Human Level Sli
-
Upload
vivek-parmar -
Category
Documents
-
view
219 -
download
0
Transcript of Human Level Sli
-
7/28/2019 Human Level Sli
1/42
John McCarthy
http://www-formal.stanford.edu/jmc
2005 November 2
THE LOGICAL ROAD TO HUMAN LEV
Will we ever reach human level AIthe main am
AI research?
Sure. Understanding intelligence is a difficult sciebut lots of difficult scientific problems have been s
nothing humans can do that humans cant make
We, or our descendants, will have smart robot se
AI research should use AI Drosophilas, domains
informative about mechanisms of intelligence, not
-
7/28/2019 Human Level Sli
2/42
Who proposed human-level AI as goaloutside o
Alan Turing was probably firstin 1947, but all th
in AI took human level as the goal. AI as an indust
with limited goals came along in the 1970s. I do
of this research aimed at short term payoff is o
human-level AI. Indeed the researchers dont claim
Is there a Moores law for AI? Ray Kurzweil seperformance doubles every two years.
No.
-
7/28/2019 Human Level Sli
3/42
When will we get human-level AI?
Maybe 5 years. Maybe 500 years.
Will more of the same do it? The next factor of
puter speed. More axioms in CYC of the same
neural nets?
No.
Most AI research today is aimed at short term pay
conceptually difficult problems.
-
7/28/2019 Human Level Sli
4/42
Most likely we need fundamental new ideas. Mor
the ideas now being pursued by hundreds of resea
limited in scope by the remnants of behaviorist
philosophywhat Steven Pinker calls the blank you my ideas, but most likely they are not enough
My article Philosophical and scientific presupposit
AI, http://www.formal.stanford.edu/jmc/phil2.ht
explains what human-level AI needs in the way of
-
7/28/2019 Human Level Sli
5/42
REQUIREMENTS FOR HUMAN-LEVE
An ontology adequate for stating the effects
amples include situations, fluents, actions and oth
functions giving the new situations that result fro
can be told facts e.g. the LCDs in a laptop ar
glass. (stated absolutely but in an implicit contex
knowledge of the common sense worldfacts
3-d flexible objects, appearance including feel and
fects of actions and other events.extendable to
-
7/28/2019 Human Level Sli
6/42
the agent as one among many It knows abou
and their likes, goals, and fears. It knows how its a
with those of other agents.
independence A human-level agent must not be d
human to revise its concepts in face of experience,
or new information. It must be at least as capabl
reasoning about its own mental state and mental
elaboration tolerance The agent must be able
account new information without having to be re
person.
-
7/28/2019 Human Level Sli
7/42
relation between appearance and reality betwe
and their 2-d projections and also with the sensa
ing them. Relation between the course of events
observe and do.
self-awareness The agent must regard itself as
as an agent and must be able to observe its own
connects reactive and deliberated action e.g
removing ones keys from a pocket.
counterfactual reasoning If another car had com
when you passed, there would have been a head
-
7/28/2019 Human Level Sli
8/42
If the cop believes it, youll be charged with re
McCarthy and Costello on useful counterfactuals
reasons with ill-defined entitiesthe purposethe welfare of a chicken, the rocks of Mount Ev
that might have come over the hill.
These requirements are independent of whether th
based or an imitation of biology, e.g. a neural net
-
7/28/2019 Human Level Sli
9/42
APPROACHES TO AI
biologicalimitate human, e.g. neural nets, shou
tually, but theyll have to take a more general app
engineeringstudy problems the world presents, s
direct programming, genetic programming.
use logic and logical reasoning The logic approaawkwardexcept for all the others that have bee
the work with fmri makes it look like the logical
approaches may soon usefully interact.
-
7/28/2019 Human Level Sli
10/42
WHY THE LOGIC ROAD?
If the logic road reaches human-level AI, we will ha
understanding of how to represent the informatioable to achieve goals. A learning or evolutionary
achieve the human-level performance without the u
Leibniz, Boole and Frege all wanted to form
sense. This requires methods beyond what worke
mathematicsfirst of all formalizing nonmonoton
Since 1958: McCarthy, Green, Nilsson, Fikes, Re
Bacchus, Sandewall, Hayes, Lifschitz, Lin, Kow
-
7/28/2019 Human Level Sli
11/42
Perlis, Kraus, Costello, Parmar, Amir, Morgenste
Doherty, Ginsberg, McIlraith . . . and others I ha
Express facts about the world, including effects other events.
