Algorithm 版 (精华区)
发信人: ssos (存在与虚无·守拙), 信区: Algorithm
标 题: Three Open Problems in AI
发信站: 哈工大紫丁香 (2003年07月27日11:15:45 星期天), 站内信件
Three Open Problems in AI
RAJ REDDY
Carnegie Mellon University
We present three open problems in AI which, if solved, would bring us
closer to achieving the goal of Human Level AI. Common to all these
systems is the ability of a system to learn from examples,
observations and errors. Indeed, one might say that these represent
necessary conditions before we can aspire to reach Human Level AI.
1. Read a Chapter in a Book and Answer the Questions at the End of the
Chapter
Reading and understanding books is a quintessentially human
activity. It is the process by which much knowledge transfer occurs
from generation to generation. For example, we test students'
understanding of a given subject by asking them to answer questions at
the end of the chapter in a textbook. In general, we are satisfied if
a student correctly answers 80 to 90% of the questions. When a
computer can do the same task, we will have arrived at a significant
milestone.
A human being who starts reading at age 4, lives to be 100 years old,
and reads a book a day every day could complete 35,000 books in a
lifetime. Bymany estimates, the total number books ever written in all
languages is under 100 million. Harvard library has around 12 million
volumes. The Library of Congress has fewer than 30 million
volumes. All the unique titles in the OCLC member libraries is under
42 million. However, US libraries do not have most of the books
published in other languages internationally, thus leading to an
approximate estimate of 100 million books ever published.
Once a computer can read one book and prove that it understands it by
answering questions about it correctly, then, in principle, it can
read all the books that have ever been written. This has led to the
speculation that once computers can read, un derstand, and share
knowledge with each other, without the limitations that biology
imposes, they will begin to exhibit super human intelligence.
For a machine to read a book, understand it and answer questions about
it, it needs mechanisms
---for converting (analog) paperbased information into machine
processable form;
---to read and understand text with all the implied ambiguity and
imprecision of natural language and interpret the intention of the
author;
---to convert the understanding into executable representation of the
knowledge;
---to interpret and represent questions into initial conditions and
the desired end goal; and
---for problem solving to apply the knowledge extracted from the
chapter and previously known (acquired) knowledge including
qualitative common sense knowledge to solve the problem at hand.
What does it mean to ``read a book''? First, the book needs to be in a
machinereadable format. Many books are ``born digital'', that is,
their content is available in machine processable form. Books
published before the advent of digital publishing would have to be
scanned resulting in a digital image (essentially a digital
photograph) of the page in the computer. Optical character recognition
systems exist today that process text, tables, graphs, and grey scale
and color images in a page and produce processable text and table with
less than 1% error. Graphs and images are usually left
uninterpreted. But these OCR systems are never perfect. Thus, any
system that attempts to read and ``understand'' must also cope with
errors and typos.
Once we have machinereadable text, the real problem begins. At
present, we do have systems that can translate from one language to
another with mixed results. This and other forms of understanding
(such as information retrieval or document classification) can be done
with only a superficial understanding of the meaning of words, phrases
and sentences. The implication of the proposed task is that to answer
question at the end of the chapter, the knowledge within has to be
distilled into an executable representation, to be used in conjunction
with reasoning and problem solving methods to solve the problems at
the end.
Herbert Simon (with Dorothea Simon) attempted to solve this problem in
the early 1970s. Their problem solving tool was a Production
System. They reasoned that if all the knowledge in a chapter of a
physics textbook could be represented as ``productions'' (also known
as ``production rules'') that then were used to solve many of the
problems at the end of the chapter, that would represent true
understanding. Lacking the tools for ``Natural Language
Understanding'', they read the chapter themselves and represented
their understanding as ``productions'', which they were able to
successfully use to solve the problems.
Having partially solved the big problem, Simon had a succession of
students work on the rest of the problem of getting a machine to read
and understand---without success. The main difficulty was ``to
understand the meaning of a sentence'' in the absence of systems with
common sense knowledge.
Since the 1970s, we have made great strides in ``Natural Language
Understanding''. These advances combined with a systematic approach to
transforming ``understanding'' into executable productions should
result in significant progress. The ``systematic approach `` we
propose is to extend the Simon paradigm of getting humans to do ``what
machines cannot do'' and creating a research agenda of all such human
interventions, to be explored by generations of graduate students
ultimately approaching ``nirvana''.
Once we can read one book and demonstrate deep understanding by
solving problems at the end, we are not at the end. There is the other
unsolved problem of assimilation and integration of knowledge from
multiple sources, that is, from all the books. That is not the end
either; we also have to integrate knowledge from other life
experiences such as auditory knowledge and visual knowledge. Once we
have mastered the concepts of visual knowledge, we will be in a
position to interpret diagrams and pictures in books along with other
tasks such as knowing how to repair a robot from observation, which is
the next open problem.
2. Remote Repair
The second open problem we propose is the task of getting a machine to
learn how to repair a Mobile Robot and successfully demonstrate the
capability by repairing one on Mars (or with appropriate simulated
time delay on earth) after observing a human being repairing a similar
Robot in the lab.
