Kenneth Forbus
Ken Forbus is a Professor of Computer Science and Education.
Before coming to Northwestern, Prof. Forbus was the head of the
Artificial Intelligence group at the Beckman Institute at the
University of Illinois at Urbana-Champaign. Prof. Forbus received
his Ph.D. from MIT in 1984 in Artificial Intelligence, received
an NSF PYI award in 1987, and was elected a AAAI Fellow in 1992.
His interest in the construction of intelligent tutoring systems
and learning environments stems in part from his experience
working on the STEAMER Project at Bolt, Beranek, and Newman in
the 1980s.
Prof. Fobus' current research interests include:
Qualitative physics. Prof. Forbus is one of the
founders of qualitative physics, the area of artificial
intelligence that develops representations and reasoning
techniques that capture the ways that people reason about the
physical world, ranging from the person on the street to
scientists and engineers. This research includes:
- Qualitative Process theory. Qualitative process
theory normalizes the notion of physical process that
seems to be crucial to human common sense reasoning and
to technical reasoning in many domains. QP theory
introduced several ideas now widely used in qualitative
physics, including the idea of using ordinal
relationships to provide a qualitative representation for
numerical values, and these of partial information about
monotonic functional dependencies to provide a
qualitative representation for functions and causal
relationships.
- Compositional modeling. This modeling methodology,
developed in collaboration with Dr. Brian Falkenhainer
(now at XeroxPARC), extends the use of (logical)
quantification in QP theory to include explicit
representations for modeling assumptions and techniques
to organize them. A central goal of compositional
modelings to help automate the model formulation process,
i.e., the creation of models for specific tasks based on
descriptions of a scenario and a large, general purpose
knowledge base. Compositional modeling has been used with
qualitative models, quantitative models, and hybrid
models, suggesting that it is a viable formalism for
organizing large bodies of knowledge for scientific and
engineering reasoning, compositional modeling is now the
leading modeling methodology in qualitative physics, with
exciting new contributions being made by laboratories
world-wide.
- Qualitative spatial reasoning. The key idea of the
eMtric Diagram/Place Vocabulary model of spatial
reasoning is that quantitative spatial representations
serve as a crucial substrate for performing qualitative
spatial reasoning. The Metric Diagram models the role of
perception in human processing. It is well-known that
people heavily rely on diagrams and other physical models
for spatial reasoning, most likely because we have
evolved with very Powerful perceptual processes. Making
computers that reason spatially as well as we do seems to
require providing a functionally equivalent Facility.
Using this quantitative input, task-specific qualitative
representations (place vocabularies) are extracted for
further reasoning. This model has been tested in several
programs, including PROB (which reasoned about motion
through space) and CLOCK (which was the first program to
automatically analyze and qualitatively predict the
behavior of fixed-axis mechanisms, such as mechanical
clocks). Systems descended from these ideas can now
analyze substantial numbers of mechanical mechanisms.
- Self-explanatory simulation. This new simulation
methodology, developed in collaboration with Dr.
Falkenhainer, combines the accuracy and speed of
numerical simulation with the explanatory capabilities of
qualitative representations. A self-explanatory simulator
incorporates the qualitative models used to originally
generate it, thus it can explain the results it produces
and can monitor itself to see when its results become
implausible.Such simulators can be generated
automatically, in polynomial time, from high-level
physical system definitions. Self-explanatory simulators
have potential applications in engineering design
(e.g.generating system-level simulators for parameter
selection & optimization) and in education and
training (e.g., generating simulated laboratory setups
and training simulators).
- Articulate virtual laboratories for science and
engineering education . This project is using the
fruits of our research on qualitative physics and
analogical processing to develop intelligent learning
environments and tutoring systems for teaching science
and engineering. Conceptual design tasks seems to offer
an excellent setting for teaching fundamental physical
principles, both to undergraduates and to trainees in
technical areas. We are building virtual laboratories
that will allow students to learn science and engineering
by designing and "building" artifacts. A
virtual laboratory should provide high-level CAD tools,
which aid students in creating and analyzing their
designs, and a manufacturing facility, which produces
simulations of the student's design that can then be
tested in simulated environments. The software must be
articulate, in that it must provide explanations for its
results, and coach students on improving their designs.
Self-explanatory simulators will be used to provide the
manufacturing facility. In collaboration with Dr. Peter
Whalley of Oxford University, we have built a prototype
conceptual CAD system for engineering thermodynamics,
called CyclePad, which is being field-tested on students
at Oxford and at Northwestern.
Cognitive simulation of analogical processing The goal
of this work is to develop a computational account of the
processes involved in reasoning and learning by analogy, and an
account of the role of similarity more generally, in human
cognition. Our programs are thus motivated and evaluated in terms
of their ability to account for psychological phenomena. This
work, which is carried out in collaboration with Prof. Dedre
Gentner, includes: