systems with general intelligence a new perspective michael...
TRANSCRIPT
![Page 1: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/1.jpg)
Systems with General Intelligence —
A New Perspective
Michael Thielscher
![Page 2: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/2.jpg)
Outline
PART I
A Grand AI Challenge General game playing
Defining your own Grand AI Challenge Systems with general intelligence
PART II
A new research agenda Combining representations, methods, systems
![Page 3: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/3.jpg)
How Intelligent are AI Systems?
The intelligence lies with the programmers—not their systems
Most existing AI systems aredesigned for a specific and narrow applicationuse tailormade algorithms
Do they, really?
AI systems are able tomake autonomous decisionsadapt flexibly to unforeseen situations
![Page 4: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/4.jpg)
Example: Chess Computers
Turk(Vienna 1770)
In the early days, chess playing was considered a key to AI
![Page 5: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/5.jpg)
Example: Chess Computers
Secret revealed(1857)
![Page 6: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/6.jpg)
Chess computers reach human level
Example: Chess Computers
Deep Blue(New York 1997)
![Page 7: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/7.jpg)
Deep Blue was a success story. But also a major leap for AI?
Chess computers are highly specialised systems
Deep Blue can't handle anything outside its 64square world
Example: Chess Computers
No:
Deep Blue's capabilities were just not general enough
![Page 8: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/8.jpg)
A Grand AI Challenge: General Game Playing
A General Game Player is a system that
understands description of arbitrary games
learns to play these games without human intervention
General Game Playing Contest @AAAI since 2005
![Page 9: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/9.jpg)
How it Works
Game Master
Player1
Player2
Playern
...
Game descriptionTime to think: 1,800 secTime per move: 45 secYour role
![Page 10: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/10.jpg)
How it Works
Game Master
Player1
Player2
Playern
...
Start
![Page 11: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/11.jpg)
How it Works
Game Master
Player1
Player2
Playern
...
Your move, please
![Page 12: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/12.jpg)
How it Works
Game Master
Player1
Player2
Playern
...
Individual moves
![Page 13: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/13.jpg)
How it Works
Game Master
Player1
Player2
Playern
...
Individual informationabout state/moves
![Page 14: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/14.jpg)
How it Works
Game Master
Player1
Player2
Playern
...
End of game
![Page 15: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/15.jpg)
Game Descriptions
Games are described by logic programs using a fewpredefined keywords
role(jane).role(rick).role(random).card(♣7). card(♣8). ... card(♣ace).
init(dealingRound).
![Page 16: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/16.jpg)
Game Descriptions (Cont'd)
legal(random,deal(C,D)) <= true(dealingRound), card(C), card(D), distinct(C,D).
sees(jane,yourCard(C)) <= does(random,deal(C,D)).sees(rick,yourCard(D)) <= does(random,deal(C,D)).
legal(jane,...) <= ...legal(rick,...) <= ...
terminal <= ...goal(P,N) <= ...
![Page 17: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/17.jpg)
Example 1
AAAI 2007
![Page 18: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/18.jpg)
Example 2
AAAI 2010
![Page 19: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/19.jpg)
History
1968 J. Pitrat: “Realization of a General Game Playing Program”
2005 First GGP Competition @AAAI
2009 First GGP Workshop @IJCAI
2010 First Technical Paper Session on GGP @AAAI
A Vibrant Reserch Area
Online repositories
games.stanford.edu (description language, competition)
generalgameplaying.de (game server, basic players, literature)
Research centers
Dresden, Edmonton, Paris, Potsdam, Reykjavik, Stanford, Sydney, ...
![Page 20: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/20.jpg)
Two Questions
Can a general game player beat Deep Blue in chess?
➔ No (but may change in the future)➔ Focus is on general players, not savants➔ There is a market for a chess computer that is weaker but can adapt to any chess variant without being reprogrammed
➔ Yes, but will change in the future
Isn't a general game player still a very special system?
![Page 21: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/21.jpg)
Some Ideas for General General Game Playing
Natural Language
➔ Systems understand game rules in (controlled) English
Vision
➔ Camera system identifies
Robotics
➔ Robotic manipulation of
new boards and pieces
(Purdue University 2010) game hardware
![Page 22: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/22.jpg)
A Continuous Scale
GeneralChess Computer
GeneralGame Player
GeneralGame Robot
![Page 23: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/23.jpg)
From Gen
eral Game Playing to General X
Systems with general intelligence
understand descriptions of new environments and tasks
adapt to these environments/tasks without human intervention
How to create your own General AI Challenge:
Define a broad—but sufficiently restricted—problem class X
Design a suitable communication/description language for X
From General Game Playing to General X
The idea behind General Game Playing can be applied to other areas, bringing today's AI systems to a new level of generality
![Page 24: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/24.jpg)
Two Random Ideas
General Trading Agents
understand new trading scenarios
trade without human intervention
General Robots
understand new tasks
adapt without human intervention
![Page 25: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/25.jpg)
Part II:
Addressing a General AI Challenge
![Page 26: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/26.jpg)
“Silver bullets” have been proposed throughout the history, eg
GOFAI (1960's)
Subsymbolic AI (1980's)
Bayesian AI (1990's)
but:
different problems may require different representations
different tasks may require different computations
A Brief History of AI
![Page 27: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/27.jpg)
AI Today
AI
Subsymbolic AISymbolic AI
Agents
BDI
SitCalc
NMR DL
Action Logics Planning
Event Calculus Fluent Calculus
......
