A Vision for
Computational Science
Pat Langley
Computational Learning Laboratory
Center for the Study of Language and Information
Stanford University, Stanford, California
http://cll.stanford.edu/~langley
A Day in the Life of a Future Scientist
Professor Jones comes into her office on Tuesday morning. Her first action
is to check the status of an experiment she submitted the night before. Her
computerized assistant reports the results for culture growth over time under
20 different conditions, displaying each curve in relation to the current
model’s predictions. The system highlights two conditions in which results
diverged from those expected.
A Day in the Life of a Future Scientist
Professor Jones comes into her office on Tuesday morning. Her first action
is to check the status of an experiment she submitted the night before. Her
computerized assistant reports the results for culture growth over time under
20 different conditions, displaying each curve in relation to the current
model’s predictions. The system highlights two conditions in which results
diverged from those expected.
Dr. Jones asks the assistant if it has any explanations to propose for the two
anomalies, and the system returns a list of ten alternatives, ranked by fits to
the data and consistency with knowledge of the field. The researcher asks
the computer aide if the literature contains other reports of either similar
results or similar hypotheses. It recommends five papers in response, and
she spends the next hour reading the two that seem most relevant.
A Day in the Life of a Future Scientist
Professor Jones comes into her office on Tuesday morning. Her first action
is to check the status of an experiment she submitted the night before. Her
computerized assistant reports the results for culture growth over time under
20 different conditions, displaying each curve in relation to the current
model’s predictions. The system highlights two conditions in which results
diverged from those expected.
Dr. Jones asks the assistant if it has any explanations to propose for the two
anomalies, and the system returns a list of ten alternatives, ranked by fits to
the data and consistency with knowledge of the field. The researcher asks
the computer aide if the literature contains other reports of either similar
results or similar hypotheses. It recommends five papers in response, and
she spends the next hour reading the two that seem most relevant.
After some thought, Dr. Jones tells her assistant to focus on the three model
revisions she feels are most plausible and asks it to design a new experiment
that will discriminate among them. The researcher makes a few changes to
its proposed conditions and submits the design for robotic execution when
the resources become available. She leaves the office in time to walk across
campus to attend a weekly committee meeting.
The Computer Revolution
In recent years, information technology has changed the way we:







communicate with each other
learn new facts
listen to music
make travel plans
shop and make purchases
bank and pay bills
prepare and give presentations
Computers support all of these activities by storing, retrieving,
processing, and interchanging information in digital form.
The Scientific Revolution
The much older scientific revolution has greatly increased our
understanding of:







the universe
the Earth
matter
life
disease
the mind
society
Science supports such advances by collecting data systematically,
stating clear theories/models, and relating them to each other.
Science and Computation
Clearly, the potential for combining these important revolutionary
movements is greater than either in isolation.
science
computation
computational
science
This approach has already produced some impressive advances in
a number of scientific fields.
Collection / Analysis of Sky Surveys
Simulation / Visualization of Fluid Dynamics
Mining Data from Earth-Observing Satellites
Analysis of Genetic Sequences
Clustering of Gene Expressions
Recording / Analysis of Brain Activity
Indexing / Retrieval of Scientific Papers
The Importance of Computational Science
These are each important contributions, but they only touch on
the full potential of computational science.
A recent report from the President’s Information Technology
Advisory Committee stated:
Universities . . . must make coordinated, fundamental, structural
changes that affirm the integral role of computational science in
addressing the 21st century’s most important problems, which are
predominantly multidisciplinary . . . and collaborative.
We need a systematic vision for computational science, a broad
research agenda, and clear plans for educating the next generation.
Facets of
Computational Science
A Broader Definition
In its broadest sense, we can define computational science as:
 the use of computational methods and metaphors
 to understand and support the scientific enterprise.
This requires that we understand the ways in which:
 content disciplines pose computational problems;
 method disciplines offer computational solutions.
Computational science attempts to relate these two areas, just as
science relates data to theory.
Content-Oriented Disciplines
Computational science’s problems come from content fields:
biology
chemistry
Earth science
materials science
medicine
computational
science
astronomy
physics
These disciplines stand to benefit from computational science,
but they also provide data, theories, and other content.
Method-Oriented Disciplines
Computational science’s techniques come from method fields:
astronomy
chemistry
Earth science
materials science
medicine
physics
computational
science
biology
computer science
mathematics
decision analysis
engineering
logic
phil. of science
statistics
These fields provide the underlying processes that computational
science uses to aid research in the content disciplines.
Additional Content Disciplines
The social sciences also have roles to play as content fields:
anthropology
education
geography
history
psychology
sociology
computational
science
economics
computer science
mathematics
decision analysis
engineering
logic
phil. of science
statistics
These disciplines stand to benefit from computational support
as much as the natural sciences.
Scientific Representations and Structures
Computational science should study the full range of content that
scientific fields must represent, including:
descriptive laws
theories / models
predictions/explanations
experimental designs
records / databases
computational
science
taxonomies / ontologies
documents / images
Scientific disciplines would benefit from the ability to encode
and store each of these structures in digital form.
