FROM HERE TO HUMAN-LEVEL AI
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
FROM HERE TO HUMAN-LEVEL AI
John McCarthy
Computer Science Department
Stanford University
Stanford, CA 94305
jmc@cs.stanford.edu
http://www-formal.stanford.edu/jmc/
Abstract clude nonmonotonic reasoning, ap-
proximate concepts, formalized con-
texts and introspection.
It is not surprising that reaching
human-level AI has proved to be dif-
ficult and progress has been slow—
though there has been important 1 What is Human-Level AI?
progress. The slowness and the de-
mand to exploit what has been dis- The first scientific discussion of human level
covered has led many to mistakenly machine intelligence was apparently by Alan
redefine AI, sometimes in ways that Turing in the lecture [Turing, 1947]. The no-
preclude human-level AI—by rele- tion was amplified as a goal in [Turing, 1950],
gating to humans parts of the task but at least the latter paper did not say what
that human-level computer programs would have to be done to achieve the goal.
would have to do. In the terminology Allen Newell and Herbert Simon in 1954 were
of this paper, it amounts to settling the first people to make a start on program-
for a bounded informatic situation in- ming computers for general intelligence. They
stead of the more general common were over-optimistic, because their idea of
sense informatic situation. what has to be done to achieve human-level in-
Overcoming the “brittleness” of telligence was inadequate. The General Prob-
present AI systems and reaching lem Solver (GPS) took general problem solv-
human-level AI requires programs ing to be the task of transforming one expres-
that deal with the common sense sion into another using an allowed set of trans-
informatic situation—in which the formations.
phenomena to be taken into account Many tasks that humans can do, humans can-
in achieving a goal are not fixed in not yet make computers do. There are two ap-
advance. proaches to human-level AI, but each presents
We discuss reaching human-level AI, difficulties. It isn’t a question of deciding be-
emphasizing logical AI and especially tween them, because each should eventually
emphasizing representation problems succeed; it is more a race.
of information and of reasoning.
Ideas for reasoning in the com- 1. If we understood enough about how the
mon sense informatic situation in- human intellect works, we could simulateit. However, we don’t have have suffi- A formal theory in the physical sciences deals
cient ability to observe ourselves or others with a bounded informatic situation. Scientists
to understand directly how our intellects decide informally in advance what phenomena
work. Understanding the human brain to take into account. For example, much ce-
well enough to imitate its function there- lestial mechanics is done within the Newtonian
fore requires theoretical and experimental gravitational theory and does not take into ac-
success in psychology and neurophysiol- count possible additional effects such as out-
ogy. 1 See [Newell and Simon, 1972] for gassing from a comet or electromagnetic forces
the beginning of the information process- exerted by the solar wind. If more phenom-
ing approach to psychology. ena are to be considered, a person must make
a new theory. Probabilistic and fuzzy uncer-
2. To the extent that we understand the tainties can still fit into a bounded informatic
problems achieving goals in the world system; it is only necessary that the set of pos-
presents to intelligence we can write intel- sibilities (sample space) be bounded.
ligent programs. That’s what this article
is about. Most AI formalisms also work only in a
bounded informatic situation. What phenom-
ena to take into account is decided by a person
What problems does the world present to in-
before the formal theory is constructed. With
telligence? More narrowly, we consider the
such restrictions, much of the reasoning can be
problems it would present to a human scale
monotonic, but such systems cannot reach hu-
robot faced with the problems humans might
man level ability. For that, the machine will
be inclined to relegate to sufficiently intelli-
have to decide for itself what information is
gent robots. The physical world of a robot
relevant. When a bounded informatic system
contains middle sized objects about which its
is appropriate, the system must construct or
sensory apparatus can obtain only partial in-
choose a limited context containing a suitable
formation quite inadequate to fully determne
theory whose predicates and functions connect
the effects of its future actions. Its mental
to the machine’s inputs and outputs in an ap-
world includes its interactions with people and
propriate way. The logical tool for this is non-
also meta-information about the information
monotonic reasoning.
it has or can obtain.
Our approach is based on what we call the 2 The Common Sense Informatic
common sense informatic situation. In order Situation
to explain the common sense informatic situ-
ation, we contrast it with the bounded infor- Contention: The key to reaching
matic situation that characterizes both formal human-level AI is making systems that
scientific theories and almost all (maybe all) operate successfully in the common
experimental work in AI done so far.2 sense informatic situation.
1
Recent work with positron emission tomography has In general a thinking human is in what we call
identified areas of the brain that consume more glucose the common sense informatic situation first
when a person is doing mental arithmetic. This knowledge
will help build AI systems only when it becomes possible discussed in3 [McCarthy, 1989]. It is more
to observe what is going on in these areas during mental general than any bounded informatic situation.
arithmetic.
2
The textbook [David Poole and Goebel, 1998] puts it The known facts are incomplete, and there is
this way. “To get human-level computational intelligence no a priori limitation on what facts are rel-
it must be the agent itself that decides how to divide up
3
the world, and which relationships to reason about. http://www-formal.stanford.edu/jmc/ailogic.htmlevant. It may not even be decided in ad- sure with altitude. However, in every case,
vance what phenomena are to be taken into the physics knowledge is embedded in com-
account. The consequences of actions cannot mon sense knowledge. Thus before one can
be fully determined. The common sense in- use Galileo’s law of falling bodies s = 21 gt2 , one
formatic situation necessitates the use of ap- needs common sense information about build-
proximate concepts that cannot be fully de- ings, their shapes and their roofs.
fined and the use of approximate theories in- Bounded informatic situations are obtained by
volving them. It also requires nonmonotonic nonmonotonically inferring that only the phe-
reasoning in reaching conclusions. nomena that somehow appear to be relevant
The common sense informatic situation also are relevant. In the barometer example, the
includes some knowledge about the system’s student was expected to infer that the barom-
mental state. eter was only to be used in the conventional
way for measuring air pressure. For example,
A nice example of the common sense infor-
matic situation is illustrated by an article in a reasoning system might do this by apply-
the American Journal of Physics some years ing circumscription to a predicate relevant in a
ago. It discussed grading answers to a physics formalism containing also metalinguistic infor-
problem. The exam problem is to find the mation, e.g. that this was a problem assigned
height of a building using a barometer. The in a physics course. Formalizing relevance in
intended solution is to measure the air pres- a useful way promises to be difficult.
sure at the top and bottom of the building Common sense facts and common sense rea-
and multiply the difference by the ratio of the soning are necessarily imprecise. The impreci-
density of mercury to the density of air. sion necessitated by the common sense infor-
However, other answers may be offered. (1) matic situation applies to computer programs
drop the barometer from the top of the build- as well as to people.
ing and measure the time before it hits the Some kinds of imprecision can be represented
ground. (2) Measure the height and length of numerically and have been explored with the
the shadow of the barometer and measure the aid of Bayesian networks, fuzzy logic and simi-
length of the shadow of the building. (3) Rap- lar formalisms. This is in addition to the study
pel down the building using the barometer as of approximation in numerical analysis and the
a measuring rod. (4) Lower the barometer on physical sciences.
a string till it reaches the ground and measure
the string. (5) Offer the barometer to the jani- 3 The Use of Mathematical Logic
tor of the building in exchange for information
about the height. (6) Ignore the barometer, What about mathematical logical languages?
count the stories of the building and multiply
by ten feet. Mathematical logic was devised to formal-
ize precise facts and correct reasoning. Its
Clearly it is not possible to bound in advance founders, Leibniz, Boole and Frege, hoped to
the common sense knowledge of the world use it for common sense facts and reasoning,
that may be relevant to grading the prob- not realizing that the imprecision of concepts
lem. Grading some of the solutions requires used in common sense language was often a
knowledge of the formalisms of physics and the necessary feature and not always a bug. The
physical facts about the earth, e.g. the law biggest success of mathematical logic was in
of falling bodies or the variation of air pres- formalizing mathematical theories. Since thecommon sense informatic situation requires AI formalism make programs reason logically.
using imprecise facts and imprecise reason- However, we have to extend logic and extend
ing, the use of mathematical logic for common the programs that use it in various ways.
sense has had limited success. This has caused
One important extension was the development
many people to give up. Gradually, extended of modal logic starting in the 1920s and using
logical languages and even extended forms of it to treat modalities like knowledge, belief and
mathematical logic are being invented and de- obligation. Modalities can be treated either
veloped. by using modal logic or by reifying concepts
It is necessary to distinguish between mathe- and sentences within the standard logic. My
matical logic and particular mathematical log- opinion is that reification in standard logic is
ical languages. Particular logical languages more powerful and will work better.
are determined by a particular choice of con-
A second extension was the formalization of
cepts and the predicate and function symbols
nonmonotonic reasoning beginning in the late
to represent them. Failure to make the dis- 1970s—with circumscription and default logic
tinction has often led to error. When a par- and their variants as the major proposals.
ticular logical language has been shown inad- Nonmonotonic logic has been studied both as
equate for some purpose, some people have pure mathematics and in application to AI
concluded that logic is inadequate. Different problems, most prominently to the formaliza-
concepts and different predicate and function tion of action and causality. Several variants
symbols might still succeed. In the words of of the major formalisms have been devised.
the drive-in movie critic of Grapevine, Texas,
“I’m surprised I have to explain this stuff.” Success so far has been moderate, and it isn’t
clear whether greater success can be obtained
The pessimists about logic or some particular by changing the the concepts and their rep-
set of predicates might try to prove a theorem resentation by predicate and function symbols
about its inadequacies for expressing common or by varying the nonmonotonic formalism. 5
sense.4
We need to distinguish the actual use of logic
Since it seems clear that humans don’t use from what Allen Newell, [Newell, 1981] and
logic as a basic internal representation formal- [Newell, 1993], calls the logic level and which
ism, maybe something else will work better was also proposed in [McCarthy, 1979]6 .
for AI. Researchers have been trying to find
this something else since the 1950s but still
4 Approximate Concepts and
haven’t succeeded in getting anything that is
Approximate Theories
ready to be applied to the common sense in-
formatic situation. Maybe they will eventually Other kinds of imprecision are more funda-
succeed. However, I think the problems listed mental for intelligence than numerical impre-
in the later sections of this article will apply cision. Many phenomena in the world are ap-
to any approach to human-level AI. propriately described in terms of approximate
Mathematical logic has been concerned with concepts. Although the concepts are impre-
how people ought to think rather than how cise, many statements using them have precise
people do think. We who use logic as a basic truth values. We offer two examples: the con-
5
One referee for KR96 foolishly and arrogantly pro-
4
Gödel’s theorem is not relevant to this, because the posed rejecting a paper on the grounds that the inadequacy
question is not one of decideability or of characterizing of circumscription for representing action was known.
6
truth. http://www-formal.stanford.edu/jmc/ascribing.htmlcept of Mount Everest and the concept of the posals for handling nonmonotonic reasoning.
welfare of a chicken. The exact pieces of rock In particular, getting from the common sense
and ice that constitute Mount Everest are un-
informatic situation to a bounded informatic
clear. For many rocks, there is no truth of the
situation needs nonmonotonic reasoning.
matter as to whether it is part of Mount Ever-
est. Nevertheless, it is true without qualifica-
6 Elaboration Tolerance
tion that Edmund Hillary and Tenzing Norgay
climbed Mount Everest in 1953 and that John
Human abilities in the common sense infor-
McCarthy never set foot on it.
matic situation also include what may be
The point of this example is that it is possi- called elaboration tolerance—the ability to
ble and even common to have a solid knowl- elaborate a statement of some facts without
edge structure from which solid conclusions having to start all over. Thus when we begin
can be inferred based on a foundation built on to think about a problem, e.g. determining
the quicksand of approximate concepts with- the height of a building, we form a bounded
out definite extensions. context and try to solve the problem within it.
As for the chicken, it is clear that feeding it However, at any time more facts can be added,
helps it and wringing its neck harms it, but e.g. about the precision with which the time
it is unclear what its welfare consists of over for the barometer to fall can be estimated us-
the course of the decade from the time of its ing a stop watch and also the possibilities of
hatching. Is it better off leading a life of poul- acquiring a stop watch.
try luxury and eventually being slaughtered Elaboration Tolerance7 discusses about 25
or would it be better off escaping the chicken elaborations of the Missionaries and Cannibals
yard and taking its chances on starvation and problem.
foxes? There is no truth of the matter to be
What I have so far said so far about ap-
determined by careful investigation of chick-
proximate concepts, nonmonotonic reasoning
ens. When a concept is inherently ap-
and elaboration tolerance is independent of
proximate, it is a waste of time to try to
whether mathematical logic, human language
give it a precise definition. Indeed differ-
or some other formalism is used.
ent efforts to define such a concept precisely
will lead to different results—if any. In my opinion, the best AI results so far have
been obtained using and extending mathemat-
Most human common sense knowledge in-
ical logic.
volves approximate concepts, and reaching
human-level AI requires a satisfactory way
of representing information involving approxi- 7 Formalization of Context
mate concepts.
A third extension of mathematical logic in-
volves formalizing the notion of context8
5 Nonmonotonic Reasoning [McCarthy, 1993]. Notice that when logical
theories are used in human communication
Common sense reasoning is also imprecise in and study, the theory is used in a context
that it draws conclusions that might not be which people can discuss from the outside. If
made if there were more information. Thus computers are to have this facility and are to
common sense reasoning is nonmonotonic. I 7
http://www-formal.stanford.edu/jmc/elaboration.html
will not go into the details of any of the pro- 8
http://www-formal.stanford.edu/jmc/context.htmlwork within logic, then the “outer” logical lan- travel, but the travel agent will not tell
guage needs names for contexts and sentences his customer to be sure and wear clothes.
giving their relations and a way of entering a
context. Clearly human-level AI requires rea- • The ramification problem concerns how to
soning about context. treat side-effects of events other than the
principal effect mentioned in the event de-
Human-level AI also requires the ability to
scription.
transcend the outermost context the system
has used so far. Besides in [McCarthy, 1993],
this is also discussed in Making Robots Each of these involves elaboration tolerance,
Conscious of their Mental States9 e.g. adding descriptions of the effects of
[McCarthy, 1996]. additional events without having to change
the descriptions of the events already de-
Further work includes [Buvač, 1996] and scribed. When I wrote about applications of
[Buvač et al., 1995]. circumscription to formalizing common
sense10 [McCarthy, 1986], I hoped that a sim-
8 Reasoning about ple abnormality theory would suffice for all of
Events—Especially Actions them. That didn’t work out when I tried it,
but I still think a common nonmonotonic rea-
Reasoning about actions has been a major AI soning mechanism will work. Tom Costello’s
activity, but this paper will not discuss my or draft “The Expressive Power of Circumscript-
other people’s current approaches, concentrat- tion” 11 argues that simple abnormality theo-
ing instead on the long range problem of reach- ries have the same expressive power as more
ing human level capability. We regard actions elaborate nonmonotonic formalisms that have
as particular kinds of events and therefore pro- been proposed.
pose subsuming reasoning about actions under
the heading of reasoning about events. Human level intelligence requires reasoning
about strategies of action, i.e. action pro-
Most reasoning about events has concerned grams. It also requires considering multiple
determining the effects of an explicitly given actors and also concurrent events and contin-
sequence of actions by a single actor. Within uous events. Clearly we have a long way to
this framework various problems have been go.
studied.
Some of these points are discussed in a draft
• The frame problem concerns not having on narrative12 [McCarthy, 1995].
to state what does not change when an
event occurs. 9 Introspection
• The qualification problem concerns not People have a limited ability to observe their
having to state all the preconditions of an own mental processes. For many intellectual
action or other event. The point is both tasks introspection is irrelevant. However, it
to limit the set of preconditions and also is at least relevant for evaluating how one is
to jump to the conclusion that unstated using one’s own thinking time. Human-level
others will be fulfilled unless there is evi- AI will require introspective ability.
dence to the contrary. For example, wear-
10
ing clothes is a precondition for airline http://www-formal.stanford.edu/jmc/applications.html
11
http://www-formal.stanford.edu/tjc/expressive.html
9 12
http://www-formal.stanford.edu/jmc/consciousness.html http://www-formal.stanford.edu/jmc/narrative.htmlThat robots also need introspection13 summarized as that of succeeding in the com-
is argued and how to do it is discussed in mon sense informatic situation.
[McCarthy, 1996].
The problems include:
10 Heuristics
common sense knowledge of the world
The largest qualitative gap between human Many important aspects of what this
performance and computer performance is in knowledge is in and how it can be
the area of heuristics, even though the gap is represented are still unsolved questions.
disguised in many applications by the millions- This is particularly true of knowledge of
fold speed advantage of computers. The gen- the effects of actions and other events.
eral purpose theorem proving programs run
very slowly, and the special purpose programs epistemologically adequate languages
are very specialized in their heuristics. These are languages for expressing
what a person or robot can ac-
I think the problem lies in our present in- tually learn about the world15
ability to give programs domain and prob- [McCarthy and Hayes, 1969].
lem dependent heuristic advice. In my Ad-
vice Taker paper14 [McCarthy, 1959] I adver- elaboration tolerance What a person
tised that the Advice Taker would express its knows can be elaborated without starting
heuristics declaratively. Maybe that will work, all over.
but neither I nor anyone else has been able to
get a start on the problem in the ensuing al- nonmonotonic reasoning Perhaps new sys-
most 40 years. Josefina Sierra-Santibanez re- tems are needed.
ports on some progress in a forthcoming arti-
cle. contexts as objects This subject is just be-
Another possibility is to express the advice in ginning. See the references of section 7.
a procedure modification language, i.e. to ex-
tend elaboration tolerance to programs. Of introspection AI systems will need to exam-
course, every kind of modularity, e.g. object ine their own internal states.
orientation, gives some elaboration tolerance,
but these devices haven’t been good enough. action The present puzzles of formalizing ac-
tion should admit a uniform solution.
Ideally, a general purpose reasoning system
would be able to accept advice permitting it
to run at a fixed ratio speed of speeds to a I doubt that a human-level intelligent program
special purpose program, e.g. at 1/20 th the will have structures corresponding to all these
speed. entities and to the others that might have been
listed. A generally intelligent logical program
11 Summary probably needs only its monotonic and non-
monotonic reasoning mechanisms plus mecha-
Conclusion: Between us and human-level in- nisms for entering and leaving contexts. The
telligence lie many problems. They can be rest are handled by particular functions and
predicates.
13
http://www-formal.stanford.edu/jmc/consciousness.html
14 15
http://www-formal.stanford.edu/jmc/mcc59.html http://www-formal.stanford.edu/jmc/mcchay69.html12 Remarks and Acknowledgements It will be much more scientifically satisfying to
understand human level artificial intelligence
1. To what extent will all these problems logically than just achieve it by a computer-
have to be faced explicitly by people ized evolutionary process that produced an in-
working with neural nets and connection- telligent but incomprehensible result. In fact,
ist systems? The systems I know about the logical approach would be worth pursuing
are too primitive for the problems even to even if the intellectually lazy evolutionary ap-
arise. However, more ambitious systems proach won the race.
will inhabit the common sense informatic
situation. They will have to be elabora- References
tion tolerant and will require some kind
of mental model of the consequences of [Buvač, 1996] Buvač, S. (1996). Quantifica-
actions. tional logic of context. In Proceedings of
the Thirteenth National Conference on Ar-
2.
tificial Intelligence.
3. I got useful suggestions from Eyal Amir,
Saša Buvač and Tom Costello. [Buvač et al., 1995] Buvač, S., Buvač, V., and
Mason, I. A. (1995). Metamathematics of
4. Some additional relevant papers are in my contexts. Fundamenta Informaticae, 23(3).
book [McCarthy, 1990] and on my Web
site16 . [David Poole and Goebel, 1998] David Poole,
A. M. and Goebel, R. (1998). Computa-
5. My understanding that I should prepare a tional Intelligence. Oxford.
printable version of this invited talk came
rather late. I expect that both the spoken [McCarthy, 1959] McCarthy, J. (1959). Pro-
version and the 1996 November Web ver- grams with Common Sense17 . In Mecha-
sion will have better explanations of the nisation of Thought Processes, Proceedings
important concepts. of the Symposium of the National Physics
Laboratory, pages 77–84, London, U.K. Her
6. This work was partly supported by ARPA Majesty’s Stationery Office. Reprinted in
(ONR) grant N00014-94-1-0775. McC90.
13 Conclusion [McCarthy, 1979] McCarthy, J. (1979). As-
cribing mental qualities to machines18 . In
Many will find dismayingly large the list of Ringle, M., editor, Philosophical Perspec-
tasks that must be accomplished in order to tives in Artificial Intelligence. Harvester
to reach human-level logical intelligence. Per- Press. Reprinted in [McCarthy, 1990].
haps fewer but more powerful ideas would sim-
plify the list. Others will claim that a sys- [McCarthy, 1986] McCarthy, J. (1986). Ap-
tem that evolves intelligence as life does will plications of Circumscription to Formaliz-
be more straightforward to build. Maybe, but ing Common Sense Knowledge19 . Artifi-
the advocates of that approach have been at it cial Intelligence, 28:89–116. Reprinted in
as long as we have and still aren’t even close. [McCarthy, 1990].
17
http://www-formal.stanford.edu/jmc/mcc59.html
So it’s a race. 18
http://www-formal.stanford.edu/jmc/ascribing.html
16 19
http://www-formal.stanford.edu/jmc/ http://www-formal.stanford.edu/jmc/applications.html[McCarthy, 1989] McCarthy, J. (1989). Ar- [Newell and Simon, 1972] Newell, A. and Si-
tificial Intelligence, Logic and Formalizing mon, H. A. (1972). Human Problem Solving.
Common Sense20 . In Thomason, R., edi- Prentice–Hall, Englewood Cliffs, NJ.
tor, Philosophical Logic and Artificial Intel-
ligence. Klüver Academic. [Turing, 1950] Turing, A. (1950). Computing
machinery and intelligence. Mind.
[McCarthy, 1990] McCarthy, J. (1990). For-
[Turing, 1947] Turing, A. M. (1947). Lec-
malization of common sense, papers by John
ture to the london mathematical society.
McCarthy edited by V. Lifschitz. Ablex.
In The Collected Works of A. M. Tur-
ing, volume Mechanical Intelligence. North-
[McCarthy, 1993] McCarthy, J. (1993).
21 Holland. This was apparently the first pub-
Notes on Formalizing Context . In
lic introduction of AI, typescript in the
IJCAI-93. Available on http://www-
King’s College archive, the book is 1992.
formal.stanford.edu/jmc/.
[McCarthy, 1995] McCarthy, J. (1995). Situa-
tion Calculus with Concurrent Events and
Narrative22 . Contents subject to change.
Reference will remain.
[McCarthy, 1996] McCarthy, J. (1996). Mak-
ing Robots Conscious of their Mental
States23 . In Muggleton, S., editor, Machine
Intelligence 15. Oxford University Press.
[McCarthy and Hayes, 1969] McCarthy, J.
and Hayes, P. J. (1969). Some Philosophical
Problems from the Standpoint of Artificial
Intelligence24 . In Meltzer, B. and Michie,
D., editors, Machine Intelligence 4, pages
463–502. Edinburgh University Press.
[Newell, 1981] Newell, A. (1981). The knowl-
edge level. AI Magazine, 2(2):1–20. Origi-
nally delivered as the Presidential Address,
American Association for Artificial Intelli-
gence, AAAI80, Stanford, CA, August 1980.
[Newell, 1993] Newell, A. (1993). Reflections
on the knowledge level. Artificial Intelli-
gence, 59(1-2):31–38.
20
http://www-formal.stanford.edu/jmc/ailogic.html
21
http://www-formal.stanford.edu/jmc/context.html
22
http://www-formal.stanford.edu/jmc/narrative.html
23
http://www-formal.stanford.edu/jmc/consciousness.html
24
http://www-formal.stanford.edu/jmc/mcchay69.htmlYou can also read