A solution to this class of problems will have significant societal
impact. Remote repair technology will spawn new service industries
such as remote mechanics, remote monitoring and diagnosis,
telemedicine and telesurgery, teleagriculture, and operations in
hazardous environments.
Systems that can successfully perform tasks in a real world
environment must understand concepts of space and time, and
approximate algorithms where re execution of a program does not
always give the same result.
To repair a mobile robot on Mars, we need
---a mobile repair platform with all the relevant tools and fixtures;
---a semiautonomous system in which a human supervisor can provide
guidance but not intimate teleoperation (Note that the time delay of
10 to 15 minutes, depending on the relative position of Earth and
Mars, implies that most of the navigation and obstacle avoidance must
be locally controlled);
---a system that can repair implies precision manipulation capability
for disassem bly and assembly of the disabled platform;
---a system that can learn from observation by looking over the
shoulder of a human operator (requires a system with 3D vision,
modeling of space, discovering and programming equivalent manipulation
operators to the human actions); and
---a system that can engage in clarification dialog with humans to
verify and validate the understanding of the observations of human
operations.
Each of the tasks stated above seems likely to yield progressive
advancement given sustained effort. Perhaps the hardest is likely to
be the learning task, which is central to progress given that we
cannot afford to program all such tasks.
At present we are not aware of any system that can perform these tasks
at the implied generality of repairing a mobile electromechanical
system. Professor Katsu Ikeuchi of CMU (currently at Tokyo University)
demonstrated in 1988, a robotic system capable of learning from
observation, a simple stacking operation in a blocksworld. Even the
simple task of observing ``picking up a block and place it on top of
another block'' and deriving (learning) the equivalent motion action
from a sequence images proved to be difficult. Professor Ikeuchi,
instead, chose to derive the robot actions required, by inferring
(planning) the sequence of actions given the beginning and end states
of the scene.
3. Encyclopedia on Demand
We are living in the age of ``Wealth of Information and Scarcity of
Human Attention'' [Simon, 1995].
As of last year the information available on the web exceeds 100
terabytes. Infor mation that is publicly available in libraries and
other copyrighted forms exceeds 100 times the information on the
web. The deep web consisting of all data on all the disks in all
corporations and households is a million times larger. We have been
facing the information glut for many years and it promises to get
worse.
The third open problem we have chosen arises from what Professor Jaime
Carbonell of CMU calls the ``Bill of Rights for the Information Age'',
namely how to get the ``Right Information'' to the ``Right People'' in
the ``Right Language'' in the ``Right Timeframe'' in the ``Right Level
of Granularity''. Specifically (and narrowly) the right level of
granularity task we propose is to ``produce a 5000 word or less,
encyclopedia style article, on a given subject, by summarizing from
the relevant information available on the web in less than 24 hours''
Given the web has 100 terabytes, just reading all the data at 100
megabytes per second (current best bandwidth of a single disk) would
take over 11 days! Finding all related information using inverted
index techniques will help to retrieve most of the discoverable
data. Much of data retrieved will be disjointed, containing duplicate
entries, and obsoleted by later web postings.
The task of creating an encyclopedia style article requires several
new technolo gies such as
---document clustering to identify a group related articles;
---synthesis of information from all the related articles into a
single merged docu ment;
---summarization of the merged information into a convenient size; and
---language generation of a natural and intuitive sentences of the
finished summary. Conventional ``googlelike'' keyword based
retrieval systems often return thousands of hits. Finding entries that
are duplicate or similar (as in reporting of the same story by
different newspapers) and grouping them together is called document
clustering. This is currently achieved defining a documentsimilarity
metric based on a vector of the content words in each document.
Synthesis of information from multiple sources requires merging
relevant information from different sources while removing duplicate
or equivalent sentences, rationalizing conflicting information (as in
the case of an evolving story on number deaths after an earthquake),
and inserting background information that is often assumed to be known
in a news story (as in information about Second World War in a story
about the Berlin Wall). This is an area of current research interest
but progress appears to be slow.
Summarization and abstraction is routinely done by trained human
experts and on a less professional level by most of us. Summarization
by machine involves selection of informationrich phrases and
sentences from an article and producing a shorter article that
hopefully captures the essence of the article. Informationrich
content words and phrases are currently assembled from the document
content and structure, and by giving extra weight to title, abstract
and conclusions, and to chapter and section headings.
Language generation or sentence synthesis based on the intended
meaning has proven to be a challenging task. It has usually been
studied in the context of language translation and is equally
important to our task. Human language provides many different ways of
saying the same thing. Human beings seem to somehow choose the most
succinct form and this fact has been called the ``principle of least
effort''. Computers that can do the same are still in their infancy.
We appear to be making progress on all these technologies and yet we
seem to be far away from being able to create an acceptable quality
encyclopedia style article on a subject. We may be able to make
substantial progress by accepting human assisted generation of
encyclopediaondemand as an intermediate goal. Removing these human
interventions can then become the research agenda for the next
generation system.
--
<<社会契约论>>是一本好书,应当多读几遍
风味的肘子味道不错,我还想再吃它
※ 来源:·哈工大紫丁香 bbs.hit.edu.cn·[FROM: 202.118.230.220]
Powered by KBS BBS 2.0 (http://dev.kcn.cn)
页面执行时间:203.178毫秒