`Initial stack s
`
`Floo r
`Goa l stac k
`
`Fig . 13. 1 A simpl e bloc k stackin g task .
`
`INITIAL STATE : ON(C,A) , O N (A , Floor) , ON(B , Floor) ,
`CLEAR (C) , CLEA R (B) , CLEA R (Floor )
`The goa l stat e i s on e i n whic h th e followin g tw o assertion s ar e true .
`GOAL ASSERTIONS : ON(A,B) , ON(B,C )
`With onl y thes e rules , th e formalizatio n o f th e bloc k stackin g worl d yield s a ver y
`"loose" semantics . (Th e tas k easil y translate s t o sortin g integer s wit h som e re
`strictions o n operations , o r t o th e "sedation " tas k o f arrangin g block s horizontall y
`in orde r o f size , o r a hos t o f others. )
`Actions transfor m th e se t o f assertion s describin g th e world . Fo r problem s o f
`realistic scale , th e representatio n o f th e tre e o f worl d state s i s a practica l problem .
`The issu e i s on e o f maintainin g severa l coexistin g "hypothetica l worlds " an d rea
`soning abou t them . Thi s i s anothe r versio n o f th e fram e proble m discusse d i n sec .
`12.1.6. On e wa y t o solv e thi s proble m i s t o giv e eac h assertio n a n extr a argument ,
`naming th e hypothetica l worl d (usuall y calle d a situatio n [Nilsso n 1980 ; McCarth y
`and Haye s 1969] ) i n whic h th e assertio n holds . The n action s ma p situation s t o situ
`ations a s wel l a s introducin g an d changin g assertions .
`An equivalen t wa y t o think abou t (an d implement ) multiple , dependent , hy
`pothetical world s i s wit h a treestructure d contextoriented data base. Thi s ide a i s a
`general on e tha t i s usefu l i n man y artificia l intelligenc e applications , no t jus t sym
`bolic planning . Suc h dat a base s ar e include d i n man y artificia l intelligenc e
`languages an d appea r i n othe r mor e traditiona l environment s a s well . A context
`oriented dat a bas e acts lik e a tre e o f dat a bases ; a t an y nod e o f th e tre e i s a se t o f
`assertions tha t make s u p th e dat a base . A ne w dat a bas e (context ) ma y b e spawne d
`from an y contex t (dat a base ) i n th e tree . Al l assertion s tha t ar e tru e i n th e spawn
`ing (ancestor ) contex t ar e initiall y tru e i n th e spawne d (descendant ) context .
`However, ne w assertion s adde d i n an y contex t o r delete d fro m i t d o no t affec t it s
`ancestor. Thu s b y goin g bac k t o th e ancestor , al l dat a bas e change s performe d i n
`the descenden t contex t disappear .
`Implementing suc h a dat a bas e i s a n interestin g exercise . Copyin g al l asser
`tions t o eac h ne w contex t i s possible , bu t ver y wastefu l i f onl y a fe w change s ar e
`made i n eac h context . Th e followin g mechanis m i s muc h mor e efficient . Th e roo t
`or initia l contex t ha s som e se t o f assertion s i n it , an d eac h descendan t contex t i s
`merely a n add list o f assertion s t o ad d t o th e dat a bas e an d a delete list o f assertion s
`to delete . The n t o se e i f a n assertio n i s tru e i n a context , d o th e following .
`
`1. I f th e contex t i s th e roo t context , loo k u p "a s usual. "
`2. Otherwise , i f th e assertio n i s o n th e add list o f thi s context , retur n true. I f th e
`assertion i s o n th e delete list o f thi s context , retur n false.
`
`4 4 0
`
`Ch. 13 Goal Achievement
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 452
`
`
`
`3. Otherwise , recursivel y appl y thi s procedur e t o th e ancesto r o f thi s context .
`In a genera l programmin g environment , context s hav e names , an d ther e i s
`the facilit y o f executin g procedure s "in " particula r contexts , movin g aroun d th e
`context tree , an d s o forth . However , i n wha t follows , onl y th e abilit y t o loo k u p
`assertions i n context s i s relevant .
`
`13.1.2 Representin g Action s
`
`Represent a n actio n a s a triple .
`ACTION :: = [PATTERN , PRECONDITIONS , POSTCONDITIONS] .
