
`
`
`
`finding ma y b e don e wit h a subgrap h isomorphis m algorithm ; henc e th e reduction .
`Several othe r cliquefindin g algorithm s exis t [Amble r e t al . 1975 ; Knode l 1968 ;
`BronandKerbosch 1973 ; Ostee n an d To u 1973] .
`
`11.2.2 Complexit y
`
`It i s o f som e practica l importanc e t o b e awar e o f th e computationa l complexit y o f
`the matchin g algorithm s propose d here ; the y ma y tak e surprisin g amount s o f com
`puter time . Ther e ar e man y accessibl e treatment s o f computationa l complexit y o f
`graphtheoreti c algorithm s [Reingol d e t al . 1977 ; Aho , Hopcrof t an d Ullma n
`1974]. Theoretica l result s usuall y describ e worstcas e o r averag e tim e complexity .
`The stat e o f knowledg e i n grap h algorithm s i s stil l improving ; som e interestin g
`worstcas e bound s hav e no t bee n established .
`A "hard " combinatoria l proble m i s on e tha t take s tim e (i n a usua l mode l o f
`computation base d o n a seria l computer ) proportiona l t o a n exponentia l functio n
`of th e lengt h o f th e input . "Polynomialtime " solution s ar e desirabl e becaus e the y
`do no t gro w a s fas t wit h th e siz e o f th e problem . Th e tim e t o find al l th e clique s o f a
`graph i s i n th e wors t cas e inherentl y exponentia l i n th e siz e o f th e inpu t graphs , be
`cause th e outpu t i s a n exponentia l numbe r o f graphs . Bot h th e singl e subgrap h iso
`morphism proble m an d th e "cliqu e problem " (doe s ther e exis t a cliqu e o f siz e £? )
`are NPcomplete; al l know n deterministi c algorithm s ru n (i n th e wors t case ) i n tim e
`exponential i n th e lengt h o f th e descriptio n o f th e graph s involve d (whic h mus t
`specify th e node s an d arcs) . No t onl y this , bu t i f eithe r o f thes e problem s (o r a hos t
`of othe r N P complet e problems ) coul d b e solve d deterministicall y i n tim e polyno
`mial^ relate d t o th e lengt h o f th e input , i t coul d b e use d t o solv e al l th e othe r N P
`problems i n polynomia l time .
`Graph isomorphism , bot h directe d an d undirected , i s a t thi s writin g i n a
`netherworld (alon g wit h man y othe r combinatoria l problems) . N o polynomial
`time deterministi c algorithm s ar e know n t o exist , bu t th e relatio n o f thes e prob
`lems t o eac h othe r i s no t a s clearcu t a s i t i s betwee n th e NPcomplet e problem . I n
`particular, finding a polynomialtim e deterministi c solutio n t o on e o f the m woul d
`not necessaril y indicat e anythin g abou t ho w t o solv e th e othe r problem s determin
`istically i n polynomia l time . Thes e problem s ar e no t mutuall y reducible . Certai n
`restrictions o n th e graphs , fo r instanc e tha t the y ar e plana r (ca n b e arrange d wit h
`their node s i n a plan e an d wit h n o arc s crossing) , ca n mak e grap h isomorphis m a n
`"easy" (polynomialtime ) problem .
`The averagecas e complexit y i s ofte n o f mor e practica l interes t tha n th e wors t
`case. Typically , suc h a measur e i s impossibl e t o determin e analyticall y an d mus t b e
`approximated throug h simulation . Fo r instance , on e algorith m t o find isomor
`phisms o f randoml y generate d graph s yield s a n averag e tim e tha t seem s no t ex
`ponential, bu t proportiona l t o N 3 , wit h /Vth e numbe r o f node s i n th e grap h [Ull
`man 1976] . Anothe r algorith m seem s t o ru n i n averag e tim e proportiona l t o N 2
`[Corneil an d Gotlie b 1970] .
