Pagi World from RPI Licato and Bringsjord

26
PAGI World: A simula1on Environment to Challenge Cogni1ve Architectures John Licato Selmer Bringsjord Rensselaer AI and Reasoning (RAIR) Lab

Transcript of Pagi World from RPI Licato and Bringsjord

PAGI  World:  A  simula1on  Environment  to  Challenge  Cogni1ve  Architectures  

John  Licato  Selmer  Bringsjord  

Rensselaer  AI  and  Reasoning  (RAIR)  Lab  

Last  week  •  John  Laird  talked  

about  “interac1ve  task  learning”  

•  Today,  we  will  present  a  simulator  to  facilitate  such  research  

Developmental  AI  –  Emerging  field  aOemp1ng  to  show  how,  using  an  agent  endowed  with  minimal  innate  capaci1es  embedded  in  a  sufficiently  

rich  environment,  higher-­‐level  cogni1ve  abili1es  can  emerge.  

What  makes  an  environment  sufficiently  rich?  

Guerin,  Frank  (2011).  Learning  like  a  baby:  A  survey  of  ar1ficial  intelligence  approaches.  The  Knowledge  Engineering  Review,  26(2),  209-­‐236.  

Guerin  (2011)’s  requirements  

A  sufficiently  rich  environment…  C1  –  is  rich  enough  to  provide  knowledge  which  would  bootstrap  the  understandings  of  concepts  rooted  in  physical  rela1onships,  e.g.:  inside  vs.  outside,  large  vs.  small,  above  vs.  below  C2  –  can  allow  for  the  modeling  and  acquisi1on  of  spa1al  knowledge  (widely  regarded  to  be  a  founda1onal  domain  of  knowledge  acquisi1on)  through  interac1on  with  the  world.  C3  –  can  support  the  crea1on  and  maintenance  of  knowledge  which  the  agent  can  verify  itself.  

Our  addi1onal  requirements  A  sufficiently  rich  environment…  C4  –  rich  enough  to  provide  much  of  the  sensory-­‐level  informa1on  accessible  to  a  real-­‐world  agent.  C5  –  allows  for  tes1ng  of  a  virtually  unlimited  number  of  tasks,  whether  they  test  low-­‐level  implicit  knowledge,  high-­‐level  explicit  knowledge,  or  any  of  the  other  areas  required  by  PAGI,  ideally  allowing  for  the  crea1on  of  new  tasks  without  substan1al  programming  effort.  C6  –  allows  a  wide  variety  of  AI  systems  based  on  vastly  different  theore1cal  approaches  to  aOempt  the  same  tasks,  thus  enabling  these  different  approaches  to  be  directly  compared.  CT  –  can  support  tasks  capable  of  verifying  AI  able  to  pass  the  Tailorability  Concern  

Tailorability  Concern  –  that  [cogni1ve  systems]  deal  almost  exclusively  with  manually  constructed  knowledge  

representa1ons,  using  toy  examples  and  source  knowledge  ocen  selected  solely  to  

display  some  par1cular  ability.  

Gentner,  Dedre  &  Forbus,  Ken  (2011).  Computa1onal  models  of  analogy.  Wiley  Interdisciplinary  Reviews:  Cogni>ve  Science,  2(3),  266-­‐276.  

Licato,  J.,  Bringsjord,  S.,  &  Govindarajulu,  N.S.  (2014).  How  models  of  crea1vity  and  analogy  need  to  answer  the  tailorability  concern.  In  Besold,  T.R.,  Kühnberger,  K.-­‐u.,  Schorlemmer,  M.,  &  Smaill,  A.  (Eds.),  

Computa>onal  Crea>vity  Research  :  Towards  Crea>ve  Machines.  Atlan1s  Press.  

Drescher  (1991):  A  star1ng  point  

•  Cell-­‐based  world  •  Simple  agent  which  occupied  one  cell  

•  Agent  had  a  “hand”  which  could  grasp  objects  in  the  world  

•  Visual  field  rela1ve  to  the  agent’s  “body”  

Drescher,  Gary  L.  (1991).  Made-­‐Up  Minds:  A  Construc>vist  Approach  to  Ar>ficial  Intelligence.  The  MIT  Press.  

Drescher  (1991):  A  star1ng  point  

•  Used  to  show  Piage1an  (construc1vist)  boOom-­‐up  crea1on  of  knowledge  

•  Simula1on  environment  was  1ghtly  coupled  with  his  schema  mechanism  

•  No  realis1c  mo1on  or  physics  

•  World  did  not  provide  rich  source  analogs  for  e.g.  inside  vs.  outside  

Drescher,  Gary  L.  (1991).  Made-­‐Up  Minds:  A  Construc>vist  Approach  to  Ar>ficial  Intelligence.  The  MIT  Press.  