`Here th e patter n give s th e nam e o f th e actio n an d name s fo r th e object s wit h whic h
`it deals—it s "forma l parameters. " Preconditions an d postcondition s ma y us e th e
`formal variable s o f th e pattern . I n a sense , th e precondition s an d postcondition s
`are th e "body " o f th e action , wit h subroutinelik e "variabl e bindings " takin g
`place whe n th e actio n i s t o b e performed . Th e precondition s giv e th e worl d state s
`in whic h th e actio n ma y b e applied . Her e th e precondition s ar e assume d simpl y t o
`be a lis t o f assertion s al l o f whic h mus t b e true . Th e postcondition s describ e th e
`world stat e tha t result s fro m performin g th e action . Th e contextoriente d dat a
`base o f hypothetica l world s ca n b e use d t o implemen t th e postconditions .
`
`POSTCONDITIONS :: = [ADDLIST , DELETELIST] .
`An actio n i s the n performe d a s follows .
`
`2.
`
`1. Bin d th e patter n variable s t o entitie s i n th e world , thu s bindin g th e associate d
`variables i n th e precondition s an d postconditions .
`I f th e precondition s ar e me t (th e boun d assertion s exis t i n th e dat a base) , d o
`the nex t step , els e exi t reportin g failure .
`3. Delet e th e assertion s i n th e delet e list , ad d thos e i n th e ad d list , an d exi t re
`porting success .
`Here i s th e Move actio n fo r ou r blockstackin g example .
`
`PATTERN
`
`Move Object Xfrom YtoZ
`PRECONDITIONS DELETELIST
`
`ADDLIST
`
`Move(X,Y,Z) CLEA R (X )
`CLEAR (Z )
`ON(X,Y)
`
`ON(X,Y )
`CLEAR(Z )
`
`ON(X,Z )
`CLEA R (Y )
`
`Here X , Y, an d Z ar e al l variable s boun d t o worl d entities . I n th e initia l stat e
`of Fig . 13.1 , Mov e (C,A, Floor ) bind s Zt o C , Yto A, Zt o Floor , an d th e precondi
`tions ar e satisfied ; th e actio n ma y proceed .
`However, notic e tw o things .
`
`Sec. 13.1 Symbolic Planning
`
`44 1
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 453
`
`
`
`1. Th e actio n give n abov e delete s th e CLEA R (Floor ) assertio n tha t alway s
`should b e true . On e mus t fix thi s somehow ; puttin g CLEA R (Floor ) i n th e
`addlis t doe s th e job , bu t i s a littl e inelegant .
`2. Wha t abou t a n actio n lik e Move(C,^(,C) ? I t meet s th e preconditions , bu t
`causes troubl e whe n th e ad d an d delet e list s ar e applied . On e fix her e i s t o kee p
`in th e dat a bas e ("worl d model" ) a se t o f assertion s suc h a s Differen t (A,B),
`Different (A, Floor) , . . .
` , an d t o ad d assertion s suc h a s Differen t {X,Z) t o th e
`preconditions o f Move .
`Such housekeepin g chore s an d detail s o f axiomatizatio n ar e inheren t i n ap
`plying basicall y syntactic , forma l solutio n method s t o proble m solving . Fo r now ,
`let u s assum e tha t CLEA R (Floor ) i s neve r deleted , an d tha t Move(A' , Y,Z) i s ap
`plied onl y i f Z i s differen t fro m Xan d Y.
`
`13.1.3 Stackin g Block s
`
`In th e blockstackin g example , th e goa l i s tw o simultaneou s assertions , ON(A,B)
`and ON(i?,C) . On e solutio n metho d proceed s b y repeatedl y pickin g a goa l t o wor k
`on, finding a n operato r tha t move s close r t o th e goal , an d applyin g it . I n thi s cas e o f
`only on e actio n th e questio n i s ho w t o appl y it—wha t t o mov e where . Thi s i s
`answered b y lookin g a t th e postcondition s o f th e actio n i n th e ligh t o f th e goal . Th e
`reasoning migh t g o lik e this : ON(Z?,C ) ca n b e mad e tru e i f X is B an d Zi s C. Tha t i s
`possible i n thi s stat e i f Yis A; al l precondition s ar e satisfied , an d th e goa l O N (i?,C )
`can b e achieve d wit h on e action .
`Part o f th e worl d stat e (o r context ) tre e th e planne r mus t searc h i s show n i n
`Fig. 13.2 , wher e state s ar e show n diagrammaticall y instea d o f throug h set s o f asser
`tions. Notic e th e followin g thing s i n Fig . 13.2 .
