Professional Documents
Culture Documents
com
www.fullinterview.com
www.chetanasprojects.com
ARTIFICIAL INTELLIGENCE
between mind and cognition, and by what phenomena these terms are meant
which posits that we can factor off an other mental states such as emotions and
objections that have been raised about mechanisms and phenomena? These two
the "Strong AI" hypothesis. Thus the questions set up two dimensions of
will explain all the interesting discuss these dimensions we will discuss
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
corresponding Mi, the transformation
function is f, and f is some kind of
are particular types of information states
homomorphism of F. A study of
the underlying process is thought of as
intelligence alone can restrict itself to a
information processing. However,
characterization of K’s and f, without
besides these knowledge states, mental
producing accounts of M’s and F. If
phenomena also include such things as
cognition is in fact separable in this
emotional states and subjective
sense, we can in principle design
consciousness. Under what conditions
machines that implement f and whose
can these other mental properties also be
states are interpretable as K’s. We can
attributed to artifacts to which we
call such machines cognitive agents, and
attribute knowledge states? Is
attribute intelligence to them. However,
intelligence separable from these other
the states of such machines are not
mental phenomena?
necessarily interpretable as complete
It is possible that intelligence can be M’s, and thus they may be denied other
explained or simulated without attributes of mental states.
necessarily explaining or simulating
other aspects of mind. A somewhat B. Dimension 2: Functional
formal way of putting this Separability versus Biological
Hypothesis is that the knowledge state The second dimension in discussions
transformation account can be factored about intelligence involves the extent to
off as a homomorphism of the mental which we need to be tied to biology for
process account. That is: If the mental understanding intelligence. Can
process can be seen as a sequence of intelligence be characterized abstractly
transformations: M1 -->M2 -->..., where as a functional capability which just
Mi is the complete mental state, and the happens to be realized more or less well
transformation function (the function by some biological organisms? If it can,
that is responsible for state changes) is then study of biological brains, of human
F, then a subprocess K1 --> K2 -->. . . psychology, or of the phenomenology of
can be identified such that each Ki is a human consciousness is not logically
knowledge state and a component of the
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
necessary for a theory of cognition and something intermediate, or something
intelligence, just as enquiries into the physicalist is still an open question.
relevant capabilities of biological
organisms are not needed for the abstract III. Architectures for
study of logic and arithmetic or for the
Intelligence
theory of flight. Of course, we may learn
something from biology
We now move to a discussion of
architectural proposals within the
about how to practically implement
information processing perspective. Our
intelligent systems, but we may feel
goal is to try to place the multiplicity of
quite free to substitute non-biological
proposals into perspective. As we review
(both in the sense of architectures which
various proposals, we will present some
are not brain-like and in the sense of
judgements of our own about relevant
being un- constrained by considerations
issues. But first, we need to review the
of human psychology) approaches for all
notion of an architecture and make some
or part of our implementation. Whether
additional distinctions.
intelligence can be characterized
abstractly as a functional capability
A. Form and Content Issues in
surely depends upon what phenomena
Architectures
we want to include in defining the
functional capability, as we discussed. In computer science, a programming
We might have different constraints on a language corresponds to a virtual
definition that needed to include emotion architecture. A specific program in that
and subjective states than one that only language describes a particular (virtual)
included knowledge states. Clearly, the machine, which then responds to various
enterprise of AI deeply depends upon inputs in ways defined by the program.
this functional view being true at some The architecture is thus what Newell
level, but whether that level is abstract calls the fixed structure of the
logical representations as in some information processor that is being
branches of AI, Darwinian neural group analyzed, and the program specifies a
selections as proposed by Edelman, variable structure within this
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
architecture. We can regard the alternative languages in which to couch
architecture as the form and the program an information processing account of
as the content, which together fully cognitive phenomena, and what it means
instantiate a particular information to take a Knowledge Level stance
processing machine. We can extend towards cognitive phenomena. We have
these intuitions to types of machines further discussed the distinction between
which are different from computers. For form and content theories in AI. We are
example, the connectionist architecture now ready to give an overview of the
can be abstractly specified as the set issues in cognitive architectures. We will
{{N}, {nI}, {nO}, {zi}, {wij}}, where assume that the reader is already familiar
{N} is a set of nodes, {nI} and {nO} are in some general way with the proposals
subsets of {N} called input and output that we discussing. Our goal is to place
nodes these ideas in perspective.