`All th e grap h problem s o f thi s sectio n ar e i n NP . Tha t is , a «o«deterministi c
`algorithm ca n solv e the m i n polynomia l time . Ther e ar e variou s way s o f visualizin g
`
`Sec. 77. 2 GraphTheoretic
`
`Algorithms
`
`359
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 370
`
`
`
`nondeterministic algorithms ; on e i s tha t th e algorith m make s certai n significan t
`"good guesses " fro m a rang e o f possibilitie s (suc h a s correctl y guessin g whic h sub
`set o f node s fro m grap h B ar e isomorphi c wit h grap h A an d the n onl y havin g t o
`worry abou t th e arcs) . Anothe r wa y i s t o imagin e parallel computation ; i n th e
`clique problem , fo r example , imagin e multipl e machine s runnin g i n parallel , eac h
`with a differen t subse t o f node s fro m th e inpu t graph . I f an y machin e discover s a
`totally connecte d subset , i t has , o f course , discovere d a clique . Checkin g whethe r
`TV node s ar e al l pairwis e connecte d i s a t mos t a polynomialtim e problem , s o al l th e
`machines wil l terminat e i n polynomia l time , eithe r wit h succes s o r not . Severa l in
`teresting processe s ca n b e implemente d wit h paralle l computations . Ullman' s algo
`rithm use s a refinemen t procedur e whic h ma y ru n i n paralle l betwee n stage s o f hi s
`tree search , an d whic h h e explain s ho w t o implemen t i n paralle l hardwar e [Ullma n
`1976].
`
`11.3 IMPLEMENTIN G GRAPHTHEORETI C ALGORITHM S
`
`11.3.1 Matchin g Metric s
`
`Matching involve s quantifiable similarities. A matc h i s no t merel y a correspon
`dence, bu t a correspondenc e tha t ha s bee n quantifie d accordin g t o it s "goodness. "
`This measur e o f goodnes s i s th e matching metric. Similarit y measure s fo r correla
`tion matchin g ar e lumpe d togethe r a s on e number . I n relationa l matchin g the y
`must tak e int o accoun t a relational , structure d for m o f dat a [Shapir o an d Haralic k
`1979].
`Most o f th e structura l matchin g metric s ma y b e explaine d wit h th e physica l
`analogy o f "template s an d springs " [Fischle r an d Elschlage r 1973] . Imagin e tha t
`the referenc e dat a compris e a structur e o n a transparen t rubbe r sheet . Th e match
`ing proces s move s thi s shee t ove r th e inpu t dat a structure , distortin g th e shee t s o
`as t o ge t th e bes t match . Th e final goodnes s o f fit depend s o n th e individua l
`matches betwee n element s o f th e inpu t an d referenc e data , an d o n th e amoun t o f
`work i t take s t o distor t th e sheet . Th e continuou s deformatio n proces s i s a prett y
`abstraction whic h mos t matchin g algorithm s d o no t implement . A computationall y
`more tractabl e for m o f th e ide a i s t o conside r th e mode l a s a se t o f rigi d "tem
`plates" connecte d b y "springs " (se e Fig . 11.5) . Th e template s ar e connecte d b y
`"springs" whos e "tension " i s als o a functio n o f th e relation s betwee n elements . A
`spring functio n ca n b e arbitraril y comple x an d nonlinear ; fo r exampl e th e "ten
`sion" i n th e sprin g ca n attai n ver y hig h o r infinit e value s fo r configuration s o f tem
`plates whic h canno t b e allowed . Nonlinearit y i s goo d fo r suc h constraint s as : i n a
`picture o f a fac e th e tw o eye s mus t b e essentiall y i n a horizonta l lin e an d mus t b e
`within fixed limit s o f distance . Th e qualit y o f th e matc h i s a functio n o f th e good
`ness o f fit o f th e template s locall y an d th e amoun t o f "energy " neede d t o stretc h
`the spring s t o forc e th e inpu t ont o th e referenc e data . Cost s ma y b e impose d fo r
`missing o r extr a elements .