`
`1. Tryin g t o achiev e O N (B, C) first i s a mistak e (Branc h 1) .
`2. Tryin g t o achiev e ON(A,B) first i s als o a mistak e fo r les s obviou s reason s
`(Branch 2) .
`3. Branche s 1 an d 2 sho w "subgoa l interaction. " Th e goal s a s state d ar e no t in
`dependent. Branc h 3 mus t b e generate d somehow , eithe r throug h backtrack
`ing o r som e intelligen t wa y o f copin g wit h interaction . I t wil l neve r b e foun d b y
`the singleminde d approac h o f (1 ) an d (2) . However , i f ON(C,Floor ) wer e
`one o f th e goa l assertions , Branc h 3 coul d b e found .
`
`Clearly, representin g worl d an d action s i s no t th e whol e stor y i n planning . In
`telligent searc h o f th e contex t i s als o necessary . Thi s searc h involve s subgoa l selec
`tion, actio n selection , an d actio n argumen t selection . Ba d choice s anywher e ca n
`mean inefficien t o r loopin g actio n sequences , o r th e generatio n o f impossibl e
`subgoals. "Intelligent " searc h implie s a metaleve l capability : th e abilit y o f a pro
`gram t o reaso n abou t it s ow n plans . "Pla n critics " ar e ofte n a par t o f sophisticate d
`planners; on e o f thei r mai n job s i s t o isolat e an d rectif y unwante d subgoa l interac
`tion [Sussma n 1975] .
`
`442
`
`Ch. 13 Goal Achievement
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 454
`
`
`
`c
`A
`
`B
`
`(jWove KB, C, f T )
`
`c
`
`A
`
`B
`
`Branch 1
`
`E 0 0
`
`Bran : h 2
`
`A
`A
`A
`
`B
`B
`B
`
`C
`C
`C
`
`Branch 3
`
`Fig. 13. 2 A stat e tre e generate d i n plannin g ho w l o slac k thre e blocks .
`
`Intelligent choic e o f action s i s th e cru x o f planning , an d i s a majo r researc h is
`sue. Severa l avenue s hav e bee n an d ar e bein g tried . Perhap s subgoal s ma y b e or
`dered b y difficult y an d achieve d i n tha t order . Perhap s plannin g shoul d procee d a t
`various level s o f detai l (lik e multiresolutio n imag e understanding) , wher e th e stra
`tegic skeleto n o f a pla n i s derive d withou t details , the n th e detail s ar e filled i n b y
`applying th e planne r i n mor e detai l t o th e subgoal s i n th e lowresolutio n plan .
`
`Sec. 73. 7 Symbolic Planning
`
`443
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 455
`
`
`
`13.1.4 Th e Fram e Proble m
`
`All plannin g i s plague d b y aspect s o f th e frame problem (introduce d i n Sectio n
`12.1.6).
`
`1.
`
`I t i s impractica l (an d boring ) t o writ e dow n i n a n actio n al l th e thing s tha t sta y
`the sam e whe n a n actio n i s applied .
`2. Similarly , i t i s impractica l t o reasser t i n th e dat a bas e al l th e thing s tha t remai n
`true whe n a n actio n i s implied .
`3. Ofte n a n actio n ha s effect s tha t canno t b e represente d wit h simpl e ad d an d
`delete lists .
`
`The ad d an d delet e lis t mechanis m an d th e contextoriente d dat a bas e
`mechanism addresse d th e first tw o problems . Th e las t proble m i s mor e trouble
`some.
`Add an d delet e list s ar e simpl e ideas , wherea s th e worl d i s a comple x place . I n
`many interestin g cases , th e ad d an d delet e list s depen d o n th e curren t stat e o f th e
`world whe n th e actio n i s applied . Thin k o f action s TURNBYU ) an d MOVEBY(Z )
`in a worl d wher e orientatio n an d locatio n ar e important . Th e orientatio n an d loca
`tion afte r a n actio n depen d no t jus t o n th e actio n bu t o n th e stat e o f th e worl d jus t
`before th e action .
`Again, th e actio n ma y hav e ver y comple x effect s i f ther e ar e comple x depen
`dencies betwee n worl d objects . Conside r th e proble m o f th e "monke y an d bana
`nas," wher e th e monke y plan s t o pus h th e bo x unde r th e banana s an d clim b o n i t
`to reac h the m (Fig . 13.3) . Implementatio n o f realisticall y powerfu l ad d an d delet e
`lists ma y i n fac t requir e arbitrar y amount s o f deductio n an d computation .
`
`/ A
`
`V
`
`V V
`
`K
`
`4 \>
`
`/
`
`\
`
`/
`
`/
`
`444
`
`Fig. 13. 3 Action s ma y hav e comple x
`effects.
`
`Ch. 7 3 Coal Achievement
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 456
`
`
`
`This quic k preci s o f symboli c plannin g doe s no t addres s man y "classical "
`topics, suc h a s learnin g o r rememberin g usefu l plans . Als o no t discusse d are : plan
`ning a t varyin g level s o f abstraction , plan s wit h uncertai n information , o r plan s
`with costs . Th e intereste d reade r shoul d consul t th e Reference s fo r mor e informa
`tion. Th e nex t sectio n addresse s plan s wit h cost s sinc e the y ar e particularl y
`relevant t o vision ; som e o f th e othe r issue s appea r i n th e Exercises .
`
`13.2 PLANNIN G WIT H COST S
`
`Decision makin g unde r uncertaint y i s a n importan t topi c i n it s ow n right , bein g o f
`interest t o policymaker s an d manager s [Raiff a 1968] . Analyti c technique s tha t ca n
`derive th e strateg y wit h th e "optima l expecte d outcome " o r "maxima l expecte d
`utility" ca n b e base d o n Bayesia n model s o f probability .
`In [Feldma n an d Sproul l 1977 ] thes e technique s ar e explore d i n th e contex t
`of actio n plannin g fo r realworl d action s an d vision . A s a n exampl e o f th e tech
`niques, the y ar e use d t o mode l a n extende d versio n o f th e "monke y an d bananas "
`problem o f th e las t section , wit h multipl e boxe s bu t withou t th e maddenin g pulle y
`arrangement. I n th e extende d problem , ther e ar e boxe s o f differen t weight s whic h
`may o r ma y no t suppor t th e monkey , an d h e ca n appl y test s (e.g. , vision ) a t som e
`cost t o determin e whethe r the y ar e usable . Pushin g weighte d boxe s cost s som e
`effort, an d th e gratificatio n o f eatin g th e banana s i s "worth " onl y som e finite
`amount o f effort . Thi s extende d se t o f consideration s i s mor e lik e everyda y deci
`sion makin g i n th e numbe r o f factor s tha t nee d balancing , i n th e uncertaint y in
`herent i n th e universe , an d i n th e richnes s o f applicabl e tests . I n fact , on e migh t
`make th e clai m tha t huma n being s alway s "maximiz e thei r expecte d utility, " an d
`if on e kne w a person' s utilit y functions , hi s behavio r woul d becom e predictable .
`The mor e intuitiv e clai m tha t human s being s pla n onl y a s fa r a s "sufficien t ex
`pected utility " ca n b e cas t a s a maximizatio n operatio n wit h nonzer o "cos t o f plan
`ning."
`The sequentia l decisionmakin g mode l o f plannin g wit h th e goa l o f maximiz
`ing th e goodnes s o f th e expecte d outcom e wa s use d i n a trave l planne r [Sproul l
`1977]. Knowledg e o f schedule s an d cost s o f variou s mode s o f transportatio n an d
`the attendan t risk s coul d b e combine d wit h persona l prejudice s an d preference s t o
`produce a n itinerar y wit h th e maximu m expecte d utility . I f unexpecte d situation s
`(canceled flights , say ) aros e en route, replannin g coul d b e initiated ; thi s incremen
`tal pla n ramificatio n i s a natura l extensio n o f sequentia l decisio n making .
`This sectio n i s concerne d wit h measurin g th e expecte d performanc e o f plan s
`using a singl e number . Althoug h on e migh t expec t on e numbe r t o b e inadequate ,
`the centra l theore m o f decisio n theor y [DeGroo t 1970 ] show s essentiall y tha t on e
`number i s enough . Usin g a numerica l measur e o f goodnes s allow s comparison s
`between normall y incomparabl e concept s t o b e mad e easily . Quit e frequentl y nu
`merical score s ar e directl y relevan t t o th e issue s a t stak e i n planning , s o the y ar e
`not obnoxiousl y reductionistic . Decisio n theor y ca n als o hel p i n th e proces s o f ap
`plying a plan—th e basi c pla n ma y b e simple , bu t it s applicatio n t o th e worl d ma y b e
`complex, i n term s o f whe n t o declar e a resul t establishe d o r a n actio n unsuccessful .