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
a subset of Turing-computable functions, thoughts include mental images as well.
and if so how the properties of the subset When people are thinking for a purpose,
should be characterized. say for problem solving, there is a sense
of directing thoughts, choosing some,
Most of AI research consists of
rejecting others, and focusing them
algorithms for specific problems that are
towards the goal. Activity of this type
associated with intelligence when
has been called "deliberation."
humans perform them. Algorithms for
Deliberation, for humans, is a coherent
diagnosis, design, planning, etc., are
goal-directed activity, lasting over
proposed, because these tasks are seen as
several seconds or longer. For many
important for an intelligent agent. But as
people thinking is the act of deliberating
a rule no effort is made to relate the
in this sense. We can contrast activities
algorithm for the specific task to a
in this time span with other cognitive
general architecture for intelligence.
phenomena, which, in humans, take
While such algorithms are useful as
under a few hundred milliseconds, such
technologies and to make the point
as real-time natural language
understanding and generation, visual
that several tasks that appear to require
perception, being reminded of things,
intelligence can be done by certain
and so on. These short time span
classes of machines, they do not give
phenomena are handled by what we will
much insight into intelligence in general.
call the subdeliberative architecture, as
C. Architectures for Deliberation we will discuss later.
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
provides a natural way to integrate out, the particular advantage of
external algorithms. deliberation is distal access to and
combination of knowledge at run-time in
In the Soar view, long term memory is
a goal-specific way. In the deliberative
just an associative memory. It has the
machine, temporary connections are
capability to "recognize" a situation and
created between pieces of knowledge
retrieve the relevant pieces of
that are not hard-coded, and that gives it
knowledge. Because of the learning
the ability to realize the knowledge level
capability of the architecture, each
potential more. A recognition
episode of problem solving gives rise to
architecture uses knowledge less
continuous improvement. As a problem
effectively: if the connections are not
comes along, some subtasks are solved
there as part of the memory element that
by external computational architectures
controls recognition, a piece of
which implement special purpose
knowledge, though potentially relevant,
algorithms, while others are directly
will not be utilized in the satisfaction of
solved by compiled knowledge in
a goal.
memory, while yet others are solved by
additional deliberation. This cycle make As an architecture for deliberation, the
the overall system increasingly more goal-subgoal view seems to us closer to
powerful. Eventually, most routine the mark than the reasoning view. As we
problems, including real-time have argued elsewhere [Chandrasekaran,
understanding and generation of natural 1991], logic seems more appropriate for
language, are solved by recognition. justification of conclusions and as the
(Recent work by Patten [Patten, et al, framework for the semantics of
1992] on the use of compiled knowledge representations than for the generative
in natural language understanding is architecture.
compatible with this view.)
AI theories of deliberation give central
Deliberation seems to be a source of importance to human-level problem
great power in humans. Why isn’t solving and reasoning. Any continuity
recognition enough? As Newell points with higher animal cognition or brain
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
structure is at the level of the recognition emerged as an interaction between the
architecture of memory, about which this goal (multiplication) and the procedural
view says little other than that it is a knowledge of the human. With a
recognition memory. For supporting different goal, the human might behave
deliberation at the human level, long like a different machine. It would be
term memory should be capable of awkward to imagine cognition to be a
storing and generating knowledge with collection of different architectures for
the full range of ontological distinctions each such task; in fact, cognition is very
that human language has. plastic and is able to emulate various
virtual machines as needed.