Reason about ill-defined entities, e.g. the welfa
Thus formulas like
Welfare(x, Result(Kill(x), s)) < W elfare(x, s) are s
even though Welfare(x, s) is often indeterminate.
-
7/28/2019 Human Level Sli
12/42
LOGIC
Describes how people thinkor how people think
The laws of deductive thought. (Boole, de M
Peirce). First order logic is complete and perhaps
Present mathematical logic doesnt cover all good
does cover all guaranteed correct reasoning.
More general correct reasoning must extend logic
monotonic reasoning and probably more. Some g
monotonic reasoning is not guaranteed to always p
conclusions.
-
7/28/2019 Human Level Sli
13/42
COMMON SENSE IN LOGICAL LANGUAGES
For every boy, theres a girl who loves only him
(b)(g)(Loves(g, b) (!b)Loves(g, b))This uses different sorts for boys and girls. There is
logical way of saying loves only him.
Block A is on Block B.
Variants: On(A, B), On(A , B , s), Holds(On(A, B), s
T op(B), V alue(Location(A), s) = V alue(T op(B), s).
Pat knows Mikes telephone number.
Knows(P at, T T elephone(MMike))
-
7/28/2019 Human Level Sli
14/42
THE COMMON SENSE INFORMATIC SIT
The common sense informatic situation is the key t
AI.
I have only partial information about myself and my
I dont even have a final set of concepts.
Objects of perception and thought are only partly often only approximately defined.
What I think I know is subject to change and elab
-
7/28/2019 Human Level Sli
15/42
There is no bound on what might be relevant. T
drosophila illustrates this common sense physics.
eter to find the height of a building.]
Sometimes we (or better it) can connect a bound
situation to an open informatic situation. Thus
blocks world can be used to control a robot stacki
A human-level reasoner must often do nonmonoto
Nevertheless, human reasoning is often very effec
Im in a world in which Im a product of evolution
-
7/28/2019 Human Level Sli
16/42
THE COMMON SENSE INFORMATIC SITUAT
The world in which common sense operates has
aspects.
1. Situations are snapshots of part of the world.
2. Events occur in time creating new situations. Aare events.
3. Agents have purposes they attempt to realize
-
7/28/2019 Human Level Sli
17/42
4. Processes are structures of events and situati
5. 3-dimensional space and objects occupy regioagents, e.g. people and physical robots are ob
can move, have mass, can come apart or com
larger objects.
6. Knowledge of the above can only be approxim
7. The csis includes mathematics, i.e. abstract
their correspondence with structures in the re
-
7/28/2019 Human Level Sli
18/42
8. Common sense can come to include facts disc
ence. Examples are conservation of mass and
of volume of a liquid.
9. Scientific information and theories are imbedd
sense information, and common sense is need
ence.
-
7/28/2019 Human Level Sli
19/42
BACKGROUND IDEAS
epistemology (what an agent can know aboutgeneral and in particular situations)
heuristics (how to use information to achieve
declarative and procedural information
situations
-
7/28/2019 Human Level Sli
20/42
SITUATION CALCULUS
Situation calculus is a formalism dating from 1964
ing the effects of actions and other events.
My current ideas are in Actions and other events in
culus - KR2002, available as www-formal.stanford.
They differ from those of Ray Reiters 2001 bo
however, been extended to the programming lang
Clear(x) Clear(l) At(x, l, Result(Move(xAt(y, l1) y = x At(y, l1, Result(Move(x,
-
7/28/2019 Human Level Sli
21/42
Going from frame axioms to explanation closure ax
oration tolerance. The new formalism is just as co
based on explanation closure but, like systems u
ioms, is additively elaboration tolerant.
The frame, qualification and ramification problem
and significantly solved in situation calculus.
There are extensions of situation calculus to conccontinuous events and actions, but the formalism
entirely satisfactory.
-
7/28/2019 Human Level Sli
22/42
CONCURRENCY AND PARALLELIS
In time. Drosophila = Junior in Europe and
york. When concurrent activities dont interactcalculus description of the joined activities nee
junction of the descriptions of the separate acthe joint theory is a conservative extension otheories. Temporal concurrency is partly done
In space. A situation is analyzed as compostions that are analyzed separately and then (iinteraction. Drosophilas are Go and the geLemmings game. Spatial parallelism is hardly
-
7/28/2019 Human Level Sli
23/42
INDIVIDUAL CONCEPTS AND PROPOS
In ordinary language concepts are objects. So be
CanSpeakW ith(p1, p2, Dials(p1, T elephone(p2), s))Knows(p1, T T elephone(pp2), s) Cank(p1, Dial(T
T elephone(Mike) = T elephone(Mary)TTelephone(MMike) = TTelephone(MMary)
Denot(MMike) = Mike Denot(MMary) = Mary(pp)(Denot(T elephone(pp)) = T elephone(Denot(ppKnows(P at, T T elephone(MMike))
Knows(P at, T T elephone(MMary))
-
7/28/2019 Human Level Sli
24/42
CONTEXTRelations among expressions evaluated in differen
C0 : V alue(T hisLecture, I) = JohnMcCarthy
C0 : Ist(U SLegalHistory, Occupation(Holmes) =C0 : Ist(U SLiteraryHistory, Occupation(HolmeC0 : F ather(V alue(U SLegalHistory, Holmes)) =V alue(U SLiteraryHistory, Holmes)
V alue(CAFdb,Price(GE610)) = V alue(CGEdb,Pri+V alue(CGEdb,Price(Spares(GE610)))
Can transcend outermost context, permitting intr
Here we use contexts as objects in a logical theory, an extension to logic. The approach hasnt beenbad.
-
7/28/2019 Human Level Sli
25/42
NONMONOTONIC REASONINGCIRCUMSP P (x . . . z)(P(x . . . z) P(x . . . z))P < P P P (P A)
Circm{E; C; P; Z} E(P, Z) (P
Z
)(E(P
, Z
)
In Circm{E; C; P; Z}, E is the axiom, C is a set oconstant, P is the predicate to be minimized, andpredicates that can be varied in minimizing P.
Ab(Aspect1(x)) flies(x)bird(x) Ab(Aspect1(x))bird(x) Ab(Aspect2(x)) flies(x)
penguin(x) Ab(Aspect2(x))penguin(x) Ab(Aspect3(x)) flies
-
7/28/2019 Human Level Sli
26/42
Let E be the conjunction of the above sentences.
Then Circum(E; {bird, penguin}; Ab; flies) implies
flies(x) bird(x) penguin(x), i.e. the things tha
birds that are not penguins.
frame, qualification and ramification problems
Conjecture: Simple abnormality theories arent en
(No matter what the language).
Inference to a bounded model
-
7/28/2019 Human Level Sli
27/42
SOME USES OF NONMONOTONIC REA1. As a communication convention. A bird may b
fly.
2. As a database convention. Flights not listed d
3. As a rule of conjecture. Only the known tools
4. As a representation of a policy. The meeting is
unless otherwise specified.
5. As a streamlined expression of probabilistic info
probabilities are near 0 or near 1. Ignore the risk o
by lightning.
-
7/28/2019 Human Level Sli
28/42
ELABORATION TOLERANCE
Drosophila = Missionaries and Cannibals: The sm
ary cannot be alone with the largest cannibal. Osionaries is Jesus Christ who can walk on water. T
that the river is too rough is 0.1.
Additive elaboration tolerance. Just add sentence
See www.formal.stanford.edu/jmc/elaboration.ht
Ambiguity tolerance
Drosophila = Law against conspiring to assault a f
-
7/28/2019 Human Level Sli
29/42
APPROXIMATE CONCEPTS AND THE
Reliable logical structures on quicksand semantic
Drosophila = {Mount Everest, welfare of a chicke
No truth value to many basic propositions.
Which rocks belong to the mountain?
Definite truth value to some compound propositio
concepts are squishy. Did Mallory and Irvine rea
Everest in 1924?
-
7/28/2019 Human Level Sli
30/42
HEURISTICS
Domain dependent heuristics for logical reasoning
Declarative expression of heuristics.
Wanted: General theory of special tricks
Goal: Programs that do no more search than hu
the 15 puzzle, Tom Costello and I got close. Sha
got closer.