Scientific Processes and Mechanisms
Computational science should also study the development and
utilization of mechanisms for:
form taxonomies
taxonomies / ontologies
theories / models
predictions/explanations
experimental designs
records / databases
documents / images
computational
science
descriptive laws
predict / simulate
explain phenomena
evaluate hyps / models
propose hyps/models
devise instruments
design experiments
record / index results
communicate results
Scientific disciplines would benefit from the ability to emulate
these processes on computers.
Contributions from Computer Science
Computational science should also draw upon key subfields of
computer science:
 data structures / knowledge representation
 computer simulation
 programming languages
 database / information retrieval
 remote sensing / sensor networks
 data mining / knowledge discovery
 human-computer interaction
Each of these areas can support essential components of the
scientific process.
Claims About
Computational Science
Science as Computation
Claim: Science can be viewed as an interconnected set of
computational processes.
According to this framework, we can understand science by:
 analyzing the tasks that arise in scientific research;
 studying the behavior of historical and modern scientists;
 creating computational artifacts that address the same tasks.
Two fields with this view – artificial intelligence and cognitive
psychology – are especially relevant to computational science.
Science as Heuristic Search
Claim: Science can be characterized as search through one or
more problem spaces.
According to this framework, we can understand science by:
 identifying knowledge states that arise in scientific research;
 specifying operators that generate new knowledge states;
 describing the organization of search through the spaces.
Again, research on heuristic search in humans and machines is
highly relevant to this perspective on scientific activity.
Numeric and Symbolic Processing
Claim: Qualitative/symbolic reasoning is just as crucial to
science as quantitative/numeric reasoning.
Many fields, like biology and psychology, rely mainly on
qualitative models and explanations.
Thus, a broadly based computational science must support:
 string and document processing
 logical deduction and abduction
 reasoning over causal models
Fortunately, computers are more than number crunchers; they
support general symbolic processing.
Computational Science and the Humanities
Claim: The humanities have central roles to play in the pursuit
of computational science as content disciplines.
art history
literature
film / television
linguistics
music
theater
computational
science
classical studies
computer science
mathematics
decision analysis
engineering
logic
phil. of science
statistics
Computational Science and the Humanities
Claim: The humanities have central roles to play in the pursuit
of computational science as method disciplines:
 logical reasoning and analysis
 textual composition and rhetoric
 visual design and composition
Moreover, philosophy of science studies the nature of scientific
knowledge and reasoning.
Each has techniques that can inform the design of computational
artifacts that support the scientific process.
Human-Computer Synergy
Claim: Science is best achieved through a mixture of computercontrolled and human-controlled processes.
Scientific research is a complex endeavor that we are unlikely
to automate completely anytime soon; instead, we should:
 determine which tasks are most tractably automated;
 determine which tasks as best done by human scientists;
 create environments that support their effective interaction.
This makes another discipline – human-computer interaction –
especially relevant to computational science.
Some Important Challenges
Claim: Despite many successes in computational science, we
need more research on methods that:
 revise existing models in response to anomalies;
 construct models in knowledge-rich, data-lean fields;
 visualize relations between data and models;
 support the incremental nature of science.
Such techniques will provide better support for science as it is
normally practiced by scientists.
Computational Science
at Dartmouth
Computational Science at Dartmouth
Dartmouth seems well suited for taking a lead in developing
computational science as a distinct field:
 ongoing computational work in specific areas
 low hurdles for cross-departmental research
 focus on high-quality liberal education
 Neukom endowment to launch an institute
Most important, the field needs such leadership and Dartmouth
is willing to serve in that role.
Creating a Dartmouth Community
Computational science at Dartmouth requires a clear sense of
community, which we can foster by:
 identifying faculty across campus with the potential and
commitment to contribute to the new field;
 organizing talks by relevant faculty to advertise each others’
work and explore opportunities for collaboration;
 establishing a student organization with an emphasis on, and
with activities in, computational science;
 hosting regular social hours at which involved parties can meet
and discuss common interests.
Such activities will improve awareness of computational science
on campus and increase excitement about its potential.
Fostering Interdisciplinary Research
Computational science at Dartmouth will need interdisciplinary
collaborations, which we can encourage by:
 supporting postdoctoral fellows to work jointly with faculty
from different departments;
 providing stipends to graduate students who work jointly with
faculty from distinct departments;
 funding seed projects that involve collaboration among faculty
across departments;
 hiring new faculty with interdisciplinary records and with clear
links to multiple departments.
Such joint research efforts will make Dartmouth a role model for
collaborative work in computational science.
Raising Funds for Computational Science
Dartmouth’s efforts in computational science would benefit
from additional funding, which we can assist by:
 organizing and submitting cross-departmental proposals for
large grants from NIH, DOE, and NSF;
 pursuing gifts from, and joint projects with, companies that
believe in computational science;
 playing an active role in government advisory boards to
encourage long-term funding for the field;
 working with elected representatives and agency officials to
develop new funding programs.
Dartmouth can play a central role in such efforts to support the
field of computational science.