`The decisiontheoreti c approac h ha s bee n use d i n severa l artificia l intelligenc e an d
`
`Sec. 13.2 Planning with Costs
`
`445
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 457
`
`
`
`vision program s [Feldma n an d Yakimovsk y 1974 ; Bolle s 1977 ; Garve y 1976 ; Bal
`lard 1978 ; Sproul l 1977] .
`
`13.2.1 Planning , Scoring , an d Thei r Interactio n
`
`For didacti c purposes , th e processe s o f pla n generatio n an d pla n scorin g ar e con
`sidered separately . I n fact , thes e processe s ma y cooperat e mor e o r les s intimately .
`The planne r produce s "sequences " o f actions fo r evaluatio n b y th e scorer . Eac h ac
`tion (computation , informatio n gathering , performin g a realworl d action ) ha s a
`cost, expressin g expenditur e o f resources , o r associate d unhappiness . A n actio n
`has a se t o f possibl e outcomes, o f whic h onl y on e wil l reall y occu r whe n th e actio n i s
`performed. A goalis a stat e o f th e worl d wit h a n associate d "happiness " o r utility.
`For th e purpose s o f uniformit y an d forma l manipulation , goal s ar e treate d a s (null )
`actions wit h n o outcomes , an d negativ e utilitie s ar e use d t o expres s costs . The n th e
`plan ha s onl y action s i n it ; the y ma y b e arrange d i n a stric t sequence , o r b e i n loops ,
`be conditiona l o n outcome s o f othe r actions , an d s o forth .
`The scoring proces s evalute s th e expected utility o f a plan . I n a n uncertai n
`world, a pla n prio r t o executio n ha s onl y a n expecte d goodness—somethin g migh t
`go wrong . Suc h a scorin g proces s typicall y i s no t o f interes t t o thos e wh o woul d us e
`planners t o solv e puzzle s o r d o proofs ; wha t i s interestin g i s th e result , no t th e
`effort. Bu t plan s tha t ar e "optimal " i n som e sens e ar e decidedl y o f interes t i n real
`world decisio n making . I n a visio n context , plan s ar e usuall y usefu l onl y i f the y
`can b e evaluate d fo r efficienc y an d efficacy .
`Scoring ca n tak e plac e o n "complete " plans , bu t i t ca n als o b e use d t o guid e
`plan generation . Th e usua l artificia l intelligenc e problemsolvin g technique s o f
`progressive deepenin g searc h an d branchandboun d prunin g ma y b e applie d t o
`planning i f scorin g happen s a s th e pla n i s generate d [Nilsso n 1980] . Scorin g ca n b e
`used t o asses s th e cos t o f plannin g an d t o monito r plannin g horizon s (ho w fa r
`ahead t o loo k an d ho w detaile d t o mak e th e plan) . Scorin g wil l penaliz e plan s tha t
`loop withou t producin g results . Pla n improvements , suc h a s replannin g upo n
`failure, ca n b e assesse d wit h scores , an d th e contributio n o f additiona l step s (sa y
`for extr a informatio n gathering ) ca n b e assesse d dynamicall y b y scoring . Scorin g
`can b e arbitraril y comple x utilit y functions , thu s reflectin g suc h concept s a s "ris k
`aversion" an d nonlinea r valu e o f resource s [Raiff a 1968] .
`
`13.2.2 Scorin g Simpl e Plan s
`
`Scoring and an Example
`A simple plan i s a tre e o f node s (ther e ar e n o loops) . Th e node s represen t ac
`tions (an d goals) . Outcome s ar e represente d b y labele d arc s i n th e tree . A probabil
`ity o f occurrenc e i s associate d wit h eac h possibl e outcome ; sinc e exactl y on e out
`come actuall y occur s pe r action , th e probabilitie s fo r th e possibl e outcome s o f an y
`action su m t o unity .
`The score o f a pla n i s it s expected utility. Th e expecte d utilit y o f an y nod e i s re
`cursively define d a s it s utilit y time s th e probabilit y o f reachin g tha t nod e i n th e
`
`446
`
`Ch. 13 Goal Achievement
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 458
`
`
`
`plan, plu s th e expecte d utilitie s o f th e action s a t it s (possible ) outcomes . Th e pro
`bability o f reachin g an y "goa l state " i n th e pla n i s th e produc t o f probabilitie s o f
`outcomes formin g a pat h fro m th e roo t o f th e pla n t o th e goa l state .
`As a n example , conside r th e pla n show n i n Fig . 13.4 . I f th e pla n o f Fig . 13. 4
`
`Table locate d
`
`Table no t locate d
`
`Find telephon e
`shape
`
`Do no t
`find telephon e
`shape
`
`Telephone
`there
`
`No telephon e
`there
`
`P13
`
`PU
`
`Fig. 13. 4 Thi s pla n t o find a telephon e i n a n offic e scen e involve s finding a tabl e first
`and lookin g ther e i n mor e detail . Th e action s an d outcome s ar e shown . Th e probabilitie s
`of outcome s ar e assigne d symbol s (P10 , etc.) . Utilitie s (denote d b y U: ) ar e give n fo r th e
`individual actions . Not e tha t negativ e utilitie s ma y b e considere d costs . I n thi s example ,
`decisionmakin g take s n o effort , imag e processin g cost s vary , an d ther e ar e variou s penal
`ties an d reward s fo r correc t an d incorrec t finding o f th e telephone .
`
`Sec. 13.2 Planning with Costs
`
`447
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 459
`
`
`
`has probabilitie s assigne d t o it s outcomes , w e ma y comput e it s expecte d utility .
`Figure 13. 5 show s th e calculation . Th e probabilit y o f correctl y findin g th e tele
`phone i s 0.34 , an d th e expecte d utilit y o f th e pla n i s 433 .
`Although th e generatio n o f a pla n ma y no t b e easy , scorin g a pla n i s a trivia l
`exercise onc e th e probabilitie s an d utilitie s ar e known . I n practice , th e assignmen t
`of probabilitie s i s usuall y a sourc e o f difficulty . Th e followin g i s a n exampl e usin g
`
`Find telephon e
`shape
`
`Do no t
`find telephon e
`shape
`
`Telephone
`there
`
`No telephon e
`there
`
`0.95
`
`Fig. 13. 5 A s fo r Fig . 13.4 . U : give s th e utilit y o f eac h action . E(U) : give s th e expecte d
`utillity o f th e action , whic h depend s o n th e outcome s belo w it . Value s fo r outcom e proba
`bilities ar e give n o n th e outcom e arcs .
`
`448
`
`Ch. 13 Coal Achievement
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 460
`
`
`
`the telephonefindin g pla n an d som e assumption s abou t th e tests . Differen t as
`sumptions yiel d differen t scores .
`Computing Outcome Probabilities: An Example
`This exampl e relie s heavil y o n Bayes ' rule :
`
`P(B\A)P(A) = P(AAB) = P(A\B)P(B).
`
`(13.1 )
`
`Let u s assum e a specifi c a prior i probabilit y tha t th e scen e contain s a tele
`phone.
`
`(13.2 )
`P\ = aprior i probabilit y o f Telephon e
`Also assum e tha t somethin g i s know n abou t th e behavio r o f th e variou s test s i n th e
`presence o f wha t the y ar e lookin g for . Thi s knowledg e ma y accru e fro m experi
`ments t o se e ho w ofte n th e tabl e tes t foun d table s whe n telephone s (o r tables )
`were an d wer e no t present . Le t u s assum e tha t th e followin g ar e know n probabili
`ties.
`
`(13.3 )
`Pi = P(tabl e located|telephon e i n scene )
`(13.4 )
`Ps = P(tabl e located|n o telephon e i n scene )
`Either ther e i s a telephon e o r ther e i s not , an d a tabl e i s locate d o r i t i s not , s o
`P2 = a prior i probabilit y o f n o telephon e = 1 — P\
`Pi, = P(n o tabl e locate d |telephon e i n scene ) = 1 — P 3
`P(, = P(n o tabl e locate d |n o telephon e i n scene ) = 1 — P 5
`Similarly wit h th e "shap e test " fo r telephones : assum e probabilitie s
`P1 = P (telephon e shap e locate d | telephone )
`P9= P(telephon e shap e locate d |n o telephone )
`
`(13.8 )
`(13.9 )
`
`(13.5 )
`(13.6 )
`(13.7 )
`
`with
`
`(13.10 )
`
`P l0=lP 9
`
`Ps=lP 7,
`as above .
`There ar e a fe w point s t o make : First , i t i s no t necessar y t o kno w exactl y thes e
`probabilities i n orde r t o scor e th e plan ; on e coul d us e relate d probabilitie s an d
`Bayes' rule . Othe r usefu l probabilitie s ar e o f th e for m
`P (telephon e | telephon e shap e located) .
