One of the primary reasons for building a statistical model is to make
predictions with it. While S-Plus is a proven tool
for building models, its facilities for predicition are limited to
small models. A complex risk analysis or decision analysis problem
may require stringing several smaller models (fit in
S-Plus) together to make a complete prediction.
For example, a reliability model for a complex system is built from
smaller models for the subsystems and components. Furthermore, a good
analyst does not make a single prediction, but rather explores the
implications of critical assumptions made in the modelling. Although
an analyst could program S-Plus to do all these
tasks, S-Plus does not support them directly.
Graphical-Belief is tool for exploring the
predictive aspects of models. It is based on the technology of
graphical models (also known as influence diagrams, belief
networks or Bayes nets) which has already become a standard in
decision analysis, statistics and artificial intelligence. These
models have been used successfully in such diverse areas as system
level reliability, medical decision making, financial planning and
operations managment. Graphical-Belief is a
complete environment for building and exploring risk models. It
provides a wide range of tools for modelling tasks including:
- The model graph takes as complex task and breaks it up into
small pieces. The modeller only specifies the direct
interactions between variables in the model;
Graphical-Belief calculates all the implied
dependencies between remote variables.
- Knowledge Engineering
- A large quantity of knowledge from many sources (both expert
opinion and data) go into a typical graphical model.
Graphical-Belief provides tools for
maintaining that knowledge. Modellers can draw from a library
of previously built knowledge fragments and generic knowledge
structures. Because Graphical-Belief uses
an object-oriented schema for storing knowledge, a change to a
single prototype rule or variable is quickly propagated to many
instances kept in the model.
- Using Graphical-Belief, modellers have a
choice of representation for relationships between variables:
probability (for uncertain relationships), logic (for certain
relationship) and belief functions (for imprecise uncertain
relationships). Graphical-Belief achieves
this flexibility with a generic inference engine which can be
simply expanded to include other representations for
relationships (including utilities and possibilities).
- Dynamic Visualization
- Graphical-Belief lets the analyst
directly manipulate the model, exploring the implications of
hypothetical scenarios and assessing the sensitivity of key
predicitions to critial assumptions.
Graphical-Belief has an very flexible
parameter system which allows the analyst to easily study the
sensitivity of a single parameter which impacts the model in
Look at some
Graphical-Belief in action.
View a list
of Graphical-Belief in publications and downloadable technical
The Graphical-Belief user
interface is implemented in Garnet.
information about obtaining Graphical-Belief (and why
it is not generally available).
the home page for Russell Almond , author
here to get to the home page for Insightful (the company that StatSci
has eventually evolved into).
I would like to thank David Madigan
for his advice and collaboration on the project (he also sponsored the
original posting of these pages).
The original Belief project was supported by Army
Research Contract DAAL03-86K-0042) at Harvard University, Arthur
Dempster, Principle Investigator. The design of
Graphical-Belief is based in part on that work.
The Graphical-Belief program at StatSci has been
partially supported by the following grants:
- NASA Phase~I SBIR "Computing Environments for Interactive
Graphical Belief Modelling", Russell Almond, P.I., NAS 9-18669
- NIH Phase~I SBIR "Computing Environment for Interactive
Graphical Belief Modelling", Russell Almond, P.I., 1 R43 RR
- NASA Phase~II SBIR "Graphical Belief Models for System
Reliability", Russell Almond, P.I., NAS 9-18908
Russell Almond, <lastname> (at) acm.org
Last modified: Fri Aug 16 14:28:35 1996