Publicizing Dartmouth’s Role
We can clarify Dartmouth’s efforts in computational science by:
 hosting a colloquium series that invites researchers from many
fields to speak on campus;
 organizing and hosting an annual symposium on timely issues
in computational science;
 establishing a book series on computational science that
includes volumes based on the symposia;
 creating a Web site that reports news in computational science
to the broader community;
 collecting on-line readings that define the field and illustrate
key problems and approaches.
Combined with educational and research efforts, these activities
will establish Dartmouth as a leader in computational science.
Fostering Interest in Computational Science
Dartmouth should encourage interest in computational science
among the campus community and the general public by:
 publishing accessible overviews of the movement in venues
like Science, Nature, and CACM;
 producing a documentary on computational science that covers
the field’s potential and challenges;
 organizing a program to involve undergraduates in ongoing
research on computational science;
 hosting evening talks by campus researchers in language
understandable to a wide audience.
These activities will further establish Dartmouth’s commitment
to computational science and its leadership in the area.
Possible Homes for Computational Science
What academic unit would serve as the most appropriate home
for computational science?
 departments in the physical, life, or social sciences?
 departments of computer science, mathematics, or statistics?
 a school of of engineering or arts and sciences?
Computational science is best supported at the campus-wide level,
but coordinated with efforts on specific problems and approaches.
A key challenge is to create a general institute of computational
science that retains close ties to these more established units.
Some Relevant Dartmouth Units
Dartmouth already has many interdisciplinary units relevant to
computational science, including:
 Bioinformatics Shared Resource
 Center for Biological and Biomedical Computing
 Center for Cognitive Neuroscience
 Center for Integrated Space Weather Modeling
 Mathematical Social Sciences
 MD/PhD Program in Computational Biology
 Molecular Biology Core Facility
 Numerical Methods Laboratory
These can support Dartmouth’s vision for computational science,
but they must become active stakeholders.
A Curriculum for
Computational Science
Research and Education
Enlightened research in computational science is not enough; we
must also edcuate the next generation of scientists.
These two central activities should travel hand in hand, with:
 research developing computational methods to support the aims
of content-oriented scientists;
 education training students to use such computational methods.
Both efforts should be grounded in specific problems from the
content-driven disciplines.
However, they require very different background / prerequisites.
A Curriculum in Computational Science
Dartmouth courses in computational science should provide their
students with an understanding of:
 structures, processes, and practices that arise in science;
 computational methods to encode these structures/processes;
 how such methods can support the scientific enterprise.
The curriculum should treat computational science as a field with
its own intellectual issues but grounded in scientific applications.
Students should acquire a broad view of science and the potential
for computational support in each component activity.
Possible Courses on Computational Science
A curriculum in computational science should include courses on:
 the scientific enterprise, including findings from the history,
philosophy, and psychology of science;
 scientific formalisms that cover different frameworks and give
practice at modeling in different disciplines;
 interactive modeling environments, including visualization
methods, that build on HCI principles;
 applications of computational linguistics, including information
retrieval/extraction, summarization, and generation;
 methods for analyzing data and constructing models, illustrated
in a variety of disciplines;
 specialized methods for modeling and data analysis for fields
like Earth science, psychology, and biology.
Challenges in Computational Science Education
The broad nature of the field raises a number of challenges:




science involves a wide range of structures and processes;
different scientific disciplines have distinct characters;
students must understand both method and content areas;
they must master general principles and specific applications.
We can best address these pedagogical issues by:
 following generalized core courses with specialized tracks;
 organizing even general classes around hands-on projects;
 using high-level software environments to lower entry barriers.
These should provide the curriculum in computational science
with the right balance of generality and specificity.
Teaching Science via Computation
Claim: Computational science is an excellent unifying theme for
teaching scientific content.
Students can gain knowledge about content disciplines by:
 developing computational models
 analyzing data with computers
 simulating these models’ behavior
 comparing predictions to observations
These activities do not require substantial training in computer
science or traditional programming languages.
They can be achieved with high-level languages and interactive
software environments.
Some Environments with Instructional Promise
We can make computational science accessible to students by
drawing on high-level software environments such as:
 STELLA and PROMETHEUS let users specify, visualize, and
simulate differential equation models (e.g., of ecosystems);
 HYBROW lets users specify, visualize, falsify, and revise
qualitative causal models (e.g., of cell biology);
 ACT-R and ICARUS let users specify, simulate, and trace
symbolic process models of human reasoning and learning.
We have used two of these environments in Stanford courses
and hope to utilize the other in the future.
Closing Remarks
Summary
Computational science, developed along appropriate lines, will:
 change the ways that we carry out scientific research;
 increase the rate of scientific and technological progress;
 improve education in the sciences and other disciplines.
However, making this enterprise successful will require:
 a broad and inclusive vision for computational science;
 a strong commitment to interdisciplinary research;
 an institution willing to play a leadership role.
Dartmouth can help transform this vision of into reality and aid
computational science to develop its full potential.
End of Presentation
Descargar

Computational Discovery of Communicable Knowledge