`In som e system s [Garve y 1976 ] thes e ar e assume d t o b e availabl e directly . Thi s
`section show s ho w t o deriv e the m fro m know n conditiona l probabilitie s tha t
`describe th e behavio r o f detector s give n certai n scen e phenomena .
`Second, notic e th e assumptio n tha t althoug h bot h th e outcom e o f th e tabl e
`test an d th e shap e tes t depen d o n th e presenc e o f telephones , the y ar e take n t o b e
`independent o f eac h other . Tha t is , havin g foun d a tabl e tell s u s nothin g abou t th e
`likelihood o f finding a telephon e shape . Independenc e assumption s suc h a s thi s ar e
`
`Sec. 13.2 Planning with Costs
`
`4 4 9
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 461
`
`
`
`useful t o limi t computation s an d dat a gathering , bu t ca n b e somewha t unrealistic .
`To accoun t fo r th e dependence , on e woul d hav e t o measur e suc h quantitie s a s
`P(telephone shap e found|tabl e located) .
`Now t o comput e som e outcom e probabilities : Conside r th e probabilit y
`Pu = P (tabl e located )
`
`(13.11 )
`
`Let u s writ e
`TL fo r Tabl e Locate d
`TNL fo r Tabl e No t Located .
`A tabl e ma y b e locate d whethe r o r no t a telephon e i s i n th e scene . I n term s o f
`known probabilities , Bayes ' rul e yield s
`Pu^PiPi + PsPi
`
`(1312 )
`
`Then
`
`(13.13 )
`
`P12 = />(TNL ) = \~P n
`Calculating P 13 show s a nea t tric k usin g Bayes ' Rule :
`(13.14 )
`Pu = P (telephon e |TNL )
`That is , P1 3 i s th e probabilit y tha t ther e i s a telephon e i n th e scen e give n tha t
`search fo r a tabl e wa s unsuccessful . Thi s probabilit y i s no t know n directly , bu t
`p _ P (telephon e an d TNL )
`13
`P(TNL )
`P (TN L an d telephone )
`Pn
`= [P (TN L | telephone ) P (telephone ) ]
`Pn
`
`, J 3 15 N
`
`[P4P\]
`Pn
`
`Then, o f cours e
`
`(13.16 )
`P u l P n
`Reasoning i n thi s wa y usin g th e conditiona l probabilitie s an d assumption s
`about thei r independenc e allow s th e completio n o f th e calculatio n o f outcom e pro
`babilities (se e th e Exercises) . On e possibl y confusin g poin t occur s i n calculatio n o f
`P ] 5 , whic h i s
`
`P\s= P (telephon e shap e foun d | tabl e located )
`(13.17 )
`By assumption , thes e event s ar e onl y indirectl y related . B y th e simplifyin g assump
`tions o f independence , th e shap e operato r an d th e tabl e operato r ar e independen t
`in thei r operation . (Suc h assumption s migh t b e fals e i f the y use d commo n imag e
`processing subroutines , fo r example. ) O f course , th e probabilit y o f succes s o f eac h
`
`4 5 0
`
`Ch. 13 Coal Achievement
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 462
`
`
`
`depends o n th e presenc e o f a telephon e i n th e scene . Therefor e thei r performanc e
`is linke d i n th e followin g wa y (se e th e Exercises) . (Writ e TS L fo r Telephon e Shap e
`Located.)
`P15 = P (TS L |TL ) P (TS L | telephone ) P (telephon e |TL )
`+P(TSL|no telephone) Pino telephone|TL )
`
`(13.18 )
`
`13.2.3 Scorin g Enhance d Plan s
`
`The plan s o f Sectio n 13.2. 2 wer e calle d "simple " becaus e o f thei r tre e structure ,
`complete orderin g o f actions , an d th e simpl e action s o f thei r nodes . Wit h a riche r
`output fro m th e symboli c planner , th e plan s ma y hav e differen t structure . Fo r ex
`ample, ther e ma y b e OR nodes , an y on e o f whos e son s wil l achiev e th e actio n a t
`the node ; AND nodes , al l o f whic h mus t b e satisfie d (i n an y order ) fo r th e actio n t o
`be satisfactoril y completed ; SEQUENCE nodes , whic h specif y a se t o f action s an d a
`particular orde r i n whic h t o achiev e them . Th e pla n ma y hav e loops , share d
`subgoal structure , o r goal s tha t depen d o n eac h other . Ho w enhance d plan s ar e in
`terpreted an d execute d depend s o n th e scorin g algorithms , th e possibilitie s o f
`parallel execution , whethe r executio n an d scorin g ar e interleaved , an d s o forth .