3. Is the Search View of
Deliberation Too Is the problem space search engine that
has been proposed for the deliberative
Narrow?
architecture is also an evanescent
as a search architecture is that it is based it is not intended for a narrow goal like
on a somewhat narrow view of the multiplication, but for all kinds of goals.
The Martian might well return to his purpose. What, however, if goal
superiors and report that the human achievement is only one of the functions
machine. We, however, know that the though it might be? At least in humans,
evanescent virtual architecture that daydream, just take in the external world
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
and enjoy it, and so on. The search with a goal and equipped with
behavior that we need for problem knowledge about what alternatives to
solving can come about simply by virtue consider. In fact, a number of other such
of the knowledge that is made available emergent architectures built on top of the
to the agent’s deliberation from long deliberative architecture have been
term memory. This knowledge is either a studied earlier in our work on Generic
solution to the problem, or a set of Task architectures [1986]. These
alternatives to consider. The agent, faced architectures were intended to capture
with the goal and a set of alternatives, the needs for specific classes of goals
simply considers the alternatives in turn, (such as classification).The above
and when additional subgoals are set, argument is not to deemphasize the
repeats the process of seeking more importance of problem space search for
knowledge. In fact, this kind of search goal achievement, but to resist the
behavior happens not only with identification of the architecture of the
individuals, but with organizations. They conscious processor with one
too explore alternatives, but yet we don’t exclusively intended for search The
see a need for a fixed search engine for problem space architecture is still
explaining organizational behavior. important as the virtual architecture for
Deliberation of course has to have the goal-achieving, since it is a common,
right sort of properties to be able to though not the only, function of
support search. Certainly adequate cognition.
working memory needs to be there, and
Of course, that cognition goes beyond
probably there are other constraints on
just goal achievement is a statement
deliberation. However, the architecture
about human cognition. This is to take a
for deliberation does not have to be
biological rather than a functional
exclusively a search architecture. Just
standard for the adequacy of an
like the multiplication machine was an
architectural proposal. If we take a
emergent architecture when the agent
functional approach and seek to specify
was faced with that task, the search
an architecture for a function called
engine could be the corresponding
intelligence which itself is defined in
emergent architecture for the agent faced
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
terms of goal achievement, then a architecture came to any of its
deliberative search architecture working conclusions.
with a long term memory of knowledge
Many people in AI and cognitive science
certainly has many attractive properties
feel that the emphasis on complex
for this function, as we have discussed.
problem solving as the door to
D. Subdeliberative Architectures understanding intelligence is misplaced,
and that theories that emphasize rational
We have made a distinction between problem solving only account for very
cognitive phenomena that take less than special cases and do not account for the
a few hundred milliseconds for general cognitive skills that are present
completion and those that evolve over in ordinary people. These researchers
longer time spans. We discussed focus almost completely on the nature of
proposals for the deliberative the subdeliberative architecture. There is
architecture to account for phenomena also a belief that the subdeliberative
taking longer time spans. Some form of architecture is directly reflected in the
subdeliberative architecture is then structure of the neural machinery in the
responsible for phenomena that occur in brain. Thus, some of the proposals for
very short time spans in humans. In the subdeliberative architecture claim to
deliberation, we have access to a number be inspired by the structure of the brain
of intermediate states in problem and claim a biological basis in that
solving. After you finished planning the sense.
New Delhi trip, I can ask you what
alternatives you considered, why you 1. Alternative Proposals
rejected taking the train, and so on, and
your answers to them will generally be The various proposals differ along a
rejecting the train option because you tasks the architecture performs, degree
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
information processing architecture, Deliberation has a serial character to it.
whether it is a symbolic one or some Almost all proposals for the
other type. subdeliberative architecture, however,
use parallelism in one way or another.