C5  –  allows  for  tes1ng  of  a  virtually  unlimited  number  of  tasks,  whether  they  test  low-­‐level  implicit  knowledge,  high-­‐level  explicit  knowledge,  or  any  of  the  other  areas  required  by  PAGI,  ideally  allowing  for  the  crea1on  of  new  tasks  without  substan1al  programming  effort.  

Controlled by PAGI-side

!!!!!!!!!!

Reflex and State

Machine

Controlled by AI-side

!!!!!

TCP/ IP

pyPAGI (optional) !!!DCEC*

extractor/convertor

Physics Engine

Task Editor

Configurable by external user

Can  be  wriOen  in  almost  any  language!  

PAGI  World  can  be  run  on:  Windows  Mac  OS  Linux  (through  Chrome  browser)  Android  and  Iphone  (in  theory)    

AI  can  be  wriCen  in:  ANY  programming  language  which  supports  TCP/IP  

C1  –  is  rich  enough  to  provide  knowledge  which  would  bootstrap  the  understandings  of  concepts  rooted  in  physical  rela1onships,  e.g.:  inside  vs.  outside,  large  vs.  strong  C2  –  can  allow  for  the  modeling  and  acquisi1on  of  spa1al  knowledge  (widely  regarded  to  be  a  founda1onal  domain  of  knowledge  acquisi1on)  through  interac1on  with  the  world.  

C3  –  can  support  the  crea1on  and  maintenance  of  knowledge  which  the  agent  can  verify  itself.  

Warning:  DCEC*  is  a  highly  expressive  computa1onal  logic  and  therefore  the  cogni1on  which  it  enables  may  or  may  not  be  within  reach  of  a  given  cogni1ve  architecture.  

 But  PAGI  World  allows  us  to  test  and  see!  

C4  –  rich  enough  to  provide  much  of  the  sensory-­‐level  informa1on  accessible  to  a  real-­‐world  agent.    C6  –  allows  a  wide  variety  of  AI  systems  based  on  vastly  different  theore1cal  approaches  to  aOempt  the  same  tasks,  thus  enabling  these  different  approaches  to  be  directly  compared.  

Controlled by PAGI-side

!!!!!!!!!!

Reflex and State

Machine

Controlled by AI-side

!!!!!

TCP/ IP

pyPAGI (optional) !!!DCEC*

extractor/convertor

Physics Engine

Task Editor

Configurable by external user

PAGI  World  allows  super  rapid  demonstra1ons  of  

cogni1ve  abili1es  

“The%Brilliant%Boardroom”:%Cogni4ve%Compu4ng%with%the%DCEC*%and%ADR%

John%Licato%%*%%Selmer%Bringsjord%Konner&Atkin&*&Maggie&Borkowski&*&Jack&Cusick&*&Kainoa&Eastlack&*&Nick&Marton&*&James&Pane;Joyce&*&Spencer&Whitehead&

Abstract%This%poster%reports%on%research%and%development%done%by%the%Rensselaer%AI%and%Reasoning%(RAIR)%Lab’s%team,%in%collabora4on%with%IBM,%on%crea4ng%framework%technologies%that%can%be%used%in%many%areas%of%cogni4ve%compu4ng.%We%here%focus%on%one%such%areaMMMthe%Brilliant%Boardroom%(BB),%in%which%a%robot%or%set%of%robots,%augmented%with%mul4modal%inputs%such%as%speech%recogni4on,%synthesis,%and%basic%vision%processing,%react%and%produc4vely%add%to%a%mee4ng%of%corporate%execu4ves.%We%infuse%the%Brilliant%Boardroom%with%two%RAIRMlabMdeveloped%technologies:%the%Deon4c%Cogni4ve%Event%Calculus%(DCEC*),%a%highly%expressive%computa4onal%framework%intended%to%formally%model%and%mechanize%humanMlevel%reasoning,%decisionMmaking,%problemMsolving,%and%natural%language%communica4on;%and%AnalogicoMDeduc4ve%Reasoning%(ADR),%a%type%of%reasoning%which%is%central%to%higher%level%humanMlike%cogni4on.%