`This treatmen t ignore s parallelis m an d limit s discussio n t o expandin g enhance d
`plans int o simpl e ones .
`It shoul d b e clea r ho w t o g o abou t convertin g man y o f thes e enhance d plan s
`to simpl e plans . Fo r instance , sequenc e node s simpl y g o t o a uniqu e pat h o f ac
`tions. Alternatively , dependin g o n assumption s abou t outcome s o f suc h action s
`(say whethe r the y ca n fail) , the y ma y b e coalesce d int o on e action , a s wa s th e
`"threshold, find blobs , an d comput e shapes " actio n i n th e telephonefindin g plan .
`Rather mor e interestin g ar e th e O R an d AN D nodes , th e orde r o f whos e
`subgoals i s unspecified . Eac h suc h nod e yield s man y simpl e plans , dependin g o n
`the orde r i n whic h th e subgoal s ar e attacked . On e wa y t o scor e suc h a pla n i s t o
`generate al l possibl e simpl e plan s an d scor e eac h one , bu t perhap s i t i s possibl e t o
`do better . Fo r example , loop s an d mutua l dependencie s i n plan s ca n b e deal t wit h
`in variou s ways . A loo p ca n b e analyze d t o mak e sur e tha t i t contain s a n exi t (suc h
`as a branc h o f a n O R nod e tha t ca n b e executed) . On e ca n mak e a d ho c assump
`tions tha t th e cos t o f executio n i s alway s mor e tha n th e cos t o f plannin g [Garve y
`1976], an d scor e th e loo p b y it s executabl e branch . Anothe r ide a i s t o pla n incre
`mentally wit h a finite horizon , expandin g th e pla n throug h som e progressiv e
`deepening, heuristi c search , o r prunin g strategy . Th e accumulate d cos t o f goin g
`around a loo p wil l soo n remov e i t fro m furthe r consideration .
`Recall (Figs . 13. 4 an d 13.5 ) tha t th e expecte d utilit y o f a pla n wa s define d a s
`the su m o f th e utilit y o f eac h lea f nod e time s th e probabilit y o f reachin g tha t node .
`However, th e utilitie s nee d no t combin e linearl y i n scoring . Differen t monotoni c
`functions o f utilit y expres s suc h differen t conception s a s "aversio n t o risk " o r
`"gambling addiction. " Thes e consideration s ar e rea l ones , an d nonlinea r utilitie s
`are th e rul e rathe r tha n th e exception . Fo r instance , th e valu e o f mone y i s notori
`ously nonlinear . Man y peopl e woul d pa y $ 5 fo r a n eve n chanc e t o wi n $15 ; no t s o
`many peopl e woul d pa y $5,00 0 fo r a n eve n chanc e t o wi n $15,000 .
`
`Sec. 13.2 Planning with Costs
`
`45 1
`
`IPR2022-00092 - LGE
`Ex. 1015 - Page 463
`
`
`
`One commo n wa y t o comput e score s base d o n utilitie s i s th e "cost/benefit "
`ratio. This , i n th e for m "cost/confidence " ratio , i s use d b y Garve y i n hi s plannin g
`vision system . Thi s measur e i s examine d i n Sectio n 13.2.5 ; roughly , hi s "cost "
`was th e effor t i n machin e cycle s t o achiev e goals , an d hi s "confidence " approxi
`mated th e probabilit y o f a goa l achievin g th e correc t outcome . Th e utilit y o f correc t
`outcomes wa s no t explicitl y encode d i n hi s planner .
`Sequential pla n elaboratio n o r partia l pla n elaboratio n ca n b e interleave d wit h
`execution an d scoring . Mos t practica l plannin g i s don e i n interactio n wit h th e
`world, an d th e pla n scorin g approac h lend s itsel f wel l t o assessin g suc h interac
`tions. I n Sectio n 13.2. 5 consider s a plannin g visio n syste m tha t use s enhance d
`plans an d a limite d replannin g capability .
`A thorn y proble m fo r decisio n makin g i s t o asses s th e cos t o f plannin g itself .
`The plannin g proces s i s give n it s ow n utilit y (cost) , an d i s carrie d onl y ou t a s fa r a s
`is indicated . O f course , th e proble m i s i n genera l infinitel y recursive , sinc e ther e i s
`also th e cos t o f assessin g th e cos t