`The templat e matc h function s an d sprin g function s ar e genera l procedures ,
`thus th e template s ma y b e mor e genera l tha n pur e iconi c templates . Further ,
`
`360
`
`Ch. 11 Matching
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 371
`
`
`
`Right
`edge
`
`Mouth
`
`Fig. 11. 5 A template s an d spring s mode l o f a face .
`
`matches ma y b e define d no t onl y betwee n node s an d othe r nodes , bu t betwee n
`nodes an d imag e dat a directly . Thu s th e templat e an d spring s formalis m i s work
`able fo r "crossrepresentational " matching . Th e mechanis m o f minimizin g th e to
`tal cos t o f th e matc h ca n tak e severa l forms ; mor e detaile d example s follo w i n Sec
`tion 11.4 .
`Equation 11. 3 a genera l for m o f th e templateandspring s metric . Tem
`plateCost measure s dissimilarit y betwee n th e inpu t an d th e templates , an d
`SpringCost measure s th e dissimilarit y betwee n th e matche d inpu t elements ' rela
`tions an d th e referenc e relation s betwee n th e templates . MissingCos t measure s th e
`penalties fo r missin g elements . F( ) i s th e mappin g fro m template s o f th e referenc e
`to element s o f th e inpu t data . F partition s th e referenc e template s int o tw o classes ,
`those foun d {FoundinRefer } an d thos e no t foun d {MissinginRefer j i n th e inpu t
`data. I f th e inpu t dat a ar e symboli c the y ma y b e similarl y partitioned . Th e genera l
`metric i s
`
`+
`
`Cost=
`
`TemplateCost U Fid))
`I
`d 6 {FoundinRefer )
`SpringCos t (Fid), Fie))
`L
`id, e) 6 {FoundinRefe r x Foundinlnput l
`+
`L
`c € {MissinginRefer l U {Missinginlnput j
`
`MissingCos t (c )
`
`(11.3)
`
`Equation 11. 3 ma y b e writte n a s on e su m o f generalize d SpringCost s i n whic h
`the templat e propertie s ar e include d (a s 1ar y relations) , a s ar e "springs " involv
`ing missin g elements .
`
`Sec. 11.3
`
`Implementing GraphTheoretic Algorithms
`
`361
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 372
`
`
`
`As wit h correlatio n metrics , ther e ar e normalizatio n issue s involve d wit h
`structural matchin g metrics . Th e numbe r o f element s matche d ma y affec t th e ulti
`mate magnitud e o f th e metric . Fo r instance , i f spring s alway s hav e a finite cost ,
`then th e mor e element s tha t ar e matched , th e highe r th e tota l sprin g energ y mus t
`be; thi s shoul d probabl y no t b e take n t o impl y tha t a matc h o f man y element s i s
`worse tha n a matc h o f a few . Conversely , suppos e tha t relation s whic h agre e ar e
`given positiv e "goodness " measures , an d a matc h i s chose n o n th e basi s o f th e to
`tal "goodness. " The n unles s on e i s careful , th e shee r numbe r o f possibl y mediocr e
`relational matche s induce d b y matchin g man y element s ma y outweig h th e "good
`ness" o f a n elegan t matc h involvin g onl y a fe w elements . O n th e othe r hand , a
`small, elegan t matc h o f a par t o f th e inpu t structur e wit h on e particula r referenc e
`object ma y leav e muc h o f th e searc h structur e unexplained . Thi s goo d "sub
`match" ma y b e les s helpfu l tha n a matc h tha t explain s mor e o f th e input . T o som e
`extent th e genera l metri c (Eq.11.3 ) cope s wit h thi s b y acknowledgin g th e "miss
`ing" categor y o f elements .
`If th e referenc e template s actuall y contai n iconi c representation s o f wha t th e
`input element s shoul d loo k lik e i n th e image , a TemplateCos t ca n b e associate d
`with a templat e an d a locatio n i n th e imag e b y
`
`TemplateCost (Template , Location )
`= ( 1 normalize d correlatio n metri c betwee n
`template shap e an d inpu t imag e a t th e location) .
