hhamburg@NOTE.NSF.GOV ("Henry J. Hamburger") (11/16/88)
NATIONAL SCIENCE FOUNDATION
---------------------------
PROGRAM in
----------
KNOWLEDGE MODELS and COGNITIVE SYSTEMS
--------------------------------------
Knowledge Models and Cognitive Systems is a relatively new name
at NSF, but the Program has significant continuity with earlier
related programs. This holds for its scientific subject matter
and also with regard to its researchers, who come principally
from computer science and the cognitive sciences, each of these
emphatically including important parts of artificial intelligence.
Many such individuals are also interested in areas supported by
other NSF programs, especially in this division -- the Division
of Information, Robotics and Intelligent Systems (IRIS) -- and in
the Division of Behavioral and Neural Sciences.
This unofficial message has two parts. The first is a top-down
description of the major areas of current Program support. There
follows a list of some particular topics in which there is strong
current activity in the Program and/or perceived future
opportunity. Anyone needing further information can contact the
Program Director, Henry Hamburger, who is also the sender of this
item. Please use e-mail if you can: hhamburg@b.nsf.gov or else
phone: 202-357-9569. To get a copy of the Summary of Awards for
this division (IRIS), call 202-357-9572
Many of you will be hearing from me with requests to review
proposals. To be sure they are of interest to you, feel free to
send me a list of topics or subfields.
MAJOR AREAS of CURRENT SUPPORT
------------------------------
The Program in Knowledge Models and Cognitive Systems supports
research fundamental to the general understanding of knowledge
and cognition, whether in humans, computers or, in principle,
other entities. Major areas currently receiving support include
(i) formal models of knowledge and information, (ii) natural
language processing and (iii) cognitive systems. Each of these
areas is described and subcategorized below.
Applicants do not classify their proposals in any official way.
Indeed their work may be relevant to two or all three of the
categories (or conceivably to none of them). In particular, it
is recognized that language is intertwined with (or part of)
cognition and that formality is a matter of degree. For work
that falls only partly within the program, the program director
may conduct the evaluation jointly with another program, within
or outside the division. Descriptions of the three areas follow.
FORMAL MODELS of KNOWLEDGE and INFORMATION:
-------------------------------------------
Recent work supported under the category Formal Models of
Knowledge and Information divides into formal models of three
things: (i) knowledge, (ii) information, and (iii) imperfections
in the two. In each case, the models may encompass both
representation and manipulation. For example, formal models of
both knowledge representation and inference are part of the
knowledge area.
The distinction between knowledge and information is that a piece
of knowledge tends to be more structured and/or comprehensive
than a piece of information. Imperfections may include
uncertainty, vagueness, incompleteness and abductive rules. Many
investigations contribute to two or all three categories, yet
emphasize one.
COGNITIVE SYSTEMS
-----------------
Four recognized areas currently receive support within Cognitive
Systems: (i) knowledge representation and inference, (ii)
highly parallel approaches, (iii) machine learning, and (iv)
computational characterization of human cognition.
The first area is characterized by symbolic representations and a
high degree of structure imposed by the programmer, in an attempt
to represent complex entities and carry out complex tasks
involving planning and reasoning. The second area may have
similar long-term goals but takes a very different approach. It
includes studies based on a high degree of parallelism among
relatively simple processing units connected according to various
patterns. The third area, machine learning, has emerged as a
distinct area of study, though the choice between symbolic and
connectionist approaches is clearly relevant. In all of the
first three areas, the research may be informed to a greater or
lesser degree by scientific knowledge of the nature of high-
level human cognition. Characterizing such knowledge in
computational form is the objective of the fourth area.
NATURAL LANGUAGE PROCESSING
---------------------------
Recent work supported under the category Natural Language
Processing is in three overlapping areas: (i) computational
aspects of syntax, semantics and the lexicon, (ii) discourse,
dialog and generation, and (iii) systems issues. The distinction
between the first two often involves such intersentential
concerns as topic, plan, and situation. Systems issues include
the interaction and unified treatment of various kinds of
modules.
TOPICS of STRONG CURRENT ACTIVITY and
-------------------------------------
OPPORTUNITY for FUTURE RESEARCH
-------------------------------
Comments on this list are welcome. It has no official status,
is subject to change, and, most important, is intended to be
suggestive, not prescriptive. The astute reader will notice that
many of these topics transcend the neat categorization above.
Reasoning and planning in the face of
imperfect information and a changing world
- reasoning about reasoning itself: the time
and resources taken, and the consequences
- use and formal understanding of
temporal and nonmonotonic logic
- integration of numerical and symbolic approaches
to uncertainty, imprecision and justification
- multi-agent planning, reasoning,
communication and coordination
Interplay of human and computational languages
- commonalities in the semantic formalisms
for human and computer languages
- extending knowledge representation systems to
support formal principles of human language
- principles of extended dialog between humans
and complex software systems, including
those of the new computational sciences
Machine Learning of Classification,
Problem-Solving and Scientific Laws
- formal analysis of what features and parameter
settings of both method and domain are
responsible for successes.
- reconciling and combining the benefits of
connectionist, genetic and symbolic approaches
- evaluating the relevance to learning of AI
tools: planning, search, and learning itself