Important decisions
typically involve many difficult tradeoffs between costs and benefits, and far
too many alternatives to properly compare. The user is often forced to rank or
weight criteria, or to do other unnatural things that oversimplify their
preferences and make it easy to miss a good but unexpected alternative.
Instead of requiring the user to articulate their preferences in advance, we
present the best alternatives to the user so that they can explore and compare
them. Our solutions offer the user a confident understanding of the range of
the worthwhile options, with the ability to narrow down to the best few
choices.
We generate
choice alternatives by assembling parts according to rules. Available schemes
include:
 | Exhaustive generation of every legal configuration of components
|
| Evolutionary algorithms that allow better alternatives by mutating and combining current ones
|
| "Growing" alternatives in a recursive breadth-first search by selecting the components that provide the required functionality
|
Our system
organizes the evaluation (usually by multiple simulations) of each generated
choice alternative such that its performance on multiple criteria may be
measured. It uses distributed computing, and is robust against the changing
availability of machines in its computation cluster.
Our filter
removes the inferior choice alternatives. It removes choices where other alternatives are at least as good in every way and better in
some way.
The user
interactively explores the decision space to
learn what types of alternatives perform better, and to see the tradeoffs
among the different performance criteria. Selections can be made amid
cross-linked displays, which include:
| Two-dimensional plots of continuous or discrete attributes
|
| Histograms of frequency of occurrence of attribute values
|
| Lists and tables
|
| Stacked one-dimensional plots showing the distribution of attribute values on each criterion of interest
|
All plots show the same set of points (choice alternatives),
colored the same way, enabling comparison. As well as allowing tradeoff
judgments to be made, the viewer lets non-performance criteria to be studied.
This allows the user to perform other data mining tasks such as sensitivity
analysis – or how dimensions of performance vary with the value of
attributes intrinsic to the specification of the alternatives.
| Feedback from the evaluation of previous alternatives affects generation
of new alternatives.
|
| When evolutionary algorithms are being used, the conventional scalar
fitness function can be replaced with the filter, which then decides what will
"die," and what will "live and breed."
|
| The generation and evaluation process can produce many
alternatives. The filter can reduce this to a much more manageable number with
guarantees of losslessness.
|
| The choice alternatives that survive the filter show the tradeoffs that exist: to win any improvement on one performance
criterion, a loss on another must be accepted. Our graphical interactive
environment is ideal for making tradeoff judgments and identifying the parts
of the space where an insignificant loss on one criterion allows a
significant gain on another.
|
| Graphical exploration can reveal regions of the decision space that are
worthy of deeper investigation by further generation and evaluation.
|
See how decision
making fits into our overall technological process with the help of an active diagram.
Read the introductory
academic paper: An Architecture for Exploring Large Design Spaces.
Josephson, J. R., Chandrasekaran, B., Carroll, M., Iyer, N., Wasacz, B.,
Rizzoni, G., Li, Q., & Erb, D. A. (1998). Proceedings of the National
Conference of the American Association for Artificial Intelligence, (pp.
143-150). Madison, Wisc. (PDF)
Read a slideshow presentation about
our MCDM technology, which uses examples from a very early version of the
graphical decision-support environment.