`If th e matc h is , fo r instance , t o matc h referenc e description s o f a chai r wit h
`an inpu t dat a structure , a typica l "spring " migh t b e tha t th e chai r sea t mus t b e sup
`ported b y it s legs . Thu s i f ./u s th e associatio n functio n mappin g referenc e element s
`such a s LE G o r TABLETO P t o inpu t elements ,
`SpringCost! (/"(LEG) , /"(TABLETOP )
`
`_ J O
`1
`
`i f F (LEG ) appear s t o suppor t F (TABLETOP) ,
`i f F (LEG ) doe s no t appea r t o suppor t F (TABLETOP) .
`
`For quantifie d relations , on e migh t hav e
`SpringCost2 = numbe r o f standar d deviation s fro m th e
`canonical mea n valu e fo r thi s relation .
`Another versio n o f SpringCost 2 i s th e followin g [Barro w an d Poppleston e
`1971].
`P
`. o M . , = SpringCost s o f propertie s (unary ) an d binar y relation s
`total numbe r o f unar y an d binar y spring s
`Empirica l Constan t
`Total numbe r o f referenc e element s matche d
`The first ter m measure s th e averag e badnes s o f matche s betwee n propertie s
`(unary relations ) an d relation s betwee n regions . Th e secon d ter m i s inversel y pro
`portional t o th e numbe r o f region s tha t ar e matched , effectivel y increasin g th e cos t
`of matche s tha t explai n les s o f th e input .
`
`/, , . x
`
`,
`
`362
`
`Ch. 7 7 Matching
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 373
`
`
`
`11.3.2 Backtrac k Searc h
`
`Backtrack searc h i s a generi c nam e fo r a typ e o f potentiall y exhaustiv e searc h or
`ganized i n stages ; eac h processin g stag e attempt s t o exten d a partia l solutio n
`derived i n th e previou s stage . Shoul d th e attemp t fail , th e searc h "backtracks " t o
`the mos t recen t partia l solution , fro m whic h a ne w extensio n i s attempted . Th e
`technique i s basic , amountin g t o a depthfirs t searc h throug h a tre e o f partia l solu
`tions (Fig . 11.6) . Backtrackin g i s a pervasiv e contro l structur e i n artificia l intelli
`
`Fig. 11. 6 Th e grap h o f (a ) i s t o b e matche d i n (b ) wit h arc s al l bein g unlabele d
`but node s havin g propertie s indicate d b y thei r shapes , (c ) i s th e tre e o f solution s
`built b y a backtrac k algorithm .
`
`Implementing GraphTheoretic Algorithms
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 374
`
`
`
`gence, an d throug h th e year s severa l genera l classe s o f technique s hav e evolve d t o
`make th e basic , bruteforc e backtrac k searc h mor e efficient .
`Example: Graph Isomorphisms
`Given tw o graphs ,
`
`*
`(V X,EX)
`Y= (Vy, Ey),
`without los s o f generality , le t V x = V Y — {1 , 2 , . . . , »} , an d le t Xb e th e referenc e
`graph, Ythe inpu t graph . Th e isomorphis m i s give n by : I f / € V x, th e correspond
`ing nod e unde r th e isomorphis m i s F(i) € V Y.
`In th e algorithm , S i s th e se t o f node s accounte d fo r i n Yby a partia l solution .
`Ogives th e curren t leve l o f th e searc h i n th e tre e o f partia l solutions , th e numbe r o f
`nodes i n th e curren t partia l solution , an d th e nod e o f X whos e matc h i n Y i s
`currently bein g sought , v i s a nod e o f Y currentl y bein g considere d t o exten d th e
`current partia l solution . A s written , th e algorith m finds al l isomorphisms . I t i s
`easily modifie d t o qui t afte r finding th e first.
`
`(VyS)
`
`Algorithm 11. 1 Backtrac k Searc h fo r Directe d Grap h Isomorphis m
`Recursive Procedure DirectedGraphlsomorphism s (S,k) ;
`begin
`if S=V Y then ReportAsIsomorphism(F )
`else
`forallvt
`do
`//Match (A: , v )
`then
`begin
`F(k):=v;
`DirectedGraphlsomorphisms (£ € {v},A:rl) ;
`end;
`
`end,
`
`ReportAsIsomorphism coul d prin t o r sav e th e curren t valu e o f F, th e globa l
`structure recordin g th e curren t solution . Match(/c,v ) i s a procedur e tha t test s
`whether v € V y ca n correspon d t o k € V x unde r th e isomorphis m s o fa r define d b y
`F. Le t X k b e th e subgrap h o f A'wit h vertice s {1,2, . . .,k). Th e procedur e "Match "
`must chec k fo r i < k, whethe r (/ , k) i s a n edg e o f X k if f (F(i), v ) i s a n edg e o f Y
`and whethe r (k, i) i s a n edg e o f AT* if f (v , F(i)) i s a n edg e o f Y.
`Improving Backtrack Search
`Several technique s ar e usefu l i n improvin g th e efficienc y o f backtrac k searc h
`[Bittner an d Reingol d 1975] :
`
`364
`
`Ch. 11 Matching
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 375
`
`
`
`1. Branch pruning. Al l technique s o f thi s variet y examin e th e curren t partia l solu
`tion an d prun e awa y descendent s tha t ar e no t viabl e continuation s o f th e solu
`tion. Shoul d non e exist , backtrackin g ca n tak e plac e immediately .
`2. Branch merging. D o no t searc h branche s o f th e solutio n tre e isomorphi c wit h
`those alread y searched .
`3. Tree rearrangement and reordering. Give n prunin g capabilities , mor e node s ar e
`likely t o b e eliminate d b y prunin g i f ther e ar e fewe r choice s t o mak e earl y i n
`the searc h (partia l solutio n node s o f lo w degre e shoul d b e hig h i n th e searc h
`tree). Similarly , searc h firs t thos e extension s t o th e curren t solutio n tha t hav e
`the fewes t alternatives .
`4. Branch and bound. I f a cos t ma y b e assigne d t o solutions , standar d technique s
`such a s heuristi c searc h an d th e A * searc h algorith m [Nilsso n 1980 ] (Sectio n
`4.4) ma y b e employe d t o allo w th e searc h t o procee d o n a "bestfirst " rathe r
`than a "depthfirst " basis .
`
`For extension s o f thes e techniques , se e [Haralic k an d Elliot t 1979] .
`
`11.3.3 Associatio n Grap h Technique s
`
`Generalized Structure Matching
`A genera l relationa l structur e "bes t match " i s les s restricte d tha n grap h iso
`morphism, becaus e node s o r arc s ma y b e missin g fro m on e o r th e othe r graph .
`Also, i t i s mor e genera l tha n subgrap h isomorphis m becaus e on e structur e ma y no t
`be exactl y isomorphi c t o a substructur e o f th e other . A mor e genera l matc h con
`sists o f a se t o f node s fro m on e structur e an d a se t o f node s fro m th e othe r an d a 1: 1
`mapping betwee n the m whic h preserve s th e compatibilitie s o f propertie s an d rela
`tions. I n othe r words , correspondin g node s (unde r th e nod e mapping ) hav e
`sufficiently simila r properties , an d correspondin g set s unde r th e mappin g hav e
`compatible relations .
`The tw o relationa l structure s ma y hav e a comple x makeu p tha t fall s outsid e
`the norma l purvie w o f grap h theory . Fo r instance , the y ma y hav e parameterize d
`properties attache d t o thei r node s an d edges . Th e definitio n o f whethe r a nod e
`matches anothe r nod e an d whethe r tw o suc h nod e matche s ar e mutuall y compati
`ble ca n b e determine d b y arbitrar y procedures , unlik e th e muc h simple r criteri a
`used i n pur e grap h isomorphis m o r subgrap h isomorphism , fo r example . Recal l
`that th e variou s grap h an d subgrap h isomorphism s rel y heavil y o n a 1: 1 match , a t
`least locally , betwee n arc s an d node s o f th e structure s t o b e matched . However , th e
`idea o f a "bes t match " ma y mak e sens e eve n i n th e absenc e o f suc h perfec t
`correspondences.
`The association graph define d i n thi s sectio n i s a n auxiliar y dat a structur e pro
`duced fro m tw o relationa l structure s t o b e matched . Th e beaut y o f th e associatio n
`graph i s tha t i t is a simple , pur e graphtheoreti c structur e whic h i s amenabl e t o
`pure graphtheoreti c algorithm s suc h a s cliqu e finding. Thi s i s usefu l fo r severa l
`reasons.
`
`Sec. 77. 3
`
`Implementing GraphTheoretic
`
`Algorithms
`
`365
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 376
`
`
`
`•
`•
`
`I t take s relationa l structur e matchin g fro m th e a d ho c t o th e classica l domain .
`I t broaden s th e bas e o f peopl e wh o ar e producin g usefu l algorithm s fo r struc
`ture matching . I f th e rathe r specialize d relationa l structur e matchin g enterpris e
`is reducibl e t o a classica l graphtheoretica l problem , the n everyon e workin g o n
`the classica l proble m i s als o workin g indirectl y o n structur e matching .
`• Knowledg e abou t th e computationa l complexit y o f classica l grap h algorithm s il
`luminates th e difficult y o f structur e matching .
`
`Clique Finding for Generalized Matching
`Let a relationa l structur e b e a se t o f element s V, a se t o f propertie s (o r mor e
`simply unar y predicates ) P define d ove r th e elements , an d a se t o f binar y relation s
`(or binar y predicates ) R define d ove r pair s o f th e elements . A n exampl e o f a grap h
`representation o f suc h a structur e i s give n i n Fig . 11.7 .
`Given tw o structure s define d b y (V\, P, R) an d (V 2, P, R), sa y tha t "simi
`lar" an d "compatible " actuall y mea n "th e same. " The n w e construc t a n associa
`tion grap h Ga s follow s [Amble r e t al . 1975] . Fo r eac h v\ i n V\ an d v 2 i n V 2, con
`struct a nod e o f G labeled (vi , v 2) i f v i an d v 2 hav e th e sam e propertie s [p(v\) if f
`p (v 2) fo r eac h p i n Pi . Thu s th e node s o f G denot e assignments , o r pair s o f nodes ,
`one eac h fro m V\ an d V 2, whic h hav e simila r properties . No w connec t tw o node s
`(vj, v 2) an d (v'i , v' 2) o f G i f the y represen t compatible assignment s accordin g t o R,
`that is , i f th e pair s satisf y th e sam e binar y predicate s [r(v\, v\) iff r(v 2, v' 2) fo r
`each r'mR].
`th e tw o relationa l structures , i s
`A matc h betwee n (F lf P, R) an d ( V2, P,R),
`just a se t o f assignment s tha t ar e al l mutuall y compatible . Th e "bes t match " coul d
`well b e take n t o b e th e larges t se t o f assignment s (nod e correspondences ) tha t
`were al l mutuall y compatibl e unde r th e relations . Bu t thi s i n th e associatio n grap h
`G i s jus t th e larges t totall y connecte d (completel y mutuall y compatible)—se t o f
`nodes. I t i s a clique. A cliqu e t o whic h n o ne w node s ma y b e adde d withou t destroy
`ing th e cliqu e propertie s i s a maximal clique . I n thi s formulatio n o f matching , large r
`cliques ar e take n t o indicat e bette r matches , sinc e the y accoun t fo r mor e nodes .