With respect to the kind of tasks the
Parallelism can bring a number of
architecture performs, we mentioned
advantages. For problems involving
Newell’s view that it is just a recognition
similar kinds of information processing
architecture. Any smartness it possesses
over somewhat distributed data (like
is a result of good abstractions and good
perception), parallelism can speed up
indexing, but architecturally, there is
processing. Ultimately, however,
nothing particularly complicated. In fact,
additional problem solving in
the good abstractions and indexing
deliberation may be required for some
themselves were the result of the
tasks.
discoveries of deliberation during
problem state search. The real solution 2. Situated Cognition
to the problem of memory, for Newell, is
to get chunking done right: the proper Real cognitive agents are in contact
level of abstraction, labeling and with the surrounding world containing
indexing is all done at the time of physical objects and other agents. A new
chunking. In contrast to the recognition school has emerged calling itself the
view are proposals that see relatively situated cognition movement which
complex problem solving activities argues that traditional AI and cognitive
going on in subdeliberative cognition. science abstract the cognitive agent too
Cognition in this picture is a much away from the environment, and
communicating collection of modular place undue emphasis on internal
agents, each of whom is simple, but representations. The traditional internal
capable of some degree of problem representation view leads, according to
solving. For example, they can use the the situated cognition perspective, to
means-ends heuristic (the goal- large amounts of internal representation
subgoaling feature of deliberation in the and complex reasoning using these
Soar architecture). representations. Real agents simply use
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
their sensory and motor systems to different constraints on the architecture.
explore the world and pick out the We reviewed a number of issues and
information needed, and get by with proposals relevant to cognitive
much smaller amounts of internal architectures. Not only are there many
representation processing. At the levels each explaining some aspect of
minimum, situated cognition is a cognition and mentality, but the levels
proposal against excessive "intellection." interact even in relatively simple
In this sense, we can simply view this cognitive phenomena.
movement as making different proposals
about what and how much needs to be
represented internally. The situated
References:-
cognition perspective clearly rejects the
former view with respect to internal • Poole, David; Mackworth, Alan;
Goebel, Randy (1998),
(sub-deliberative) processes, but accepts
Computational Intelligence: A
the fact deliberation does contain and Logical Approach, Oxford
University Press., ISBN 0-19-
use knowledge. Thus the Knowledge
510270-3,
Level description could be useful to http://www.cs.ubc.ca/~poole/ci.h
tml
describe the content of agent’s
• Samuel, Arthur L. (July 1959),
deliberation. "Some studies in machine
learning using the game of
checkers", IBM Journal of
V. Concluding Remarks Research and Development 3 (3):
210−219,
http://domino.research.ibm.com/t
We started by asking how far chjr/journalindex.nsf/600cc5649e
intelligence or cognition can be 2871db852568150060213c/39a8
70213169f45685256bfa00683d7
separated from mental phenomena in 4?OpenDocument, retrieved on
general. We suggested that the problem 2007-08-20
• Searle, John (1980), "Minds,
of an architecture for cognition is not Brains and Programs",
really well-posed, since, depending upon Behavioral and Brain Sciences 3
(3): 417−457,
what aspects of the behavior of http://www.bbsonline.org/docum
biological agents are included in the ents/a/00/00/04/84/bbs00000484-
00/bbs.searle2.html
functional specification, there can be
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com
• Simon, H. A.; Newell, Allen
(1958), "Heuristic Problem
Solving: The Next Advance in
Operations Research",
Operations Research 6: 1,
doi:10.1287/opre.6.1.1
• Simon, H. A. (1965), The Shape
of Automation for Men and
Management, New York: Harper
& Row
• Turing, Alan (1936-37), "On
Computable Numbers, with an
Application to the
Entscheidungsproblem",
Proceedings of the London
Mathematical Society, 2 (42):
230–265,
http://www.abelard.org/turpap2/t
p2-ie.asp
• Turing, Alan (October 1950),
"Computing machinery and
intelligence", Mind LIX (236):
433−60,
doi:10.1093/mind/LIX.236.433,
http://loebner.net/Prizef/TuringA
rticle.html
• Weizenbaum, Joseph (1976),
Computer Power and Human
Reason, W.H. Freeman &
Company
www.1000projects.com
www.fullinterview.com
www.chetanasprojects.com