-
7/28/2019 Human Level Sli
31/42
LEARNING AND DISCOVERY
Learning - what can be learned is limited by whatsented.Drosophila = chess
Creative solutions to problems.Drosophila = mutilated checkerboard
Declarative information about heuristics.
Domain dependent reasoning strategiesDrosophilas = {geometry, blocks world}
Strategy in 3-d world.Drosophila = Lemmings
-
7/28/2019 Human Level Sli
32/42
Learning classifications is a very limited kind of lea
Learn about reality from appearance, e.g 3-d re
appearance. See
www-formal.stanford.edu/jmc/appearance.html fo
zle.
Learn new concepts. Stephen Muggletons induc
gramming is a good start.
-
7/28/2019 Human Level Sli
33/42
ALL APPROACHES TO AI FACE SIMILAR P
Like humans AI systems must communicate in fac
grams or in objects. To communicate requires ve
edge of the mental state of the recipient.
Succeeding in the common sense informatic situ
elaboration tolerance.
It must infer reality from appearance.
Living with approximate concepts is essential
-
7/28/2019 Human Level Sli
34/42
Transcending outermost context, introspection.
Nonmonotonic reasoning
-
7/28/2019 Human Level Sli
35/42
INTUITIONS AND ARGUMENTS AGAINST
In 1975 Marvin Minsky argued that logic didnt h
tonic reasoning. Nonmonotonic extensions of logi
The connectionist argument of 1980: Logical AI h
human-level intelligence. Therefore, our way must
years have elapsed, and connectionism hasnt don
Your logical language cant express X. Hence
quate. Extend the language. Getting a univers
unsolvedrequires metamathematics in the langu
-
7/28/2019 Human Level Sli
36/42
People dont reason logically, e.g. Kahneman
examples. When people reason in opposition to
mistaken. Formal logic, starting with Aristotle
vented for communication among people and to imreasoning.
Present general first order logic programs do
on problems expressed in first order logic. Be
are neededincluding metamathematical reasonin
tirely on resolution was a mistake.
Godel showed incompleteness of first order arithm
ing showed undecideability of the halting problem.
-
7/28/2019 Human Level Sli
37/42
around these limitationswhich also apply to h
ing. As Turing (1930s), Gentzen (1930s) and Fe
showed, strengthening arithmetic is possible, but
complicated. Some very smart people, e.g. Penrosget it wrong, perhaps because of philosophical and
-
7/28/2019 Human Level Sli
38/42
QUESTIONS
What can humans do that humans cant make co
What is built into newborn babies that we havto build into computer programs? Semi-permaneobjects.
Is there a general theory of heuristics?
First order logic is universal. Is there a general figuage? Is set theory universal enough?
What must be built in before an AI system can leaand by questioning people?
-
7/28/2019 Human Level Sli
39/42
CAN WE MAKE A PLAN FOR HUMAN LE
Study relation between appearance and reality.
www-formal.stanford.edu/jmc/appearance.html
Extend sitcalc to full concurrency and continuo
Extend sitcalc to include strategies
Mental sitcalc
Reasoning within and about contexts, transcend
-
7/28/2019 Human Level Sli
40/42
Concepts as objectsas an elaboration of a t
concepts. Denot(TTelephone(MMike)) = T elephon
Uncertainty with and without numerical probabilit
of a proposition as an elaboration.
Heavy duty axiomatic set theory. ZF with abbre
defining sets. Programs will need to invent the E
the comprehension set former {x , . . . |E{x , . . .}}.
Reasoning program controllable by declaratively e
tics. Instead of domain dependent or reasoning st
-
7/28/2019 Human Level Sli
41/42
logics use general logic with set theory control
dependent advice to a general reasoning program
All this will be difficult and needs someone youngedgeable, and independent of the fashions in AI.
For the rest of us: Ask oneself: Where is my wo
to human-level AI?
-
7/28/2019 Human Level Sli
42/42
AI-HARD PROBLEMSadapted from Fanya
Used to describe problems or subproblems in AI, t
the solution presupposes a solution to the stron(that is, the synthesis of a human-level intelligenc
that is AI-hard is, in other words, just too hard.
Examples of AI-hard problems are The Vision Pr
ing a system that can see as well as a human) and
Language Problem (building a system that can uspeak a natural language as well as a human). T
pear to be modular, but all attempts so far (1996)
have foundered on the amount of context inform
telligence they seem to require.