`
`F I R . 11. 7 A grap h representatio n o f a
`relational structure . Element s (nodes ) v\
`and V 3 hav e propert y p i , 1 2 an d i> 4 hav e
`property p2 , an d th e arc s betwee n node s
`indicate tha t th e relatio n r l hold s
`between i> | an d i> 2and betwee n v 2 an d
`vh an d r 2 hold s betwee n v 3 an d v 4 an d
`between v 4and v j .
`
`366
`
`Ch. 11 Matching
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 377
`
`
`
`Thus th e bes t matche s ar e determine d b y th e larges t maxima l clique s i n th e associ
`ation graph . Figur e 11. 8 show s a n example : Certai n subfeature s o f th e object s hav e
`been selecte d a s "primitiv e elements " o f th e objects , an d appea r a s node s (ele
`ments) i n th e relationa l structures . T o thes e node s ar e attache d properties , an d
`between the m ca n exis t relations . Th e choic e o f primitives , properties , an d rela
`tions i s u p t o th e designe r o f th e representation . Her e th e primitive s o f th e
`representation correspon d t o edge s an d corner s o f th e shape .
`The associatio n grap h i s show n i n 11.8e . It s node s correspon d t o pair s o f
`nodes, on e eac h fro m A an d B , whos e propertie s ar e similar . [Notic e tha t ther e i s
`no nod e i n th e associatio n grap h fo r (6,6')] . Th e arc s o f th e associatio n grap h indi
`cate tha t th e endpoint s o f th e ar c represen t compatibl e associations . Maxima l
`cliques i n th e associatio n grap h (show n a s set s o f node s wit h th e sam e shape ) indi
`cate set s o f consisten t associations . Th e larges t maxima l cliqu e provide s th e nod e
`pairings o f th e "bes t match. "
`In th e exampl e construction , th e associatio n grap h i s forme d b y associatin g
`nodes wit h exactl y th e sam e propertie s (actuall y unar y predicates) , an d b y allowin g
`as compatibl e association s onl y thos e wit h exactl y th e sam e relation s (actuall y
`binary predicates) . Thes e condition s ar e eas y t o state , bu t the y ma y no t b e exactl y
`what i s needed . I n particular , i f th e propertie s an d relation s ma y tak e o n range s o f
`values greate r tha n th e binar y "exists " an d "doe s no t exist, " the n a measur e o f
`similarity mus t b e introduce d t o defin e whe n nod e propertie s ar e simila r enoug h
`for association , and~whe n relation s ar e "simila r enough~fo r ccjnpaTibilityTArbitraril y
`complex function s ca n decid e whethe r propertie s an d relation s ar e similar . A s lon g
`as th e functio n answer s "yes " o r "no, " th e complexit y o f it s computation s i s ir
`relevant t o th e matchin g algorithm .
`The followin g recursiv e cliquefindin g algorith m build s u p clique s a nod e a t a
`time [Amble r e t al . 1975] . Th e searc h tre e i t generate s ha s state s tha t ar e ordere d
`pairs (se t o f node s chose n fo r a clique , se t o f node s availabl e fo r inclusio n i n th e
`clique). Th e roo t o f th e tre e i s th e stat e (0 , al l grap h nodes) , an d a t eac h branc h a
`choice i s mad e whethe r t o includ e o r no t t o includ e a n eligibl e nod e i n th e clique .
`(If a nod e i s eligibl e fo r inclusio n i n cliqu e X, the n each cliqu e includin g A"mus t ei
`ther includ e th e nod e o r exclud e it) .
`
`Algorithm 11.2 : CliqueFindin g Algorith m
`Comment Nodes i s th e se t o f node s i n th e inpu t graph .
`Comment
`Cliques (X, Y) take s a s argument s a cliqu e X, an d F , a se t o f node s tha t include s
`X. I t return s al l clique s tha t includ e Jan d ar e include d i n Y.
`Cliques (0,Nodes ) finds al l clique s i n th e graph .
`CliquesiXJ) : =
`// n o nod e i n Y—X\s connecte d t o al l element s o f X
`then {X\
`else
`CliquesUriJ {y],V) \J Clique s (X, Y{y})
`wherey i s connecte d