... NLP KR UAI ...
Individual theories cater for individual aspects of intelligence
![Page 28: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/28.jpg)
Today, there exist a variety of
wellunderstood approaches—for many individual aspects of AI
highly optimized algorithmic solutions—to many specific problems
Specialization: Pro
Focusing on a single, narrow AI problem allows to
use a tailormade representation
gain a deeper understanding of the fundamental and computational issues related to this particular aspect of AI
![Page 29: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/29.jpg)
➔ Challenge 1: combine different representations➔ Challenge 2: integrate different implementations
Specialization: Cons
There is a danger to fiddle with minor details
AI Challenges require to address a range of aspects together
![Page 30: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/30.jpg)
Systems with General Intelligence
Programs or robots with general intelligence (GI) must exhibit many facets of intelligence
need to integrate successful AI methods
BottomUp
Choose and combinerepresentation formalismsalgorithmic solutionsimplementations
TopDown
Take welldefined GI challengeidentify subtaskschoose methods to combinebuild integrated system
![Page 31: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/31.jpg)
TopDown Combinations (Example)
—
FLUXPLAYER
![Page 32: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/32.jpg)
General Game Playing Systems
A General Game Player requires methods from
Knowledge Representation and Reasoning
Planning and Search
Computer Game Playing
Learning
![Page 33: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/33.jpg)
Our General Game Player FLUXPLAYER combines
Reasoning about Actions (“FLUX”, to understand the game rules)
Planning and Search
Automated Theorem Proving (to generate knowledge about a game)
Fuzzy Logic (to evaluate intermediate positions)
Neural Nets (to improve parameter settings of evaluation functions)
FLUXPLAYER's performance in all previous GGP Championships
AAAI: 2005 Semifinal, 2006 Winner, 2007 Second, 2008 Semifinal
IJCAI: 2009 Second
FLUXPLAYER
![Page 34: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/34.jpg)
Two examples of research output from this Grand Challenge
Answer Set Programming for verification of dynamic systems
Combining Neural Networks with Symbolic Logic (Michulke & T, ECML 2009)
(Schiffel & T, IJCAI 2009; T & Voigt, AAAI 2010)
FLUXPLAYER
![Page 35: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/35.jpg)
BottomUp Combination: Example
—
BDIBased Agent Programs&
Action Logics
![Page 36: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/36.jpg)
AI
Subsymbolic AISymbolic AI
Agents
SitCalc
NMR DL
Action Logics Planning
Event Calculus Fluent Calculus
......
... NLP KR Bayesian ...
Combining Formalisms
BDI
![Page 37: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/37.jpg)
Action Logics
since late 1960'stheory of cognitive agents
BDIbased Programming
since early 1990'sto build cognitive agents
Two Distinct Areas with a Similar Goal
![Page 38: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/38.jpg)
Action Logics
+ rich action model– barely used in practice
BDIbased Programming
+ practical programming– simplistic action model
Similar Goal—Different Strengths
![Page 39: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/39.jpg)
Action Logics
+ rich action model– barely used in practice
BDIbased Programming
+ practical programming– simplistic action model
Why Combine the Two?
![Page 40: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/40.jpg)
Need to Align Representations
Agent programs are collections of reactive behaviors
+!capture(X) : have(X) | !nextto(X); get(X); !at(home)
Action knowledge is given in form of logical formulas
poss(get(X),S) holds(nextto(X),S)
holds(have(X),do(A,S)) A = get(X) ∨ holds(have(X),S)
Main issue: two methods based on different representations
Reactive programs come with operational semantics, based on the (Beliefs, Desires, Intentions)model of agents
Action theories have declarative semantics, based on logic
![Page 41: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/41.jpg)
Solution
A bridging language helps aligning the two representations
Agent Logic Programs
Resulting integration
➔ extend logic programs (Prolog) by actions➔ come with an operational semantics➔ and with a declarative semantics
➔ provides declarative semantics for BDIbased languages➔ provides formal underpinnings for combining implementations➔ is correct—provided 8(!) assumptions and conditions are met
(MT, KR 2010)
![Page 42: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/42.jpg)
Conclusion
![Page 43: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/43.jpg)
First Demonstration of AI
Turk(Vienna 1770)
![Page 44: Systems with General Intelligence A New Perspective Michael …cgi.cse.unsw.edu.au/~mit/Slides/AAAI10-Invited-Slides.pdf · Example: Chess Computers Turk (Vienna 1770) In the early](https://reader033.vdocuments.site/reader033/viewer/2022060703/606feb34eb215366872ea85a/html5/thumbnails/44.jpg)
Future Demonstrations of AI
To do so,the technology is out therebut combining AI methods can be a challenge of its own
When built, these systemsprovide impressive demonstrations of AI's potentiallift a specific AI field to a new level
Systems with general intelligenceunderstand descriptions of radically new environments/tasksadapt to these environments/tasks without human intervention