Syntax

S ::=Object | Agent | Self @ Agent | ActionType | Action v Event |

Moment | Boolean | Fluent | Numeric

f ::=

action : Agent⇥ActionType ! Action

initially : Fluent ! Boolean

holds : Fluent⇥Moment ! Boolean

happens : Event⇥Moment ! Boolean

clipped : Moment⇥Fluent⇥Moment ! Boolean

initiates : Event⇥Fluent⇥Moment ! Boolean

terminates : Event⇥Fluent⇥Moment ! Boolean

prior : Moment⇥Moment ! Boolean

interval : Moment⇥Boolean

⇤ : Agent ! Self

payoff : Agent⇥ActionType⇥Moment ! Numeric

t ::= x : S | c : S | f (t1 , . . . , tn)

f ::=

t : Boolean | ¬f | f^y | f_y | 8x : S. f | 9x : S. f

P(a, t,f) | K(a, t,f) | C(t,f) | S(a,b, t,f) | S(a, t,f)

B(a, t,f) | D(a, t,holds( f , t0)) | I(a, t,happens(action(a⇤ ,a), t0))

O(a, t,f,happens(action(a⇤ ,a), t0))

Rules of Inference

C(t,P(a, t,f)! K(a, t,f))[R1 ] C(t,K(a, t,f)! B(a, t,f))

[R2 ]

C(t,f) t t1 . . . t t

n

K(a1 , t1 , . . .K(an

, tn

,f) . . .)[R3 ]

K(a, t,f)

f[R4 ]

C(t,K(a, t1 ,f1 ! f2)! K(a, t2 ,f1)! K(a, t3 ,f3))[R5 ]

C(t,B(a, t1 ,f1 ! f2)! B(a, t2 ,f1)! B(a, t3 ,f3))[R6 ]

C(t,C(t1 ,f1 ! f2)! C(t2 ,f1)! C(t3 ,f3))[R7 ]

C(t,8x. f ! f[x 7! t])[R8 ] C(t,f1 $ f2 ! ¬f2 ! ¬f1)

[R9 ]

C(t, [f1 ^ . . .^fn

! f]! [f1 ! . . .! fn

! y])[R10 ]

B(a, t,f) B(a, t,f ! y)

B(a, t,y)[R11a

]B(a, t,f) B(a, t,y)

B(a, t,y^f)[R11b

]

S(s,h, t,f)

B(h, t,B(s, t,f))[R12 ]

I(a, t,happens(action(a⇤ ,a), t0))

P(a, t,happens(action(a⇤ ,a), t))[R13 ]

B(a, t,f) B(a, t,O(a⇤ , t,f,happens(action(a⇤ ,a), t0)))

O(a, t,f,happens(action(a⇤ ,a), t0))

K(a, t,I(a⇤ , t,happens(action(a⇤ ,a), t0)))[R14 ]

f $ y

O(a, t,f,g)$ O(a, t,y,g)[R15 ]

1

DCEC*:%The%Deon4c%Cogni4ve%Event%Calculus%

%The%DCEC*,%pictured%in%Figure%1,%is%a%highly%expressive%framework%that%has%been%used%for%the%mechaniza4on%of%humanMlevel%reasoning,%automated%decisionMmaking,%natural%language%parsing%and%genera4on,%and%many%other%applica4ons.%Because%it%allows%ar4ficial%agents%to%represent%arbitrarily%nested%beliefs%and%knowledge%(e.g.%that%the%execu4ve%in%chair%1%believes%that%the%execu4ve%2%in%chair%believes%that%the%execu4ve%in%chair%1%is%lying),%it%can%perform%reasoning%far%beyond%that%of%many%other%formalisms%proposed%to%represent%commonsense%knowledge.%This%sort%of%ability%is%extremely%important%in%situa4ons%where%an%ar4ficial%agent%is%asked%to%exist%in%a%complex%social%environment,%much%less%one%that%may%require%the%agent%to%provide%jus4fica4ons%for%its%conclusions%(as%the%robo4c%agent%in%our%demonstra4on%was%made%to%do).%%The%DCEC*%also%lends%itself%to%social%environments%because%of%its%inherent%capturing%of%deon4c%no4ons.%It%has%operators%such%as%O%(for%obliga4on),%which%is%treated%carefully%by%a%set%of%inference%rules%(see%Figure%1),%themselves%chosen%to%help%ensure%that%commonsense%no4ons%of%what%it%means%to%be%obliged%to%do%something%can%be%captured%through%straighWorward%applica4ons%of%deduc4ve%reasoning.%These%inference%rules%are%constantly%being%refined%and%explored%through%RAIR%lab%R&D.%%Of%course,%deduc4ve%reasoning%alone%may%be%insufficient%to%capture%the%sort%of%reasoning%expected%by%an%ar4ficial%agent%in%a%boardroom;%therefore%we%augment%our%system%with%ADR,%which%is%another%major%research%focus%of%our%lab.%

ADR:%AnalogicoMDeduc4ve%Reasoning%%Although%analogical%and%deduc4ve%reasoning%can%interact%in%a%myriad%of%different%combina4ons,%the%par4cular%intersec4on%between%hypothe4coMdeduc4ve%and%analogical%reasoning,%which%we%call%ADR,%has%been%shown%to%be%par4cularly%useful%to%human%reasoners%from%young%children%performing%Piage4an%experiments%to%groundbreaking%mathema4cal%logicians%like%Gödel.%In%its%simplest%form,%ADR%involves%using%analogical%processes%to%select%poten4ally%relevant%source%analogs,%match%them%to%the%target%domain,%and%produce%hypotheses%about%the%target%domain.%However,%because%these%hypotheses%are%prone%to%error,%deduc4ve%reasoning%is%invoked%to%verify,%support,%or%refute%the%hypotheses%before%they%are%incorporated%into%a%knowledge%base.%%%In%our%demonstra4on,%the%BB%(personified%by%the%Aldebaran%NAO%Bot%pictured%in%Figure%2)%u4lized%ADR%to%answer%a%ques4on%about%how%one%of%the%boardroom%mee4ng’s%par4cipants%might%get%access%to%some%sales%dataMMMthe%correct%answer%was%to%ask%Mr.%Smith,%which%is%knowledge%that%the%robot%did%not%previously%have.%It%inferred%this%by%drawing%an%analogy%to%a%previous%instance%in%which%a%mee4ng%par4cipant%obtained%similar%sales%data%by%asking%Mr.%Johnson,%who%at%the%4me%held%the%office%currently%held%by%Mr.%Smith.%The%deduc4ve%step%did%not%find%any%contradic4ons,%and%so%the%robot%reported%its%findings.%

Rensselaer*AI*and*Reasoning*(RAIR)*Lab*Rensselaer*Polytechnic*Ins9tute,*Troy,*NY*

Conclusion%/%Future%Work%%The%coming%of%Cogni4ve%Compu4ng%raises%many%interes4ng%ques4ons%about%what%it%means%to%be%cogni4ve%in%the%first%place.%But%we%must%also%ask%what%we%want%our%ar4ficial%cogni4ve%companions%to%do,%even%when%those%things%may%not%be%cogni4vely%plausible.%Here%we%will%see%at%least%two%concerns:%First,%that%cogni4ve%agents%should%be%able%to%reason%ethically;%and%second,%that%these%agents%should%be%able%to%provide%jus4fica4ons%for%their%ac4ons%(in%part%to%ensure%that%the%first%concern%is%met).%Again,%the%DCEC*%and%ADR%offer%results%in%this%direc4on.%Although%it%may%turn%out%that%this%pair%of%technologies%is%not%all%that%is%needed%to%ensure%that%our%cogni4ve%companions%behave%correctly,%they%represent%a%line%of%research%that%takes%the%concerns%we%have%raised%here%seriously%and%cons4tute%a%larger%effort%that%con4nues%to%be%a%focus%of%RAIR%lab%R&D.%

Figure%1.%The%Deon4c%Cogni4ve%Event%Calculus%(DCEC*).%

Figure%2.%The%robot%used%as%the%personifica4on%of%the%BB.%

RPI%Sugges4on%and%Jus4fica4on%Service%

User%ADR%

Module% Local%KB%

DCEC*%Reasoner%

DBPedia%

Figure%3.%DCEC*%and%ADR%was%recently%used%in%a%demonstra4on%of%another%service,%hosted%in%RPI’s%“red%zone,”%accessed%from%services%hosted%on%IBM’s%“blue%zone.”%%

Theorem 3: There is a way to satisfy both obligations.

From the Licato presentation in IBM’s Cognitive Systems Institute Lecture Series: “PAGI World: A Simulation Environment to Challenge Cognitive Architectures". For more information, visit https://www.linkedin.com/groups/Cognitive-Systems-Institute-6729452

PAGI  World  is  a  challenge  to  AI  and  cogni1ve  architecture  researchers  

Let’s  create  tasks,  AI  systems  to  solve  them,  

compare  the  approaches,  and  repeat-­‐-­‐-­‐and  keep  this  field  moving  forward!  

How  to  access  PAGI  World  (beta  version)    

Email  John  Licato  at  [email protected]